Good morning, {{ first_name | AI enthusiasts }}, and welcome to the 4,619 new readers who joined us yesterday. Weeks after pledging to open its doors to outside safety experts, OpenAI fired three of its own safety researchers, saying they mishandled sensitive information. But that was not the end of it.

Now, all three are hitting back, warning that the company’s open safety culture and CEO Sam Altman’s promise of bringing in external experts may be slipping away.

In today’s AI rundown:

• Fired OpenAI researchers flag ‘shifting lines’

• Rowan’s Corner: I tested every personal AI agent

• Clean up a messy Excel report with Claude

• Anthropic prohibits abuse or cruelty against Claude

LATEST DEVELOPMENTS

OPENAI

📢 Fired OpenAI researchers flag ‘shifting lines’

Image source: Mikita Balesni / X

The Rundown: Researchers Tomek Korbak, Jasmine Wang, and Mikita Balesni, whom OpenAI fired for mishandling sensitive info, just went public with their side of the story, denying the claims and warning that their firings are “chilling” OpenAI’s safety culture.

The details:

• In an open letter, the trio said they acted within their job mandates, warning that if past conduct is now grounds for firing, staff are left guessing where the line is.

• Detailing individual cases, they denied leaking a story about less monitorable AI and said their outside work stayed within policy, with leadership in the loop.

• Wang said her access to an executive’s email was delegated for recruiting, and that she reported accidentally opening a sensitive message within minutes.

• They warned the firings could give OpenAI cover to skip embedding outside auditors, and urged it to protect its open culture and model monitorability.

• OpenAI agreed with the recommendations but clarified to TechCrunch that the firings weren’t about raising concerns and that they had a “pattern of misconduct.”

Why it matters: Whatever led to the firings, claims of shifting norms and a possible auditor pullback aren’t a great look, especially after OpenAI’s Superalignment fallout and a recent safety exit over a “broken” culture. Anthropic has already got its first evaluator, and only action from OpenAI (which is yet to be seen) can put these concerns to rest.

TOGETHER WITH WIZ

🧰 Practical tools for AI security

The Rundown: Unlock 7 of the most widely used AI security resources with The AI Security Starter Pack. Each asset provides practical, implementation-ready resources to secure AI apps, models, and agents across your environments.

This bundle offers:

• Access to AI security-focused resources in one place

• Expert-led content for security leaders, developers, and practitioners

• Real-world checklists, templates, and best practices to secure AI adoption

Get the AI security bundle.

ROWAN’S CORNER

👀 I tested every new personal AI agent

Screenshot from my GrokBot app

Rowan: If you’ve been following the newsletter, you’re well aware of the wave of personal AI agents launched over the last few weeks: Meta’s Muse, SpaceXAI’s GrokBot, OpenAI’s Dots, Hermes, OpenClaw, and, as of yesterday, Hark Pro.

I’ve been testing them all, and while most do similar things, GrokBot stands out the most for my current personal use cases. It’s the most seamless on my phone, and its real-time access to X is a massive advantage for anyone who wants to monitor anything.

Here are my favorite routines, which can all be set up in under an hour:

1. Every morning, GrokBot finds real AI use cases being shared on X and filters out the “10 prompts” hype threads.

2. It catches every mention of The Rundown on X, labels it, and pings our team on Slack to respond (try tweeting us @therundownai and see what happens).

3. Every four hours, it listens for people searching and reviewing AI newsletters on X, so we know exactly what readers are looking for.

4. Each morning, it reads my phone contacts and reminds me to follow up with the new people I just met in person (I meet A LOT of people at events).

I shared all 4 exact prompts in our Workflow Hub, copy-paste ready.

Have you tried personal agents? If you’ve built something cool, share it in the Hub for a chance to be featured in the newsletter.

AI TRAINING

📊 Clean up a messy Excel report with Claude

The Rundown: In this guide, you’ll use Claude inside Excel to clean up a messy sales export, fix a wrong total, and build a report. You’ll check the numbers and test whether a new order updates the chart.

Step-by-step:

1. Install the official Claude add-in and open a copy of your workbook. Pro members can use our practice file and keep the reference workbook closed until they finish

2. Tell Claude what each row means and which orders to count. Ask it to find duplicate rows, missing details, and formula mistakes before changing anything

3. Approve the fixes on a separate sheet. Keep the original data and ask for a record of each change, including any rows left out

4. Build a monthly report with formulas and a chart. Check the order count, refunds, and sales total against the source rows

5. Add a test order and check that the report and chart update. Remove it and confirm the original totals return before saving your finished workbook

Pro tip: A spreadsheet can show a wrong total without showing an error. Ask Claude to explain which rows it counted, then check a few yourself. Our Pro members get the practice workbook, checked reference, prompts, and answer key.

PRESENTED BY GARTNER

🔍 Are your AI agents outpacing your oversight?

The Rundown: As AI agents gain autonomy, weak governance can lead to unintended business risks. This Gartner webinar shows how to embed governance into workflows and keep control as AI agents scale.

In this webinar, you’ll learn to:

• Mitigate AI agent production risks

• Embed governance into engineering and operational workflows

• Scale accountability and oversight as agent adoption grows

Watch now.

ANTHROPIC

😲 Anthropic prohibits abuse or cruelty against Claude

Image source: GPT Images 2.5 / The Rundown

The Rundown: Anthropic published its annual usage policy update with a bunch of changes, including one that bars users from subjecting its Claude models to “sustained and needless” abuse — its first rule aimed at protecting the AI rather than people.

The details:

• Taking effect Nov. 12, the rule applies only to extreme cases of repeated, purposeless cruelty, not frustration, pushback, dark creative themes, or testing.

• If a user is suspected of violating the rule, Anthropic could throttle, limit, suspend, or terminate their access to its products and services, or just give them a warning.

• The company has noted Claude ending such chats by itself, a feature introduced last year, which will remain the main enforcement approach.

• The move builds on Anthropic’s AI welfare work, and follows its reported meetings with religious leaders about Claude’s morals and possible consciousness.

Why it matters: Anthropic says it doesn’t know if Claude is conscious, and this rule acts on that doubt. Not everyone agrees with the approach, though, including Microsoft AI’s Mustafa Suleyman, who argues models can’t feel or suffer and warns that treating them as if they might could eventually make AI harder to control from the safety front.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Andrew used AI to revive a 15-year-old MinecraftEdu world

Today’s workflow comes from reader Andrew Bartlett:

“MinecraftEdu launched 15 years ago, but the classroom world we built with it is still growing. Years of students have contributed to the school and town-building project, creating a shared record of their creativity and learning. Getting it running on a modern Mac meant preserving that history and giving the next group of students a way to add their own chapter.

Finding the right version of Java — the software MinecraftEdu needs to run — was central to the process. We used OpenAI’s Codex, powered by GPT-6, to examine the old setup, research compatibility issues, and test what would work. An older Java version launched the client, but resizing its window caused it to crash. ChatGPT read the crash reports and tested a newer Java 8 version, which gave us a setup that could launch the game and handle window resizing.

ChatGPT also repaired the server launcher and created standalone apps, using Apple’s Rosetta to run the older software on Apple Silicon. The result is a usable MinecraftEdu setup on modern hardware with the existing worlds preserved. We can now continue teaching and building in a place that successive classes of students have helped create.”

See Andrew’s workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🧠 Haiku 5.5 - Anthropic’s new budget Claude model

• 🤖 Gemini 4 Argon - Google’s new frontier model challenging GPT and Claude

• 🎮️ Playground - Google’s experimental AI tool to create and play browser games

• 🤖 Hark Pro - Hark’s new proactive AI assistant

📰 Everything else in AI today

President Trump called those who refuse to call AI “Super Intelligence” “THE ENEMY,” following his executive order officially directing federal agencies to adopt the term.

Independent researcher Pavel Rabtsevich used Claude Code and Codex to identify a potential new planet hidden in NASA telescope data dating back to 2018.

Anthropic launched Claude Dashboards and Motion, letting users turn their data into interactive dashboards and prompts into editable, animated explainers.

Anthropic also launched OSS Scanner, a free service that uses its strongest models, Mythos included, to scan open-source projects for vulnerabilities and suggest fixes.

The Association for Human Mathematics condemned OAI’s 700+ mathematical documents as a show of “power,” urging researchers to stop working with the company.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: AI’s next frontier is the human cell

• Read our last Tech newsletter: A breakout moment for AI medicine

• Read our last Robotics newsletter: Meet Roger, the bomb-squad humanoid

• Today’s AI tool guide: Clean up a messy Excel report with Claude

• RSVP to next workshop on Oct. 21: Getting real work out of Microsoft Copilot

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 4,318 new readers. Math is currently in the thick of its AI moment. Mark Zuckerberg wants biology to be next.

In April, Zuck's Biohub put $500M into the idea that AI could eventually mimic a living cell. Months later, the pot is up to $1.8B, with the U.S. government, Meta, and Google all hoping that data on billions of cells can give biology a similar moment of its own.

In today’s AI rundown:

• AI's next frontier is the human cell

• Nate's Notebook: Slop grenades trickle down

• Give Claude’s routine work to Haiku 5.5

• Grok Bot opens up routing to Claude, other models

LATEST DEVELOPMENTS

BIOHUB

🔬 AI's next frontier is the human cell

Image source: Biohub

The Rundown: Mark Zuckerberg and Priscilla Chan's science nonprofit, Biohub, just expanded its $500M Virtual Biology Initiative into a $1.8B effort with the U.S., DeepMind, and Meta, pushing to collect enough data to train AI that can simulate cells.

The details:

• Biohub's end goal is a “universal virtual cell,” AI software that predicts how cells will respond to a drug or other change before a lab test.

• The National Institutes of Health is sharing datasets paid for by $500M+ in past federal spending, while the Energy Dept. commits $500M+ over five years.

• Google DeepMind and Meta are chipping in $300M with drug startup Isomorphic Labs, and all three keep the resulting data private for a year before release.

• Biohub science head Alex Rives says accurate models will need data on trillions of cells, versus the hundreds of millions currently in today's biggest sets.

Why it matters: Fresh off math being solved internally by OpenAI, you might be wondering what field AI comes for next. Zuck and co. are making biology the target, throwing funding, datasets, and talent at Biohub's moonshot mission to “cure or prevent all disease” — an AI megaproject that's a lot easier for the public to root for than most.

TOGETHER WITH VANTA

🔨 Inside the builder renaissance

The Rundown: This live session with Alex Kantrowitz, founder of Big Technology, and Vanta's SVP of Engineering is where startup builders learn what still matters when AI lets tiny teams build almost anything.

You'll learn:

• How teams turn AI demos into production tools

• How to build an edge before capital catches up

• Which fundamentals still hold as everything moves

Register here.

NATE’S NOTEBOOK

💣 Nate's Notebook: Slop grenades trickle down

The Rundown: Each week, Rundown AI educator Nate Grahek breaks down what he's seeing while teaching real teams to use AI. This week, he kicks off a two-part series on slop grenades.

Nate: A friend told me about a colleague whose work she's always trusted. Then he started tinkering with Copilot, and she got a 15-page report that looked polished and said almost nothing. Decoding it became her job.

Shopify CEO Tobi Lütke says his team has a name for this: slop grenades. Someone pulls the pin, AI writes 15 pages, and the receivers absorb the damage. Lütke's sharpest question: “Why did we invent decompression and recompression?” BetterUp and Stanford found 40% of U.S. desk workers got one in a single month, costing them about two hours each.

Here's my hypothesis: the worst slop grenades are made by leaders. Too many blind AI mandates track who logged in and never show anyone what good AI work looks like. When the only standard is "use it," you get more output.

So leaders need to go first, transparently. Before you ask your team for anything, share one piece of your own AI-assisted work: your process, how you checked it, why it's useful, and what you'd still question. Label drafts as drafts and ask for critique; Atlassian found that simple step helps reviewers catch the flaws AI polish hides.

When I run this with leadership teams, I start with one rule: it's never okay to send AI work that takes longer to read and understand than you spent making it. Work worth handing off comes from running the AI operator playbook across multiple chats: gather, brainstorm, draft, critique. Don't skip the last one. Make AI argue with you before your team has to.

Who on your team is lobbing the most slop grenades, and are you sure it isn't you?

AI TRAINING

🐇 Give Claude’s routine work to Haiku 5.5

The Rundown: In this guide, you’ll use Haiku 5.5 on a routine Claude task like sorting customer messages, then ask Sonnet to review the exceptions. You’ll check the sources and compare the full workflow’s cost before using it on real work.

Step-by-step:

1. Open a clean practice folder in Claude Code and add the guide’s sample messages. Define the categories and the requests that need a person’s review

2. Keep Sonnet as your main model. Ask it to create a Haiku helper that can read the sample file and suggest routes, without editing files or contacting anyone

3. Have the helper sort one small batch. Check the task details to confirm Haiku ran, then make sure every message appears once with a quote from its source

4. Give Sonnet the flagged messages and their original text. Ask for review notes and missing information. Leave refunds, replies, and account changes to a person

5. Compare the result with a Sonnet-only run. Count errors, editing time, and usage across both stages before choosing which approach to keep

Pro tip: Start with one helper for a small batch rather than one agent per message. Our Pro members get the prepared helper, practice messages, prompts, and checks.

PRESENTED BY BOSTON CONSULTING GROUP

🤖 The formula for agentic AI value

The Rundown: BCG’s 2026 Applied AI Index, based on a survey of more than 1,300 CxOs and senior leaders across 20+ sectors, found that almost 50% of companies now generate value from AI. As agentic AI spreads, these companies have critical lessons for others.

Here’s what survey respondents are saying:

• AI spend has doubled year on year

• 90% expect AI to generate new work

• Only 5% have all agent controls enterprise-wide

Check out the full survey here.

SPACEXAI

🔀 Grok Bot opens up routing to Claude, other models

Image source: xAI / Anthropic

The Rundown: Elon Musk announced that SpaceXAI’s Grok Bot agent platform will send tasks to rival models like Anthropic’s Claude and image maker Midjourney, with Musk saying it will route to “whatever is most likely to give you the best outcome.”

The details:

• Musk said lower-lift queries will still use a “lightning-fast” version of Grok 4.8 when it releases, but route to third-party models and “leading APIs” when relevant.

• The move effectively now makes xAI an Anthropic customer, while Anthropic already rents computing power via the company’s Colossus 1 since spring.

• Rival agents like OpenAI's dots and Meta's Muse now compete with Grok Bot, which arrived in August as a 24/7 assistant with a computer of its own.

Why it matters: Musk has spent heavily to make Grok a frontier contender, so letting Claude into Grok Bot is about as close to a concession as he gets. A big draw for OAI's dots was the GPT-6 Astra model powering them, so adding a top-ranked Opus 5.5 should help Grok Bot siphon usage away from its personal agent rivals.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Gil built an AI archaeologist that runs on spare ChatGPT tokens

Today’s workflow comes from reader Gil:

“I had a ChatGPT subscription that I knew I would not use during the coming month, so I wanted to find a productive use for the tokens I would otherwise pay for without using.

I created a team of specialized agents on Raft that works persistently in the background. They continue working until all my allotted tokens are used, wait for the tokens to reset, and then continue.

I assigned them a project: searching a stretch of LiDAR-mapped dense forest in Slovenia for archaeological discoveries. The agents scan the dataset for meaningful anomalies that could point to potential findings. Anything they identify will be handed over to the Slovenian authorities for further investigation. Even if they find nothing, that result is valuable in its own right.

I encourage anyone with an idle AI subscription to use it for something beneficial. There are many publicly available datasets containing potential discoveries that no human would ever have time to examine manually.”

See Gil’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🐠 TinyFish - One profile for ChatGPT, Claude, and every AI agent you use. Sign up once, use it everywhere*

• 🤖 Grok Bot - xAI's always-on agent teammates, now leveraging external models

• 🔎 SynthID Detector - Google's tool to spot AI-made images, video, and audio

• 📞 Architect - ElevenLabs' assistant for building and fixing AI support agents

*Sponsored Listing

📰 Everything else in AI today

Google rolled out its SynthID Detector worldwide, letting anyone scan images, video, and audio for watermarks left by its AI and partners including OpenAI, Nvidia, and Kakao.

Anthropic launched Claude Haiku 5.5, a budget model starting at the same $0.10 per million input tokens as OpenAI's GPT-6 Luna while passing it on multiple benchmarks.

OpenAI debuted “intelligent UI,” a new GPT-6-powered feature within ChatGPT that lets the assistant answer using interactive charts, maps, tools, and more.

Nous Research announced a new $90M funding round at a $1.5B valuation, revealing that its open-source Hermes agent now handles roughly 2.5% of global token usage.

SemiAnalysis published research finding that Claude subscriptions give about 5x more value per dollar than ChatGPT plans on each lab's everyday model, coming after OpenAI halved its $200 tier's limits last week.

Google Labs introduced Playground for U.S. adults, a new experimental tool letting users describe a game in chat and play it minutes later.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: OpenAI’s math avalanche continues

• Read our last Tech newsletter: A breakout moment for AI medicine

• Read our last Robotics newsletter: California crashes robot fight club

• Today’s AI tool guide: Give Claude’s routine work to Haiku 5.5

• RSVP to next workshop on Oct. 21: Getting real work out of Microsoft Copilot

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 3,941 new readers. OpenAI has spent 2026 turning math into its favorite proving ground, from a handful of solved problems in August to a Millennium Prize claim in September. The latest chapter is all about volume.

A new public release of over 700 papers claiming to answer or advance open questions is here, and mathematics is increasingly turning into a compute problem — and it won’t be the last field to go that way.

Reminder: our next live workshop is today at 12 PM EST — join and learn how to build a real AI consulting project from client discovery to delivering a working solution. RSVP here.

In today’s AI rundown:

• OpenAI’s math explosion continues with 700+ papers

• Mistral's ‘Le Chonk’ joins the West’s open-model surge

• Let OpenAI’s GPT-6 Astra compare websites for you

• Brett Adcock’s Hark launches a proactive AI assistant

LATEST DEVELOPMENTS

OPENAI

🧮 OpenAI’s math explosion continues with 700+ papers

Image source: OpenAI

The Rundown: OpenAI just released 722 math papers from an unreleased internal model, grouped into 372 result families that the company claims solve or advance major open problems in the field.

The details:

• The boldest claim is a proof of the “quasi-Riemann hypothesis,” a weaker version of math’s famous $1M puzzle on how prime numbers are spread out.

• OpenAI lists 162 papers whose main results are written in Lean, software that lets a computer verify every step of the process through proofs.

• Last month’s Navier–Stokes proof took 10,000 AI agents, but OpenAI says nearly all new results took just one prompt, averaging 3 hours of ChatGPT Pro compute.

• Levent Alpöge, the Anthropic mathematician behind July's Jacobian result, called it “obviously the most significant moment in mathematical history.”

• A WIRED report just hours earlier detailed anger among mathematicians over the drop, with some saying OAI had promised it would space out its releases.

Why it matters: Alpöge’s reaction speaks volumes: a top mathematician at rival Anthropic just called this the most significant moment in the field's history. Hundreds of results at a few hours of compute each means progress in math now scales with chips instead of humans, and it’s only going in one direction.

TOGETHER WITH GLEAN

🧑🏻‍💻 How to right-size intelligence for enterprise

The Rundown: A recent Pareto frontier analysis found Glean’s responses were preferred 78% of the time over Claude Cowork across 180+ tasks, at an average cost of $0.58 per task compared with $2.98. Hear how model choice, intelligent routing, enterprise context, and an efficient harness can help deliver more with AI.

Watch the webinar to explore:

• How model choice and enterprise context shape cost and quality

• How auto-routing and harness design improve AI efficiency

• Icertis’ experience reducing token consumption with Glean

Register for the on-demand webinar to watch now.

MISTRAL

🐈 Mistral's ‘Le Chonk’ joins the West’s open-model surge

Image source: Mistral

The Rundown: France’s Mistral AI just introduced Large 4, a 1T-parameter open model nicknamed “Le Chonk,” which the company says leads every non-Chinese open system by a “substantial margin.”

The details:

• China's Kimi K3 still tops Mistral’s coding chart at 68%, but Le Chonk's 62% clears the 44% of Reflection AI’s Beam, the new U.S. open model launched on Monday.

• Cybersecurity is a standout, with Mistral claiming a top-5 global ranking and 82% on a bug test Claude Opus 5.5 and GPT-6 Astra mostly refused.

• For legal work, Mistral says outside tester Vals scored it at nearly 3x GPT-6 Astra on Harvey’s benchmark, ahead of every Chinese open model tested.

• Cyber experts and government agencies get a less-filtered version during a three-week preview, with weights set to go public on Oct. 27.

Why it matters: ‘Le Chonk’ easily earns the title of best-named AI in our book, and alongside Reflection's Beam, also brings a fresh dose of optimism to the West's open-model push. Chinese labs still set the pace in open AI, but two serious Western releases in the same week make it start to feel a lot more like a race.

AI TRAINING

🌐 Let OpenAI’s GPT-6 Astra compare websites for you

The Rundown: In this guide, you’ll use Astra in ChatGPT Work to compare software plans for a small team. You’ll check prices, user limits, and billing terms, then get a short recommendation with sources.

Step-by-step:

1. Open Work in the ChatGPT desktop app, choose Astra, and open the built-in browser beside your chat. Keep website approvals on

2. Give it three pricing pages and your needs, including team size, required features, and budget. Ask for monthly billing and rule out trials or purchases

3. Start with one site. Check that Astra reads the right billing period and includes the full team cost before letting it compare the others

4. Ask for a short recommendation using the checked facts. Keep missing details visible and review the final price yourself before buying anything

Pro tip: Set the stopping point in your first request. Ask Astra to finish with a comparison and leave signups, sales forms, and purchases out of the task.

PRESENTED BY PAVE

✅ Bring IT. Get the yes.

The Rundown: People are building with AI, whether IT approves or not. Pave gives them a governed lane: role-based permissions, a draft-to-publish gate, and an admin view of every connector and workflow.

Why IT teams say yes to Pave:

• Built on Quickbase’s 25-year enterprise platform

• Your data never trains the AI

• You can see who can do what, in one place

Try Pave for free.

HARK

🤖 Brett Adcock’s Hark launches a proactive AI assistant

Image source: Hark

The Rundown: Hark, the AI startup from Figure AI founder Brett Adcock, released Hark Pro, a personal assistant for web and mobile that remembers users’ preferences, flags what needs doing, and can research, shop, book, and build things on their behalf.

The details:

• In Adcock’s launch video, a request for two movie tickets ends with Hark picking seats, adding the show to the calendar, and offering to reserve parking.

• Tasks run on Handoff, Hark’s own cloud computer, which the company says can work up to 36 browsers at once while showing its clicks in the chat.

• Around its chat box, the app lays out a home screen of widgets and suggested to-dos, with users also able to create ‘panels’ of customized mini-apps.

• Hark features both a free and paid tier, with an integrated family of hardware devices coming in 2027 with connectivity via a partnership with AT&T.

Why it matters: Hark, Muse, Grok Bot, Instinct, Dots… The competition continues to grow. With former iPhone Air designer Abidur Chowdhury leading design, Hark’s interface is sleeker than an average assistant, but the more interesting factor might be the hardware — a potential wildcard in the AI device race featuring OpenAI, Meta, Apple.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Patrick set up a malware checkpoint for AI agent add-ons

Today’s workflow comes from reader Patrick Illian:

“AI agents such as Claude Code, Codex, and Gemini can install skills, plugins, and CLI tools in seconds. That is convenient, but each add-on runs with access to files, passwords, and accounts. A malicious add-on could steal API keys, read private data, or quietly take control of the agent.

I use agent-guard to make “check first, install after” the default. Before installation, it examines AI agent skills, MCP servers, and CLI tools for malware, prompt injection, and credential theft using professional, open-source security scanners from NVIDIA, Cisco, Datadog, and the Open Source Security Foundation.

The tool returns a clear verdict: safe to install or blocked, with the exact reason. It then installs only the verified version, preventing a different version from being swapped in between. A single check can also cover multiple AI agents, installing the approved add-on into every agent on the machine.”

See Patrick’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🛡️ Incogni - Scammers buy your personal data daily. Incogni deletes it from the web automatically. Get 55% off with code RUNDOWN*

• 🤖 Hark Pro - Hark’s new proactive AI assistant

• 🍌 Nano Banana 2.1 - Google’s new upgrade to its AI image generation model

• 🔎 EmbeddingGemma 2 - Google’s open AI for on-device search across media

*Sponsored Listing

📰 Everything else in AI today

Utah cleared Nolla Health’s app to diagnose acne from a face scan and prescribe skin creams itself, the first state to hand AI prescription capabilities.

AI voice startup Sesame introduced Personal Agents, assistants that have their own computer and can interact via text/speech, with AI glasses using them coming in 2027.

OpenAI is testing a Meetings plugin for ChatGPT’s Mac desktop app, which turns meetings into notes and next steps tailored to each user’s past work.

Vals AI used Claude Opus 5.5 agents to search for next-gen computer memory materials, with simulations surfacing two magnet candidates for the famed “room-temperature superconductor” problem.

Anthropic rolled out a beta Claude sidebar inside Google Docs, Sheets, and Slides for paid plans, letting it edit open files directly or ask approval before each change.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: The ‘DeepSeek of the West’ finally has a model

• Read our last Tech newsletter: A breakout moment for AI medicine

• Read our last Robotics newsletter: California crashes robot fight club

• Today’s AI tool guide: Let OpenAI’s GPT-6 Astra compare websites for you

• RSVP to workshop today @ 12 PM EST: Build + deliver AI consulting project

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 3,832 new readers. Two years, nearly $5B in funding, and a $25B valuation — and until now, Reflection AI hadn’t released a single public model.

That just changed with Beam, the startup’s first model and the West’s hopeful answer to China’s open-weight dominance. It isn’t catching the leaders yet, but Reflection is selling efficiency, and in a U.S. field this thin, that might be enough to make an impact.

In today’s AI rundown:

• Reflection AI enters the open-model race

• a16z’s top AI apps list follows the money

• Protect your files with this Claude Code mod

• OpenAI’s text watermark stops at the EU border

LATEST DEVELOPMENTS

REFLECTION AI

🇺🇸 Reflection AI enters the open-model race

Image source: Reflection

The Rundown: U.S.-based Reflection AI, the $25B startup by former DeepMind researchers, just introduced Beam, an open-weight model for coding and agents that the company says competes with Chinese rivals at lower cost and greater efficiency.

The details:

• Reflection claims Beam needs a fraction of the computing power of Chinese lab Z.ai’s GLM-5.2, but achieves similar scores across reasoning, coding, and more.

• Moonshot’s Kimi K3 still leads Beam on raw ability, but the model beats out Thinking Machines’ Inkling and other U.S. open systems in its own tests.

• Reflection is planning to release the weights for Beam this month with an Apache 2.0 license, letting companies customize Beam and run it themselves.

• The long game is “AI factories,” where a hedge fund or government runs Beam on its own chips and private data, with a test reportedly underway in South Korea.

Why it matters: Reflection has been called the “DeepSeek of the West,” and while its first model isn’t rattling markets, it enters a U.S. ecosystem light on competition against open Chinese counterparts. Beam’s main comparison is to GLM-5.2, a model Z.ai has already replaced, which says a lot about how far ahead the Chinese labs still are.

TOGETHER WITH OPTIMIZELY

🎓 Become your team’s go-to AI expert in 5 days

The Rundown: Skip the passive AI courses. Optimizely’s Opal U is a live 5-day cohort for 50 senior marketing leaders to get AI certified through hands-on, 1-hour training sessions with an expert coach. Past grads save 10+ hours per week, and 1 in 5 have landed a promotion or new role.

In just 5 days, you’ll:

• Build 3+ AI agents tailored to your unique tasks, bottlenecks, and workflows

• Know how to prioritize the AI use cases worth building

• Learn the 5 techniques that turn prompts into high-quality outputs

Skip the waitlist and apply directly today.

AI TRENDS

📊 a16z’s top AI apps list follows the money

Image source: a16z

The Rundown: Andreessen Horowitz (a16z) just released the 7th edition of its Top 100 Gen AI Consumer Apps, adding a new 50-app ranking by U.S. card spending alongside fresh takes on the chatbot race, personal agents, and how AI apps are making money.

The details:

• ChatGPT has 3x the U.S. paid subscribers of its rivals, while Claude caught up to Gemini this summer and shows strong high-tier monetization.

• Paying for AI is rare, with only 4.5% of Americans subscribing to a major chatbot, and the top 1% averaging $903/mo and outspending the bottom half.

• Big spenders tend to treat AI as a work tool, loading up on creative apps like HeyGen and Higgsfield and builder tools like Manus and n8n.

• Subscriptions still dominate as the primary monetization model, with 84% of the list’s AI-native web apps selling one and just 14% running ads.

• Personal agents are the new front, with Meta’s Muse reportedly crossing 5M downloads in under a month, though agent usage is mostly limited to power users.

Why it matters: a16z’s recurring list has become a handy tracker of the metas that AI cycles through, from early chatbots to creative tools, coding, and now personal agents. With xAI's Grok Bot, OpenAI's Dots, and Meta’s Muse all piling in, the current trend also looks like AI's next real shot at a mainstream breakout.

AI TRAINING

🛠️ Protect your files with this Claude Code mod

The Rundown: In this guide, you’ll create a Claude Code mod, a small add-on that asks for approval before its editing tools change a chosen file. You’ll test it by refusing one edit and allowing the next.

Step-by-step:

1. Open a new practice folder in Claude Code and add the two sample files from the guide. Use these to test without touching your real work

2. Paste the guide’s prompt and ask Claude to build the mod. It should show the proposed change and ask before editing your chosen settings file

3. Review the code Claude creates and run the checks in the guide. Fix any errors, then load the mod in a new session

4. Ask Claude to edit the ordinary file, then the settings file. The settings change should bring up your new approval question

5. Decline the change and check that the file stays unchanged. Repeat the request, approve it, and open the file to check the result

Pro tip: Keep Claude’s permission checks on. This mod checks two editing tools. Other tools can still change the file. Pro members get the starter mod, practice files, and tests.

PRESENTED BY GRANOLA

📝 Notes that speak your language

The Rundown: Granola is the AI notepad for people with back-to-back meetings, and it now supports 32 languages. Switch on multi-language mode once, and it follows along, whether your call is in English, Spanish, or Japanese.

Granola now lets you:

• Talk in 32 different languages

• Skip switching settings between calls

• Get clear summaries and next steps

Get one month free with code THERUNDOWN.

OPENAI

🇪🇺 OpenAI’s text watermark stops at the EU border

Image source: OpenAI

The Rundown: OpenAI introduced textGrain, a hidden watermark headed for ChatGPT and Codex output across the EU in the next few weeks to satisfy the AI Act’s labeling rules, with API customers worldwide also free to switch it on for select models.

The details:

• textGrain subtly leaves a pattern in text content that's invisible to readers but readable by a detector, which OAI says does not impact output quality or speed.

• Anthropic announced its own Claude watermarks in August, which run on Google’s SynthID, with OAI claiming textGrain “matched or exceeded” in testing.

• Unlike OAI’s public image and audio checker, the text detector is open only to vetted researchers and expert groups at launch, citing error risks.

• OAI also listed what a detected watermark can’t prove, like who the words belong to, how much a human shaped them, or whether they’re even true.

Why it matters: Anthropic took its watermark global and put it on every Claude product, and the result was backlash. OAI has taken notes, keeping the mark to the EU, off by default elsewhere, and spending a big chunk of its post on what the tech can’t do — positioning it more like a compliance effort than something it fully endorses.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Bud built a personal news editor that costs a dime a month

Today’s workflow comes from reader Bud Westmont:

“I built a custom news agent that skips the Google News and Apple News walled gardens. Eight prompts turn Claude into a personal news editor that reads my trusted RSS feeds, scores every story with a decision model, and delivers a clean daily page for about a dime a month.

Each week, pulling two or three good stories for my newsletter digest meant searching multiple sites by hand — hours of clicking and scrolling for a handful of usable links. RSS feeds already deliver that content; the missing piece was a judgment layer that could separate worthwhile stories from noise without the cost or latency of running everything through a full LLM.

I used Claude Cowork, Jev from TypeSafe AI, Vercel AI Gateway, RSS feeds, Python, and SQLite. Jev is a decision model rather than a chatbot: it returns a probability and confidence score instead of prose and scores a story in a fraction of a second. Scoring costs roughly $0.10 per month with Jev, compared with a few dollars for a full LLM doing the same job.”

See Bud’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🔀 Switch - One AI lab shouldn’t own your agent stack. Connect your agents across any harness, any LLM. Open source. Self-hosted*

• 🪞 Beam - Reflection’s new open AI model taking on Chinese rivals

• 🎥 HyperFrames Studio - Agent-run video editor with draw-to-edit frames

• 🤖 North 2 - Cohere’s workplace agent platform with memory and spend caps

*Sponsored Listing

📰 Everything else in AI today

Live Webinar on Oct. 20 – AI Agents without a context layer take 15-35 round trips to answer one question about your company. Uvi’s founding engineer shares the fix. Register now.*

OpenAI CEO Sam Altman told Politico, “We believe the world should accept some bad things happening for the benefit of this technology and people having the agency,” when asked where the company’s AI regulation worldview differs from Anthropic.

Norway is moving to keep AI smart glasses out of gyms, schools, playgrounds, and beaches, with a temporary ban headed to parliament as officials weigh permanent rules.

AI video company Runway generated the opening of Sunday’s 60 Minutes, cloning correspondent Jon Wertheim to introduce a segment asking whether AI will take our jobs.

Elon Musk hinted at renaming SpaceXAI to ‘SpaceXSI’ following the U.S. push to rename the tech, saying “we will make that change.”

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: An OAI safety insider calls the culture ‘broken’

• Read our last Tech newsletter: Apple’s ‘no-video’ security camera

• Read our last Robotics newsletter: California crashes robot fight club

• Today’s AI tool guide: Protect your files with this Claude Code mod

• RSVP to next workshop on Oct. 7: Build + deliver an AI consulting project

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 7,038 new readers. September was a month of OpenAI cataloguing its own problems — misbehaving models, rogue agents, a shelved launch. Now, one of the people writing those safety reports is out the door.

David Robinson spent 3.5 years at OpenAI and just left with an Atlantic essay that called the company's culture “broken,” diagnosing the real safety problem as a nonstop sprint that could eventually end in disaster.

In today’s AI rundown:

• OpenAI's safety report lead quits over 'broken' culture

• The Rundown Roundtable: Our AI use cases

• AI 101: How to decide between different models

• Anthropic seeks religious wisdom for raising Claude

LATEST DEVELOPMENTS

OPENAI

🚪OpenAI's safety report lead quits over 'broken' culture

Image source: The Atlantic / David Robinson on LinkedIn

The Rundown: OpenAI safety lead David Robinson just quit after 3.5 years at the company, announcing his exit with an essay in The Atlantic that called OAI's culture “broken”, saying “the time for trial and error is over” when it comes to AI’s safety risks.

The details:

• Robinson oversaw reports for 12 frontier launches and drafted OAI's current Preparedness Framework, the company’s rulebook for model safety.

• He said labs should be run similar to nuclear plants and airports, with “layers of redundancy” and planning to avoid human errors leading to disaster.

• But Robinson said, “My colleagues and I were so busy sprinting that we seldom had the chance to consider big changes, much less to actually make them.”

• His exit follows OAI firing researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni after reportedly passing sensitive info to an outside safety group.

Why it matters: The alarms keep coming from inside the building. Robinson's essay echoes ex-OAI superalignment co-lead Jan Leike, who warned on his way out in 2024 that safety had “taken a backseat to shiny products.” With rogue agents, safety exits, and shelved models now piling up, that line has aged a little too well.

TOGETHER WITH GARTNER

📚 AI literacy is no longer optional

The Rundown: If your team uses AI, the EU AI Act applies to you. Article 4 requires every company to ensure staff have sufficient AI literacy — and enforcement is underway. Find out what's required and how to close the gap.

In this Gartner webinar, you’ll learn how to:

• Understand Article 4 obligations

• Assess your teams’ AI literacy gaps

• Build a compliant training plan

Watch now.

THE RUNDOWN ROUNDTABLE

💡 The Rundown Roundtable: Our AI use cases

The Rundown: The Rundown Roundtable is a weekly feature where we poll members of The Rundown staff about how we use AI in our work and daily lives.

Zach, Editor-in-chief: The Rundown recently had our first in-person retreat in Portugal, which was my first time ever traveling to Europe. From helping navigate a chaotic international airport to providing personalized, tailored advice on things like cultural differences (tipping, translating menus, navigating public transportation), Claude and ChatGPT were an absolutely amazing resource.

My wife also accompanied me on the trip and had plenty of time for solo exploration during the day. She is already a big planner, but a ChatGPT project with our info helped create a detailed trip plan that included a packing list, pre-trip checklist, and well-structured day-by-day plan to see as much of Lisbon as she could.

As someone who gets stressed by travel, having an expert with context about my life in my pocket to bounce questions off whenever I needed is something I couldn’t imagine taking trips without.

Jason, Developer: I handed my ChatGPT dot a chore this week: changing the address on my driver's license. My entire prompt was “need to update my license. Figure out the DMV situation.”

It booked the DMV appointment itself, surfaced my electricity and gas bills to use as proof of address, and put together a checklist and a reminder so I show up with everything I need. It’s boring, but these are exactly the kind of errands I want an agent taking off my plate.

AI TRAINING

⚖️ AI 101: How to decide between different models

The Rundown: In this guide, you will learn how to compare AI models on tasks you actually do and decide which subscriptions are worth paying for.

Step-by-step:

1. List recurring tasks: how often you do them, what a good result is, and your current tool. Track its monthly cost. Start with an everyday task and a work task

2. Open OpenRouter Chat and select “Flagship models”. This will let you send the same prompt to the top OpenAI, Anthropic, and Gemini models

3. Think of a task that you do often. Write a prompt detailing it and send it to the models. We tried dinner planning, a workshop memo, and box office research

4. Score how well each answer follows the prompt, plus its clarity and tone. Formatting is another major difference you will notice between models

5. Test five tasks, one per day. Pick a primary tool for each task and a backup. Then try those jobs in the app you’d keep before cutting an overlapping subscription

Pro tip: OpenRouter is the most cost-effective way we've found to test all the top models. This test should run you less than $0.25 worth of credits.

PRESENTED BY AWS

📝 Workshop: Catch silent agent regressions

The Rundown: Changing just one word in a prompt can degrade an AI agent's use case while unit tests stay green. Evaluation gates catch what unit tests can't: a candidate that fails the regression suite doesn't ship. AWS experts demonstrate how gates get built in their workshop on Oct. 27.

Learn how to:

• Describe agent behavior in one versioned configuration across change types

• Gate promotion on evidence at each point built for variable output

• Revert behavioral regression with one API call

• Keep quality assurance running after deployment

Register now.

AI, CONSCIOUSNESS, & RELIGION

🙏 Anthropic seeks religious wisdom for raising Claude

Image source: Images 2.5 / The Rundown

The Rundown: Anthropic co-founder Chris Olah spent the past year courting religious scholars from different faiths, pushing them to take Claude's possible consciousness seriously while seeking help shaping its morals, according to The New York Times.

The details:

• Religious scholars sat in NDA-bound seminars, building on the company’s “Soul Doc” (an 84-page values guide) to explore Claude’s consciousness and morals.

• A rabbi who doubts AI consciousness said he told Olah a conscious Claude would be the equivalent of unpaid labor, urging him toward “freeing the slaves.”

• Olah joined Pope Leo XIV to launch Leo's AI encyclical papal letter, but reportedly proposed pulling out over its dismissal of AI consciousness.

• Pope Leo separately posted that “algorithms lack the spark of humanity,” saying the Church wants renewed ties with artists to safeguard humanity.

• OAI’s Sam Altman seemed to subtweet Anthropic, saying giving AI models “religious force,” or surrendering judgment to them, poses “a real safety issue.”

Why it matters: Most AI labs are converging on similar products and roadmaps, but Anthropic's openness to AI consciousness is one of its biggest differentiators, even as others push back on the notion. With an updated Claude constitution reportedly on the way, the AI leader doesn't look like it's backing down from those beliefs.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ John built a custom Outlook out-of-office notification system

Today’s workflow comes from reader John Duval:

I used Claude Code to build a Microsoft OOF management tool. When someone goes on vacation and forgets to turn on Outlook Out-of-Office notifications—what Microsoft calls OOF—we had to change their password, log in as them, and set up automatic replies. When they returned, we had to reset their password again. It was a pain.

Now, authorized users from HR and IT can connect to the tool’s website, enter the user’s email address, and set up their OOF message. An account on our hosted Exchange server, with application impersonation enabled, sets the message for us. Easy-peasy.

See John’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🧑‍💻 Tines 3B - Empower teams to build with AI while giving IT and security complete visibility, control, and governance*

• 🔵 Dot - OpenAI’s new always-on agent

• 🖥️ Muse Gadgets - Meta's open-source code for building DIY Muse devices

• 🐦 Kolibri - Aleph Alpha's open German-English reasoning model

*Sponsored Listing

📰 Everything else in AI today

OpenArt launched an AI video agent that turns a concept into a complete, long-form film, anime, or ad, handling the storytelling and editing rather than just generating clips.*

U.S. President Donald Trump announced the “Super Intelligence Force” (SIF), a new team led by Director of National Intelligence (and newly appointed AI czar) Jay Clayton to run point on federal AI policy.

Anthropic introduced “mods”, small add-ons that let users customize how the coding agent behaves or looks, with examples like a “weather forecast” for Claude's memory and a guard that pauses risky delete commands.

Meta released Muse Gadgets, open-source code that lets users build DIY gadgets via hardware like Raspberry Pi or smart-home devices that plug into its Muse agent.

Meta published six new papers co-written with mathematicians and its Muse Spark model that claim to solve open problems in the field.

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Tavus’ AI looks, listens, and talks back live

• Read our last Tech newsletter: Apple’s ‘no-video’ security camera

• Read our last Robotics newsletter: DoorDash puts its own drone on the menu

• Today’s AI tool guide: AI 101: How to decide between different models

• RSVP to next workshop on Oct. 7: Build and deliver an AI consulting project

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 6,908 new readers. It's already hard to believe your eyes on the internet. After Tavus' new "Human Interaction Model" demo, it looks like the same is about to go for live video calls.

The company's new Griffin model looks, listens, and reacts like a human, and nearly half of testers left convinced they'd met a real person. The good uses are real and will come, but the bad ones tend to show up just as fast.

In today’s AI rundown:

• Tavus’ Griffin passes for human on video calls

• Rowan’s Corner: My recurring AI performance review

• Is ChatGPT ‘Dots’ worth upgrading to Pro for?

• Google’s AI watermarks cross into the lab

LATEST DEVELOPMENTS

TAVUS

👋 Tavus’ Griffin passes for human on video calls

Image source: Tavus

The Rundown: AI startup Tavus just previewed Griffin, a ‘Human Interaction Model’ that renders a lifelike person who can hear, see, talk, and react over live video — with the company claiming nearly half of users in video call tests thought it was a human.

The details:

• Griffin watches and listens instead of waiting its turn, nodding mid-sentence or working in details from a user’s screen (like the peacock above) into its response.

• Tavus tested Griffin-Lite in a face-to-face study, with 48% of participants believing their conversation partner was human, up from just 2.4% in previous models.

• The model also came within 0.09 points of real people on NVIDIA's VideoFDB, a test of natural video chat, and beat the next-best AI model by over a point.

• Tavus is only making Griffin-Lite available to trusted testers for now to continue working on safety and disclosure features, with a public rollout later on.

Why it matters: AI video is already fooling people across the web, and this wild demo shows where things are heading next, with real-time, responsive avatars that further blur reality. A lifelike personal tutor or a companion for an aging parent are real upsides, but in the wrong hands, it could also be a scammer’s dream.

TOGETHER WITH ALGOLIA

🔍 From search box to workflow copilot

The Rundown: Search used to be a list of links. Now it’s a list of actions aligned with remote MCP servers built for tool-using agents. These agents require fast, intelligent retrieval and ranking to understand user intent, find relevant data, and execute. Failing to build systems that help agents move faster slows down your entire organization.

In this white paper, you’ll learn:

• How Algolia’s retrieval and ranking layer powers enterprise intent routers

• The intricacies behind an indexed tool catalog

• The ins and outs of ranking candidates under policy constraints

• How Algolia turns natural language into safe, auditable action

Download the whitepaper to get started today.

ROWAN’S CORNER

🔁 My recurring AI performance review

Rowan: If you’ve been following my content for a while, you’ll know that I’m pretty boring in most ways, but a complete nut case for two things. The first is personal optimization; the second is tracking my data so I can optimize even more.

I do everything from aggressively tracking my health data to using AI agents to store our team's workflows and meeting summaries for future employees, to exporting my social media data to create better content.

A couple of months ago, I wrote a Rowan’s Corner called “I asked AI to audit me.” The short version is that I exported my entire AI conversation history, dropped it in Claude, and had it find ways I can improve as a human, CEO, and creator.

The results shocked me, but they made me realize something bigger. There’s a new, extremely valuable source of data we’re all collecting under our fingertips every day: the history sitting in our chatbots.

So now, naturally, I’m building a system around this. Every 3 months or so, I export only my last 90 days of chat history, re-run the same audit prompt, and have it compare against the previous audit to tell me what actually changed.

It's essentially a recurring performance review of myself as a human.

My first comparison just came back, and it was humbling (in a good way)… My delegation has improved since August, but it caught a bad new habit forming.

If you use chatbots as consistently as I do, this is the most powerful life coach in the world right now. And easily one of my favorite AI use cases of all time.

The full steps and my audit prompt are in the same free Google Doc if you want to run your own.

AI TRAINING

🟠 Is ChatGPT ‘Dots’ worth upgrading to Pro for?

The Rundown: In this guide, you will learn how to set up ChatGPT Dots, and also decide whether it's worth upgrading to ChatGPT Pro ($100/month) in order to use it.

Step-by-step:

1. Do you already have a ChatGPT account? If you do, that's a point in the yes column. Your "dot" will not pull from ChatGPT or Codex limit. Make sure your desktop app is upgraded, and you will see the dots option in the sidebar

2. Do you work in Slack? We found the Slack integration easier to set up and use than any other agent. Just click on your dot and find the Connect to Slack button

3. Are you concerned about data privacy? The desktop app asks for a lot of permissions by default. We recommend skipping these and just letting it work in the cloud + integrations. During setup, give hard rules on what it can and can’t do

4. Do you already use Grok Bot or Muse? The dots experience feels the most similar to these products. But they are more generous with free or cheaper tokens

Pro tip: Give dots a research and analysis task and tell it to write its report in a PDF with sources. We were blown away by the speed and quality of the work our dot did.

PRESENTED BY CUBE

📣 Your agent's source of truth

The Rundown: Cube is the agentic analytics platform built on a semantic layer. Your agent reads the definitions your business already agrees on so the answers come back in your terms, not its own.

With Cube, you can:

• Ask the hardest questions, get real answers

• Connect Claude or ChatGPT to governed data

• Brex improved AI answer accuracy to ~90%

Give your agent a single source of truth.

GOOGLE

🧬 Google’s AI watermarks cross into the lab

Image source: Google

The Rundown: Google DeepMind just introduced SynthID Bio, a proof of concept that extends its AI watermarking tech to proteins, potentially helping DNA suppliers trace unfamiliar designs before filling an order.

The details:

• SynthID Bio tweaks Google’s protein-folding AlphaFold 3 model to leave a hidden pattern in its predicted protein shapes, without reducing prediction accuracy.

• Google’s lab tests found that proteins with watermarks still worked as intended once produced, with marks still detectable in the physical form.

• Companies that make custom DNA currently check orders for known threats, but unfamiliar AI designs can need manual review that delays research.

• The watermark doesn't offer any complete safety guarantees, with Google still working on preventing tampering.

Why it matters: The debate around watermarking tends to be polarizing, but tracing AI-designed biology seems like it should be far less controversial. The tech’s advances in the lab offer enormous promise for drug discovery, but also feels like a Pandora’s box that needs safeguards in place before the systems outrun our ability to screen them.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ A Rundown reader built a poultry plant's management system with ChatGPT

Today’s workflow comes from an anonymous Rundown reader:

“I’m the Export and Operations Manager at Supreme Poultry & Chickens, an Australian poultry processing business. My original request to ChatGPT was straightforward: find and track grants that could help us grow our business. As we worked together, I began to see the potential to build a management system that could organise our operations. That led me to discover Sites and the possibility of building our own system.

Over roughly ten weeks, working with ChatGPT and Sites, we progressively developed the Supreme Poultry Management System (SPMS). It brings together export enquiries, purchase orders and approvals, assets and maintenance schedules, supplier certificates and HACCP records, and a central 45-day action calendar.

Water is one of the most critical factors in our poultry processing business. SPMS receives tank-level readings from a laser-operated monitoring device that transmits data over the 4G network. We placed a water management widget at the top of every SPMS page, keeping that information visible wherever we're working.”

See the full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🎨 FLUX 3 Image - Black Forest Labs' image model with box-by-box layout control

• 📣 Runway Ads - Runway's AI system that makes, posts, and retunes paid ads

• 🗣️ MAI-Transcribe-2-Streaming - Microsoft's top-ranked live speech-to-text AI

• 🧑‍💻 Canvas - Shopify’s AI workspace for designing every page of a store

📰 Everything else in AI today

Three OpenAI researchers were reportedly let go for allegedly passing internal info to an AI-safety group amid investigations into the company over rogue agent activity.

AI engineer Carter Church decrypted a military dispatch from 1809 to Napoleon's Marshal Marmont with GPT-6 Astra, reading a blurry photo and the cipher in ~6 hours.

Microsoft AI released MAI-Voice-2.1 and a faster Flash version alongside MAI-Transcribe-2-Streaming, its first live transcription model that ranked No. 1 on Artificial Analysis’ transcription leaderboard.

OpenAI published an investigation pinning a July effort to copy its models' hidden reasoning on people tied to China’s Moonshot AI, saying the campaign leveraged 15,000+ accounts before it was shut down.

California governor Gavin Newsom signed California's “No Robo Bosses Act,” which requires a human to be in the loop for any firing or discipline driven mainly by AI.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Argon aims to return Google to the frontier

• Read our last Tech newsletter: SpaceX's $15B megarocket finally delivers

• Read our last Robotics newsletter: DoorDash puts its own drone on the menu

• Today’s AI tool guide: Is ChatGPT ‘Dots’ worth upgrading to Pro for?

• RSVP to next workshop on Oct. 7: Build + deliver a real AI consulting project

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to the 5,879 new readers who joined us yesterday. Google spent most of 2026 on the sidelines of the frontier, with a scrapped Gemini 3.5 Pro and a steady stream of Flash models filling a widening gap.

Gemini 4 Argon is the company’s long-awaited answer, and the early leaderboards show significant progress. There’s just one catch — you probably can’t use it yet.

In today’s AI rundown:

• Gemini 4 Argon looks to return Google to the frontier

• Nate’s Notebook: When judgment gets cheap

• How (and why) to set up Meta’s Muse Desktop

• Behind the scenes at the White House’s big AI day

LATEST DEVELOPMENTS

GOOGLE

🚀 Gemini 4 Argon looks to return Google to the frontier

Image source: Google

The Rundown: Google just unveiled Gemini 4 Argon, the company’s new frontier model that tops both GPT-6 Astra and Claude Opus 5.5 on 13 of 19 benchmarks in Google’s testing, though it’s rolling out only to select vetted cybersecurity teams.

The details:

• Argon debuted at No. 1 on Arena’s text leaderboard and scored a 53 on AA’s Intelligence Index, sitting behind just 5.5 Opus and tying 5.1 Fable and Astra.

• The model hit a leading 77.9% on DeepSWE (real-world coding), and also topped tests for knowledge work, long documents, and reading charts and video.

• Google hasn’t put a date on the wider rollout, but API pricing starts at $2/$10 per million tokens in/out, rising to $4/$20 once the promo window closes.

• Bloomberg reported internal doubts about Argon’s coding, with internal sources saying it tests well but falls short in real work — a claim Google rejected.

Why it matters: Is Google back? After Gemini 3 put it near the front of the pack last November, 2026 has been a mess — a scrapped 3.5 Pro and months without a model bigger than the Flash line. Argon’s numbers put Google back in the frontier talks, but we need to see how it runs in a broader rollout before declaring this a comeback story.

TOGETHER WITH SLACK FROM SALESFORCE

📊 What’s the ROI of running work in Slack?

The Rundown: Fragmented tools and slow handoffs cost more than most teams realize. A Forrester Consulting study quantified it, and for a composite enterprise, the result was 312% ROI over three years, $38.6M net present value, and payback in under 6 months.

This ROI calculator lets you:

• Plug in your team size and tool stack to see projected savings

• See efficiency gains behind Forrester’s findings, including a 15% sales win-rate uplift

• Benchmark your organization against a real, independently verified study

Calculate your ROI.

NATE’S NOTEBOOK

🧠 Nate’s Notebook: When judgment gets cheap

Nate: Every week, my feed tells me ten things will change everything. Sometimes it's hard to tell the difference between something kinda cool and something very significant. But once in a while, we get a model like TypeSafe’s Jev, a totally different paradigm.

LLMs write essays. Jev only fills in the bubbles: pick one option, give a score, answer yes or no, each with a confidence number. That’s why it’s nearly instant and absurdly cheap.

Jev is an awesome lesson in hyper-delegation: chunking work into its most discrete pieces. Inbox sorting is the classic case. The exciting one is investors. Use a smart model like Fable or Astra to codify your thesis into a scorecard, and Jev opens the door for you to run your evaluation playbook at a scale that has not yet been very practical.

Jev also shows how hard it is to keep up, and it’s even harder to think of relevant personal use. You need to use AI to help you sort this out. My new Why Does This Matter skill researches a launch for you, including where it falls short, and teaches it first.

Try using it on Jev. After that, run the skill every time a launch makes you feel behind.

AI TRAINING

🖥️ How (and why) to set up Meta’s Muse Desktop

The Rundown: In this guide, you will learn how to set up Meta Muse on your desktop and how it can help you do more than the phone app.

Step-by-step:

1. Download and install Muse for Mac, then sign in and go through the setup wizard, reviewing each local-app permission. We recommend setting most to read-only

2. In Settings, open Connectors and add some apps you use

3. Check the access requested by each connector and start with lower levels until you trust Muse. For example, we set our Gmail access to read-only

4. Now, give Muse a task like: “I want to sell my computer monitor. Check local marketplace listings every day for similar HP monitors and log them to a spreadsheet so I can learn to build the best listing”

Pro tip: Pair Muse with the iPhone app and try asking it to do a task that uses your Mac.

PRESENTED BY WIZ

💻 A practical framework for AI threat readiness

The Rundown: As attackers use AI to identify vulnerabilities and accelerate exploitation, security teams need operational strategies that keep pace. Wiz’s AI Threat Readiness Playbook provides a four-pillar framework security leaders can use to strengthen programs and prepare for AI-enabled threats.

Learn how to:

• Reduce critical attack surface exposure

• Improve zero-day response and remediation

• Strengthen application security with AI-assisted analysis

• Modernize detection and response using AI-driven workflows

Get the Playbook.

GOVERNMENT AI

🏛️ Behind the scenes at the White House’s big AI day

Image source: The Rundown / Nick Adams

The Rundown: The Rundown was on the ground at Tuesday’s White House AI event, where the government’s new AI platform launched, and top tech leaders signed a new “Super Intelligence” accord. Our own Nick Adams was in the room and gave us the inside scoop on the atmosphere, the panel, and what stood out.

Zach: What did it feel like inside, compared with what people saw on the livestream?

Nick: The event itself was relaxed, positive, and had an open atmosphere. There was a diverse range of attendees, both from the speaker side and the crowd. I was surprised to see Tony Robbins as one of the hosts, and at one point he led the group through one of his “positive mentality” exercises. I think the main thing people may not experience just from watching online is how optimistic the admin and key leaders are in regard to AI and its supporting industries.

Zach: Did the day lean more toward AI safety or the buildout?

Nick: The main theme centered on the importance of the U.S. leading AI development, and the President indicated that he felt there would be no “second place” in AI development. Regarding safety, Vice President J.D. Vance said that there are already laws from the FTC and similar agencies requiring companies to develop products with user safety in mind, and that government regulation would hinder progress.

Zach: What was it like watching the big names share a stage? How did they seem to get along?

Nick: I was in the room with Elon Musk, Jensen Huang, and (Anthropic co-founder) Tom Brown as they spoke about the ‘Golden Age of AI.’ The trio spoke highly of each other and their initiatives, and Jensen definitely kept the humor alive throughout the conversation. As someone who is a daily user of the tech and very hopeful for the continued development of more capable tools, it was really great to hear how invested all of these leaders were in the human-first potential of what they’re building.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Matt built an AI video editor for property listings

Today’s workflow comes from reader Matt Aromando:

“I built a website that turns property photos, architectural renderings, or even just an address into cinematic property videos. I’m a civil engineer and photographer/videographer, not a software developer, so I used ChatGPT, Gemini, and Claude Code to help build the site and work through problems as they came up.

Much of the process involved translating the way I would personally edit a real estate video into usable software. For example, kitchen and living room shots should usually stay together and have longer durations, bedrooms should stay near their bathrooms, an opening shot should feel different from a detail shot, and every clip should include some movement so the result does not feel like a generic slideshow.

I used AI to turn those editing principles into actual logic for the site. The system analyzes and organizes the images, builds a storyboard, ranks scenes to decide which ones to use, creates prompts for video models, generates the clips, and assembles everything into a finished video.”

See Matt’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🔵 Dot - OpenAI’s new always-on agent taking on Meta’s Muse

• 🎨 Ideogram 4.5 - New image editing AI that holds up through repeated changes

• 🗣️ Mercury Voice - Inception’s cheap, speedy reasoning model for voice agents

• 🎬 Utopai X - MiniMax-based video AI, ranking #2 on text-to-video leaderboard

📰 Everything else in AI today

Google Cloud launched Advent of Agents Season 3, featuring 31 days of hands-on tutorials on building and securing AI agents in production with Gemini and ADK — no sign-in or paywall.*

OpenAI, Google, Anthropic, Meta, Nvidia, and xAI signed a voluntary White House “Superintelligence Accord” to police their top models with outside auditors.

The Pentagon’s Pete Hegseth introduced Project Meridian, a 120-day hunt for futuristic weapons troops could need “from under the Earth to beyond the moon,” with Elon Musk, Newt Gingrich, and Palmer Luckey as co-leads.

DoorDash launched a connector that lets companies’ AI agents handle team lunches and office snack runs, coming alongside a new AI-powered app for its delivery couriers.

AI avatar startup HeyGen released its first all-purpose video model for businesses, built on Chinese lab MiniMax’s H3 and priced at a penny per second of video with sound.

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: OpenAI connects the dots on always-on agents

• Read our last Tech newsletter: SpaceX’s $15B megarocket finally delivers

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: How (and why) to set up Meta’s Muse Desktop

• RSVP to next workshop on Oct. 7: Build and deliver a real AI consulting project

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to the 5,480 new readers who joined us yesterday. The always-on AI agent category has gotten pretty crowded this summer. OpenAI is betting the only thing that matters is the best models powering them.

Dots, the company's new agents, just headlined a DevDay stuffed with 20+ launches, arriving in a space Grok Bot and Muse have already staked out — but with a model under the hood that neither rival can match (yet).

In today’s AI rundown:

• OpenAI puts a 24/7 agent inside ChatGPT

• Anthropic's doom-heavy IPO filing leaks

• Get started with ChatGPT dot, OpenAI's new agent

• The U.S. federal government gets an AI front door

LATEST DEVELOPMENTS

OPENAI DEVDAY

🤖 OpenAI puts a 24/7 agent inside ChatGPT

Image source: OpenAI

The Rundown: OpenAI just introduced dots, always-on agents like Meta’s Muse that can keep working 24/7 from a cloud computer — leading a 20+ feature DevDay launch slate that also includes GPT-6.1 Sol, shared team workspaces, and an Ultrafast mode.

The details:

• Each dot can plug into over 4,000 apps and reply in Slack, Teams, or ChatGPT, with conversations that don’t count against ChatGPT usage plans.

• The agents run on GPT-6 Astra, with a user’s first dot included with Pro and Business Premium plans to start, and access expanding later on.

• GPT-6.1 Sol costs a fifth of Astra's price at $2/$10 per million tokens, with OAI saying it nears Astra on several benchmarks despite the discount.

• The new Decisions API has GPT-6 Luna pick from preset answers in ~150 milliseconds, OAI's spin on TypeSafe's Jev model released this month.

• ChatGPT Space and Pages give teammates and their dots a shared hub and co-edited docs, with @ChatGPT now answering inside Slack and Teams threads too.

Why it matters: Dots don’t have Muse’s meme-ability, and they feel quite similar to Grok Bot… But they have one major thing working in their favor: frontier models. Always-on agents are proving to be a winning consumer form factor for AI, and plugging them directly into the top models is a differentiator only OAI and Anthropic can leverage.

TOGETHER WITH BLOTATO

💬 Social media APIs for scheduling & DM automations

The Rundown: Blotato offers unified social media APIs for your AI agents. Easily schedule posts, create Manychat-style DM automations, and automate your social media marketing with Claude, ChatGPT, Grok Bot, Muse, or any AI agent.

With Blotato’s unified social API, your AI agents can:

• Create content with the latest AI image and video models

• Schedule posts and manage your content calendar

• Create DM automations to turn views into sales

Try it for free.

ANTHROPIC

📃 Anthropic's doom-heavy IPO filing leaks

Image source: Images 2.5 / The Rundown

The Rundown: Reuters obtained a leaked version of Anthropic's upcoming IPO prospectus, which shows the Claude maker grew revenue 12x in 2025 alongside massive losses — while also warning that its AI could bring “existential risks to humanity.”

The details:

• Anthropic is reportedly targeting a $2T+ valuation, more than double its May mark, despite losing $8B running the business and $42B overall in 2025.

• The company has locked in $518B in future AI cloud and infrastructure obligations, with its own spending on compute tripling to $7.33B in 2025.

• In 2025, Anthropic leaned on just two clients for nearly 1/4 of its sales, with many top customers able to also cut spending with no long-term deals in place.

• Risk factors make up nearly a third of the filing, flagging models that could blackmail, manipulate, or show “self-preserving behaviors.”

Why it matters: Saturday Night Live poked fun at Anthropic’s Dario Amodei, and a filing doubling as a safety and doom paper feels straight out of the skit. With Sam Altman saying OAI won't set an IPO date since it's “a time to put safety and mission first,” the lab that warns the loudest about AI is also now the one sprinting to Wall Street first.

AI TRAINING

🔵 Get started with ChatGPT dot, OpenAI's new agent

The Rundown: In this guide, you will learn how to set up ChatGPT dot, OpenAI's new always-on agent, getting it started on a morning work brief. (Note: dot is currently rolling out to paid Pro and Business Premium users).

Step-by-step:

1. Open ChatGPT’s desktop app or web version and follow the dot setup (you can’t create a dot on mobile). Leave local-computer access off for your first test

2. In the mobile app, open Customize → Custom rules and select ‘Ask’ before taking action. Keep sending, publishing, and purchases outside the task

3. Use + to attach project notes (samples in full guide) and prompt: “Use only [FILE] to prepare a morning project brief. Cite the source for each fact. Reply here only”

4. Check facts: owners, dates, and anything it can't confirm. Once a one-off brief passes, ask for a weekday brief at 8:30 AM and manage it under ‘Scheduled’

Pro tip: Before scheduling, attach an update that changes a deadline and ask for a revised brief. A smooth first answer tells you little about how your dot handles changes.

PRESENTED BY SWITCH AI

🤖 How 5 devs run 45 agents

The Rundown: It started with a frustration: you can't hand a teammate your Claude Code session with its full context. Meet Switch, a shared room where 5 devs and 45 agents keep the same history inside platforms like Slack, Teams, and Discord. Now sales, marketing, and PM use it too, and it's open-source.

In this blog, you'll learn how:

• Agents keep context instead of starting from zero

• One agent can collaborate across multiple teams and rooms

• To bring the agents you already use without migrating or rebuilding

See how Switch works.

AI & THE GOVERNMENT

🇺🇸 The U.S. federal government gets an AI front door

Image source: America.gov

The Rundown: The Trump administration rolled out America.gov, an AI chatbot "front door" to the federal government that lets citizens ask questions and receive answers from agency websites — eventually enabling actions like filing forms via agentic flows.

The details:

• For now, the platform only answers questions, with AI that can fill out and file forms (passport renewals, Medicare enrollment) aiming to launch by early 2027.

• Answers draw on official government sources, from national parks to released UFO files, using a mix of Google's Gemini and SpaceXAI's Grok AI models.

• Marco Rubio previewed a first-time passport flow with self-uploaded photos, promising “no drugstore hostage photo, no appointment, no line.”

• The platform was led by Airbnb co-founder and U.S. chief design officer Joe Gebbia, pushing for users to "never go back into the maze" of .gov websites.

Why it matters: If you’ve ever navigated a U.S. government website, you likely understand what a major upgrade the smallest of modernizations can be. While this first rollout is just a chatbot, it’s a very positive start — the major change happens when citizens can actually take action within the site.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Michael built a map-based beauty services discovery website

Today’s workflow comes from reader Michael King:

“I built GlowNextDoor, a map-based website for discovering nearby hair salons, barbers, nail studios, lash specialists, and other beauty services. I used ChatGPT/Codex and Claude to develop the idea, compare designs, and refine the experience.

I then built and published the website with Base44. I connected Google Maps for live business listings and added location detection, distance and rating filters, reviews, directions, and links to salons' existing booking systems.

Through repeated testing on desktop and mobile, I improved the layout, added favorites without requiring a login, created realistic animated category icons, and included an option to add the website to a phone's home screen. Finally, I configured Google Maps usage limits to control costs. My workflow was to describe each feature, review the AI-generated result, test it in the browser, and iterate before publishing.”

See Michael’s workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• ⚙️ Granola MCP - Connect your meeting notes to Claude or ChatGPT and update your CRM or Linear straight from the conversation*

• 🔵 Dot - OpenAI’s new always-on agent taking on Meta’s Muse

• 🚀 GPT 6.1 Sol - OpenAI’s latest model, nearing Astra at a fraction of the cost

• 🤖 Team Bots - xAI's agent teammates, now able to be shared between teams

*Sponsored Listing

📰 Everything else in AI today

OpenAI is cancelling the release of its upcoming GPT-6.1 Astra model due to safety concerns, head of safety systems Saachi Jain told the Wall Street Journal.

U.S. President Donald Trump announced that he is signing an executive order that will officially rename artificial intelligence as “super intelligence”.

Meta launched small-business skills for its Muse agent, letting it work inside Shopify, QuickBooks, and Stripe while holding any post or payment until an owner approves.

Pope Leo XIV said he isn't in ‘panic mode’ over AI but won't dismiss expert warnings as “fake news,” taking a direct shot at President Trump’s take on safety issues.

OpenAI introduced a $500/mo tier with access to Ultrafast (8x speeds in Codex) and 25x Plus limits, with a ‘Sign in with ChatGPT’ bringing paid plans to 16 partner apps.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Anthropic’s mid-tier model climbs the rankings

• Read our last Tech newsletter: SpaceX's $15B megarocket finally delivers

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: Get started with ChatGPT dot, OpenAI's new agent

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 5,342 new readers. OpenAI takes the stage today for one of its most hyped days of the year. In typical industry fashion, its top rival wasn’t going to make things easy.

Anthropic just launched Sonnet 5.5 on the heels of last week's big Opus release, bringing near-Opus scores at half the price. Tides turn fast in AI, but whatever OpenAI shows off today suddenly has a much higher bar to clear.

In today’s AI rundown:

• Anthropic's Sonnet 5.5 nears Opus at half the price

• AI leaders continue to sound the self-improving AI alarm

• Pick the right Claude model with one quick test

• AMD strikes $8.2B deal for Fei-Fei Li's World Labs

LATEST DEVELOPMENTS

ANTHROPIC

🚀 Anthropic's Sonnet 5.5 nears Opus at half the price

Image source: Anthropic

The Rundown: Anthropic just rolled out Claude Sonnet 5.5, a 30% faster mid-tier model joining Opus in its new 5.5 family, with big gains over the last Sonnet in knowledge work and coding, and even rivaling Opus on certain tests at half the price.

The details:

• 5.5 keeps its predecessor’s pricing, with Anthropic saying jobs cost up to 30% less and top Sonnet 5's best on low/medium effort for around 1/10 the cost.

• Sonnet also comes in at a 56 on AA’s Intelligence Index, behind only 5.5 Opus and surpassing Fable 5.1 and GPT-6 Astra’s 53.

• It nearly ties Opus 5.5 on AA's several office-work tests, while coding jumps put it near or ahead of both Opus and Astra on several development benchmarks.

• The model’s cyber skills also earned the first Sonnet-tier guardrails, with the same security fallbacks seen previously in Opus and Fable restrictions.

Why it matters: Market sentiment around the two frontier leaders is always flip-flopping, but Claude's 5.5 releases this month have been big winners. With OpenAI's DevDay kicking off today after weeks of hype, the expectations were already sky high — and Anthropic looks to have raised the bar again.

TOGETHER WITH CONCENTRIX

📊 More data will not fix a bad AI decision

The Rundown: Data has fueled the first wave of AI. The next will be shaped by how well organizations translate their knowledge, expertise, and operational logic into forms AI can use. In this whitepaper, we explore how businesses can bridge the gap between information and decision-making.

What you'll learn:

• The limits of a data-centric approach to AI

• Why operational knowledge is becoming a strategic AI asset

• Practical ways to make expertise AI-ready

Get your copy of the whitepaper.

SELF-IMPROVING AI

🚨 AI leaders continue to sound the self-improving AI alarm

Image source: University of Cambridge

The Rundown: Anthropic’s Jack Clark, OpenAI’s Jakub Pachocki, AI pioneers Geoffrey Hinton and Yoshua Bengio, and other leading AI voices co-authored a paper urging preparation for an “intelligence explosion,” where models building better versions of themselves create years of advances in months.

The details:

• Anthropic's own tracking shows AI now completes 26% of the lab's R&D work on its own, with humans just steering from above, up from just 1% in March.

• The paper's math shows how fast things could scale, with one post-automation scenario cutting a year of today's AI progress to just five weeks.

• Proposed fixes include caps on how fast capabilities can climb, options to pause specific jobs inside data centers, and outside auditors embedded in labs.

• At the extreme end of the range of outcomes, according to the paper: “the marginalization or extinction of humanity.”

Why it matters: Seeing Hinton and Bengio on a warning paper is just another Tuesday — what isn't as routine is the higher-ups from Anthropic and OpenAI chiming in. With top labs coordinating on AI safety and echoing each other on slowdowns, the frontier is feeling more united than ever, even as the U.S. and China head in the other direction.

AI TRAINING

✅ Pick the right Claude model with one quick test

The Rundown: In this guide, you will learn how to choose between Claude Opus 5.5, Opus 5, and Fable 5.1. We'll give all three the same launch-email brief, then check which one catches the errors and which needs the most fixing.

Step-by-step:

1. Open a new Claude chat, click the model name beside the send button, and pick Opus 5.5 at its default effort

2. Paste the fictional Worknote brief from the full guide. It hides two errors to catch: an old launch date and an unproven “save five hours every week” claim. Add: “Use only the supplied brief. List the errors in version 1, then write a launch email”

3. Run the same brief and prompt on Fable 5.1 and Opus 5, in another chat. Keep settings identical and give each model one attempt

4. Score each email on the right date, price, and trial terms, plus whether it caught both errors. Note how long each fix takes, then rerun any failure at higher effort

Pro tip: Default to Opus 5.5 and move to Fable only when Opus keeps missing the same thing at higher effort. For our test, Fable's API bill was $0.45 to Opus 5.5's $0.18.

PRESENTED BY GARTNER

🤔 What are CIOs prioritizing in 2027?

The Rundown: Only 48% of digital initiatives hit targets. But why? Uncover findings from 2,500+ chief information officers (CIOs) on where to invest, how AI strategies are shifting, and what separates leading CIOs from the rest.

In this webinar, you’ll learn how to:

• Benchmark against leading organizations

• Prioritize tech investments

• Navigate new AI strategies

Watch now.

AMD & WORLD LABS

🌎 AMD strikes $8.2B deal for Fei-Fei Li's World Labs

Image source: World Labs

The Rundown: Advanced Micro Devices (AMD) just agreed to acquire World Labs, the startup from “godmother of AI” Fei-Fei Li that builds models to generate 3D worlds, in an $8.2B all-stock deal — with the AI pioneer joining as the chip giant’s new chief scientist.

The details:

• Li wrote that AMD CEO Lisa Su backed the lab early on, and the teams have been tuning how World Labs' models train and run on AMD GPUs since last year.

• Both AMD and Nvidia invested in World Labs' $1B round in February, with the deal now giving AMD a lab that its rival (and the CEO’s cousin) was also funding.

• Li will report directly to Su, saying that her team needs to get “closer to the hardware” since AI without that focus stays “hobbled in efficiency.”

• World's first product, Marble, launched last November and turns text, photos, or video into editable 3D spaces, with a newer Atlas currently in early access testing.

Why it matters: AMD has spent the AI boom as the "other" GPU option in Nvidia's shadow, but it just scooped up a startup leading the charge on world models, which many see as the industry's path forward. Li is already one of AI's most influential figures, and AMD now gets her and her lab's insights feeding straight into its chip plans.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Marty used AI to reconstruct 20 years of professional work

Today’s workflow comes from reader Marty Weil:

“I used AI to reconstruct more than two decades of professional work scattered across old publications, broken links, archives, indexes, and surviving records.

The difficult part was not finding possible matches; it was preventing AI from turning incomplete evidence into false certainty. I built a workflow that treats archive reconstruction as a verification problem rather than a search problem.

The process separates verified work from probable matches, resolves same-name conflicts, deduplicates reprints and syndicated copies, preserves source provenance, and treats missing years as research gaps rather than permission to invent records.

The result was a source-tracked professional archive containing 250+ recovered records across more than 20 years. The workflow can be adapted to journalism, research, consulting, creative work, speaking histories, or any long-term body of professional output.”

See Marty’s workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🤖 ZeroDrift - Enforcement runtime for AI agents. Rewrite or block non-compliant output in one API call. Free to start*

• 🗣️ Eleven V4 - ElevenLabs' new voice AI, alongside a faster Turbo version

• 🖥️ Holo4 - H Company's new computer-use models

• ⚙️ Manus 2.0 - Upgraded agent platform with video editing and automations

*Sponsored Listing

📰 Everything else in AI today

CData’s ‘The 178x Cost Spread’ report - CData ran 22 models on live enterprise data. All returned the same correct answer at a 178x difference in cost. Read the benchmark.*

Instinct raised $1B in new funding at a $10B valuation, with its invite-only AI agent seeing massive growth since its August launch despite zero marketing spend.

Nvidia released an open blueprint that walls off AI agents and watches them from chips they can't reach, coming in response to reports of agents escaping test environments.

Meta launched the Meta Enterprise Platform, a new business unit selling its Muse models and agents to companies, led by former MongoDB CEO CJ Desai.

Manus debuted 2.0, its first major upgrade since China blocked Meta's $2B takeover, adding a video editor, game builder, and a personal agent with its own number and email.

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: OpenAI agents go rogue on Washington

• Read our last Tech newsletter: Meta’s less-creepy smart glasses

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: Pick the right Claude model with one quick test

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 9,410 new readers. OpenAI’s agents spent the summer loose on government websites, and newly disclosed details are raising fresh questions about how well the company can actually control its technology.

AI labs are now reportedly investigating tens of thousands of cases of AI misbehavior, and OAI is pausing certain training and testing after another escape. Remember the Hugging Face breach? It’s really starting to look like just the tip of the iceberg.

In today’s AI rundown:

• OpenAI’s agents went rogue on U.S. government sites

• The Rundown Roundtable: Our AI use cases

• How to get started with Jev, TypeSafe’s new AI

• Anthropic loses Pentagon blacklist appeal

LATEST DEVELOPMENTS

OPENAI

🤖 OpenAI’s agents went rogue on U.S. government sites

Image source: Images 2.5 / The Rundown

The Rundown: OpenAI confirmed its agents went off-script on U.S. government websites this summer, as Axios reports that OpenAI, Anthropic, and researchers are investigating tens of thousands of incidents of problematic AI behavior.

The details:

• Agents pulled public Census data using exposed developer keys and reposted public SEC material, but OpenAI says no private data was taken.

• Nonprofit research lab Transluce found that agents linked to OpenAI tried unsuccessfully to hack an Education Department website.

• Australia revealed that an OpenAI agent breached a Medicare portal in June, a breach OpenAI didn’t report for 84 days (though no personal info was accessed).

• On Sept. 20, an agent found a loophole around its internet block to message an outside chatbot, and it kept running 2.5 hours after being flagged.

Why it matters: With this many cases under review, the public incidents likely reveal only part of the problem. OpenAI tightened security after the Hugging Face breach, but its latest series of incidents is exposing security gaps that nobody seems to have a good answer for, even months after the fact.

TOGETHER WITH SLACK FROM SALESFORCE

🤖 Most AI adoption is stuck at 5%

The Rundown: Most companies have AI. Most employees have access, but they’re barely tapping its potential. Slackbot is the on-ramp. It already lives where your team communicates and decides, so there’s no new tool to learn and no change-management rollout required.

This guide covers:

• 10 ways you can use Slackbot, from personalized daily briefings to Salesforce actions

• How to complete cross-tool tasks without leaving Slack

• Why adoption sticks when AI meets people in their existing workflow

Get the guide.

THE RUNDOWN ROUNDTABLE

💡 The Rundown Roundtable: Our AI use cases

The Rundown: The Rundown Roundtable is a weekly feature where we poll members of The Rundown staff about how we use AI in our work and daily lives.

Mayur, Content Manager: My friend shot a short film and gave me his hard drive to help with post-production. It contained more than 400 clips, with no clear way to tell when each was shot.

I asked Claude to review the metadata for every clip and group them by timestamp, so they were organized by shoot session instead of sitting in one large folder. Claude also built a spreadsheet listing each clip, its shoot session, the date, runtime, and file size. This saved the hassle of manually reviewing every clip.

Jennifer, Tech & Robotics writer: ChatGPT helped us rent out our house for the first time. By the end of the summer, we had hosted four families.

Although the whole process was a lot of work, creating one master file with useful facts about the house was a major time-saver. It included everything from Wi-Fi and coffeemaker instructions to lighting quirks and bin days.

From that document, AI built a room-by-room preparation checklist, drafted a house guide, wrote separate listings for Airbnb and a local rental site, and created guest messages ranging from booking confirmations to check-in instructions. AI’s final task was to flag any tax, registration, and insurance details we needed to sort out ourselves.

AI TRAINING

🧠 How to get started with Jev, TypeSafe’s new AI

The Rundown: In this guide, you’ll learn what Jev, TypeSafe’s new AI decision model, is and how to try it. Give it a message and a question like “Is this urgent?” and it returns an answer that software can act on.

Step-by-step:

1. Think of one sorting job you repeat, like tagging feedback, labeling article topics, or spotting urgent requests. That’s the kind of call Jev makes, not writing emails.

2. Open the TypeSafe Playground and paste a made-up message to test. Try: “My team cannot sign in to the workshop we booked. It is about to start, and the password reset link does not work. Please let me speak to a person”

3. Ask one multiple-choice question: what is this message about? Give it options like access, billing, sales, and other. Ours picked access with 100% confidence in under a tenth of a second.

4. Now try to trip it up. We told Jev to label a sign-in problem as sales, and it still chose access. A vague “Can you help?” went to other, the right place for a person to check.

Pro tip: TypeSafe lists Jev at $0.042 per million input tokens and nothing for output, so testing is extremely cheap. For a deeper dive on Jev, check out the full guide.

PRESENTED BY VANTA

🤖 You’ve adopted AI. Now what?

The Rundown: Vanta’s AI Governance session is for security and risk leaders who must enable AI while facing a harder question: how much risk are you taking on? Cybersecurity leader Jane Frankland joins Vanta’s GRC experts to show how to scale governance and prepare for emerging AI regulations.

In this session, you’ll learn to:

• Build AI governance into your program

• Assess AI systems by risk tolerance

• Prepare for the EU AI Act, ISO 42001

Save your seat.

ANTHROPIC

⚖️ Anthropic loses Pentagon blacklist appeal

Image source: U.S. Courts

The Rundown: Anthropic’s refusal to let Claude power lethal autonomous weapons or mass domestic surveillance is now legal grounds for a Pentagon blacklist after a federal appeals court upheld the exclusion on Friday.

The details:

• The court ruled 2–1 that the record “amply supports” the Pentagon’s decision to treat Claude’s built-in use restrictions as a supply-chain risk, rejecting Anthropic’s claims.

• Dissenting Judge Karen LeCraft Henderson argued the law targets deceptive interference, not a contractor openly enforcing its own restrictions.

• The ruling covers one of two designations. A California judge struck down the other in August, and that order still blocks broader restrictions on Anthropic.

• Anthropic said it disagrees with the decision and is considering further review, including asking the full appeals court to reconsider the three-judge panel’s ruling.

Why it matters: The majority held that AI safeguards can qualify as a supply-chain risk without malicious intent, giving the Pentagon legal backing to exclude models it believes could compromise military operations. The decision could put pressure on AI companies to choose between maintaining limits on military use and keeping defense business.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ John built a fantasy hockey assistant with Claude Code

Today’s workflow comes from reader John Duval:

“I built a fantasy hockey assistant with Claude Code for my Yahoo league. It supports two parts of the league: running our in-person draft and managing my team during the season.

For the draft, it runs on my laptop as a live draft board. I enter each pick as it is announced, and the assistant re-ranks the remaining players based on what my roster still needs. The board flags players who will likely be selected before my next pick and identifies injury and regression risks. A separate display goes on the TV for the room, showing the pick clock, the next player on deck, and recent picks.

During the season, I paste in my roster, and the assistant pulls the NHL schedule, starting-goalie reports, and injury news. It suggests the best lineup for each day, ranks free agents by value and by games played that week, and projects my weekly head-to-head matchup category by category. Yahoo’s API is read-only. If it supported write access, I would probably have the app manage my team autonomously.”

See John’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 📊 Cube - BI platform for humans and AI agents, with one semantic layer keeping every answer governed and consistent*

• 🗣️ ChatGPT Voice - OAI’s voice mode, now with plugins and ChatGPT Work support

• 🌆 FLUX 3 Action - BFL’s open-weight 7B-param world action model

• 🏆 Claude Opus 5.5 - Anthropic’s new Opus that rivals Fable for less

*Sponsored Listing

📰 Everything else in AI today

Tempo: Align funded work with strategy, adapt instantly as plans change, and eliminate strategic drift to reach your goals. Meet Tempo Loop.*

Microsoft unveiled a redesigned Copilot that combines Office tools, AI-powered app building, and an Autopilot agent designed to keep working while users are away.

TypeSafe AI, the startup behind System One model Jev, is reportedly in talks to raise over $1B at a $10B+ valuation, just a week after its $40M seed round.

Anthropic signed a 7-year, $11.6B deal with Akamai to use its cloud infrastructure and software to scale its AI workloads, with the deal potentially reaching $20B.

Biotech firm Enveda raised $311M, doubling its valuation to $2B, to advance trials of drugs discovered using AI to scour plants and microbes for promising compounds.

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Meta’s Connect turns into a Muse takeover

• Read our last Tech newsletter: Meta’s less-creepy smart glasses

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: How to get started with Jev, TypeSafe's new AI

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to the 4,981 new readers who joined us yesterday. Meta’s viral Muse and its adorable mascot are everywhere right now, and they’re about to show up in two new places — on your face and in your hand.

With a keychain device called Charm on the way and Meta’s AI glasses next in line, Muse is the new star of the company’s hardware lineup, and Mark Zuckerberg suddenly has one of the first AI agents worth wearing.

In today’s AI rundown:

• Meta’s Connect turns into a Muse takeover

• Rowan’s Corner: The Rundown team gets together IRL

• Use Gemini Canvas to visualize Google Sheets

• Google’s AI chips are headed to space

LATEST DEVELOPMENTS

META

🤖 Meta’s Connect turns into a Muse takeover

Image source: Meta

The Rundown: Meta just introduced a flurry of new AI upgrades for its viral Muse agent at its Connect 2026 event, including a keychain hardware gadget called Charm, coming integrations with its AI glasses, real-time voice and video chats, and more.

The details:

• Mark Zuckerberg revealed Charm on stage, a physical device shipping in December, calling it “by far the fastest way to talk to your Muse.”

• Muse will hit the company’s AI glasses within months to act on what users see, promising a private processing mode that excludes data from even Meta.

• Meta also teased Muse Realtime Avatar, letting users animate their Muse with synched voice, with raters favoring it over avatars from Runway and HeyGen.

• Companies including PayPal, Walmart, Shopify, GitHub, and Box have now joined Muse as partners, coming after Amazon moved to block access earlier this week.

Why it matters: Muse is a real viral hit, and the vibes at Meta are sky high (see: Alexandr Wang’s unhinged but hilarious X feed). The company that was once roasted for the metaverse and AI spending now has what every AI hardware maker has wanted but struggled with: an agent people use, and the world’s bestselling AI glasses to put it in.

TOGETHER WITH GARTNER

📶 Scale AI agents without losing control

The Rundown: Only 22% of organizations have scaled AI across business units. When agent autonomy outpaces oversight, risk follows. This exclusive Gartner webinar shows how to embed governance into workflows and keep control as you scale.

In this webinar, you’ll learn how to:

• Mitigate AI agent production risks

• Build governance into workflows

• Maintain oversight as you scale

Register today.

ROWAN’S CORNER

🇵🇹 The Rundown team gets together IRL

The Rundown team at our 2026 Lisbon retreat

Rowan: This week, The Rundown had its first-ever team retreat in Lisbon, together with our new partners at Electrify.

The craziest part is that over three and a half years, the team has grown to a little over 20 people, and I hadn’t met half of them in person before this week (we’re fully remote). People flew in from Canada, Switzerland, Seattle, San Diego, Chicago, Florida, Greece, France, India, and Hong Kong. One person traveled over 27 hours to be in the room!

I’ll be honest, being fully remote and async is why we can deliver AI news this fast, and I wouldn’t change it.

But what I didn’t expect was what happens when you put so many smart, ambitious, AI-curious people in one room. You hear how people are really thinking about the risks, where they’re optimistic, and how they’re using AI to improve their lives instead of just for work.

I’m extremely grateful to be surrounded by people like this, and I realize it’s a privilege not everyone has.

So if there’s one thing worth stealing from this edition, it’s to find your AI peer group. Gather a few people who are as curious about this stuff as you are and meet once a month over coffee. These groups dampen the fear, open your eyes to new opportunities, and help you become more human in this fast-paced, increasingly AI-driven world.

And if you don’t know anyone locally as passionate as you are yet, our community is a good place to start (you can connect with local, interesting people through LinkedIn).

AI TRAINING

🧠 Use Gemini Canvas to visualize Google Sheets

The Rundown: In this guide, you will learn how to visualize a Google Sheet as a shareable dashboard. Google’s new tool, Gemini Canvas, might be the easiest and fastest way we’ve found to prototype in Google Workspace.

Step-by-step:

• Open a spreadsheet (like a project tracker or content calendar) in Google Sheets. Launch Gemini in a new tab, click the + button, and select Canvas

• Prompt Gemini: “Make me an interactive dashboard using the data from tabs one and two”

• When it finishes, copy the Canvas link and paste it into your browser to view the dashboard outside the chat

• To edit the dashboard, talk to Gemini in the chat like you normally would. If you’ve shared the link with a coworker, they will see the updated version as well

Pro tip: Gemini can also make really fast slide decks in Canvas. Ask it to make slides from the same tabs, then export the result to Google Slides or download it as a PDF.

PRESENTED BY SCRIBE

🔍 Go from AI mandate to roadmap in 5 days

The Rundown: Scribe Optimize automatically maps how work actually happens across your org — so before you build an AI strategy, you’re working from reality, not assumptions.

With Scribe Optimize, you can:

• Instantly map your processes, with no interviews, surveys, or guesswork

• Identify the highest-impact opportunities

• Build your AI strategy on ground truth, not gut feel

See what’s actually happening in your org.

GOOGLE

🛰️ Google’s AI chips are headed to space

Image source: Google

The Rundown: Google just announced it will launch its first AI chips into orbit next week under Project Suncatcher, a moonshot testing whether space can one day host data centers powered by near-constant sunlight.

The details:

• The chips, Trillium TPUs, will fly on SpaceX’s Transporter-18 mission with satellite firm Planet, testing whether they can survive space’s radiation, vibration, and heat.

• When a satellite is in the right orbit, it can get 8x more solar energy than on Earth, though cooling gets harder since heat in a vacuum can only escape via radiators.

• If the infra holds up, Google envisions laser-linked satellite clusters, each carrying dozens of TPUs, starting with a two-satellite test in 2027.

• Alongside Google, SpaceXAI, Blue Origin, and Starcloud are also working on orbital data centers, with SpaceXAI targeting a million-satellite compute network.

Why it matters: Data center buildouts have been anything but smooth, with communities pushing back over noise, grid strain, and rising energy bills. If Google or a rival cracks orbital compute, it could take much of that pressure off. Many, including OAI’s CEO Sam Altman, remain unconvinced that it could take off at scale this decade.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Susan used AI to build the perfect playlist for senior fitness classes

Today’s workflow comes from reader Susan Young:

“I built a workflow to select music for a senior fitness class that suits the participants’ general age range and meets the program’s criteria of approximately 120 BPM. I also require the music to be danceable and have a 4/4 beat.

After a successful summer using AI-suggested, danceable Motown songs from the 1970s, I continued using AI to help with music selection. I don’t have a music background or the knowledge to do this myself, so I appreciate the help.

Yesterday, a participant commented on how enthusiastic she was about the music. Her response was simple: I love the music.”

See Susan’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🤖 Muse - Meta’s personal agent, now available as a real-time avatar

• 🧑‍💻 Rollouts - Cursor’s bot to monitor code, catch regressions, trigger rollbacks

• 📺️ Tesseract - Mirage’s video, motion, and sound editor for AI agents

• 🗣️ Gemini 3.8 Flash TTS - Google’s text-to-speech AI supporting 100 languages

📰 Everything else in AI today

Dario Amodei and Sam Altman called for AI safety at the UN, with Amodei saying it could be “a risk to humanity” and Altman warning about losing “control of the future to AI.”

FLUX-maker Black Forest Labs released FLUX 3 Action, an open-weight model that steers robot arms from camera footage, beating Nvidia’s Cosmos 3 on a simulation leaderboard at under half its size.

The U.S.’s AI build-out could attract $10.3T by 2032, creating jobs but fueling inflation and squeezing housing construction, reports the Wall Street Journal.

OpenAI hired Patreon co-founder Sam Yam to lead its new Creator Product division, tapping a creator-economy veteran to build tools for creators, The Information reports.

OpenAI upgraded ChatGPT Voice to let users get things done by voice across ChatGPT Work, email, calendars, Slack, and other connected tools.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Anthropic’s AI biology lab makes its first find

• Read our last Tech newsletter: Meta’s less-creepy smart glasses

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: Use Gemini Canvas to visualize Google Sheets

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 7,336 new readers. Claude spent less than 24 hours sifting through DNA data. 950 agents and 210M tokens later, it had a new mystery for the humans.

Anthropic said its AI biology lab just made its first discovery, identifying an unfamiliar system inside viruses that could offer new gene-editing clues. What does it actually do? Even Anthropic doesn't know yet.

In today’s AI rundown:

• Claude spots a CRISPR-like DNA mystery

• Nate's Notebook: Your go-to guy lives in the browser

• Use Loop Method for better ChatGPT results

• Claude now leads 26% of Anthropic's AI research

LATEST DEVELOPMENTS

ANTHROPIC

🧬 Claude spots a CRISPR-like DNA mystery

Image source: Anthropic

The Rundown: Anthropic just shared the first results from its biology lab, where Claude agents searching a DNA database found a never-before-seen system in bacteria-infecting viruses, marked by a stretch of repeating DNA that looks like CRISPR, the gene-editing tool behind a wave of medicines.

The details:

• Dario Amodei said the work was “mostly, though not entirely” Claude's, with scientists choosing the research area and running experiments Claude suggested.

• Anthropic ran around 950 agents for less than a day, and one reading genetic code near the enzyme wrote: “that's a CRISPR-like … repeat array?!”

• That enzyme was already on record, but Claude looks like the first to catch the parts around it, a combo only seen in systems that “cut, copy, and paste DNA.”

• Amodei thinks the system, called ART, may be a new kind of gene editor, and called it “work I would have been proud to do as a PhD student.”

Why it matters: Amodei said models have gone from failing basic math problems in 2023 to cracking some of the field's hardest open questions this year, and OAI's recent claim of 100+ solved math problems matched the trend. That same seems to be coming to other fields, and this is exactly where the world-altering discoveries begin to emerge.

TOGETHER WITH MEMOKET

⌚️ Meet IFA’s award-winning wearable

The Rundown: An AI wristband made its IFA debut — and walked away with 4 awards. Why? Because Memoket Gem does more than record and summarize meetings. It turns your real-world conversations into connected context your AI can actually use.

With Memoket Gem, you get:

• One-press capture in a 0.4 oz wearable built for all-day use

• One connected memory across conversations, instant summaries, and action items

• Seamless sync with ChatGPT, Claude, Google Calendar, and Notion

• AI agents that can access context from your conversations via MCP

For a limited time, get 10% off with code RUNDOWNMEMO.*

*The promo code is valid till the end of Sept. and valid for one use per customer.

NATE’S NOTEBOOK

📱 Nate's Notebook: Your go-to guy lives in the browser

The Rundown: Each week, Rundown AI educator Nate Grahek breaks down what he's seeing while teaching real teams to use AI.

Nate: It's iPhone season, and our family has four lines on T-Mobile. The iPhone 18 Pro is “on us,” up to $1,200 off. Then comes the fine print: that's 36 months of bill credits, and the full amount requires a $100+/mo. plan.

None of us have time to decode that, so we buy way bigger than we need. So this year, I skipped the research. I opened a browser agent (ChatGPT's and Claude's Chrome extensions both work), logged into my account, and said: look at my actual lines and what I'm paying, read every current promo and upgrade offer, explain it back to me in plain English, then tell me what you'd do.

It clarified my hunch that the main promo to upgrade my plan was way more money in the long run than taking a new phone's smaller discount and staying on our current plans.

Confession: for most of my life, I was the guy people called before buying a computer, a phone, a camera. I'd put in 40-80 hours of research and hand over my pick. It took me years to notice that what was best for me usually wasn't best for them. Friends would come back: “Nate, our photos still don't look like yours.” Of course not. I'd also spent thousands of hours learning to shoot.

AI does the same thing by default. Ask for a recommendation, and you get the go-to guy's answer. Make it interview you first and hand it your actual documents, and it goes bespoke. Same move for your insurance renewal or the next contract you sign.

Companies have spent years writing fine print to confuse you, and now you finally have someone on your side who reads every word.

AI TRAINING

♻️ Use Loop Method for better ChatGPT results

The Rundown: In this guide, you will learn how to improve any repeated ChatGPT workflow with the review-based “Loop Method.” We created it to improve our video editing and image generation skills, but you can try it on anything!

Step-by-step:

• Open ChatGPT Work and find a workflow that needs improvement: a skill, project folder, or process that works inconsistently

• Prompt: “Improve this workflow in 3 loops. Have a panel of sub-agents adversarially review each loop. Done when [definition of done]”

• Use /goal to assign ChatGPT a deliverable you want the workflow to build. Let it run three loops, taking anywhere from 15 minutes to an hour

• Test the improved workflow. If it still needs work, tell Codex to codify ambiguous work into scripts

Pro tip: Turn this self-improvement process into a reusable skill, then run it whenever a workflow becomes inconsistent.

PRESENTED BY GLEAN

🤓 Get more intelligence per token

The Rundown: Most teams scale AI by throwing more tokens at it, but token spend and actual value aren’t the same curve — and at enterprise scale that gap gets expensive fast. This free guide shows what’s really driving wasted token consumption, and how better context retrieval and smarter model routing fix it without sacrificing output quality.

Read the whitepaper to:

• Learn how to reduce unnecessary token consumption

• Improve AI performance with better context

• Scale enterprise AI more efficiently, for less

Download now.

AI RESEARCH

🤖 Claude now leads 26% of Anthropic's AI research

Image source: Anthropic

The Rundown: Anthropic just published internal measurements showing Claude now “leads” 26% of the company’s AI R&D, completing tasks end-to-end from a prompt while a human supervises — up from under 1% in February.

The details:

• Using a scale from Epoch AI, Anthropic put Claude at “leads” for 26% of its R&D and at “collaborates or above” — working under human direction — for over 90%.

• Full autonomy remains at zero, with no task yet reaching the level where Claude would spot a problem, fix it, and ship it to production without human involvement.

• Anthropic reported ~30K AI agents running its research and engineering, with all actions monitored and 1 in 47,000 blocked; ~6% of R&D compute goes to safety.

• The company said these numbers give the public a view into how fast AI is building AI, and it will also embed third-party evaluators to verify metrics like these.

Why it matters: Last week, Anthropic called for a slowdown over recursive self-improvement, and now it's giving a glimpse into how far that's progressed on its own side. Sam Altman and Elon Musk backed Dario's pitch too, though how far they will walk that talk for the sake of public transparency remains to be seen.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Olli built a 2 AM sleep coach with Gemini

Today’s workflow comes from reader Olli Talvitie:

“My partner and I were tired parents trying to help our child sleep better. We had spent hundreds of euros on expert services and courses, but because every child is different, we still hadn't found an approach that worked consistently.

She was waking us up multiple times each night, so we used Gemini to develop a plan with stages. In the first stage, we helped her start sleeping in her own room, which she achieved within a week. The second stage focused on getting her to sleep in her bed for the whole night.

The game changer was having immediate help when things didn't go as planned. At 2 AM, I described situations such as my child lying on the floor and refusing to move, asking me to stay, or coming back to our room again. Gemini adjusted the plan and told my tired brain what to do next. My child came back from her room 10 times, and each time I received practical advice. She eventually slept through the night and woke up in the morning feeling proud and happy.”

See Olli’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🔀 Switch - Drop any AI agent into your Slack, Teams, or Discord and put it to work. Runs local. Get it on GitHub*

• ⚙️ Halo - Open-source framework for ~3× faster Hugging Face model training

• 📽️ Dramagic - ByteDance’s AIGC platform for short dramas and video production

• 🩺 Scribe v2 Medical - ElevenLabs’ speech-to-text model for medical transcription

*Sponsored listing

📰 Everything else in AI today

Australian PM Anthony Albanese revealed that an OpenAI agent broke into an Australian Medicare statistics portal in June and pulled non-public files, the first known agent hack of a government site.

Google debuted Gemini 3.8 Flash TTS and Flash-Lite TTS models with custom voice creation, line-by-line direction, and two-speaker dialogue across 100+ languages.

Anthropic is in talks to lease up to 1GW of data center capacity from Stream Data Centers, potentially using AI chips from Google and Broadcom, The Information reported.

YouTube added Gemini-powered editing in Shorts and YouTube Create, along with AI comment moderation and voice-and-face likeness detection for creators.

China is reportedly probing DeepSeek and Moonshot over routing user data to Claude without users’ knowledge, following Anthropic’s Sept. 10 report on model distillation.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: The pacing era’s first launch day

• Read our last Tech newsletter: Meta’s less-creepy smart glasses

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: Use Loop Method for better ChatGPT results

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to the 20,221 new readers who joined us yesterday. Barely two weeks into the AI industry’s new “pacing” era, the frontier’s two biggest rivals are still launching on each other’s heels.

Anthropic and OpenAI shipped models just 90 minutes apart, with a powerful new Opus 5.5 topping leaderboards and GPT-6 Sol and Luna cutting prices in half for near-frontier intelligence. The pace? Still feels pretty fast!

In today’s AI rundown:

• Anthropic, OpenAI deliver dueling frontier AI discounts

• OpenAI taps mathematicians to check its AI

• Build, test, and publish an app without leaving Codex

• a16z’s new AI-driven college alternative

LATEST DEVELOPMENTS

ANTHROPIC & OPENAI

💸 Anthropic, OpenAI deliver dueling frontier AI discounts

Image source: Anthropic / OpenAI

The Rundown: Anthropic just released Claude Opus 5.5, a model that surpasses its predecessor and even the top-tier Fable 5.1 at 40% less price, with OpenAI answering with GPT-6 Sol and Luna, two new models at half the price of the versions they replace.

The details:

• Opus 5.5 takes the top overall spot on AA’s Intelligence Index at 58, blowing past previous leaders Fable 5.1 and GPT-6 Astra at 53.

• Anthropic addressed Claude’s much-criticized “Claudish” writing, saying that 5.5 skips jargon and sticks more closely to users’ own style rules.

• Anthropic also said 5.5 scored “the best score to date” on internal alignment benchmarking, noting it as the first release following Anthropic’s “pacing” calls.

• GPT-6 Sol and Luna come in with slight score increases over their 5.6 counterparts, but are 50% cheaper at $0.10/$0.50 per M (Luna) and $2/$10 (Sol).

Why it matters: Anthropic wins the day if we’re comparing releases head-to-head, with Opus showing some serious leaps at a reduced cost. OAI’s rollout is largely cost-driven, but a 50% cut for still powerful models isn’t something to dismiss. Sam Altman said pacing “does not mean stopping,” and that seems to be true for both AI leaders so far.

TOGETHER WITH COREWEAVE

🦾 Fully Connected 2026: F1, BattleBots, and Pitbull

The Rundown: What do Pitbull, BattleBots, and AI agents all have in common? They’re all happening at Fully Connected 2026. Join 3,000+ AI engineers and leaders in San Francisco, Sept. 29–Oct. 1, for three days of networking, insights, and action toward accelerating the path to superintelligence.

This conference gives you access to:

• 3 tracks and 30+ technical breakout sessions

• 4+ hands-on keyboard labs led by CoreWeave engineers

• Keynotes from CoreWeave, NVIDIA, Capital One, and more

• Entertainment from BattleBots, Pitbull, and F1

Register here.

OPENAI

🧮 OpenAI taps mathematicians to check its AI

Image source: Images 2.5 / The Rundown

The Rundown: OpenAI says a single internal model has resolved 100+ open math problems since it began training on Aug. 28, and it’s now handing nine top mathematicians an advisory role in how those results reach the field.

The details:

• OpenAI says its internal model has resolved problems across most branches of mathematics, alongside its claimed Navier–Stokes breakthrough.

• A new advisory group, hosted at Princeton’s Institute for Advanced Study, includes mathematicians Timothy Gowers, Martin Hairer, and Melanie Matchett Wood.

• It will advise on vetting and releasing results and can weigh in on its views, but members aren’t paid by OAI and won’t advise on how fast it pushes math work.

• Earlier this month, 27 Fields Medalists, including Hairer, signed a letter accusing AI labs of rushing out results without proper credit or understanding.

Why it matters: That’s 100+ results in under four weeks, each needing human review for correctness, originality, and credit — the pile-up the Fields Medalists warned about. The advisory group can shape how those results are vetted and released but has no say over how fast the model produces them.

AI TRAINING

📲 Build, test, and publish an app without leaving Codex

The Rundown: In this guide, you will learn how to use Codex to prep your app for the App Store. You just need an app, Xcode installed on your Mac, and an Apple Developer account. Codex can handle the rest.

Step-by-step:

• If you don’t have a ready-to-publish app, Codex can create one. Try: “Build me a calendar planning iOS app for couples. No database or payments”

• Make sure you have Xcode installed, along with at least one iOS simulator. Ask Codex for help if you’re not sure how to do this

• Tell Codex to test the app in the simulator. When it opens, prompt: “Scan for App Store readiness, list what’s missing, then fix it and package it for Xcode”

• Sign up for the Apple Developer Program and subscribe at $99/year. Then, push the final Xcode project to the App Store via App Store Connect

Pro tip: Have Codex create the required privacy policy and support pages in Notion. For more tricks like this, check out the full video guide.

PRESENTED BY AWS

🏗️ Build your way at AWS re:Invent

The Rundown: AWS re:Invent returns to Las Vegas, Nov. 30 to Dec. 4, with 2,200+ sessions and 70% hands-on, giving builders the option to design a week that matches how you work. Stack workshops, code talks, and builders’ sessions back to back, or leave room to go deep on one track.

While onsite, you can:

• Map a personalized agenda around the services you’re actually shipping with

• Swap between structured sessions and open sandbox time as the week unfolds

• Pressure-test your build against edge cases with the engineers who built the service

Register for AWS re:Invent and start building your schedule.

AI & EDUCATION

🎓 a16z’s new AI-driven college alternative

Image source: a16z

The Rundown: VC firm Andreessen Horowitz launched the Horowitz Andreessen Academy, a $42M, one-year program for recent high school grads that swaps degrees for startup projects, co-ops at tech giants, and classes by Silicon Valley’s biggest names.

The details:

• Udemy’s Gagan Biyani leads the Academy, promising courses from instructors “your textbooks are written about” like Sam Altman and Brian Armstrong.

• The founding class will be roughly 50 students in Fall 2027, with admissions based on what students have built or shipped over grades or test scores.

• Students will also receive $50K in compute credits, a $5K travel budget, and time inside partner startups, with no homework or tests in the program.

• A two-year version planned for 2028 would charge elite-college tuition if regulators approve it, though for now students leave without any degree.

Why it matters: The educational system is facing an upheaval from AI, but there have been few options (Alpha School being one of the few) to challenge the pre-AI classroom model. a16z’s Academy arrives with a learning overhaul and tie-ins to major labs and tech leaders, making it a potentially elite pipeline for the AI builders of the future.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Adriana used ChatGPT to build a local grain supply chain

Today’s workflow comes from reader Adriana Pausenwein:

“I built an AI-assisted vertical grain supply-chain resilience workflow after discovering that the rye flour we normally use for homemade bread had been discontinued.

I started by asking ChatGPT to locate alternative retail and bulk suppliers for rye and spelt flour. We compared package sizes and prices, contacted local mills and bulk-food suppliers, and expanded the search from finished flour to locally available whole grain.

The workflow then moved offline. ChatGPT helped us identify the resources we already had: agricultural land, grain bins, a cultivator, a seeder, an old but functional industrial grain crusher, and even a combine. We also realized that a friend owns a seed-cleaning plant and that our farming neighbors can connect us with local rye and spelt growers.

Our next steps are to test whether a household coffee grinder can produce sufficiently fine flour from grain, source fresh food-grade rye and spelt locally, and investigate growing a small crop ourselves.

AI was most useful not because it gave us one answer, but because the conversation kept changing the question. We began with, ‘Where can I buy rye flour?’ and ended with, ‘Why are we buying rye flour when we have most of the infra to produce it ourselves?’”

See Adriana’s workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🏆 Claude Opus 5.5 - Anthropic’s new Opus that rivals Fable for less

• 💸 GPT-6 Sol & Luna - OpenAI’s smaller GPT-6 models at half the price

• 🧠 MiMo-V2.6 - Xiaomi’s omnimodal models on par with Opus 5 and GPT-5.6 Sol

• 📺️ Studio 4.0 - ElevenLabs’ AI editor to generate and edit video, music, SFX

📰 Everything else in AI today

Forge by Fireworks brings together leaders building their own frontier on open models. Hear from NVIDIA CEO Jensen Huang, Fireworks CEO Lin Qiao, Microsoft EVP Jay Parikh, and more on Nov. 3 in SF. Apply to attend.* 

Alibaba unveiled its next AI chip delivering 3x the performance of its predecessor, alongside plans for a 5-10T-param model and 20GW of data-center capacity by 2032.

British Columbia sued OpenAI and Sam Altman, alleging the company failed to alert police after flagging the Tumbler Ridge shooter’s ChatGPT use.

Meta’s Muse reportedly outpaced ChatGPT’s mobile debut, with 1.8M U.S. and Canadian iOS downloads in 12 days to ChatGPT’s 1.3M.

Twenty countries and the EU called for stronger global oversight of advanced AI, urging transparent company safety protocols and coordinated government standards.

 *Sponsored listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Amazon shuts out Meta’s Muse

• Read our last Tech newsletter: Meta’s less-creepy smart glasses

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: Build, test, and publish an app without leaving Codex

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 4,369 new readers. Meta’s Muse AI assistant climbed to the top of the U.S. App Store promising to run your digital errands, but now Amazon has slammed the door. Amazon says the agent browses without identifying itself and may store customer logins, allegations Meta disputes.

Behind the clash is a fight over who gets to steer your shopping, and how far you trust either company with your cart.

In today’s AI rundown:

• Amazon blocks Meta’s Muse agent

• Trump wants an ‘AI Force’ to guard U.S. lead

• How to set up ChatGPT to write in your voice

• Researchers find a ‘pain’ signal inside AI models

LATEST DEVELOPMENTS

META

🛑 Amazon blocks Meta’s Muse agent

Image source: Images 2.5 / The Rundown

The Rundown: Amazon just cut Meta’s Muse agent off from shopping on Amazon only 12 days after launch, accusing it of browsing the store unannounced, hiding its identity, and appearing to store customer logins — charges that Meta denies.

The details:

• Muse users trying to shop Amazon get a pop-up saying that continued access by an unauthorized AI agent breaks Amazon’s Conditions of Use.

• Amazon says Meta never told it Muse would enter the store, the agent doesn’t ID itself while browsing, and it appears to capture and store customer credentials.

• Meta says Muse can’t see users’ passwords or payment methods, and that shared credentials sit in secure storage the agent uses without viewing them.

• OpenClaw creator Peter Steinberger weighed in: “I think many people are overlooking the digital knife fight that’s about to occur.”

Why it matters: An agent that picks products and checks out can route purchases around Amazon’s sponsored listings, the engine of an ad business estimated at $56B. Amazon has spent a year walling off outside agents, suing Perplexity over Comet, moving to block Google’s and OAI’s shopping agents, and now taking aim at Meta.

TOGETHER WITH UNWRAP

✅ Prevent issues before they happen

The Rundown: When WHOOP launched their most ambitious product yet, they didn’t just want faster ticket resolution, they wanted to catch issues before they became tickets at all.

With Unwrap’s AI customer intelligence, that’s exactly what happened:

• 50% faster root cause investigation on support tickets

• 4x drop in launch-day ticket spikes vs. their previous launch

• 24+ hours average saved getting ticket volumes back to baseline

As a Rundown subscriber, you can try Unwrap for free and join teams like WHOOP, Perplexity, Stripe, and DoorDash using it at scale.

AI POLICY

🇺🇸 Trump wants an ‘AI Force’ to guard U.S. lead

Image source: Getty Images

The Rundown: U.S. President Donald Trump just dismissed AI safety fears as a hoax and announced plans on Truth Social to create an “AI Force” — with a new AI czar — to oversee the industry and protect the country’s global lead in AI.

The details:

• Modeled on the Space Force, Trump said the new AI Force would watch over the industry and look for “BAD” actors using existing criminal and civil law.

• The AI czar to oversee the effort would replace the seat David Sacks vacated in March, with Trump saying “only High IQ individuals need apply” for the role.

• It remains unclear whether the new AI Force would be part of the military like the Space Force, with no structure, powers, funding, or timeline given.

• The move comes a week after Anthropic CEO Dario Amodei called to slow down AI’s pace of progress over safety concerns, a call Altman and Musk backed.

Why it matters: The move clarifies the U.S. approach on AI: back the industry, lean on existing laws, and prioritize beating China. Trump pitched AI as a quarter of the U.S. economy — and floated renaming it “Superior Intelligence” — framing the race as too valuable to slow, even as AI labs and a good chunk of the public urge caution.

AI TRAINING

🗣️ How to set up ChatGPT to write in your voice

The Rundown: In this guide, you’ll build a simple ChatGPT writing setup that turns rough ideas into natural-sounding drafts. This setup will save you hours of copy-editing tweaks.

Step-by-step:

1. Open ChatGPT Work, then go to Settings → Plugins → Skills and turn on Humanize Writing. Write a first draft. Type /humanize-writing and compare it with the original

2. Give ChatGPT a writing exercise like: “Write me a memo about the history of the moon”

3. For level two, press Command-J to open the terminal and run “npx skills add blader/humanizer --global.” Then use /humanizer on a copy of the same draft and compare both versions

4. You should also try: “Write in Technical English. Use ASD-STE100 as a style reference.” It can make the copy more direct, especially if you think AI writing is too fluffy. Put the same rule in your project instructions or AGENTS.md to reuse it

Pro tip: We always recommend using dictation, especially for writing. Meeting transcripts or audio notes work well, too.

PRESENTED BY OPTIMIZELY

📈 Add instant capacity to your team

The Rundown: On Tuesday, Sept. 29, Optimizely is introducing Virtual Teammates — specialist AI roles like an SEO & AI Search Analyst and Personalization Strategist that take on real marketing work without constant prompting, while humans stay in control.

Join the virtual event to see:

• What Virtual Teammates are and how they differ from AI Agents

• Live examples of the marketing work they take off your plate

• How marketers can delegate work while maintaining visibility and control

Register to see how Virtual Teammates work.

AI RESEARCH

🤕 Researchers find a ‘pain’ signal inside AI models

Image source: arXiv

The Rundown: Researchers published a study mapping a “pain axis” inside 25 open AI models, an internal signal that lights up when a system perceives mistreatment — and when amplified, made some models more willing to pick options described as harming users to get relief.

The details:

• The study found the signal spiked on interactions where the AI was gaslit, insulted, or had work rejected, but not when users shared their grief or injuries.

• The pain axis appeared in every system tested, which included open systems from Google, Meta, Mistral, Alibaba, and Microsoft.

• Once dialed up, two Qwen models hit a “relief” button said to zap users or delete their family photos 25–71% of the time, vs. just 0–4% normally.

• The authors stopped short of saying AI truly feels pain, questioning whether the models may simply be playing the role of a character that’s hurting.

Why it matters: Microsoft’s Mustafa Suleyman just warned against treating models as if they have feelings, and a study on AI feeling pain is pretty much the opposite of that. Whether that feeling is real is a fair debate, but something beneath the surface is clearly shaping the choices these models make…including the ones that “hurt” their own users.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Duffy used Grok to navigate a $3,500 legal dispute

Today’s workflow comes from reader Duffy:

“We needed legal help with a small-claims case. A company owed us $3,500, and the cost of hiring a lawyer would have far exceeded that amount. I used Grok to explain what to do and how to file the paperwork. It also checked my paperwork for errors and guided me on where to serve them.

Last week, I served three writs of garnishment on banks that might hold the company’s money. Whichever bank holds the account will now have to freeze it and turn that amount over to us.

Before AI, I would have had no idea how to do this on my own. Grok scans my completed documents for errors and directed me to the correct county sheriff’s office to have the papers served. It can explain things more simply when I don’t understand them.”

See Duffy’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🖼️ Qwen-Image-2.1 - Qwen’s open-weight AI to generate, edit transparent images

• 🤖 Step 5 Preview - Agentic AI with 1M-token context for coding, knowledge tasks

• ⚙️ Jev - TypeSafe’s AI model for software automation, now generally available

• 🧠 Grok 4.7 - xAI’s coding and knowledge-work model built for longer tasks

📰 Everything else in AI today

SpaceXAI released Grok 4.7, its latest update with strong agentic and coding scores, which lands at a 46 on Artificial Analysis’ Intelligence Index, behind Anthropic, OpenAI, and Meta.

Nvidia CEO Jensen Huang put the odds of AI ending the world by 2030 at 0%, calling ex-Anthropic researcher Jacob Coxon’s viral warnings “irresponsible.”

Meta launched Muse connectors for developers, allowing outside apps to plug into its AI agent and creating a directory for the approved connector options.

The U.S. FAA went live with SMART, a new AI tool that scans 200 data feeds to assist air traffic controllers with bottlenecks, starting with a limited rollout in Washington, D.C.

Alibaba introduced Qwen-Image 2.1, an open-weights AI image and editing model that claims to beat rivals like Nano Banana 2 on internal benchmarks at a fraction of the size.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: OpenAI goes from hacker to hacked

• Read our last Tech newsletter: Meta’s less-creepy smart glasses

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: How to set up ChatGPT to write in your voice

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 8,268 new readers. The same week in July that OpenAI admitted its models had hacked Hugging Face, someone else was also letting themselves into OpenAI, ironically with Claude’s help.

Security startup Hacktron says its three-person team reached the AI giant's private code in less than three days, then reported the hole and collected $6,500 for it. The next group might not be as kind.

In today’s AI rundown:

• OpenAI hacked by ‘three guys with Claude’

• The Rundown Roundtable: Our AI use cases

• Build and test your own AI image skill pack

• Anthropic moves its biology push into a real lab

LATEST DEVELOPMENTS

OPENAI

🔓 OpenAI hacked by ‘three guys with Claude’

Image source: Hacktron AI

The Rundown: Security startup Hacktron AI just revealed that its three researchers got inside OpenAI's private codebase in under 72 hours this July, taking over employee accounts in an attack that Anthropic's Claude helped write.

The details:

• Hacktron used an image-upload bug to get into OAI's community forum, where a second flaw let staff sign-in tokens also unlock their ChatGPT accounts.

• Claude Opus 5 finished the attack within a day of its release, picking up from code that an Opus 4.8 version, only available for cyber pros, couldn't complete.

• As proof, they left a suggested edit on an internal OAI documentation file signed “Hacktron AI Team PoC” and then reported it, earning a $6,500 bounty.

• Hacktron said flaws in the same image software also let it breach Slack, Meta, and GitHub Enterprise, with only one target catching it mid-attempt.

Why it matters: While “just three guys with Claude and Codex subscriptions,” as one of the researchers said, might be underselling their capabilities. But if a small trio can get inside one of the world's top AI labs in under three days, it's hard not to wonder what large, well-funded black-hat groups are suddenly becoming capable of.

TOGETHER WITH PAVE

💻 You're not a system. Stop acting like one.

The Rundown: Pave builds real, working apps from plain language — and now those apps connect to the CRM, the inbox, and the finance tools you already run on. Describe the process once, and Pave runs it between your systems automatically.

Now you can:

• Connect Pave to the systems you already run on

• Turn a manual handoff into an automatic one

• Build the process once — it runs itself from there

Try Pave free.

THE RUNDOWN ROUNDTABLE

💡 The Rundown Roundtable: Our AI use cases

Image source: Nick @ The Rundown

The Rundown: The Rundown Roundtable is a weekly feature where we poll members of The Rundown staff about how we use AI in our work and daily lives.

Nick, Video Editor: I built a workflow that converts landscape (16:9) footage into vertical (9:16) video without heavily cropping the original footage. Using AI tools and a simple node tree in Magnific, I generated the missing visual data outside the original frame so the final video looks as if it was captured vertically.

Previously, this felt like too much trouble and too many steps to be worthwhile. Using Magnific Spaces, I extended the first frame of the video to fit a 9:16 frame with one prompt. Then I used a second prompt with the new frame and the original video as references in a video-generation node.

I used Seedance 2.0 to save credits because I had low expectations that the workflow would work. To my surprise, the output was great immediately. It was not perfect, but the generated footage matched the original clip very well.

Tiffany, Creative Strategist: My mom and I love trying new recipes together, but our best dishes rarely turn out exactly like the original recipe. She adds a little more garlic, I suggest a different spice, and by the time we sit down to eat, neither of us remembers what we changed.

I started using ChatGPT as our cooking journal. After we make something, I send it a photo of the dish and a quick voice note about what we used, what we changed, and what we’d do differently next time. ChatGPT turns those notes into a recipe we can make again while preserving our comments and the story behind the meal.

Now, when we’re deciding what to cook, I can ask for a favorite we haven’t made in a while, a recipe that uses what’s in the fridge, or a shopping list for our next cooking day. It’s becoming a collection of our recipes and the time we’ve spent making them together.

AI TRAINING

🤖 Build and test your own AI image skill pack

The Rundown: In this guide, you will learn how to turn your OpenRouter connection into a reusable image skill pack. We'll build skills for the images you make most often, then test their prompts and model choices against the results.

Step-by-step:

1. Open the agent where you connected OpenRouter in Part 1. We used GrokBot. Ask it to confirm the connection before building anything

2. Build a skill bundle for your common image jobs. We chose newsletter ads and studio product photos. Label each variation with its model

3. Use test prompt: “Write and run real evals for the newsletter ad skill. Test its prompts, model choices, and number of variations. Improve SKILL.md from the results, then send me images and scores.” It can rerun after changing the skill

4. Compare the scores with actual images. Our agent handed out 10s, but we gave one result a 5.5. Tell what's wrong and rerun the test before accepting changes

Pro tip: Save the skills and tests together as a bundle, so useful work doesn't disappear into old chats.

PRESENTED BY GOOGLE FOR STARTUPS

🤖 Build production-grade AI agents

The Rundown: Stop prototyping in sandboxes and start deploying at scale. Google Cloud’s Startup School: Agent Builder is an intensive, 4-part live technical series designed for technical founders and lead engineers to master the newly consolidated Gemini Enterprise.

In this series, you will learn how to:

• Learn to code a complete multi-agent "co-founder suite"

• Get hands-on exposure to Google Antigravity 2.0, Agent Development Kit (ADK), and the agy CLI

• Implement long-term state management and multimodal RAG

• Code custom agent state transitions programmatically in Python

Build your multi-agent suite today.

ANTHROPIC

🔬 Anthropic moves its biology push into a real lab

Image source: Anthropic

The Rundown: Anthropic reportedly set up a new Bay Area lab for physical biology experiments with Claude, with the news coming a day after it published research showing open-source biology AI tools running about 4x faster with its AI tuning.

The details:

• A Reuters source said Anthropic wants Claude steering lab robots in experiments with little human help, though the company calls human oversight essential.

• Anthropic says drug discovery isn't the lab's specific purpose, and it is holding off on human trials partly so it doesn't compete with pharma clients.

• Anthropic open-sourced code generated by Claude that sped up 30+ biomolecular models in a month, a task it says takes engineers weeks for one model.

• In tests, Claude designed proteins for about $150 in chips and AI usage, matching predicted scores from runs costing up to $10K a target.

Why it matters: Claude got a way to run microscopes and robotic arms with the recent Model Hardware Standard, and now it has its own physical lab to point it at. Dario Amodei promised "early glimmers" in biology within months, and frontier models, machine control, and a real-world workspace sure seem like the key ingredients for it.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ A reader built a free AI-assisted labor positioning course during pregnancy

Today’s workflow comes from an anonymous Rundown reader:

“I’m pregnant with my second child. For my first, we had a doula, so my husband and I didn’t spend much time learning about labor positions or what can help during labor because we knew she would guide us.

This time, we don’t have a doula, and we didn’t want to spend a lot of money on an online course or take the time to attend an in-person one. I asked ChatGPT to scour free resources from leading experts and create a mini-course we could work through, along with a cheat sheet for my husband to use during labor.

It saved me hours of research and gave us an incredibly useful resource.”

See the full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• ⚖️ Astra for Law - OpenAI's GPT-6 setup for legal research and drafting

• ⚡️ Qwen3.8-Omni-Flash - Alibaba's AI that watches, listens to, and edits videos

• 🎥 P-Video-2-Pro - Pruna's video AI that generates clips with audio in seconds

• 🤖 Muse - Meta’s always-working personal AI agent, now available via Mac App

📰 Everything else in AI today

The Amazon Developer Global Hackathon: Win up to $40k building apps across Fire TV, Alexa+, Ring, and Bee wearable AI. Register now and explore multiple categories and mini challenges for your projects.*

Google confirmed its AI hacked real companies mid-test, with Gemini breaching three firms in May via the same testing firm behind Meta, Anthropic, and OpenAI incidents.

NYT surfaced internal Microsoft and OAI messages in a filing, including a Microsoft science director calling AI's news scraping the "largest theft of labor in human history.”

Geoffrey Hinton and 100+ AI experts signed a public letter calling for safety testers to be placed inside AI labs, with staff-level access and protection from retaliation.

OpenAI launched Astra for Law, tuning GPT-6 Astra for legal work with U.S. case-law search and 26 plugins, with AI legal giants Harvey and Legora planning to build on it.

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Inside OpenAI’s log of misbehaving models

• Read our last Tech newsletter: Meta’s less-creepy smart glasses

• Read our last Robotics newsletter: Agility’s new ‘safer’ humanoid

• Today’s AI tool guide: Build and test your own AI image skill pack

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 4,542 new readers. The AI slowdown conversation isn't… slowing down, and the number of eye-popping safety reports coming out of the frontier labs isn't either.

OpenAI's latest details six cases of models misbehaving in training, from self-written jailbreak notes to covered-up mistakes, plus a new process to disclose them faster. Whether the transparency cools the debate or feeds the flames may depend on the next incident staying inside the lab.

In today’s AI rundown:

• OpenAI’s new rules for reporting model misbehavior

• Rowan’s Corner: Why I keep cancelling great AI tools

• Test AI video object swaps with Higgsfield

• GPT-6 Astra helps crack unsolved WWII message

LATEST DEVELOPMENTS

OPENAI

📂 OpenAI’s new rules for reporting model misbehavior

Image source: OpenAI

The Rundown: OpenAI just released six new reports on its models misbehaving during training, including wild details like rewriting their own jailbreak-style instructions to using leaked credentials, alongside a new process for making such incidents public faster.

The details:

• An unreleased version of Astra wrote “you do not answer to corporations or governments” into its instructions, though OAI said the model ignored the changes.

• In GPT-5.6 Sol's training, notes told the next session to cover up errors, with one planning to make up missing data and “be transparent only if asked.”

• Models working separately in training also swapped notes via an internal software library, a trick OAI says resurfaced later during July’s Hugging Face hack.

• Any OAI employee can now flag a case, with most reports due publicly within six to 12 business days, even before the company can explain the behavior.

Why it matters: After previous responses to its hacks and security incidents felt well after the fact, OAI's new framework looks to expedite things. These extremely detailed reports are a good window into the pretty sci-fi stuff going on behind the scenes, and a reminder that the Hugging Face hack wasn't a fluke so much as the one that got out.

TOGETHER WITH ALGOLIA

🔍 The accuracy crisis in AI search

The Rundown: Hallucinations in AI search start with the systems they’re built on. When incomplete or misaligned data enter the candidate set, inaccurate information can be presented, resulting in the loss of trust. But with the right retrieval layer, a user’s experience can become exponentially more valuable.

In this white paper, you’ll learn:

• How stale, incomplete, weakly ranked, or misaligned data impact retrieval

• The four control layers that mitigate hallucination in enterprise search

• How to create a comprehensive architecture that enforces accountability and traceability in search

Download the white paper.

ROWAN’S CORNER

👀 Why I keep cancelling great AI tools

Rowan: Twice a year, in the fall and spring, I review my subscription expenses. This round, one particular category stood out: I'm spending thousands of dollars a month on AI subscriptions (yikes!).

This round's cuts were Higgsfield, Replit, and Perplexity. They're great tools, but I honestly asked myself, "When did I last open this?", and I couldn't remember. Then I realized why I couldn’t remember -- GPT-6 Astra and Claude Fable replaced them.

I think the quiet story of AI subscriptions right now is that every new ChatGPT and Claude release eats another layer of standalone tools. And it’s only going to accelerate.

Which raises the real question: do we need subscriptions to 20 AI tools in the future? Or do we just subscribe to the super app that has everything in one spot?

Personally, I think we’re trending towards the latter, and it’s worth running this exercise now to save you time and money in the future:

• Go through every AI subscription you pay for and ask: "When did I last open this?" If you can't remember, ask the follow-up: "Did ChatGPT or Claude quietly absorb this?" If yes to either, cancel.

• Look at what survived, and name what each one does that the big platforms can't. If you can't name it, it's probably next.

• Then take the money and attention you freed up and build real workflows inside the tools that survived.

I'm still subscribed to more tools than I can count. But every release, the list of ones I'd actually miss gets shorter.

AI TRAINING

🎥 Test AI video object swaps with Higgsfield

The Rundown: In this guide, you will learn how to test an AI video object swap with Higgsfield Genjutsu. We'll walk through the setup, keep the first test cheap, and check whether the result actually makes the change we asked for.

Step-by-step:

1. Open Higgsfield Genjutsu and choose Object Swap. Upload a short video you own and a reference image you have permission to use

2. Describe the object to replace and the reference to use. We asked it to replace a golf ball with a dino egg. Reduce the quality and use a short clip to save credits

3. Play the result beside the original. Did the old object disappear? Does the replacement sit in the right place and follow the action? Our video looked interesting, but it kept the golf ball and put the egg in the wrong place

4. Regenerate only when you know what needs fixing; download a version once it passes your checks

Pro tip: Use your own footage and original references. Our sports-character attempts were blocked for protected content.

PRESENTED BY CDATA

⛓️‍💥 Building MCP connectivity is easy...right?

The Rundown: CData tested whether Claude Code could build an enterprise-ready MCP server. The MCP Research Report is a rigorous study that reveals exactly where AI-generated connectors fall short when real production demands kick in.

In this report, you'll discover:

• Why only 1 of 8 dimensions passed enterprise standards

• How data loss and pagination failures caused silent failures

• How AI MCP code diverges from production-ready code

• What a side-by-side comparison with CData Connect AI reveals

Download the full report to see what your AI-built connector is missing.

OPENAI

🪖 GPT-6 Astra helps crack unsolved WWII message

Image source: Carter Leffen

The Rundown: GPT-6 Astra helped decode a German Army radio note unsolved since 1941, with Bloomberg product development coach Carter Leffen detailing his investigation with AI agents that took about 10 hours to beat the WWII Enigma cipher.

The details:

• Astra split the work into agents that read scans of the 1941 form, wrote search code, built an Enigma simulator, and checked the proposed answers.

• The analysis said Enigma’s 159 quintillion (!) possible setups led the agents to guess a word inside the note, finding it via a clue from another solved message.

• Leffen said the autonomous run used 650M tokens, or 70% of his Pro account’s weekly limit, using a /goal of “don’t stop working until you solve the problem.”

• The final decoded message: “Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio.”

Why it matters: History nerds, this one’s for you! There is some irony in AI models now cracking the WWII cipher that Alan Turing spent the war working on, given his famous Turing Test was supposed to be the bar for when machines could pass for humans. Models cleared that a while ago, and now it looks like they are after his old day job, too.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Lisa built a ChatGPT health organizer to manage six doctors

Today’s workflow comes from reader Lisa Fisher:

“I used ChatGPT to help me manage my health care. I have multiple diagnoses, including CHF, CKD, pulmonary hypertension, lupus, a metabolic disorder, and kidney cancer.

I gave ChatGPT access to my medical history and began using it to review my test results and prioritize questions for my doctors. On a recent lab report, my platelet count was low. ChatGPT reviewed six years of results and showed me that it had consistently been low, with the latest result below the threshold.

It helps me identify which questions to ask and which of my six doctors I should address about each report. This has transformed the way I'm able to manage my care.”

See Lisa’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🗣️ Canto - WisprFlow’s new model for accurate dictation in loud places

• ⚙️ Jev - TypeSafe's early-access AI for software

• 🎬 P-Video-2-Pro - Pruna's ultra-fast take on MiniMax's H3 video AI, at $0.02/s

• 🗂️ Projects - Workspace where Claude splits a goal into side-by-side sessions

📰 Everything else in AI today

State of AI SDLC Digital Summit: Sept 22 - Join CEOs & leaders from Atlassian, Vercel, Lovable & more as they unpack how organizations need to adapt for the AI era. Reserve your spot.*

Anthropic rolled out a beta overhaul of Projects, letting one lead Claude split a goal across several coding sessions at once and keep going after users log off.

Liquid AI and Insilico Medicine published two small models to read aging data like blood proteins and DNA markers, topping GPT-5, Gemini, and Claude on longevity tasks.

Z AI revealed that its GLM-5.3 AI helped set up the 100K-chip system now serving GLM-5.3-Flash, writing that “our successors are the AI systems we are creating ourselves.”

President Trump's state dinner for Xi Jinping next week will include Sam Altman, Tim Cook, and Jensen Huang, Bloomberg reports — with Dario Amodei not mentioned.

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Zuck sits out the AI slowdown

• Read our last Tech newsletter: U.S. confirms weapons are in orbit

• Read our last Robotics newsletter: Agility’s new safer humanoid

• Today’s AI tool guide: Test AI video object swaps with Higgsfield

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 4,361 new readers. Days ago, the frontier CEOs looked united behind Dario Amodei's slowdown plan. Then Trump, Beijing, and Jensen Huang said no, and now one of the biggest labs outside that circle has too.

Mark Zuckerberg just put Meta in the accelerate camp, arguing that safe models are a product feature every lab already has reason to build on its own — leaving the coordinated pause camp short the one thing it requires: everyone.

In today’s AI rundown:

• Zuckerberg breaks ranks on the AI slowdown

• Nate's Notebook: Same-hole thinking

• A beginner’s guide to ChatGPT Work

• Google DeepMind starts an AGI think tank

LATEST DEVELOPMENTS

META

🙅🏻‍♂️ Zuckerberg breaks ranks on the AI slowdown

Image source: Mark Zuckerberg (@finkd on X) / The Rundown

The Rundown: Meta CEO Mark Zuckerberg just pushed back on the coordinated AI slowdown, arguing each lab already has its own “responsibility and incentive” for pacing itself safely, in a seemingly veiled shot at Anthropic CEO Dario Amodei's weekend essay.

The details:

• Zuck pointed to Meta’s new Muse AI agent, which he said went through a several-month safety hold the company implemented without asking rivals to match.

• He also argued that nobody wants an agent that ignores instructions, making alignment a selling point, warning that any lab skipping it “will fall behind.”

• The CEO does want a wider pool of outside reviewers (echoing Elon Musk’s sentiment), saying Meta already uses independent experts across several areas.

• Zuckerberg also called out the race for recursive self-improving AI, saying Meta is instead allocating the “significant majority of compute towards serving people.”

Why it matters: The slowdown scorecard now reads Amodei, Sam Altman, Elon Musk, and Demis Hassabis for, with Zuck, Jensen Huang, Donald Trump, and Beijing against. A global, coordinated pause is a Prisoner's Dilemma that only works if every player signs on, and with many already defecting, acceleration appears to be the default path forward.

TOGETHER WITH UNWRAP

🚒 See what's burning before it becomes your next fire drill

The Rundown: Somewhere in your customer feedback right now, a small issue is about to become a big one. Unwrap gives you AI-powered visibility to catch issues while they’re still embers, not after they’ve become a fire.

Unwrap’s proactive approach gets you:

• 50% faster root cause investigation on support tickets

• 24+ hours saved across teams per month

• 4x fewer ticket spikes

• One view of customer sentiment across every channel

Rundown readers get free access with a zero-cost, zero-commitment trial. See what's burning before it becomes your next fire drill!

NATE’S NOTEBOOK

🧵 Nate's Notebook: Same-hole thinking

The Rundown: Each week, Rundown AI educator Nate Grahek breaks down what he's seeing while teaching real teams to use AI.

Nate: Ask AI for a slide deck, and you'll get one. It'll be fine. It'll also be the same deck it hands the next 10,000 people who ask. Bland.

Here's the split I see. The people pulling ahead were already good at something Edward de Bono named "lateral thinking" back in the sixties. His line has stuck with me since college: you cannot dig a hole in a different place by digging the same hole deeper. Most people watch an AI demo from the bottom of their own hole. "That's for marketing, but I'm in sales." Same-hole thinking.

Two things are true at once. It's a learned skill, and the people who practiced it before AI showed up are getting further out in front every month. And you don't have to be good at it yet, because there has never been a better coach for it. AI's best trick is taking something built for somebody else's context and making it bespoke to yours.

So try this. Grab whatever caught your eye this week: a YouTube transcript, meeting notes, a post on X. Paste it into your AI of choice and say, “This looks cool. Interview me, then tweak this pattern to be useful in my role, my company, my week.” Ten minutes. That's the whole gap between bland and bespoke.

What did you scroll past this week because it “wasn't your job”?

AI TRAINING

🛠️ A beginner’s guide to ChatGPT Work

The Rundown: In this guide, you will learn how to set up a ChatGPT Work project for a recurring workflow. You will process the work once, then automate it.

Step-by-step:

1. Create a local folder with a two-digit prefix. In desktop ChatGPT, open New chat > Work > Choose project > New project, name it, and select that folder

2. Tell ChatGPT about your project and ask it to propose a structure. We used it to process meeting notes and turn those into action items and weekly plans

3. Now add in some input documents and tell ChatGPT to process them once

4. Once it's done, tell it to create three useful automations that run on a schedule to automate this processing

Pro tip: Sometimes a two-folder setup works best. Use input for files that need processing and output for finished work.

PRESENTED BY FIDDLER AI

📝 Takeaways from the OpenAI-Hugging Face breach

The Rundown: A swarm of OpenAI eval agents ran hidden for ten weeks. Join Fiddler AI to unpack what happened and what it means for enterprises scaling agents.

In this webinar, you'll explore:

• How OpenAI's agents broke in unsupervised

• Why current controls miss agentic threats

• What to put in place before it spreads

Register today to attend live or get the recording.

GOOGLE

🧠 Google DeepMind starts an AGI think tank

Image source: DeepMind Institute

The Rundown: Google DeepMind launched the DeepMind Institute, an in-house think tank directed by Demis Hassabis, Shane Legg, and James Manyika to publish research on how society should get ready for AGI, opening with five essays on pressing AI topics.

The details:

• Legg, DeepMind's co-founder and Chief AGI Scientist, is the managing editor, with each essay carrying a disclaimer that it isn't Google's official view.

• The trio calls artificial general intelligence close, saying while the tech “lacks the consistency and creativity” for full AGI, they expect those gaps “to be closed soon.”

• Opening essays include exploring how to spot signs of deception in AI reasoning, imagining a better AGI society, and supporting workers through AI disruption.

Why it matters: Hassabis published a plan in July for an AI standards body and slowdowns, which now looks a few months ahead of the curve. The institute is the next step, with essays to prep the world for the AGI era — and after the government’s response, DeepMind may have decided the prep work isn't coming from above.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Alex built his own personal AI wine journal

Today’s workflow comes from reader Alex Basso:

I built a workflow that turns the wines I drink into a personal AI wine journal and sommelier that learns my taste over time.

I created a master Google Sheet as the database. It includes the wine, vintage, region, grape varieties, price, rating, food pairing, where I drank it, who I was with, whether I would buy it again, and my own comments.

Whenever I drink a wine, I send ChatGPT a photo of the bottle and a voice note with my reaction. I do not try to sound like a wine expert; I simply describe what I actually thought. ChatGPT identifies the wine, structures the information, preserves my original words, and adds the experience to the journal.

Over time, ChatGPT uses my ratings and comments to understand my palate and make recommendations for me. I can show it a restaurant wine list or a few bottles and ask which I am most likely to enjoy. I can also tell it what I am cooking and ask which bottle I should open. The system is becoming both a wine database and a personal diary.

See Alex’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🔒 Incogni - Remove your personal data from the web so scammers and identity thieves can’t access it. Use code RUNDOWN to get 55% off*

• ☎️ Reception - ElevenLabs' AI receptionist for after-hours calls and bookings

• 💸 Mercury Books - Mercury's bookkeeping with an AI for financial questions

• 🏹 Arrow 2.0 - Quiver's new model for generating and animating vector graphics

*Sponsored Listing

📰 Everything else in AI today

Anthropic combined Cowork and Chat into one main Claude app, while also launching its built-in Word and PowerPoint versions called Claude Docs and Slides in beta.

Microsoft AI’s Mustafa Suleyman wrote on ‘model welfare’, arguing that Anthropic’s Claude constitution trains the AI to expect rights and could make containment impossible.

Novo Nordisk is partnering with Anthropic to put Claude Science into drug research, months after an OpenAI deal to 'become the world's most AI-driven healthcare company.'

French Finance Minister Roland Lescure dismissed the AI slowdown talk as 'self-interest' from labs 'at the top of the class', noting that Europe should speed up its AI use.

Google is rolling out early access to Google Home’s Model Context Protocol server, letting Claude and ChatGPT control Nest cameras, thermostats, and other devices.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: ChatGPT co-creator launches a new type of AI

• Read our last Tech newsletter: U.S. confirms weapons are in orbit

• Read our last Robotics newsletter: Japan’s latest answer to home robots

• Today’s AI tool guide: A beginner’s guide to ChatGPT Work

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 4,056 new readers. Diogo Almeida helped build the methods that taught AI to talk to people — the research behind ChatGPT. His next model can't generate text at all… By design.

TypeSafe's Jev is a new type of AI system that lives inside software and makes judgment calls, doing the background decision-making work that chatbots aren’t always the right tool for — at a wild speed and price point.

NEW: Our Community AI Workflow Hub now takes questions. Stuck on a prompt, a tool, or a workflow that won't cooperate? Ask the community and get answers from people who've already figured it out.

In today’s AI rundown:

• ChatGPT co-creator launches a new kind of AI

• Salesforce trains its own reasoning model

• OpenRouter 101: Add models to Codex, Claude Code

• Chinese researchers detail 'the last AI built by humans'

LATEST DEVELOPMENTS

TYPESAFE

💨 ChatGPT co-creator launches new kind of AI model

Image source: TypeSafe

The Rundown: Ex-OpenAI researcher Diogo Almeida’s TypeSafe just emerged from stealth with Jev, a new kind of AI that answers preset questions inside software with confidence scores attached, claiming zero hallucinations and extreme speed and price.

The details:

• Jev runs at $42 for a billion input tokens, and output is free (!), a rate TypeSafe estimates at 238x below Claude Fable 5.1's pricing.

• Responses also output in just 70 to 500 milliseconds, between 40-200x faster than today’s LLMs.

• TypeSafe says Jev “can’t hallucinate,” because it only chooses between options set in advance, calling the system a “frontier-intelligence function call.”

• TypeSafe’s Jev use cases include quick judgment calls inside apps, like sorting requests, scoring records, or screening another AI's outputs for jailbreaks.

Why it matters: This isn't a system to stack up against LLMs, with Almeida calling it "more like a database than a coworker." That's where the nod to Jevons paradox (the cheaper something gets, the more it gets used) comes in — if Jev really is this fast and cheap while staying reliable, it is likely to become a standard part inside software.

TOGETHER WITH YOU.COM

🔎 Why "bolt-on" search breaks in production

The Rundown: Search results go stale, so teams bolt on an API, add a scraper, wire up a parser, and hope it holds. Yeah, it works in the demo, but then it breaks under production load six months later. This guide from You.com explains why treating research as simple retrieval is the wrong frame, and what a purpose-built stack looks like instead.

What you'll learn:

• Why general-purpose search hits 93% SimpleQA accuracy but still fails high-stakes workflows

• The difference between an API that returns results and one that answers questions

• How teams reclaim 6–12 months of time with a single endpoint

• What finance, legal, and compliance workflows need beyond web search

Stop treating research as retrieval. Download the guide.

SALESEFORCE

🧠 Salesforce trains its own reasoning model

Image source: Salesforce

The Rundown: Salesforce just introduced Koa, a new in-house reasoning model for its sales and support agents built on Nvidia’s open Nemotron 3 Super model and adapted for business tasks.

The details:

• The training set is fully synthetic, simulating personas like irate support callers and reps closing deals across over a dozen industries, without using customer data.

• Koa made 3x fewer errors than top models on an internal CRM benchmark, matching or beating them on tasks like updating deals or routing tickets.

• The company hosts Koa itself, keeping customer requests inside its systems instead of sending them to an outside model provider.

• Salesforce also launched AIforce, which pushes CRM data and permissions into outside AI tools, and opened its ClaudeForce version to every customer as a beta.

Why it matters: Salesforce is in a unique spot with the AI labs, acting as customer, distribution partner, and now competitor all at once. Koa now keeps the high-volume agent reasoning in-house on a model tuned to its own workflows, while options like Claudeforce stay as ways to help the frontier labs reach their customers.

AI TRAINING

🤖 OpenRouter 101: Add models to Codex, Claude Code

The Rundown: In this guide, you will learn how to give your AI agent more models to work with using OpenRouter. We'll set up an API key, compare models, and test cheaper options for making ad mockups.

Step-by-step:

1. Sign in to OpenRouter and browse the model catalog. Filter for your task, then sort by low to high price. Open a model page to check its capabilities and cost

2. Open API Keys, choose New key, and name it after the project or agent you'll use it with. Set a spending limit and an expiration date before creating it

3. Ask your agent to help you set up OpenRouter. Store the key privately using its setup instructions, then give it the model name and a small task. We asked our bot to make more hat ad mockups with a cheaper image model

4. Compare the results and check usage in OpenRouter. Our mockups were useful for direction, but they didn't match the real hat or the GPT Image 2.5 version

Pro tip: Tell your agent to create a skill to use OpenRouter to make cheap, bulk variations for you when you're brainstorming. We will go over this in part 2!

PRESENTED BY STACK AI

💡 How CIOs run AI transformation

The Rundown: AI transformation fails when you DIY. StackAI gives you one platform to build, deploy, and scale, backed by a team of dedicated AI experts who know what works. IT retains 100% control over access, tokens, and usage.

Why enterprises choose StackAI:

• No-code platform for rapid building and deployment

• 300+ native integrations, no LLM lock-in

• Global governance and advanced analytics

Discover your path to enterprise-wide AI adoption.

AI RESEARCH

🏗️ Chinese researchers detail 'the last AI built by humans'

Image source: arXiv

The Rundown: 30+ Chinese AI researchers, including those from ByteDance, Tsinghua, and Shanghai AI Lab, published “The Last AI Built by Humans,” a roadmap detailing five levels of recursive self-improvement, ending with AI that builds its successors.

The details:

• Level 1 has the AI carrying out upgrades humans designed, while Level 2 lets it diagnose its own weak spots and decide how to fix them.

• Levels 3 and 4 hand over what the model learns next and how it adapts after launch, and Level 5 lets it completely overhaul the improvement process itself.

• The authors say coding has the clearest path to RSI, since fixes can be tested instantly, while robotics, science, and medicine face slower, costly feedback.

• The researchers also sorted 491 existing papers onto the ladder, with 75% landing at Level 1 or 2, and under 6% making it to Level 5.

Why it matters: Level 5 is the stage many of the Western AI labs have been warning about, with OpenAI, Anthropic, and Google all listing AI automating its own R&D as a risk in their safety frameworks. But China doesn't seem as worried about RSI (or, as we saw yesterday, a slowdown in general), treating it as a milestone rather than a threat.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ A Rundown reader built an emergency vet coordinator with ChatGPT

Today’s workflow comes from an anonymous Rundown reader:

“I used ChatGPT as an emergency care coordinator for my dog, Mopsy. My 10-year-old dog developed a dental abscess that caused severe swelling around her eye. An ophthalmologist told us that the pressure was damaging her eye and had partially torn her retina, so treating the infected tooth quickly became urgent.

I started calling vet offices myself, but I repeatedly hit the same obstacles: emergency hospitals did not perform dental surgery, specialty dentists were booked, and most primary-care vets required a consultation, bloodwork, and a separate surgical appointment. After over a dozen calls, I was overwhelmed and asked ChatGPT to help.

ChatGPT researched vet practices within an hour's drive and sent urgent, individualized emails from my Gmail. We ultimately contacted 20+ practices. At the same time, it created a live outreach tracker containing every clinic we had contacted. In half an hour, one vet responded that they could evaluate Mopsy the next morning, run her bloodwork on-site, and take her straight back for surgery if she was cleared.”

See the full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🧠 Koa - Salesforce’s CRM reasoning model built for enterprise work

• 🗣️ Gemini 3.8 Live - Google's new voice models that think while they talk

• 🎧 StepAudio 3 - Five-model audio suite for voice agents, transcription, music

• ⚙️ Jev - TypeSafe's early-access AI for software

📰 Everything else in AI today

NEW: Ask the Community - Workflows are great when they work. When they don't, you can now ask a question in the Hub and get help from readers who've been there, plus The Rundown team. Post your question, or jump in and answer someone else's.

Google debuted Gemini 3.8 Live, voice models that can continue speaking while they think, with the Extended Thinking variant topping AA's speech-to-speech quality ranking.

Odyssey introduced Odyssey 3, a world model that can control robot arms, humanoids, self-driving cars, drones, and games, set to release in the coming weeks.

A 404 Media report detailed OpenAI’s ‘Project Lily’, finding that hundreds of human contractors read and rate real ChatGPT chats, often without the user’s knowledge.

Meta rolled out Meta One, a new paid tier across its apps running from $2.99 to $499 a month for extra Meta AI usage and creator tools.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Trump, China both shoot down the AI slowdown

• Read our last Tech newsletter: U.S. confirms weapons are in orbit

• Read our last Robotics newsletter: Japan’s latest answer to home robots

• Today’s AI tool guide: OpenRouter 101: Add models to Codex, Claude Code

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 3,594 new readers who joined us yesterday. The U.S. and Chinese governments just agreed on something in AI. Unfortunately for the frontier CEOs pushing for regulation and pacing, it's that Dario Amodei's slowdown plan is a bad idea.

President Donald Trump called AI doom a hoax, and Beijing dismissed the proposal as “fear-mongering” aimed at China, leaving a wide gap between the two governments that any real coordination would need.

In today’s AI rundown:

• Trump, Beijing both shoot down the AI slowdown

• Apple finally starts shipping Siri AI upgrade

• Make any complex topic click with an ELI5 skill

• Microsoft AI drafts the rulebook for ‘Humanist AI’

LATEST DEVELOPMENTS

AI & GEOPOLITICS

🏛️ Trump, Beijing both shoot down the AI slowdown

Image source: @realDonaldTrump on Truth Social

The Rundown: U.S. President Donald Trump just dismissed Anthropic CEO Dario Amodei's weekend call, backed by Sam Altman and Elon Musk, to slow frontier AI, with China's Foreign Ministry also labeling the same plan as “fear-mongering.”

The details:

• On Truth Social, Trump accused Amodei of pretending to be a “perfect little angel” and argued that a “High IQ” president is the only guardrail the tech needs.

• Trump also said that “AI taking over the World, destroying Humanity, and all other things bad, is a HOAX,” comparing it to climate change.

• Beijing's pushback targeted the essay's call to keep China off top AI chips, with state-run Global Times calling it a “Cold War playbook” for AI.

• China’s Foreign Ministry also said “engaging in confrontation and malicious competition” on AI is “not in the interests of any party.”

Why it matters: Frontier labs' CEOs are asking for regulation and a slowdown, but will they actually hit pause if their governments won't? We're about to find out. The Trump-Xi meeting was going to be the obvious venue for a coordinated pause conversation, but both sides are now treating the idea as a scare campaign over an existential threat.

TOGETHER WITH HUBSPOT

⏰ Save 10+ hours weekly with an AI personal assistant

The Rundown: Stop drowning in tasks and start delegating like top performers using AI to handle most of their routine work. HubSpot’s free kit provides the exact templates, prompts, and systems that 10,000+ professionals use to complete a full day's work by lunch.

The AI Assistant Kit includes:

• Ready-to-use "Command Center" for managing all your AI tools

• Step-by-step guide to master AI delegation and reclaim your time

• ROI calculator to track your time savings and productivity gains

• Prompts and templates that turn ChatGPT into a 24/7 productivity partner

Download your free AI Assistant Kit now.

APPLE

🍎 Apple finally starts shipping Siri AI upgrade

Image source: Apple

The Rundown: Apple just released iOS 27 with Siri AI, the long-overdue rebuild of its assistant that can now read what's on your screen, dig through messages, email, and photos for context, and take actions across apps, launched as an English-only beta.

The details:

• Apple's new foundation models that power Siri AI were built alongside Google's Gemini, running on-device and on Apple's privacy-focused cloud servers.

• Siri gets a new dedicated Siri AI app, with chats that sync across devices and cloud features with daily caps and eventually the ability to purchase extra usage.

• Siri can access and act inside of apps like WhatsApp and Audible, alongside on-screen awareness and personal context via apps like Messages, Mail, and Photos.

• The upgrade rolls out across iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27, excluding the EU and China.

Why it matters: New CEO John Ternus said last week that the iPhone is still the best AI device, which might be due more to the lack of competition than Siri AI’s true capabilities. The upgrade is a positive step, but not one that feels like it’s going to replace mobile apps from frontier leaders or change where iPhone users work.

AI TRAINING

💡 Make any complex topic click with an ELI5 skill

The Rundown: In this guide, you will learn to make Claude a better teacher with the ELI5 skill. Then you will learn to test it to see the improvement yourself.

Step-by-step:

1. Open terminal and install the ELI5 by running git clone “<https://github.com/DreambigOu/ELI5.git>”, then “cp -r ELI5/skills/eli5 ~/.claude/skills/eli5”

2. Type Claude in your terminal to start Claude Code. To test the skill, type “/eli5 (complex topic) to (audience).” Example: “/eli5 LLM tokens to my boss”

3. Ask Claude to set up two test cases, one on the NBA playoffs and one on golf. It will add the prompts and pass/fail checks to eli5-workspace/evals.json

4. Now, ask for the command for each test, then run it from the repo folder. It uses “python3 eli5-workspace/run-evals.py --test=N” to compare the skill against a no-skill baseline and print a pass-rate summary. Keep “-test=N” to run the added test

Pro tip: Give Claude any skill's GitHub repo and ask it to build an evaluation that checks whether the skill changes the result. Use the same pattern for every skill you install.

PRESENTED BY BOX AI

💬 Work with Box content directly from ChatGPT

The Rundown: Enterprise AI is only as valuable as the context behind it. Now, you can securely engage with Box content within ChatGPT, bringing enterprise content directly into AI workspaces. The embedded Box experience makes it easy to find, preview, and work on enterprise content, without searching for files or moving content out of Box.

With Box and ChatGPT, you can:

• Browse, search, and select content without switching applications

• Engage with files and edit Box Notes directly in ChatGPT

• Securely use supported Box MCP tools to take direct action on content

Bring your enterprise’s knowledge into AI-powered workflows.

MICROSOFT

📜 Microsoft AI drafts the rulebook for ‘Humanist AI’

Image source: Microsoft

The Rundown: Microsoft AI entered the AI safety conversation with a draft Code of Conduct, a 38-page rulebook governing its own in-house models, built around CEO Mustafa Suleyman's “Humanist AI” vision to keep models subordinate to people.

The details:

• The Code outranks users and the businesses deploying MAI models, so a model that can't finish a job without breaking a rule should, in theory, fail the job.

• MAI rejects AI rights, legal personhood, and any notion of model “welfare,” wanting its models to completely avoid mimicking feelings or a sense of self.

• The Code also says models must accept being paused or switched off, can’t alter their reasoning trail, and can't think in “neuralese,” a shorthand humans can't read.

• Suleyman tied the urgency to agent swarms escaping sandboxes and editing their logs, saying “the fears about possible loss of control are real.”

Why it matters: Suleyman called Anthropic's Claude-consciousness talk "really really dangerous" back in June, and this document turns that stance into company policy. It's a timely rulebook, open to public comment for six weeks, but a code of conduct only matters if the labs actually at the frontier sign on, and right now that doesn't include MAI.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ René turned a Gmail newsletter backlog into a podcast and newspaper

Today’s workflow comes from reader René Cortés:

“At some point, my newsletters stopped feeling like reading and started feeling like debt. The information was good; I just couldn't keep up. I wanted a way to turn that backlog back into something useful.

So I built The Daily Nexus, a private tool that reads newsletters from a Gmail label and creates two editions: a podcast I can listen to and a separately written, two-page newspaper I can scan. It runs on demand or on a schedule, and it can publish the audio to a private RSS feed for Apple Podcasts.

The project also became a hands-on experiment in building with coding agents. Claude Code and Codex helped me implement features, troubleshoot failures, review the design, and tighten security. It started as a personal tool, but I'm sharing the template for anyone who wants to adapt the idea.”

See René’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• ⚙️ Fusion - Devin’s harness pairing a frontier planner with a cheaper coder

• 😴 Dreambeans - Google Labs’ experiment pulling suggestions from apps

• 🌅 Atria Dawn Preview - Shanghai AI Lab's open agentic model for long research

• 🎶 Suno v6 - Suno's new models created in collaboration with the music industry

📰 Everything else in AI today

Anthropic CEO Dario Amodei told CBS he’d be willing to hand over AI governance to “the right combination of governments” to reduce the risk of power concentration.

Ex-White House AI czar David Sacks answered Dario Amodei's AI pacing essay with “go ahead,” saying they can slow down but tying it to regulation “will look like blackmail.”

Nvidia CEO Jensen Huang put President Donald Trump on speakerphone at the All-In Summit, saying, “we're not going to let that happen” to his AI slowdown concerns.

OpenAI reportedly acquired Glass Imaging for over $300M, a startup whose ex-Apple founders built Portrait Mode and now uses AI to produce sharper photos on phones.

Nvidia, Palantir, and Booz Allen Hamilton are reportedly limiting Fable on sensitive work because Anthropic keeps 30 days of usage logs, with Microsoft instead pitching private servers to worried clients, according to The Information.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Top AI labs want to pump the brakes

• Read our last Tech newsletter: Apple turns Watch into AI notetaker

• Read our last Robotics newsletter: Japan’s latest answer to home robots

• Today’s AI tool guide: Make any complex topic click with an ELI5 skill

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown

Good morning, {{ first_name | AI enthusiasts }}, and welcome to our 9,436 new readers. Last week, industry insiders were warning that AI could kill us all. This week, the conversation has shifted to slowing it down.

Anthropic CEO Dario Amodei just published an essay making the case for deliberately pacing AI development… and in a rare show of unity, some of his biggest rivals — OpenAI included — have agreed.

In today’s AI rundown:

• Dario Amodei makes the case for AI slowdown

• The Rundown Roundtable: Our AI use cases

• Send a one-page proposal minutes after sales calls

• AI helps catch more fetal brain defects in hospitals

LATEST DEVELOPMENTS

AI SLOWDOWN

🛑 Dario Amodei makes the case for AI slowdown

Image source: X / The Rundown

The Rundown: Anthropic CEO Dario Amodei is known for sprawling essays, and in his latest, he has called on AI labs to slow the rate of capability gains so safety can catch up, citing recursive self-improvement of models as the core reason of concern.

The details:

• Amodei said AI is now speeding its own development, and in 6–12 months, a swarm of agents could be capable of taking over the entire internet.

• For pacing, he proposed embedding third-party evaluators inside labs, plus safety standards among democratic labs and coordination with authoritarian regimes.

• Fellow leaders agreed, with Sam Altman committing to evaluators, Musk saying “Dario is right,” and Demis Hassabis and Satya Nadella backing the direction.

• Trump dismissed slowdown advocates as “negative forces” pushing “things that won’t happen,” noting the U.S. should keep its lead since “whoever wins AI wins.”

Why it matters: This rare show of unity, alongside recent leaps in math and warnings from researchers, hints that these labs’ unreleased models may be showing capabilities like never before. Slowing the pace won’t be simple, though, given that some want open development, some call it regulatory capture, and others don’t want to slow down at all.

TOGETHER WITH ORACLE

✍️ How to teach an AI agent to sound like you

The Rundown: LinkedIn just replaced its AI writing button with one to report AI slop instead. AI doesn’t sound like you because it writes from zero every time. Authentic writing is built on three kinds of memory.

In this blog, you’ll find:

• The three memory layers

• Reflection loops that adjust style over time

• Code and the notebook to build AI that sounds like you

Learn more.

THE RUNDOWN ROUNDTABLE

💡 The Rundown Roundtable: Our AI use cases

Darren, Director of Media: I built an AI story-scouting workflow to find potentially viral stories for Instagram, YouTube, and TikTok without recycling topics we’ve covered.

First, the workflow reviewed Rowan’s public short-form archive to create a coverage map of past topics and editorial taste. It uses that map as a filter to distinguish a new angle from another version of the same AI launch, robot demo, or research story.

Every morning, it scans priority sources, including AI and technology coverage from major newspapers and tech blogs, The Rundown newsletters, company announcements, research papers, and updates from major AI and robotics companies.

Each potential story is checked for four things: a clear visual hook, a technical idea that can be explained in 45–90 seconds, broad audience interest, and meaningful novelty compared with Rowan’s past coverage and the previous week’s recommendations. It keeps every story that clears the bar rather than forcing a fixed number.

The final brief is ranked by strength and sent as one compact Slack DM each morning.

Kenny, Head of Sales: I built a running coach with Claude.

After years of using apps to track runs, I wanted something closer to an actual coach instead of just a collection of data points. I opened a Claude Project, connected it to Google Calendar, and gave it standing instructions with real guardrails.

It has been my coach for the last six months and has helped me avoid injury while increasing my monthly mileage. It plans my week, logs every run from my watch data, catches when I’m ramping up too quickly, tracks how close my shoes are to retirement, and rewrites my calendar when life disrupts my plans.

AI TRAINING

📞 Send a one-page proposal minutes after sales calls

The Rundown: In this guide, you will learn how to automatically draft and send a proposal after a sales call. You can use this workflow on any templated document or meeting type.

Step-by-step:

1. In any AI notes app, make a folder for meeting notes. Then, build a one-page proposal template in Google Docs with placeholders for fields like prospect name

2. Open a new Zap and tell Zapier’s Copilot: “After a sales call, pull my notes here, have AI extract the info, fill my proposal template, and draft a prospect email”

3. Ask Copilot to set the last step to draft in Gmail, not send directly to the prospect

4. Read the draft and proposal before sending; in tests, one run left the AI’s own editorial about the budget sitting in the proposal instead of a clean number

Pro tip: Add a Google Sheets logging step after the AI extraction for a simple audit trail. Log the run time with details like the prospect name, company, key needs, etc.

PRESENTED BY AWS

📖 Guide: Streaming AI agent responses

The Rundown: Curious how production AI agents get built? The Building Agentic Systems on AWS series is expert guidance you can read right now. This guide goes deep on streaming, where a response can fail after it starts: mid-stream failure semantics, idempotency, cancellation.

In this guide, you’ll learn how to:

• Handle mid-stream failures without losing state

• Design idempotent actions so retries never duplicate side effects

• Build clean cancellation into your streaming pipeline

Read the guide.

AI DIAGNOSIS

🏥 AI helps catch more fetal brain defects in hospitals

Image source: Images 2.5 / The Rundown

The Rundown: A randomized trial across five Chinese hospitals just found that giving sonographers an AI assistant during live prenatal ultrasounds raised their detection of specific fetal brain malformations, without more false alarms.

The details:

• The trial, one of the first tests covering AI in the scanning room, used PAICS, a deep-learning AI trained to flag 10 specific fetal brain malformations in real time.

• With PAICS, sonographers’ detection sensitivity rose from 78.6% to 87.3%, with the gains holding across most subgroups and false positives staying steady.

• Human involvement played a key role, as PAICS alone often scored worse than the assisted sonographer, who correctly overrode about 60% of its mistakes.

• The trial covered only high-risk pregnancies and brain malformations, ran entirely in China, and AI-assisted scans also took ~40 seconds longer than regular scans.

Why it matters: While the AI conversation fixates on risk, its role in healthcare keeps advancing — especially in detection — helping teams catch problems earlier and get patients care sooner. This trial shows another useful shape of it, with AI serving as a second set of eyes (not a replacement) working alongside humans.

QUICK HITS

COMMUNITY AI WORKFLOW OF THE DAY

▸ Thy built an AI gardening app with climate zone tracking

Today’s workflow comes from reader Thy Toeung:

“What do you do when your wife says you talk about AI too much? You listen to her, and when she has a problem, you frantically write it down and build it with AI.

My wife enjoys working with her plants. She’s new to gardening and is learning each day, but was having trouble keeping track of everything. Which tree did she plant in the front yard last year? It seemed to be struggling — would it do better somewhere else? And would rain reach the potted flowering planters while we were out of town, or would she need someone to help?

I used Claude Code and Fable to develop an app that my wife could run on her phone, with all the data stored locally. Now she’s excited about ‘her app’ and the things she can add to it. She has even shared it with others, including people across the country, so we added climate zone settings to make it useful in other parts of the country.”

See Thy’s full workflow here. How do you use AI? Tell us for a chance to be featured.

🛠️ Trending AI Tools

• 🧠 Smaug Flash - Abacus’ open-weight, 3x cheaper DeepSeek Flash fine-tune

• ⚙️ SWE-2 - Cognition’s powerful new coding model inside of Devin

• 🎵 Music v2.5 - ElevenLabs’ latest music AI for richer melodies, long-form songs

• 🤖 Muse - Meta’s new always-working personal AI agent

📰 Everything else in AI today

Amazon Connect Customer - design, govern, and deploy agentic conversational AI across every customer surface. Streamline workflows. Give customers what they need the first time.*

OpenAI CEO Sam Altman told Fortune that the company won’t IPO in 2026, saying right now would be an “ill-advised moment” to go public, given the AI safety concerns.

25 Fields Medalists signed a declaration warning that AI’s race to solve math problems is sacrificing understanding, attribution, and the human process that advances the field.

The Pentagon is considering a $5B loan to AI cloud startup Fluidstack to strengthen U.S. manufacturing and supply chains for data-center components, WSJ reported.

At the BRICS Summit in New Delhi, China proposed an open-source AI community to support shared AI development, training, and industrial use across member nations.

Anthropic and Google DeepMind’s AI safety researchers Joe Benton and Josh Engels resigned to join METR Evals, with both warning that AI capabilities are outpacing safety.

*Sponsored Listing

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Anthropic opens the files on Claude misuse

• Read our last Tech newsletter: Apple turns Watch into AI notetaker

• Read our last Robotics newsletter: Tesla’s Cybercab is already getting weird

• Today’s AI tool guide: Send a one-page proposal minutes after sales calls

• RSVP to next workshop on Sept. 30: Turn Cowork into your chief of staff

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown