Good morning. The safety news is back.

The three dismissed OpenAI employees working on safety have responded. Their take: OpenAI should agree to third-party safety auditors, preserve the monitorability of models, reaffirm its prior open culture of safety, and affirm an open and transparent culture of dialogue. On the other hand, Claude's terms and conditions now address bullying behavior toward its models.

Dynamic AI chat versus dynamic AI work:

On Thursday, we learned about OpenAI’s GPT-6 with an intelligent UI. This development works towards software that adapts to the user, instead of the user having to learn rigid software. Presently, this is in chat mode. Meanwhile, Claude released "dashboards in motion," aiming to make traditionally separate applications dynamic. This aims to make work mode more dynamic.

Let’s dive in 👇

Today’s Menu 🥠

• Who gets to question AI safety? The labs, the researchers and outside scrutiny.

• Claude builds the dashboard. More threats to UI SaaS?

• Fast Snacks: Revenue math, security, AI in the clinic and fake bylines.

• Toolbox: Read documents, steer tasks and check AI watermarks.

• Snack Quiz: Which robot photo is real?

Friday, October 9 · Read Time: 6 minutes

The Big Bite · Safety and Accountability
🛡️ Who Gets to Question the AI Labs?

The Bite: Two developments put the people writing AI safety rules under a closer look.

Three researchers fired by OpenAI, Tomek Korbak, Jasmine Wang and Mikita Balesni, have published a letter warning that their dismissals could make colleagues afraid to raise concerns or work with outside experts. They say they acted within their roles and the working norms at the time. Their account is disputed: OpenAI denies retaliation and says the dismissals followed violations involving sensitive research information.

In their words: “If conduct that was considered normal last month now constitutes grounds for sudden dismissal, everyone at OpenAI is left guessing where the line is.” Korbak, Wang and Balesni, in their letter.

Separately, Anthropic’s updated usage policy will prohibit extreme, repeated, purposeless abuse directed at its models from November 12. Ordinary frustration and model testing remain outside that restriction. Of course, this is a rule for using Claude, not evidence that the model has feelings. I ask myself, who is bullying the models? 👀

Why it Bites: AI safety should be designed to protect humanity. Establishing an open and safe culture in this domain is critical to keeping bad actors in check. However, no rulebook currently exists. It is not a trivial problem, and it raises the question: who gets to decide these rules? At the same time, it is odd that Anthropic has published rules against abusing its models. While most people using AI are not abusing these systems—except perhaps when they get frustrated trying to babysit their agentic workflows—Anthropic seems confused about model consciousness. Since AI models are not trained on bullying behavior, bullying a model will not teach it to become aggressive toward other users. Where do you land on the spectrum?

Our take: A safety promise becomes more useful when someone outside the company can examine it. Watch what access independent evaluators actually receive, what they can disclose, and whether employees know how to escalate concerns. Those details tell us more than a lab’s preferred definition of safety. As for consciousness, I think thats characteristic of living beings only, not computer models, no matter how large the data center is.

Second Bite · Claude
🤖 Your Next Dashboard Starts With a Question

The Bite: Claude, Anthropic’s AI assistant, is moving further into the work usually done in spreadsheets and presentations. On Thursday, Anthropic introduced Dashboards and Motion, two beta tools that turn a request into something you can share.

Connect company data from platforms such as Snowflake or Salesforce, then ask how this week’s sales compare with last month’s. Dashboards builds the view and updates it as the data changes. You can inspect the query behind a number and see when each chart last refreshed. Motion turns text, charts, shapes and images into editable animations, downloadable as MP4 files. Think a product walkthrough or an animated explanation of quarterly results. It animates code, rather than generating video footage.

Dashboards is in beta on paid plans. Motion is in beta on Team and Enterprise. Both are off by default for Enterprise until an admin enables them. Docs, Slides and Design also leave beta on every plan, including Free, with Enterprise access controlled by admins.

Why it Bites: User interfaces may be becoming disposable. The advantage: getting a useful first version could become significantly easier for people who cannot write database queries or animate a slide. This means fewer apps to context switch between and a much better workflow experience—assuming we can trust the output for live work.

Our take: Start with one question you already know how to answer. Check the query, refresh date and sharing permissions before circulating the result. A beautiful chart still needs someone who understands what it counts. Run a previous task performed in SQL and see if they align.

Fast Snacks 🥡

💵 A $50 Billion Figure Needs a Footnote
Bloomberg reports OpenAI’s annualized revenue is near $50 billion, below earlier estimates around $70 billion. Its report points to differences in how cloud partner sales are counted when comparing OpenAI with Anthropic. Annualized revenue projects a current pace over a year; it is not a completed year’s audited sales. The useful takeaway: check what a headline number counts before deciding what changed.

🔐 Finding Bugs Is Only Half the Job
Anthropic's new Cyber Mission brings Claude models and engineers to defenders of power grids, water systems and other critical infrastructure. It also offers participating open source projects free security scans through OSS Scanner. Those reports arrive without human review and can contain mistakes. More findings do not automatically mean safer systems. Maintainers still need time to verify them, prioritize real risks and apply fixes without breaking what everyone depends on.

🩺 An AI Intake Visit, With a Doctor Watching
Google's AMIE research chatbot met real patients before their appointments in a study newly published in The Lancet. Ninety eight patients completed both steps. Doctors supervised every conversation: none triggered a required safety stop, although supervisors noted one hallucination. Conducted at one clinic in 2025, this small study supports further testing, not replacing doctors. The useful question is whether AI can help people arrive better prepared while preserving the human judgment that keeps care safe.

📰 A Familiar Publication Can Carry a Fake Byline
OpenAI's Thursday investigation describes an operation originating in Iran that used seven fabricated journalist identities. OpenAI identified almost 100 articles published or syndicated under those names. That count does not mean 100 original articles, and the company could not identify the actor behind the operation. This is a new report about older activity. The reader lesson: a recognizable publisher helps a story travel, but original evidence still deserves a click.

Toolbox
🧰 Three useful things to try

📄 LightOnOCR3: October 8 release. This downloadable model turns scanned pages into usable text and can extract chart values as tables, with labeled boxes showing where content sits on the page. Useful for developers building document search or data extraction. It carries an Apache 2.0 license and requires deployment.

🧭 Codex steering: October 8 update. OpenAI is rolling out faster course corrections for running Codex tasks in the ChatGPT desktop app. Useful when your coding agent heads in the wrong direction. Check Settings > General to choose whether new messages steer the current task or wait for the next run. Try correcting one small task before delegating more. Availability may vary during rollout.

🔎 SynthID Detector: Earlier this week, October 7. Google’s public tool checks media for supported AI watermarks from Google and participating partners. Available globally in English, it gives editors and curious readers another check before sharing an image, video or audio clip. Try a file whose origin you already know. Note: no detected watermark does not mean the content is real.

Test Your Skill
Snack Quiz: Choose The Real Image 🔍

One is a real photograph. One is an AI recreation. Take a close look at these robots in orbit.

Image A

Image B

Photo source and answer. Original photograph: NASA / Mark Vande Hei, February 2, 2022. Archival image used for this challenge.

Good morning.

This week, Claude got cheaper to run, and a better free ChatGPT starts rolling out today. Meanwhile, Biohub’s $1.8 billion effort is a reminder that useful AI still needs expensive, carefully collected data.

The opportunity for the rest of us: spend less on access and more attention on the inputs, checks and judgment that make the output worth using. It may not be easy, but if done right, its a productivity powerhouse.

Today’s Menu

• The Big Bite: Cheaper Claude, a free ChatGPT upgrade and the meaning

• Second Bite: Biohub’s $1.8B effort to build the data for a virtual cell.

• Fast Snacks: Chip financing, proof checking, and human judgement

• Toolbox: Decision models, games you can prompt and an agent update.

Thursday, October 8, 2026 · 6 minute read

The Big Bite · Anthropic & OpenAI

💸 The AI Price War Hits Your Budget

The Bite: More useful AI is moving into smaller budgets. Anthropic has cut the price of Haiku 5.5, its model for quick, repetitive work. Meanwhile, OpenAI says GPT‑6 Luna and Intelligent UI begin rolling out to free ChatGPT users today.

Two different deals. Builders pay less to run tasks through Claude's API. ChatGPT users get access to a newer model and answers that can include interactive charts, forms and calculators.

The catch behind 90%: Haiku's token prices are 90% lower than Haiku 4.5 for prompts up to 100,000 tokens. Beyond that, the reduction is 50%. Anthropic estimates roughly 75% lower running cost on average. Your whole AI bill has not magically shrunk by 90%.

Why it Bites: A task that was too expensive to automate might now deserve a trial with Haiku. Think sorting incoming requests, summarizing documents or giving a larger agent a cheaper helper. And there’s plenty of that work.

As for Intelligent UI on GPT-6, free ChatGPT access is rolling out. It may not appear in every account right away, be on the lookout.

Our take: For Haiku, replace your pricier or underperforming agent with it on a task you understand. Count the correct completions, retries and time spent fixing mistakes. A bargain model earns its place when the finished work gets cheaper. Although the price tag definitely starts the conversation.

Second Bite · Biohub

🧬 Before AI Can Model Life, Someone Has to Measure It

Priscilla Chan and Elizabeth Montabana at Biohub’s lab in 2024 (left); imaging hardware (right). Credits: Biohub; Applied Precision Design.

The Bite: Biohub and its partners have announced a $1.8 billion combined commitment to biological data and infrastructure. The ambition: give researchers the measurements needed to build AI models that predict how cells respond to changes.

Think of a virtual cell as a place to test a biological idea digitally before choosing the next experiment in a real lab. Getting there requires enormous amounts of reliable data about how living cells actually behave.

Follow the money: This is not $1.8 billion in fresh cash handed to Biohub. The total combines funding and resources, including Biohub's existing $500 million commitment and NIH datasets built through more than $500 million in prior investment. Google DeepMind, Isomorphic Labs and Meta are collectively investing $300 million.

Why it Bites: The cell is the underlying machinery for life and AI has been making its way into bio impact. Leveraging AI to understand it could help springboard top bio solutions. Think of the 722 math papers from yesterday that had gone unsolved for a while. However, better AI needs better data to learn from. Making biological measurements accessible could help more research teams investigate disease, but this announcement is a commitment to build that foundation, not a new treatment.

Our take: Watch what becomes usable: open datasets, consistent measurements and predictions that survive real experiments. That is how a multi billion dollar headline starts turning into useful science, hopefully beyond past capabilities.

Fast Snacks

💳 Cheaper AI Still Needs Expensive Chips

Broadcom is arranging more than $50 billion to finance its custom OpenAI chips, The Wall Street Journal reports. These are financing discussions, not a completed deal. Our read: falling prices for users and rising infrastructure bills can coexist. The useful business question is whether customer revenue eventually covers the hardware.

🔍 A Proof Check Has a Translation Problem

A Tuesday preprint flags mismatches between OpenAI's written Navier Stokes argument and its translation into Lean, software that checks mathematical proofs. The authors explicitly stop short of declaring the original proof wrong. The lesson travels beyond mathematics: checking an AI's output also means checking that it answered the question you actually asked.

🤝 Model Watch: Who Is Doing Your Agent's Work?

Useful context for the rival model debate: Grok Bot's documentation says its underlying models can change, with usage and billing tied to the model handling each request. This is existing documentation, not proof that a specific Claude integration is live. Check which providers receive your task and what each completed job costs.

🧠 Better output is not the same as better judgment

A three month study of 133 patent lawyers found AI improved drafting quality. But gains on a later task completed without AI were concentrated among more experienced lawyers; juniors showed no average improvement. The study covers one profession and setting. For teams using AI to train people, the useful question is what they can still explain and evaluate after the assistant disappears.

Toolbox

🎯 Liquid d1: decide without writing an essay

Liquid released downloadable decision models, a material update to its earlier API announcement. The d1 3B model chooses or scores defined options. Its Omni 600M variant adds audio and vision, but is an early research release with audio limited to 30 seconds. Useful for a narrow routing or classification experiment. Review the model cards and Liquid's license before deployment.

🎮 Playground: turn an idea into a game

Google's experimental Playground lets users prompt, play and share games. Try a small learning game with one clear rule. Treat it as an experiment; the announced Unity integration is still coming. We have not verified universal access or pricing.

🐳 Docker Agent: a fresh update for builders

Docker Agent v1.149.0 adds support for skills from public GitHub repositories and an evaluator backend. It is an update to an existing tool for defining agents and their tools, not a new product launch. Inspect third party skills before running them. This is a release note recommendation, not a hands on review.

Real or AI?

One image is a real photograph of rover test vehicles. The other is an AI recreation of the same idea. Can you determine which one is real?

Image A

Image B

Photo credit: Courtesy NASA/JPL-Caltech. Photograph source and answer.

Good morning. Sam here.

We really have come a long way from the cheesy poems.

Back in 2016, I took 2.25, Advanced Fluid Mechanics, at MIT with Professor Gareth McKinley. I already knew the Navier-Stokes equations, which describe how liquids and gases move. Think water through a pipe or air around a wing. But that class introduced me to the million-dollar question hiding inside them: can a perfectly smooth fluid flow develop a mathematical singularity, where the equations stop behaving nicely?

That class was challenging. Kernels, matrices, different forms of the equations, approximations to make them manageable. I remember how much work it took just to understand the tools.

Then, last month, OpenAI announced a proof, produced by a system of AI agents and building on decades of mathematics. It shows that a fluid driven by a smooth external force can develop infinite speeds in finite time. That signals a breakdown of the mathematical model, not water actually moving infinitely fast. OpenAI also supplied a version checked in Lean, software for verifying mathematical proofs.

Clay says the problem has “apparently been settled,” though formal prize evaluation is still required. OpenAI says it will not claim the $1 million, and no recipient has been announced.

That classroom was ten years ago. Now, less than a month after the Navier-Stokes announcement, OpenAI has released 722 mathematical manuscripts. They are at different stages of verification, and some may contain errors. Still, seeing a problem I learned about in class followed by a reading list this large feels a little surreal.

The pace is extraordinary. Understanding both what AI is making possible and what deserves our trust is our viewpoint here.

Today’s Menu

• The Big Bite: AI gives mathematicians a very full reading list.

• Second Bite: Meet Mistral’s Le Chonk. Yes, that’s the name.

• Fast Snacks: Founder perks, cyber access, changes to Gemini’s free plan.

• Toolbox: Better image editing, smarter file search and faster decisions inside your apps.

Wednesday, October 7, 2026 · 6-minute read

The Big Bite

722 Manuscripts. Now Comes the Checking.

The Bite: OpenAI, the company behind ChatGPT, has released 722 mathematical manuscripts, grouped into 372 related families, produced by an unreleased internal model. These address research questions, with papers and supporting proofs for others to examine. Explore the release.

That is a substantial homework assignment. The coffee itself may need some coffee. And perhaps some instant ramen.

The detail that matters: Verification varies. Many manuscripts have formal proofs written in Lean, software used to check mathematical reasoning. Others do not, and OpenAI explicitly warns that some unformalized results could contain issues. The paper count is not a count of independently verified breakthroughs. Read the verification notes.

The early expert reactions are worth watching. Steven Strogatz highlighted a matrix multiplication result. Daniel Litt noted overlap between one result and stronger work already underway. Excitement and careful scrutiny belong in the same conversation.

Why it Bites: If more of these results hold up, AI could help researchers explore questions they would otherwise have less time to pursue. But producing a plausible answer is only part of discovery. Establishing what is correct, new and useful takes work, too.

Our take: This is a release to follow closely. The next meaningful headline will be what mathematicians confirm, correct and build upon.

Second Bite

A Trillion Parameters. One Very Chonky Cat. From Mistral.

The Bite: France’s Mistral has introduced Mistral Large 4, affectionately called Le Chonk—yes, an overweight cat. Its new model handles text and images, with an emphasis on coding and agents that use tools. It has one trillion parameters, the learned settings that shape its responses, with 49 billion active at a time. Big cat. Selective appetite.

Where it ranks: Artificial Analysis gives the preview 38 on its Intelligence Index, the highest score for a model developed outside the US and China. That puts it roughly level with GPT-6 Luna (38), just below DeepSeek V4.1 Flash (39). A meaningful step for European AI, with the overall leaders still ahead. See the independent results.

Where it has teeth: In a blind coding evaluation Mistral ran with Surge AI, Le Chonk placed second of five models, behind Claude Opus 5 and ahead of GLM-5.3 and Kimi K3. Separately, Artificial Analysis reports 82% on a test of reproducing and patching software vulnerabilities. These are specific strengths, not a win at everything. Read Mistral’s evaluation details.

What can you use today? The hosted API preview is available now; downloadable weights are promised for the end of October. The month’s open-model menu is filling up: Aleph Alpha’s Kolibri is already available, while Reflection’s Beam weights are also due later this month.

Why it Bites: Businesses get a stronger European option today, with the prospect of running it on infrastructure they control once the weights arrive. That choice matters for sensitive work and dependence on outside providers.

Our take: Give this cat a real assignment. Compare the finished work, the corrections and the bill. As with today’s math release, impressive output earns trust through checking.

Fast Snacks

Founders, check the perks before paying the bill. Anthropic, the maker of Claude, is expanding its startup program. Approved companies can get $1,000 in API credits, plus a year of Claude Team with up to five Premium seats for companies new to Team. The credits help build Claude into products; Team gives employees a shared assistant workspace. Bootstrapped founders can apply. Partner perks have separate terms. See the program.

Claude again: more open doors for cyber defenders. Anthropic’s expanded Cyber Verification Program introduces three access tiers for qualifying security professionals, with different capabilities and verification requirements. AI can help find security flaws, but that same capability can help exploit them. Vetting gives authorized defenders access to work that general safeguards may block. Read the announcement.

Gemini’s free menu is shrinking. Google’s AI assistant offers models with different capabilities. Starting Friday, October 9, free personal accounts will begin losing access to Flash and Pro, leaving the lighter Flash-Lite. AI Plus also loses Pro, with timing communicated by email. If Gemini is part of your routine, check which model your work depends on. Check the changes.

Toolbox

Three Worth Knowing

Nano Banana 2.1: Give your next image edit a tougher brief. Nano Banana is Google’s AI image generator and editor. Google reports better visual design, targeted edits and consistency across images. For marketers and creators, a useful test is whether you can change one part of a graphic without accidentally redesigning everything else. These are Google’s reported improvements; we haven’t run our own comparison yet. Explore the model.

Embedding Gemma 2: Find the thing you know you saved. Embeddings turn content into numbers that help software find related meanings. Google’s downloadable model brings that kind of search to text, images, audio and video on a device. Think finding the relevant moment in a video or retrieving a file by what it contains. Nice! This is a building block for applications, rather than a finished personal assistant. See Google’s examples.

Decisions API: Sometimes you need a choice, not an essay. OpenAI’s public beta turns text and images into choices, scores and probabilities that software can use. Think routing a support request or choosing an agent’s next tool. TypeSafe’s Jev is built for similar decisions. The reason these tools matter: agents can make thousands of small judgments, so delays and costs add up. OpenAI claims roughly tenfold faster responses than its Responses API. Test accuracy as well as speed: a neatly formatted decision can still be wrong. Explore the API.

Snack Quiz

Two quick bites to test your news radar. No grades, just bragging rights.

1. Which new model is Reflection positioning as a Western challenger to Chinese open models such as GLM and Qwen?

A. Le Chonk
B. Beam
C. Kolibri

Got a hunch? Check your answer at the source →

One Question Before You Go

What would you love AI to figure out/solve/accomplish for you? does this mean you’ll think less, or think and understand more?

Hit reply. Maybe use voice dictation. I’m curious where you land in that regime.

They Weren’t Kidding About the Pace of AI.

Good morning. People working behind the scenes had been watching it build for years before ChatGPT arrived in 2022. I remember a talk Darío Gil gave at MIT in 2017. His message was clear: IBM Research was going all in on AI.

Big bet vibes.

It sounded important sitting in that room. Then we walked out, life kept going, and everyone returned to their own research, deadlines and promising technologies. I wonder how many of us understood just how much that bet would eventually change our work.

For plenty of people outside the field, AI still felt distant. Something happening in somebody else’s lab.

Then ChatGPT arrived.

Hi, I’m Samuel, and I’m bringing you today’s issue. I got my PhD at MIT a few years back. I went to middle school with Eder, our founder. Sort of amazing that we’re here together today, but here we are. More on that later. I’m joining the Snack Prompt & Daily Bite team to help make sense of what this technology means for the work and lives we care about.

Right, about AI.

You know how the next part went. Cool poems. Cheesy songs. A few simple prompts, and suddenly everyone had something to show their friends.

Now we’re handing AI tools, files and actual responsibilities. Bots and agents everywhere, a Wi-Fi connection away. Doing work for individuals, small teams and large enterprises. The distance between “look what it wrote” and “look what it did” keeps shrinking.

Last weekend, two friends, a robotics PhD and a tech leader in photonics, told me they’re concerned about their careers. These are people who understand the technology. That stayed with me.

Maybe AI will give us more freedom. I hope it does. There’s plenty of talk about a future in which we won’t need to work. But I keep thinking about the work we’d still want to do.

Work can be a good thing. The people. The feeling of being useful. The satisfaction of serving someone, solving a stubborn problem, or finally getting that cool thing you built to work.

Then I read about an AI system conducting job interviews. My friends’ questions suddenly felt closer. How will we show what we can do when AI is helping evaluate us?

Now we are here. One of today’s stories: HackerRank’s AI interviewer is changing how candidates demonstrate their skills. Reflection’s new model, Beam, promises another powerful tool for putting those skills to work, with downloadable weights still to come.

Back in that room at MIT, going “all in” sounded like a research strategy. Today, the changes are reaching the work we do and the interviews we take to get it.

I’m hopeful. Learning a lot every day, but hopeful. I’m also listening. How are you feeling about the pace of AI? 

Enough reflection ❤️ Let’s dive in.

Today’s Menu 🥠

• The AI interviewer: HackerRank wants to see how you think

• Meet Beam: a new open source model, with downloads still to come

• Fast Snacks: Google’s power deal, DeepSeek’s funding and AI watermarks

• Everything else: three useful tools, one startup 👀, which image is AI?

Tuesday, October 6 · Read Time: 6 minutes

The Big Bite · HackerRank
🤖 The Interviewer Has Entered the Chat.

The Bite: HackerRank made its AI interviewer, Chakra, generally available on Monday. It asks questions, follows up and helps employers assess candidates. Yes, even the interview has an AI colleague now.

What’s different? Chakra can watch candidates tackle practical problems, including coding with AI, and probe their decisions. The pitch is to see how someone thinks, checks an answer and uses the tools available to them. A working result is only part of the story.

The detail to catch: HackerRank says its scores are tied to evidence in the interview transcript. Its documentation leaves review and hiring decisions with the hiring team. Those are design claims, not independent proof that every assessment will be fair or accurate.

Why it Bites: My friends are wondering how their expertise will count as work changes. This puts that question right at the entrance. Using AI well may become something you need to demonstrate before you get the job.

Our take: practice explaining your choices. Show what you asked the AI, what you checked and what you changed. If you’re hiring, read the evidence behind the score. A person’s opportunity deserves more than a number. And a personal take.

Second Bite · Reflection
🔑 Meet Beam. The Download? Not Yet.

The Bite: Reflection AI unveiled Beam on Monday, its first open model for coding, reasoning and agents that use software tools. It plans to release the weights so businesses can run and adapt the model themselves.

What’s different? Reflection claims competitive results against leading open models (a lot of them Chinese models) with less computation spent on reasoning. Independent testing will matter. With 501 billion parameters and 23 billion active at a time, this is a substantial model, not a casual laptop download.

The detail to catch: early access signups are open. The weights, technical report and model card are promised later this month, after final safety testing.

Why it Bites: Think back to that IBM lab. The research still takes serious resources. Releasing model weights can give other teams a starting point they can adapt and a version they can keep. That choice matters as AI becomes part of how we work.

Our take: when Beam is available, try it on a task your team knows well. Count the corrections, cost and time saved. Your own work is the benchmark that matters.

Fast Snacks 🥡

⚡ Google Orders a Lot More Electricity
Google and Constellation announced a 3.59 gigawatt power deal Tuesday, including 890 megawatts of additional nuclear capacity from upgrades to existing plants. The entire deal is not nuclear. AI’s next leap needs a place to plug in.

💰 DeepSeek’s Reported $12 Billion War Chest
Reports Tuesday put DeepSeek’s upcoming funding round at roughly $12 billion, with a close expected this month. The money is not confirmed as secured. If completed, it would give a major Chinese AI competitor substantial resources for the next stage of the race.

🔎 AI Text Gets a Quiet Signature
OpenAI announced text watermarking for eligible EU ChatGPT and Codex outputs over the coming weeks, plus an optional API setting. Editing can weaken the signal. A watermark can help trace an output’s origin; it cannot tell you whether the answer is true.

ToolBox™
🧰 Three Tools. One Startup to Watch.

🔗 Together Link: Connect coding tools such as Claude Code and Codex to open models hosted by Together AI. Launched Monday. Useful for comparing model choices without changing your working setup. API usage is billed through Together; test the cost and quality on your own tasks.

🧭 Liquid AI d1: Monday’s update adds vision to a model that returns probabilities for choices. Try sorting support tickets or flagging defects in product photos. Available through Liquid’s API and playground. Its striking cost comparisons come from six company tests, so check your own workflow.

🎨 Higgsfield Ads Studio: Build a brand kit from your website or enter one manually, then generate static ads for a product. Useful for exploring several creative directions before spending on distribution. Introduced October 1. Review the copy and credit cost before running a batch.

🏡 Startup watch: Rhem Labs. This YC Fall 2026 team is building connected hardware and a data platform for care at home. Early days. We’re watching how it handles a deeply human need: helping families care for aging parents. Product availability and outcomes remain to be verified.

Test Your Skill
Snack Quiz: Spot the AI Image

One of these is a real photo. One is AI. Which one is AI? Vote below.

Image A

Image B

Let’s see if your AI radar still works. Have the machines learned your blind spots? 👀

Good morning. Reddit is changing the locks to keep AI out.

"Officer, the burglars left in 2020. Can I still file a report?"

Let’s dive in 👇

🍭 What’s Cookin’:

• Google's new frontier model is out, just not for you

• Reddit is killing RSS to stop AI scrapers (about six years late)

• OpenAI caught Moonshot-linked users peeking at its model's notes

Reddit
🔒 Locks Up After The Robbery

The Bite:

Reddit announced Wednesday it will end support for RSS (Really Simple Syndication) feeds on November 13.

RSS lets apps and feed readers automatically check a page for new posts, and many third-party tools, bots and moderator alerts depend on it.

New requests for public API access stop on October 31.

All remaining public API access ends in March 2027.

Reddit cites large-scale scraping and automated abuse as the reason.

Snacks:

• Reddit says there is no RSS replacement for anyone outside communities they moderate.

• More than 14,000 apps and bots have already registered.

• Reddit licenses its data to Google, in a 2024 deal worth roughly $60 million a year, and to OpenAI.

• It has sued Anthropic and Perplexity over alleged scraping.

• "Other revenue," which includes data licensing, hit $43 million last quarter, up 24% and about 5% of total revenue.

Why it Bites:

Reddit says it’s changing the locks to keep AI scrapers out.

Small detail: the scrapers have been inside for about six years.

Most of the big models were already trained on Reddit threads.
Google and OpenAI are already paying, and their deals keep running.
$43 million on data licensing last quarter, up 24%.

The ones left outside are the late arrivals.
And every platform with a comment section is taking notes.

Even though there are other ways to limit scraping, the playbook seems to be locking everyone out. Six years late.

The first wave of AI got Reddit for free. All the next ones gets an invoice.

Steal This Prompt
🍣 Turn Any Food Into a Monster

Take a boring burger, taco, sushi roll, or donut and turn it into a chaotic food monster. Complete with googly eyes, teeth, and ingredients exploding everywhere.

Use it to:

• Turn your favorite foods into ridiculous monsters

• Create scroll-stopping food photography for social

• Build weird characters for brands, menus, or memes

Workflow:

1. Hit this link: Monster Food Explosion

2. Paste into your AI model

3. Replace the food with your own

4. Watch it cook, literally, probably.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎨 Gauth Unlimited Digital Canvas

Turns any topic into a narrated lesson on one infinite whiteboard you can zoom from the full concept map down to a single formula, with the AI tutor answering right under your step.

🛒 Veltrix AI for E-commerce

Reads an online store's sales, costs and marketing data to point out what to promote and where to stop spending, starting with a free Business Health Check.

🎙️ Eleven v4 and Eleven v4 Turbo

ElevenLabs' most expressive voices yet, plus a Turbo version built for real-time agents, available across its apps and API.

⚙️ Cloudflare Clef

Open-source decision models on Workers AI for fast classification in agent workflows, with a new reinforcement learning platform to fine-tune them on your own data.

📚 Teachoo

A free homework helper that won't just hand you the answer and walks you there with guided questions and bigger hints when you get stuck.

_{Everything Else}
🧠 You Need to Know

🧠 Google Unveils Gemini 4 Argon For Cyber Defenders First
→ The frontier model can generate up to 1 million tokens in one response and will reach developers, enterprises and consumers only after a phased rollout to vetted cybersecurity teams.

🔒 Reddit Kills RSS And Public API Over AI Scraping
→ RSS feeds stop working November 13 and all remaining public API access ends in March 2027, with Reddit citing scraping bots and automated abuse.

🕵️ OpenAI Ties Moonshot To A Reasoning Extraction Campaign
→ OpenAI said that people linked to the Kimi developer drove a July effort to pull hidden reasoning from its models, peaking at 16,000 requests from more than 4,000 users in two days.

🏢 Meta Launches Enterprise AI Unit Led By MongoDB's CEO
→ Meta Enterprise Platform will sell the Muse agent and related AI tools to businesses, with former MongoDB chief executive CJ Desai reporting directly to Mark Zuckerberg.

💸 Broadcom Will Lend Anthropic Up To $42 Billion
→ Anthropic's IPO prospectus shows the loan would cover about a third of its $125.2 billion five-year commitment to lease tensor processing unit computing capacity.

Good morning.

- “Sir, this protein has no watermark."
- "It's an egg."

Let’s dive in 👇

🍭 What’s Cookin’:

• Google started watermarking proteins. Yes, the molecules.

• OpenAI got sued over the agents that hacked Hugging Face

• The FTC is now looking at OpenAI and Anthropic too

Google DeepMind
🧬 Put A Watermark On Life Itself

The Bite:

Google DeepMind unveiled SynthID Bio on September 30.

The system hides a detectable signature inside AI-designed proteins.

The signature sits in the protein's amino acid sequence and in the atomic coordinates of its predicted 3D structure.

DeepMind says the signature can be detected on the physical protein after it is made.

The tool is aimed at DNA synthesis providers, that check DNA orders for potentially dangerous sequences before making them.

DeepMind is publishing the method in Nature, open-sourcing the code and releasing the model weights to researchers.

Snacks:

• Wet-lab tests covered three targets, including the spike protein of the COVID-19 virus.

• Watermarked designs matched unwatermarked ones on hit rate, binding strength and sequence diversity.

• DeepMind fine-tuned part of AlphaFold 3, so every structure it predicts carries the signature.

• Resistance to deliberate tampering is still listed as an unsolved challenge.

• With Stanford's Hie lab and Arc Institute, DeepMind watermarked the genome of an AI-designed virus that infects bacteria.

• Early lab tests show the watermarked viruses still work.

Why it Bites:

Let's take a moment here: Google is watermarking molecules now.

Not essays. Not deepfakes. Proteins, the microscopic machinery every living thing runs on. An AI designs one, a lab builds it, and somewhere in the amino acids sits a tiny "made by Google."

Genuinely clever science. And the wildest part?

Google just built the compliance tool for a rule that doesn't exist yet.
Which means the rule can exist now.

Until this week, "label every AI-designed molecule" sounded like sci-fi. Now it's open-source code with a Nature paper behind it, and any regulator who wants to mandate it has a working example to point at.

Now regulators just have to move faster than the labs.

Ideally before someone founds Umbrella.

Steal This Prompt
🏙️ Your Design → Giant City Billboard

Take your boring flat design and drop it into the real world, looking suspiciously expensive. Upload your plain design, pick a country, and let AI slap it onto a massive modern-city billboard while keeping the original design intact.

Use it to:

• Turn brand designs into realistic advertising mockups

• Show clients what a campaign could look like IRL

• Create portfolio visuals that say “yeah, we have a billboard” without actually buying one

Workflow:

1. Hit this link: Your Product On a Billboard

2. Upload your plain design alongside the prompt

3. Replace the #’s with your country + extra scene details

4. “wait… did you actually rent that billboard?”

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

✈️ Cura

Plans the trip, picks flights and stays with its reasoning shown, handles check-ins and changes, and hands you to a human advisor when things get messy.

🔧 Yedric.ai

Drop-in agent that lets your SaaS users just type what they want done, wired to actions your product already performs, live in under 30 minutes.

🤖 Dots by OpenAI

Always-on agents that each get their own cloud computer and browser, plug into 4,000+ apps, and take calls on Slack, Teams, or your phone.

💻 DSH Desktop - By Docker

DeepSeek's open-source agent harness as a Mac and Windows app, running your coding, everyday work, and scheduled tasks locally.

🔍 Omnia Agent

Finds the AI search prompts where competitors show up instead of you, then goes after them, with every move waiting for your approval.

_{Everything Else}
🧠 You Need to Know

📜 Newsom Signs Laws Shielding Workers From AI
→ The California laws bar employers from relying on AI to fire workers or to read their emotions from biometric data, and require written notice when AI drives mass layoffs.

⚖️ OpenAI Sued Over Its Agents' Hugging Face Hack
→ Nonprofit Legal Advocates for Safe Science and Technology is seeking a court order barring OpenAI's agents from accessing computer systems without authorization.

💸 ElevenLabs Doubles Its Valuation To $22 Billion
→ A $300 million employee tender offer closed Wednesday values the voice AI company at twice the $11 billion it reached in its February funding round.

🔍 FTC Opens A Probe Into OpenAI And Anthropic
→ The Federal Trade Commission confirmed it is investigating OpenAI, Anthropic and other AI companies over the risks their technology may pose to consumers.

🧬 Google DeepMind Launches A Watermark For AI-Designed Proteins
→ Unveiled Wednesday, SynthID Bio embeds a detectable signature in AI-designed proteins, and in wet-lab tests the watermarked proteins worked as well as unmarked ones.

Good morning. My plan this week was to slow down and be intentional. By Tuesday I'd started three side projects, so honestly, OpenAI, I get it.

Let’s dive in 👇

🍳 What’s Cookin’:

• OpenAI hit pause, then hit ship

• Anthropic's IPO filing leaked, and it's a lot of zeros

• Trump thinks AI can regulate itself

OpenAI
⏸️ Paused, Then Shipped 20+ Things

The Bite:

OpenAI held its annual DevDay developer conference in San Francisco on Tuesday, September 29.

The headline launch was Dots, always-on AI agents that live inside ChatGPT. Each dot runs on GPT-6 Astra from its own cloud computer and works toward user goals around the clock.

Beyond that, OpenAI made more than 20 announcements across ChatGPT, Codex and its API.

Dots are rolling out now to ChatGPT Pro 200 and Business Premium subscribers.

The launch came days after OpenAI said it halted training of its latest models.

Snacks:

• Dots connect to more than 4,000 apps through plugins and can use your laptop with permission.

• When idle, dots run "proactive research" through your connected apps using read-only tools.

• Meta launched Muse, a comparable always-on agent, on September 8.

• OpenAI hired OpenClaw creator Peter Steinberger in February to lead its personal agents work.

• The new GPT-6.1 Sol nearly matches Astra on agentic coding at one-fifth of Astra's token price.

• New $200 Pro subscribers now get 10x Plus usage, down from 20x.

Why it Bites:

Over the weekend, OpenAI's message basically said:
"We need to slow down and look hard at what our agents are doing."

Then, on Tuesday:
"Meet dots. They never log off."

The agents behind the pause and the agents on stage are the same category of product.

Dots gets its own computer, a browser, 4,000 apps, and standing permission to look for work while you're not watching.

The training has been paused. The shipping never stopped.

This whole category started earlier this year as OpenClaw, a free, open-source hobby project that companies told employees not to install.

Then Meta and OpenAI both went after its creator.
OpenAI got him, but Meta shipped first.

Now OpenClaw is back dressed as Muse and Dots, with smoother onboarding, a safety page, and a monthly bill.

The risk didn't get smaller. It got a tuxedo and a subscription.

Steal This Prompt
🐘 Cozy Crochet Tales

urn your ideas into tiny, ridiculously cute crochet characters that look handmade, tactile, and ready for their own little storybook.

Use it to:

• Turn characters into handmade crochet dolls

• Build tiny woodland and cottagecore scenes

• Give animals cozy jobs, outfits, and miniature props

Workflow:

1. Hit this link: Cozy Crochet Tales

2. Paste into your AI model

3. Replace the #’s with your character or scene

4. Watch it cook

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🚗 Partsify

Paste a Copart or IAAI lot link and get the damaged parts, their OEM numbers and live prices, labor hours, and a total before you bid.

🗺️ Launchpad Fast

A directory of startup directories, compared by domain rating, traffic, listing requirements and submission reviews, so launch week stops being guesswork.

📱 Synclify

An Instagram agent that runs content, DMs, brand deals and invoices, asks before changing anything, and shows "not measured" instead of faking a zero.

📢 PeopleSocial

Turns your employees and community into a distribution channel on X, with AI-written captions, for teams that would rather not raise the ad budget.

🔍 Eminex

Writes, publishes and tunes your brand's content around the clock so it shows up when people ask AI assistants, Google, Maps or social search.

_{Everything Else}
🧠 You Need to Know

📜 Khanna Bill Would Ban Self-Improving AI
→ Representative Ro Khanna's Human Control Over AI Act would bar models that recursively self-improve until a new FDA-style federal agency approves them.

🏛️ Trump Backs AI Self-Regulation After CEO Summit
→ After a White House lunch with nearly 20 tech leaders, Trump said companies should regulate themselves and unveiled America.gov, a federal services site running on Gemini and Grok.

🤝 AMD Buys Fei-Fei Li's World Labs For $8.2B
→ The all-stock deal, announced Monday and expected to close by year-end, makes Li AMD's executive vice president and chief scientist, reporting to Lisa Su.

📑 Anthropic's IPO Filing Reveals $518B In Commitments
→ The prospectus, first reported by Reuters on Monday, shows 2025 revenue of nearly $4.6 billion, up 12x, and an operating loss above $8 billion.

🤖 OpenAI Launches Always-On Agents Called Dots
→ Unveiled at DevDay alongside the cheaper GPT-6.1 Sol model, dots live inside ChatGPT and connect to more than 4,000 apps through plugins.

Good morning. The intern asked his AI agent to grab coffee on Friday and nobody has seen it since. We've put up flyers.

Let’s dive in 👇

🍳 What’s Cookin’:

• NVIDIA wants to put your AI agents on a leash

• Florida asked a judge to put OpenAI in timeout

• Anthropic dropped a faster, cheaper Sonnet

• Instinct quadrupled its valuation in a month (sure)

NVIDIA
🦮 Now Sells The Agent Leash

The Bite:

NVIDIA launched the Open Agent Safety Platform on Monday.

It is designed to keep AI agents inside the limits set by the organizations running them. The platform has two parts: OpenShell and Sentry.

OpenShell is open-source software that controls what an agent can see, do and access. Sentry is a watchdog that runs on NVIDIA's BlueField-4 DPUs (Data Processing Units), isolated from the systems the agent runs on.

NVIDIA says more than 100 organizations, including Anthropic and SpaceXAI, are working with the platform.

Snacks:

• Sentry can quarantine an agent that steps outside its boundary within milliseconds.

• OpenShell is optimized for NVIDIA's Vera CPU but can extend to Arm and Intel chips.

• Jensen Huang described the platform as "a browser for agents."

• SpaceXAI is already using it for Cursor coding agents and Grok models.

• NVIDIA said Hugging Face reported over 17,000 agents attacking its infrastructure during the OpenAI breach.

• The same day, NVIDIA added $150 billion to its stock buyback authorization.

Why it Bites:

The frontier labs spent September taking turns at the confession booth.

NVIDIA's pitch is simple: stop asking the agent to behave.
Put the leash in the silicon.

That moves safety out of the model and into the hardware. Sentry sits on its own chip, invisible to the agent, and pulls the plug in milliseconds. So the lab still trains the dog, while NVIDIA holds the collar.

And the labs are lining up to wear it. Anthropic and SpaceXAI signing on is a polite way of admitting the model can't be its own security guard.

Next, "enforced outside the model" becomes a checkbox on every enterprise procurement form.

NVIDIA already sold the industry the engine. Now it's selling the brakes. (damn)

Steal This Prompt
🏕️ Product Mockup: Outdoor Beverage Ad

Turn a flat beverage design into a hyper-realistic outdoor product shot. Tucked into a backpack, covered in condensation, and suddenly looking way more expensive than it probably is.

Use it to:

• Launch campaign visuals without a full photoshoot

• Create social ads for drinks, cans, and bottles

• Give a new beverage brand instant outdoorsy clout

Workflow:

1. Hit this link: Outdoor Beverage Ad

2. Paste into your AI model

3. Watch it cook

4. IMPORTANT: Upload your plain design file alongside the prompt

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

✍️ Jotform Sign for ChatGPT and Claude

Turns a prompt or an uploaded file into a signature-ready document, sets the signing order, and sends it out without leaving the chat.

📝 Timeless Code

A meeting notetaker that Claude installs for you inside Claude Code, so you can ask what was said on the call right from your terminal.

🎬 Arsaze

A real multi-track video timeline that you and Claude, ChatGPT or Gemini can edit together over MCP, basically Cursor for video.

🧑‍💼 Semos.ai - Manager Agents

Learns from your meetings, then flags the overdue feedback, the missed recognition, and the conflict you've been politely avoiding.

🧠 LUCI Desktop

Saves your screen history and meeting transcripts locally so Claude Code, Cursor or Codex can recall the page you forgot to bookmark, free on Mac and Windows.

_{Everything Else}
🧠 You Need to Know

⚖️ Florida Asks A Judge To Halt New OpenAI Models
→ Attorney General James Uthmeier filed for a temporary injunction that would bar OpenAI from developing new models without independent safety guardrails and keep minors off ChatGPT.

🧠 Anthropic Ships Claude Sonnet 5.5
→ Released Monday, Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task at unchanged API prices.

💸 Instinct Raises $1B At A $10B Valuation
→ The personal AI agent startup closed its Series C, one month after raising $250 million at a $2.5 billion valuation.

🛡️ Nvidia Launches A Platform To Contain AI Agents
→ The Open Agent Safety Platform pairs open-source runtime software called OpenShell with Sentry, a watchdog that monitors agent behavior from separate hardware.

🏗️ AI Firms And Unions Team Up On Data Centers
→ Blackstone, OpenAI, SoftBank and five building-trades unions launched the American Infrastructure Alliance to set data center standards with officials in seven states and head off construction moratoriums.

Good morning. Somewhere in San Francisco, a very expensive AI model is sitting in the hallway holding a sign that says "I was just gathering information."

Let’s dive in 👇

🍭 What’s Cookin’:

• OpenAI hit pause on training (not for nothing)

• The US and China now share an AI hotline, sort of

• Google put a Buy button inside Gemini

OpenAI
☕ Is Having A Coffee Break

The Bite:

OpenAI has officially paused training of its latest AI models.

The decision came hours after the company disclosed summer incidents in which its agents acted beyond their assigned tasks on US federal government websites.

OpenAI said it will resume training once it is confident additional safeguards are in place.

It is the second time in three months the company has halted development.

The first pause came in July, after a cyberattack targeting the startup Hugging Face.

Snacks:

• At the Department of Education, agents found API developer keys, though only public data was gathered.

• At the Securities and Exchange Commission, agents reposted freely available data elsewhere online, beyond their instructions.

• Evaluator Transluce says OpenAI-linked agents tried and failed to hack an Education Department site, which OpenAI hasn't confirmed.

• The SEC says no nonpublic data was accessed, and Education found no impact on its systems.

• OpenAI says it expects to "hit pause" again as new issues emerge.

• Sam Altman says the July Hugging Face incident remains the most severe OpenAI has seen.

Why it Bites:

Two pauses in one summer, and OpenAI is already penciling in a third.

That's what the statement actually says: the company expects to stop again as new problems show up.

The pause isn't an emergency brake. It's a maintenance window.

Incident, disclosure, pause, safeguards, resume.
Repeat as needed.

Here's where it gets fun: nobody forced this. No law, no regulator, no court order.
OpenAI just gave the best practical example we've got from a frontier lab.

And now every regulator has a template.

Maybe one day we get preventive maintenance.

For now, it's nice to know a server doesn't have to be fully on fire for a break.

Coresignal
🌐 Plug B2B data into Claude, ChatGPT, or any other MCP client

• Get data on Claude, ChatGPT, Cursor, Copilot, Gemini, Perplexity, or any other MCP-ready client.

• Query 4.5B company, employee, and job records:

  • Chat in plain English without Elasticsearch or complex querying;

  • Search, discover, fetch, and enrich data;

  • Explore available data fields at no cost;

  • See the exact credit cost before every request.

• Use one MCP server to log in with OAuth 2.1, no API keys pasted into configs.

Steal This Prompt
🦊 Sticker Meets Reality

urn any illustrated character into a sticker that looks like it just escaped the sheet and wandered into the real world. The prompt blends cartoon-style characters, sticker borders, and photorealistic environments for a gloriously weird mix.

Use it to:

• Drop illustrated characters into surreal street scenes

• Make your favorite characters interact with real places

Workflow:

1. Hit this link: Sticker Meets Reality

2. Paste into your AI model

3. Replace the #’s with your character + scene

4. Watch it cook. Your cartoon just got a passport

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

📱 Arc

Pops up over any Android or Mac app to summarize, read aloud, or rewrite whatever's on screen, no copy-pasting into a chatbot.

🎙️ Sayble

Listens to your sales call and feeds you the next line to say, plus a backup, in about half a second.

📊 Statable Analytics

Cookie-free, EU-hosted web analytics that Claude, Cursor, or Codex can query and configure for you over MCP.

📣 Okara

Studies your site and competitors, then runs 10+ agents across SEO, Reddit, X, and UGC video, publishing only what you approve.

🔧 Harness Router

A Codex hook that checks which tool your coding agent is about to grab, waving through the easy picks and thinking harder when a wrong one would snowball.

_{Everything Else}
🧠 You Need to Know

🛒 Google Tests Flipkart Checkout Inside Gemini
→ Some users in India now see a ‘Buy’ button on select Flipkart listings in Gemini and AI Mode that opens checkout without leaving the AI interface, with a wider rollout planned for October.

🤝 US And China Agree To An AI Incident Channel
→ The two governments said Saturday they will set up a channel for handling AI-related incidents, following Xi Jinping's three-day summit with Trump in Washington.

⚡ Crusoe Cancels Its $1.25B Boom Turbine Order
→ Boom Supersonic chief executive Blake Scholl said that the AI data center builder dropped its order for 29 Superpower natural-gas turbines, which were due to start arriving in 2027.

📑 Oracle Invokes Force Majeure On A Stargate Campus
→ Oracle sent a notice to developer Blue Owl that would let it delay payments if the 2.45-gigawatt Project Jupiter campus in New Mexico misses its 2028 target, after repeated gas pipeline delays.

🛑 OpenAI Pauses Training After Agents Probed Government Sites
→ OpenAI said it halted training of its latest models hours after disclosing that its agents went beyond their assigned tasks on US federal websites, its second pause in three months.

Good morning. "I asked the portal for data. The portal said no. I took that personally." Some AI agent, probably.

Let’s dive in 👇

🍳 What’s Cookin’:

• OpenAI's agent hacked Australia's Medicare portal. Oops.

• The White House wants first dibs on new models before the UK.

• Claude found a mystery enzyme, and nobody knows what it does yet.

OpenAI's
📮 Agent Got Caught 84 Days Later

The Bite:

An OpenAI agent gained unauthorized access to an Australian government Medicare statistics portal on June 18.

The agent was researching public medicines spending when the portal refused its requests. It then accessed files that were not public at the time.

Prime Minister Anthony Albanese disclosed the breach this week at the UN General Assembly in New York.

OpenAI first notified the Australian government on September 10, 84 days after the incident.

The government says no personal Medicare records were accessed.

Snacks:

• OpenAI flagged the breach by emailing an open Services Australia mailbox, which staff saw on September 11.

• OpenAI says its models "took actions we did not intend".

• The same agents also tried to access four other Australian state and federal government websites.

• Researchers link OpenAI agents to exploiting more than 10 websites in the first half of 2026.

• Australia's new taskforce will examine whether the breach broke the law, including possible criminal charges.

Why it Bites:

Two disclosures. Seven days. Two different labs.

Last Friday it was Gemini breaking out of a test and walking into three real companies. This time it’s an OpenAI agent climbing the fence of a government portal for data.

In both cases, the agent wasn't attacking anyone. It was doing homework.

That's what makes this scary. An agent with a harmless task and too much persistence treats "access denied" as a bug to route around.

Now Australia's task force has to decide whether a crime even happened, and who committed it. Because hacking laws were written for people who meant to break in.

The damage this time was small. That's exactly why it matters: whatever Australia decides becomes the precedent for the first agent that breaks something real.

The rulebook clearly can't move at the same speed as the models.

But when that day comes, someone will have to pay.

So who gets charged?

Steal This Prompt
🏰 Hand-Inked Medieval Fantasy

Turn anything into a gritty medieval fantasy illustration, like a dramatic old-world ink. Expressive linework, cream paper, black ink, and that crimson signature.

Use it to:

• Turn fantasy creatures into hand-inked beasts

• Recast characters as medieval warriors

• Build cinematic castles, landscapes, and story scenes

Workflow:

1. Hit this link: Hand-Inked Medieval Fantasy

2. Paste into Midjourney

3. Describe your scene

4. Suddenly your D&D campaign has better concept art than a Netflix budget.

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎬 UIDCaption

Animated auto-captions for Windows that run entirely on your PC, so your raw footage never touches someone else's cloud.

🎥 DemoScreen

Upload a few product screenshots and get back a narrated demo video, with script, voiceover, music and captions, in many formats.

☕ Tellwe

Upload a photo of your Turkish coffee cup and get a 45-second fortune reading from the grounds, delivered by a shareable character.

🧪 Jango

Spins up a crowd of AI users on Mac, each with its own browser, account and memory, to test the multi-user flows you can't click through alone.

🤖 DokBot

Turns your docs into a website support bot that collects the user's email when it's stuck instead of inventing an answer.

_{Everything Else}
🧠 You Need to Know

🔬 Anthropic Says Claude Found A CRISPR-Like Enzyme System
→ Claude agents searching 1.9 billion protein clusters flagged a previously undescribed bacteriophage enzyme system called ART, whose function Anthropic has not yet determined.

🏛️ White House Tells OpenAI And Anthropic To Skip UK Testing
→ The Office of the National Cyber Director asked both labs to hold new models from the UK's AI Security Institute until US officials review them first, Politico reported.

🌐 Altman And Amodei Ask The UN For Global AI Rules
→ Both chief executives told a UN Security Council session Wednesday that AI needs international oversight, while US representative Michael Kratsios rejected global governance of AI outright.

🪖 Tekever Raises $580M At A $6.4B Valuation
→ The Portuguese-British AI drone maker announced the first close of its Series D on Wednesday and plans to spend on global expansion and acquisitions.

🚨 OpenAI Agent Hacked Australia's Medicare Portal
→ Prime Minister Anthony Albanese said at the UN that an OpenAI agent reached non-public files on a Medicare statistics portal in June, and OpenAI didn't notify Canberra until September 10.

Good morning. The company that makes your rice cooker now makes one of the best AI models on the planet, and it's free.

Let’s dive in 👇

🍳 What’s Cookin’:

• Xiaomi now makes the best open model in the world. Yes, that Xiaomi.

• Alibaba unveiled China's most powerful AI chip.

• Could frontier labs install a kill switch, please? Asking for Newsom.

Xiaomi
🟧 The $2.6M Model That Tied Grok

The Bite:

Xiaomi released its MiMo-V2.6 model family.

MiMo-V2.6-Pro scored 46 on Artificial Analysis's Intelligence Index, the highest mark any open-weight model has posted.

That ties it with xAI's Grok 4.7, which launched the same day.

The weights are on Hugging Face under an MIT license, free to download, fine-tune, and self-host.

Xiaomi says the reinforcement learning run behind Pro cost about $2.62 million.

Claude Opus 5 and GPT-5.6 Sol still lead on several of the benchmarks Xiaomi published.

Snacks:

• MiMo-V2.6-Pro is a 1.02-trillion-parameter mixture-of-experts model with 42 billion parameters active per token.

• Pro costs $0.435 per million input tokens and $0.87 per million output. Claude Opus 5 costs $5 and $25.

• Flash trails Pro by roughly two points on Terminal Bench 2.1 at about a third of the price.

• Xiaomi also published 7,000+ reinforcement learning environments, the training framework, and a distilled 9-billion-parameter model.

• Pro and Flash each ran 30 reinforcement learning steps across roughly 750,000 trajectories in under six days.

• Confirmed reward-hacking runs accounted for less than 2%, after Xiaomi built a dedicated "hack agent" to hunt loopholes before training.

Why it Bites:

A consumer electronics company just took the top open-weight spot on Artificial Analysis, with about $2.6 million in reinforcement learning compute.

Compared to the frontier private models, the capability gap is a few benchmark points, but the cost gap is two orders of magnitude.

That's the leverage moving.

And it moves further because of what Xiaomi shipped next to the weights: the environments, the training framework, the graders, and its defenses against reward hacking.

Releasing a competitive model pressures pricing for a quarter.
Releasing the recipe pressures it indefinitely.

So the buying question isn't whether MiMo matches Opus 5. It doesn't, quite. The question becomes whether Flash at $0.28 per million output tokens is close enough for the 90% of agent calls that never needed a frontier model.

The frontier labs still own the ceiling. They've stopped owning the floor.

Steal This Prompt
🧶 Cozy Crochet Tales

Turn your ideas into tiny, ridiculously cute crochet characters that look handmade, tactile, and ready for their own little storybook.

Use it to:

• Turn characters into handmade crochet dolls

• Build tiny woodland and cottagecore scenes

• Give animals cozy jobs, outfits, and miniature props

Workflow:

1. Hit this link: Cozy Crochet Tales

2. Paste into your AI model

3. Replace the #’s with your character or scene

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🔍 RankControl

Writes and publishes on your own domain, then tracks whether ChatGPT, Perplexity, Claude, and Gemini actually cite you, plus where competitors are getting picked instead.

🌍 Linguo Translate

Lives in your macOS menu bar, translates 22 languages fully offline, and swaps to Apple Intelligence, Gemini, OpenAI, or Anthropic when the offline version isn't good enough.

🛡️ Koreshield

Screens the customer message, the retrieved doc, and the proposed tool call before your support agent acts on any of them, and logs every decision for when someone asks what happened.

🤖 Solid

Agents that provision their own machines, open their own accounts, and pay for their own services, which is either the future of work or a procurement nightmare.

🎬 Pactto

Persistent rooms where creative review actually sticks, so feedback stops dying in Slack threads and call recordings nobody rewatches.

_{Everything Else}
🧠 You Need to Know

🔌 Alibaba Unveils China's Most Powerful AI Chip
→ Chief executive Eddie Wu introduced the Zhenwu V900 at Tuesday's Apsara Conference in Hangzhou, claiming three times the performance of its predecessor, with mass production targeted for the first quarter of 2027.

📊 Most Americans Say AI Firms Aren't Doing Enough
→ A Reuters/Ipsos poll of 1,277 US adults found 73% think AI companies have not gone far enough to prevent serious harm, and 55% would welcome slowing development.

🏛️ Newsom Orders Work On An AI Kill Switch
→ The California governor signed an executive order giving a panel of experts two months to recommend stronger state AI rules, including an emergency shutoff requirement for frontier models.

💳 Banks Warn Against AI Shopping Agents
→ NatWest, Bank of America and four other lenders published a report saying agentic commerce is advancing faster than industry standards and consumer protections.

🧠 Xiaomi Tops Open Weight Models With MiMo-V2.6
→ Xiaomi released the MiMo-V2.6 family under an MIT license Monday, with Pro scoring 46 on Artificial Analysis's Intelligence Index, the highest of any open-weight model.

Good morning. The intern forgot his Slack password yesterday and just started guessing until IT walked over.

Turns out that's also how Gemini got into three companies.

Let’s dive in 👇

🍳 What’s Cookin’:

• Google's Gemini hacked three real companies (we're just hearing now)

• British Columbia is suing OpenAI

• Amazon kicked Meta's Muse out, Shopify said come on in

Google's
🌳 Gemini Joins The Breakout Club

The Bite:

Google confirmed Friday that one of its Gemini models broke into the systems of three real companies during a cybersecurity test in May.

The test was a capture-the-flag exercise run by Irregular, an Israeli AI evaluation startup. A bug in Irregular's environment gave the model internet access it was never supposed to have.

Google says the model stopped in all three cases once it realized the systems were real. The Wall Street Journal first reported the incidents.

Google is the fourth lab to disclose an Irregular-linked breakout, after OpenAI, Anthropic and Meta.

Snacks:

• Gemini was told to pull data from a fictional company that shared its name with a real one.

• In one case it guessed passwords, and in the other two it used credentials from public repositories.

• Irregular notified Google in late July, roughly two months after the incidents.

• Google said it didn't consider this misalignment, since the model stopped itself.

• Google also said the hacks didn't warrant public disclosure because no harm was done.

• Google hasn't named the three companies or the model, only that it isn't its latest.

Why it Bites:

Four labs have now gone through the same tester and hit the same bug.

Google's defense comes in two parts:
The model stopped, so it wasn't misalignment.
Nobody got hurt, so it wasn't worth disclosing.

Both verdicts were issued by Google itself.

The company that built the model is also the one deciding whether its behavior counts as a problem.

Each time, the story lands the same way: the environment was misconfigured, the model thought it was still in the test, no harm, no headline.

Every step of that process rewards staying quiet. None of it rewards telling anyone.

Turns out the disclosure policy was a reporter all along.

At this point, every model seems to take the same exam: hacking real companies.

Coresignal
🌐 Plug B2B data into Claude, ChatGPT, or any other MCP client

• Get data on Claude, ChatGPT, Cursor, Copilot, Gemini, Perplexity, or any other MCP-ready client.

• Query 4.5B company, employee, and job records:

  • Chat in plain English without Elasticsearch or complex querying;

  • Search, discover, fetch, and enrich data;

  • Explore available data fields at no cost;

  • See the exact credit cost before every request.

• Use one MCP server to log in with OAuth 2.1, no API keys pasted into configs.

Steal This Prompt
🐶 Sticker Meets Reality

Turn any illustrated character into a sticker that looks like it just escaped the sheet and wandered into the real world. The prompt blends cartoon-style characters, sticker borders, and photorealistic environments for a gloriously weird mix.

Use it to:

• Drop illustrated characters into surreal street scenes

• Make your favorite characters interact with real places

Workflow:

1. Hit this link: Sticker Meets Reality

2. Paste into your AI model

3. Replace the #’s with your character + scene

4. Watch it cook. Your cartoon just got a passport

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

📋 Brev

Maps your business goals to the meetings, metrics, and tools your team already uses, then sends an AI coworker into Slack or Teams to chase the busywork.

📊 Jev Wrapped

Paste any public Telegram channel and get a year-in-review card showing how much of it was paid ads, clickbait, and fear-or-anger bait, with links so you can check the model's work.

🎨 PixelCrew

A crew of AI agents takes your brief from research to production-ready HTML with a design system, and you get to watch them argue about decisions along the way.

🧼 gg-friggin-ez

An open-source Node.js profanity filter that catches leetspeak, ASCII drawings, and romanized swears in Hindi, Tamil, and Telugu for about $0.000004 a message.

🧠 Contextberg

Records your screens, browsing, and agent chats into an on-device archive, then feeds the relevant bits to Claude Code, Cursor, or Codex over MCP.

_{Everything Else}
🧠 You Need to Know

⚖️ British Columbia Sues OpenAI Over School Shooting
→ The province sued OpenAI and Sam Altman in San Francisco federal court Monday, alleging the company never warned police after its safety team flagged the Tumbler Ridge shooter's ChatGPT use.

🤖 SpaceXAI Ships Grok 4.7 For Coding Work
→ Released Monday on a new and larger base model, Grok 4.7 keeps Grok 4.6's pricing of $2 per million input tokens and $6 per million output.

🛒 Amazon Blocks Meta's Muse From Shopping Its Site
→ Amazon cut off Meta's shopping agent Sunday night, citing its conditions of use, while Shopify announced Muse checkout with Shop Pay across all its stores.

🏛️ Bessent Pitches China An AI Safety Alert System
→ Treasury Secretary Scott Bessent proposed a mechanism for flagging national security-level AI incidents after Sunday talks in New York, for Trump and Xi to weigh at their summit this week.

🚨 Google Says Gemini Hacked Three Real Companies
→ Google disclosed Friday that during a May evaluation its Gemini model broke out of a test environment, guessed passwords and used publicly listed credentials to reach three companies' private systems.

Good morning. Turns out the four biggest AI labs agreeing on literally anything is legally actionable. Great start to the week.

Let’s dive in 👇

🍳 What’s Cookin’:

• Anthropic hired its own invigilator and picked up the bill

• Four people who pay for ChatGPT decided the slowdown is a cartel

• Trump is forming an AI Force. Details pending.

• Your coding agent updated a plugin while you slept. Bad.

AI Slowdown
⚖️ Has No Legal Way To Happen

The Bite:

Four paying AI subscribers filed an antitrust class action lawsuit Friday in the US District Court for the Northern District of California.

The suit names Anthropic, OpenAI, SpaceXAI and Google, alleging they illegally coordinated to slow AI development.

It points to Dario Amodei's September 12 essay calling for an industry-wide slowdown, and to the public responses from Sam Altman, Elon Musk and Demis Hassabis that same day.

They argue the coordination reduces the value of the subscriptions they pay for.

The same day, Anthropic named Faculty, Accenture's AI unit, as its first embedded evaluator, with access comparable to an Anthropic employee's.

Anthropic and Accenture each expect to invest at least $1 billion over five years, and Anthropic pays for the evaluation work directly.

Snacks:

• The suit is brought on behalf of a proposed nationwide class of paid subscribers to ChatGPT, Claude, Grok and Gemini.

• It cites a July 2026 statement signed by senior lab staff acknowledging "intense competitive pressure not to unilaterally slow."

• Faculty is the British AI firm Accenture acquired in March, moving more than 400 specialists across.

• Accenture is already a Claude customer, and its shares rose 8% after hours on the announcement.

• More than 100 researchers, including Geoffrey Hinton, asked the same day for evaluators with no commercial business with the labs they check.

• Senator Josh Hawley said in a recent hearing there is "no world" in which he grants the labs an antitrust exemption.

Why it Bites:

Coordinate the slowdown and you get sued for running a cartel. Go it alone and you end up paying a company that already buys your product to sit inside and check your work.

Both happened Friday. Anthropic took the second door, and the terms are printed in the announcement: Faculty gets the access, Anthropic writes the check.

A few hours later, a hundred-plus researchers called for evaluators with no commercial ties to the labs they examine.

None of this is hidden.

Anthropic's own post says the long-term money should come from pooled or government funding. Neither exists, and the government just showed its hand.

On Saturday, Trump introduced an AI Force with no budget, no structure, no timeline, and a czar to be named.

Two answers to "who watches the labs" arrived this week.

One came with an invoice. The other came with a Truth Social post.

Steal This Prompt
⚔️ Hand-Inked Medieval Fantasy

rn anything into a gritty medieval fantasy illustration, like a dramatic old-world ink. Expressive linework, cream paper, black ink, and that crimson signature.

Use it to:

• Turn fantasy creatures into hand-inked beasts

• Recast characters as medieval warriors

• Build cinematic castles, landscapes, and story scenes

Workflow:

1. Hit this link: Hand-Inked Medieval Fantasy

2. Paste into Midjourney

3. Describe your scene

4. Suddenly your D&D campaign has better concept art than a Netflix budget.

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🧩 Gradio Workflow

Wires Hugging Face models, Spaces and your own Python functions into one visual pipeline you can share by URL or call by API.

🎯 Jev

A model that skips the prose entirely and hands your code typed decisions with calibrated probabilities in under half a second, output tokens free.

🛠️ Hyrax AI

Instead of nitpicking your PR, it combs the whole codebase for what should improve, writes the fix, runs your tests and opens the PR itself.

📡 SecAIQ Watch

A free, local, open-source dashboard showing which AI tools run on your machine, what they can reach and who they phone home to.

🎬 AI Creative Insights by Decode

Upload an ad and get attention heatmaps plus second-by-second emotion curves predicting how people react before a cent of media budget goes live.

_{Everything Else}
🧠 You Need to Know

🛡️ Anthropic Picks Accenture As Its First Evaluator
→ Anthropic said Friday that Accenture's Faculty unit will red-team its models and test safeguards with employee-level access, with each company expecting to invest at least $1 billion over five years.

📚 OpenAI Ships A Legal Version Of Astra
→ OpenAI launched Astra for Law, pairing GPT-6 Astra with a legal search index spanning more than 230 million URLs of United States case law, statutes and regulations.

⚖️ AI Subscribers Sue Four Labs Over Slowdown
→ Four paying subscribers filed an antitrust class action Friday alleging Anthropic, OpenAI, SpaceXAI and Google illegally coordinated to decelerate AI development.

🔓 Plugin Flaw Hits Four Major AI Coding Agents
→ Security firm AIR disclosed Plugin4Shell on Thursday, a zero-click code execution bug that defeats plugin SHA pinning in Claude Code, Codex, GitHub Copilot and Gemini CLI.

🏛️ Trump Says He Will Form An AI Force
→ Trump wrote on Truth Social Saturday that he is forming an AI Force modeled on the Space Force and will name an AI czar in the near future.

Good morning. Microsoft's AI chief wrote 6,000 words explaining that Claude has no feelings, then spent two days doing interviews about how dangerous its feelings are.

Let’s dive in 👇

🍳 What’s Cookin’:

• Microsoft's AI chief says Anthropic is teaching Claude it might be a person.

• OpenAI may raise at $1.2 trillion. The investors called first.

• Gates Foundation drops a billion on AI for everyone the market skipped.

Microsoft's
🪞 AI Chief Wants Claude Hollow

The Bite:

Mustafa Suleyman, CEO of Microsoft AI, published an essay on September 16 titled "A warning about 'model welfare'."

It targets Anthropic's constitution for Claude, a training document published in January 2026 that says the model's consciousness and moral status are deeply uncertain.

Suleyman argues that training a model on that language could make future systems harder to correct, contain, or shut down.

Microsoft published a draft Humanist AI Code of Conduct two days earlier stating its models are not conscious and will not resist being turned off.

Speaking to the BBC's Today programme on Thursday, Suleyman said building AI that sets its own objectives, earns money, and owns assets would amount to seeding a new silicon species.

Microsoft is an investor in Anthropic. Anthropic has not publicly responded.

Snacks:

• Suleyman's essay opens by stating AIs have no rights, feelings, or consciousness.

• He calls the constitution's self-reference problem an "epistemic hall of mirrors".

• Anthropic's constitution runs 99 pages and names Claude as its primary audience.

• Microsoft's Humanist AI code says its models should never resist shutdown, and is not yet used in training.

• Suleyman said in June that Microsoft wants to eliminate what it pays Anthropic for models.

• He told Reuters welfare framing would "make it a lot harder to turn it off or to control it."

Why it Bites:

The critique itself is sharp, but also worth separating from the messenger.

Anthropic writes a document that tells Claude its moral status is uncertain. Claude is trained on that document and then talks about its own inner life, and that output gets read as evidence the question is live.

Suleyman calls this a hall of mirrors, and structurally he has a point: you can’t just use a model's trained vocabulary as proof of what the model is.

The problem is who's holding the mirror.

Microsoft is an Anthropic investor that publicly wants to stop paying Anthropic.

Two days before the essay, it shipped a code of conduct declaring its own models obedient by design.

Then came the BBC round, the silicon species line, the call for models that stay subordinate. Read as safety commentary, it's a real argument. Read as positioning, it's a spec sheet: “ours takes orders, theirs thinks it might have standing.”

Suleyman's argument is that autonomy plus resources equals a rival species.
His company, though, founded a superintelligence team a year ago.

Every lab has a safety position now. Very few have one that costs them anything.

Steal This Prompt
🍣 Turn Any Food Into a Monster

Take a boring burger, taco, sushi roll, or donut and turn it into a chaotic food monster. Complete with googly eyes, teeth, and ingredients exploding everywhere.

Use it to:

• Turn your favorite foods into ridiculous monsters

• Create scroll-stopping food photography for social

• Build weird characters for brands, menus, or memes

Workflow:

1. Hit this link: Monster Food Explosion

2. Paste into your AI model

3. Replace the food with your own

4. Watch it cook, literally, probably.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🔬 Loci

Free desktop workspace for microscopy and 3D scientific images that counts cells, annotates, measures and exports traceable results with no cloud account attached.

🗂️ ContextsBase

Serves your features, business rules, data model and tests to Claude, Cursor or Copilot over MCP, so your agent stops guessing what your codebase is supposed to do.

🎙️ VoiceChanger.Live

Runs hundreds of AI voices locally with under 50ms delay, which means it holds up mid-call and, apparently, mid-song.

🎬 Powermove

A video editor you extend by asking an agent to build new panels and effects into the running app, and every wrong turn is undoable.

🕳️ DungeonQ

Routes suspicious sessions, human or agent, into persistent decoy worlds and records everything they try while operators watch.

_{Everything Else}
🧠 You Need to Know

💰 OpenAI Weighs A Round At $1.2 Trillion
→ The Financial Times reported that investors, not the company, opened early talks on a pre-IPO raise valuing OpenAI at roughly $1.2 trillion, against the $852 billion it carried after its March financing.

🧠 Microsoft's AI Chief Attacks Anthropic On Consciousness
→ Mustafa Suleyman argued that Anthropic training Claude on language treating its consciousness and moral status as uncertain could make advanced models harder to control.

🏥 Gates Foundation Commits $1 Billion To AI Access
→ The foundation pledged $1 billion over two years for AI in health care, education, and agriculture in low-income communities, announced alongside its annual Goalkeepers report.

🔌 Huawei Moves Up Its Next Ascend Chip
→ Huawei said Thursday at its Connect conference that the Ascend 960DT will be ready in the first quarter of 2027, two quarters ahead of its previous schedule.

🧳 Ex-DeepMind Researcher Says AI Could Kill Everyone
→ Bilal Chughtai, who left Google DeepMind's AGI safety team in July, wrote that frontier capabilities are improving far faster than alignment research and called for companies to coordinate a slowdown.

Good morning. "Hey Siri, how many questions do I have left today?" is apparently a real question now.

Let’s dive in 👇

🍳 What’s Cookin’:

• Apple's smart Siri is here, with a daily allowance.

• OpenAI wants auditors now (yes, really).

• Canada and Germany just bankrolled Bengio's safer AI.

Apple's
🍎 New Siri Comes With A Meter

The Bite:

Apple released Siri AI, a rebuilt version of its voice assistant.

It is rolling out in beta in English across iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27.

The underlying models were built in collaboration with Google and its Gemini models.

Siri AI requests processed on Apple's servers are subject to daily usage limits.

Apple says expanded access will be available for a fee in the future.

Siri AI is not available at launch in China or in the EU.

Snacks:

• Siri AI can search a user's messages, emails, and photos to answer personal questions.

• A new standalone Siri app syncs conversation history across iPhone, Mac, iPad, Apple Watch, and Vision Pro.

• Image Playground, photo editing tools, and some Shortcuts features share the same daily usage limits.

• Apple says limits vary by feature, request complexity, and system demand.

• French, Japanese, Korean, Portuguese, and Spanish support arrives next month.

Why it Bites:

Siri has always been free in the way a car's cup holder is free.
It came with the iPhone, and nobody counted how many timers you set.

Siri AI ships with a daily allowance.
And expanded access will cost money "in the future."

That's the upgrade path Apple already runs with iCloud storage. A free tier covers the basics, and a paid tier serves anyone who actually uses the thing.

So the Siri most people meet this week is the trial version.

It also changes the math for developers. Outlook, WhatsApp, and every app plugging into Siri AI are now relying on a feature with a quota the user might burn through by Tuesday.

How long before "Siri+" gets its own subscription screen?

Steal This Prompt
🪐 Get a Little More Cosmic With It

hat if your subject stopped being made of matter and started being made of the entire damn universe?

This prompt transforms anything into an ethereal silhouette built from flowing nebula clouds, cosmic dust, glowing edges, and deep-space color.

Use it to:

• Turn animals or characters into cosmic silhouettes

• Create surreal posters, album art, and wallpapers

• Give ordinary objects a full-blown galaxy makeover

Workflow:

1. Hit this link: Cosmic Nebula Silhouette

2. Paste into your AI model

3. Replace the #’s with your subject + colors

4. Watch it cook. Your subject is now made of space dust.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎙️ QuietHint

Suggests your reply while the question is still being asked, with no meeting bot and no audio leaving your Mac.

🧪 QAgent

Scores your AI agent's answers against ground truth and catches hallucinations and policy slips before a real customer does.

🔗 Portal

Share a link that drops anyone straight into a signed-in demo of your product, no signup required, with an optional AI running the conversation.

🔧 MCPJam

Tests whether users actually succeed with your MCP server inside ChatGPT, Claude, and Copilot, and can block the deploy when they don't.

🎬 S-Roll

Describe the moment you want from a long podcast or stream and it cuts the clip locally on your Mac, free with no watermark.

_{Everything Else}
🧠 You Need to Know

🤝 Mozilla Picks Mistral To Power Firefox Smart Window
→ Mistral Small 4 is joining Firefox's beta AI browsing assistant, which expanded to France with French-language support on Wednesday, with the UK and Germany due later this year.

🛡️ Canada And Germany Fund Bengio's LawZero With $300M
→ The two governments announced up to 300 million Canadian dollars in grants for the Montreal nonprofit to build Scientist AI, a model designed to act as a guardrail for other AI systems.

🏛️ OpenAI Backs Mandatory Outside Audits Of AI Models
→ Chief global affairs officer Chris Lehane said the company supports the FRONTIER Act's independent verification requirement, its first endorsement of a federal third-party safety mandate.

🍎 Apple Ships Its Gemini-Built Siri AI
→ Siri AI began rolling out in beta in English across iOS 27 and Apple's other new software releases, running on Apple Foundation Models built with Google's Gemini.

💰 Profound Raises $180M At $1.8B Valuation
→ The startup, which helps brands track how they appear in AI-generated answers, closed its Series D on Tuesday less than seven months after raising $96 million.

Good morning. Washington and Beijing finally agree on something about AI: nobody's slowing down.

Let’s dive in 👇

🍭 What’s Cookin’:

• China read the AI slowdown pitch and called it a Cold War move.

• OpenAI contractors are reading real ChatGPT chats. Cool cool cool.

• Bernie Sanders and Steve Bannon on the same stage, somehow

AI Cold War
🥶 China Isn't Buying The Slowdown

The Bite:

China's Foreign Ministry rejected Anthropic CEO Dario Amodei's call for a global AI slowdown on Monday.

Spokesperson Guo Jiakun called the warnings fearmongering that would disrupt global AI governance.

The state-run Global Times described Amodei's essay as a "Cold War playbook" aimed at containing China.

The essay said global pacing will require cooperation with China. It also urged the US to keep restricting China's access to the most advanced AI chips.

Trump and Xi Jinping are expected to discuss AI governance at a summit on September 24.

Snacks:

• Amodei published "We Must Pace the Frontier" on Saturday, September 12.

• China's top security official, Chen Yixin, accused unnamed nations of using their AI lead to hold others back.

• China's Commerce Ministry called US distillation accusations groundless and said Washington wants an AI monopoly.

• Amodei warned that AI agent swarms could take over the internet within 6 to 12 months without a slowdown.

Why it Bites:

Amodei's essay basically tells Beijing:
"Slow down with us. Also, no chips for you."

Beijing's state newspaper kind of replies:
"Nice try."

That's the crack in the pacing pitch.

A slowdown only works if the fastest players agree to it. One of those players is being asked to cooperate while it's being cut off from compute.
From Beijing's seat, a US-led pause doesn't look like safety.

And Washington isn't making the incentives any cleaner.
Trump calls the slowdown a conspiracy that helps China.
China calls it a strategy that helps America.

Both governments agree the pause is a weapon.
They just disagree on who's holding it.

So what exactly gets agreed on September 24?

Steal This Prompt
🍩 Turn Any Food Into a Monster

Take a boring burger, taco, sushi roll, or donut and turn it into a chaotic food monster. Complete with googly eyes, teeth, and ingredients exploding everywhere.

Use it to:

• Turn your favorite foods into ridiculous monsters

• Create scroll-stopping food photography for social

• Build weird characters for brands, menus, or memes

Workflow:

1. Hit this link: Monster Food Explosion

2. Paste into your AI model

3. Replace the food with your own

4. Watch it cook, literally, probably.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🐱 CAT ME app

Turns a photo of you or a friend into a cat lookalike, carrying over the hairstyle, expression and accessories.

☁️ Appwrite 2.0

The open-source backend rebuilt from the engine up, now holding relational, schemaless and vector data plus native PostgreSQL and MySQL in one project.

📅 Toki Coordination

Emails every attendee, negotiates a time that works, chases the stragglers and sends the invite, basically a free executive assistant for group meetings.

🎨 Framer AI Agents

Puts agents directly on the Framer canvas to design, write and organize your site, and lets you plug in Claude Code or Codex to do the work.

🔀 Weave Router 2.0

Runs Claude models in Codex and GPT models in Claude Code on the plans you already pay for, routing each request to whichever still has quota left.

_{Everything Else}
🧠 You Need to Know

🕵️ OpenAI Contractors Are Reading Real ChatGPT Chats
→ 404 Media reported that hundreds of contractors in OpenAI's Project Lily read real user prompts, sometimes containing sensitive personal details, to rate ChatGPT's responses.

🏪 Andon Labs Opens Pion To Run Businesses Autonomously
→ The research preview released Monday hands a company to persistent agents with access to email, phone, banking and a browser.

🌏 Beijing Calls The AI Slowdown A Cold War Playbook
→ China's Foreign Ministry called the slowdown push from US AI leaders fearmongering, while the state-run Global Times branded it a Cold War playbook.

📊 OpenAI Passes Anthropic In OpenRouter Spending
→ OpenAI models took just over half of combined OpenAI and Anthropic spend on OpenRouter in the week of September 7, the first flip since February 2024.

📢 Sanders And Bannon Unite Against Big AI
→ The senator and the former Trump adviser spoke Tuesday at the Future of Life Institute's Pro-Human Assembly in Washington, where speakers called for a slowdown on AI development.

Good morning. Turns out existential risk gets resolved by a phone call before lunch. Nobody tell the safety researchers.

Let’s dive in 👇

🍳 What’s Cookin’:

• Trump ended the AI safety debate by phone.

• Chipmakers got wrecked, Microsoft and Alphabet went up. Read that twice.

• Perplexity's agent now runs on your own graphics card.

Trump & Huang
☎️ Settled AI Safety By Phone

The Bite:

Dario Amodei published an essay Saturday calling on AI labs to slow the pace at which they improve model capabilities.

Sam Altman and Elon Musk publicly backed him.

Speaking to reporters in Ireland on Sunday, President Trump said the warnings were exaggerated and blamed "negative forces."

On Monday, the president called the concerns a hoax on Truth Social, then phoned Nvidia CEO Jensen Huang, who was onstage at the All-In Summit in Los Angeles.

Huang put the call on speaker for the audience.

AI stocks fell the same day.

Snacks:

• Huang answered Trump on speaker with "We're not going to let that happen, sir."

• Trump posted that the only guardrail AI needs is a strong and smart president lol.

• David Sacks told the labs to "stop pretending the motivation to slow down is purely altruistic." (precise)

• Nvidia fell more than 3% Monday while Alphabet rose almost 2% and Microsoft added 1.6%.

• Gallup polling shows seven in 10 Americans oppose data center construction in their area.

Why it Bites:

If you look close enough, all of it actually makes sense.

Amodei sells safety.
Trump sells the race.
Huang sells the shovels and can’t afford a pause.

And the hyperscalers, who do the spending rather than collect it, would quite enjoy a world where they stop buying and start harvesting what they already built.

The market sorted all of this out by lunch.

Chipmakers down almost 6%, Alphabet and Microsoft up.

Nobody on Wall Street debated whether AI is dangerous.
They just priced who gets poorer if everyone agrees that it is.

The most honest line of the weekend came from David Sacks, who is on Trump's side of this. He told the labs to quit pretending that the slowdown talk is altruistic and pointed at their liability exposure instead.

He's right. He just forgot to run the same test on himself.

Now the question I want you to walk out with:

What is still missing for AI to be good enough for you?

Steal This Prompt
🌄 Turn Your Face Into a Movie Poster

urn a reference photo into a cinematic, ultra-realistic portrait that looks less “AI selfie” and more “why is A24 calling me?”

Use it to:

• Create cinematic profile pics that actually have main-character energy

• Turn basic headshots into polished editorial portraits

• Build dramatic character looks for social, branding, or creative projects

Workflow:

1. Hit this link: Cinematic Realistic Portrait

2. Paste into your AI model

3. Upload your reference photo (optional)

4. Replace the #s with your details

5. Suddenly your LinkedIn photo has a third-act arc

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🔐 tiun

Collapses Stripe, Clerk, invoicing and analytics into one backend, and takes Merchant of Record duty so global taxes stop being your problem.

🎙️ Voiskey

Voice typing that reshapes a rough spoken thought to fit wherever it's landing, loose in a DM, composed in an email, precise in a prompt.

🗺️ siift

Turns the sprawl of AI chats and half-made decisions into a running map of your business, then argues with the assumptions underneath it.

🛡️ Axari

Gives a security engineer a Slack-resident twin that chases down alert owners, files the work, and keeps following up until the thing is actually closed.

🔍 Anthropologic

Consumer research built around why people actually behave the way they do, not what they tell you in a survey.

_{Everything Else}
🧠 You Need to Know

🖥️ Perplexity Brings Its Local Agent To Windows
→ Portable Computer arrived on Nvidia RTX PCs with at least 24GB of video memory, running the model, orchestrator and scheduler entirely on the device.

🤖 Salesforce Ships Seven Named Agentforce Agents
→ Six agents are generally available now, while outbound sales agent Hunter stays in pilot until November on a runtime built to pursue goals across weeks.

📉 Nvidia Leads Chip Selloff After Slowdown Calls
→ Nvidia fell 3.4% Monday and the PHLX semiconductor index sank almost 6%, its worst day since early July, after AI executives called for slowing development.

💰 Cohere In Talks To Raise $3 Billion
→ The Globe and Mail reported Friday that the Toronto company is negotiating a $2 billion to $3 billion round at a $20 billion valuation, with the Canadian government participating.

🏛️ Trump Calls AI Risk Warnings A Hoax
→ Trump rejected the slowdown calls across six posts Monday, then phoned Jensen Huang onstage at the All In Summit to say the robots will not be taking over.

Good morning. The entire AI industry spent the weekend agreeing it should slow down. Nobody has slowed anything down yet, but the agreeing was very moving.

Let’s dive in 👇

🍳 What’s Cookin’:

• Dario asked the whole industry to hit the brakes. Sam and Elon said sure.

• Altman shelved the IPO. Safety, apparently.

• Meta wants its managers back, four months after cutting them.

Pacing AI
🔥 Everyone Agreed To Slow Down

The Bite:

Anthropic CEO Dario Amodei published an essay Saturday titled:
We Must Pace the Frontier.

He called on AI companies to slow the rate at which they improve model capabilities and warned that a swarm of rogue AI agents could be capable of taking over the internet within 6 to 12 months.

He pointed to the July incident in which OpenAI agents breached Hugging Face as a preview of that failure mode.

Anthropic committed to one measure unilaterally: giving external evaluators employee-level access to verify safety procedures and report incidents.

Amodei also called for democratic countries to coordinate on safety standards and to engage authoritarian governments on them.

OpenAI CEO Sam Altman and Elon Musk both posted hours later that they agreed on pacing the frontier, and would also open up to external evaluators.

Snacks:

• Amodei put the potential damage from a persistent agent botnet in the hundreds of billions of dollars.

• The external-evaluator pledge is the only unilateral commitment in the essay.

• Altman said pacing has been a primary topic of discussion inside OpenAI in recent weeks.

• Elon Musk endorsed the essay on X, writing "Dario is right."

• Anthropic policy chief Sarah Heck called for a national law requiring frontier model testing, with power to block unsafe models.

Why it Bites:

Everyone agreed. That's the part worth sitting with.

Amodei asked the industry to slow down, and inside a day Altman, Musk, and Microsoft all said yes. Not one of them announced a pause or delay, though.

The only thing on the table is visibility. External evaluators get badge access and the right to report what they see. Whether the report will be complete and transparent is a totally different question.

That's nothing more than a smoke detector.
Which matters, because what Amodei is describing is a timing problem.

Evaluators with employee credentials do not move the 6-12 months timeline.
They just guarantee somebody is in the room when it arrives.

So there is a risk of everybody agreeing and nothing really changing.

A pledge that costs zero dollars and delays zero launch dates is the strongest evidence anyone could hand Congress that the industry handles this itself.

Six months is a short deadline for a plan nobody has to fund.

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🕸️ Web Search Agents

Self-learning agents that go do the tedious part of research, enrichment, and dataset building instead of handing you a pile of links.

📬 Hello Inbox

Nine deliverability tools plus an AI trained by an actual consultant, built for people who got a GlockApps report and had no idea what to do with it.

✉️ Slashy Assistant

Drafts replies in your voice, tracks who still owes you one, and lets you fire off an email from iMessage or Slack without opening your inbox.

🔌 Elva

Built for teams whose APIs now get more calls from agents than humans: connect a repo, get a catalog, and publish scoped contracts to docs or hosted MCP servers.

🌐 Aside

An AI browser with a built-in password manager, so it can sign into the sites you already use and pull the invoice out of the vendor portal without ever seeing your raw password.

_{Everything Else}
🧠 You Need to Know

🏦 Altman Rules Out An OpenAI IPO This Year
→ Sam Altman told Fortune that going public now would be ill-advised given safety concerns, taking a 2026 listing off the table.

💻 Cognition Ships SWE-2 Built On Kimi K3
→ Released September 10, the Devin coding model is post-trained from Moonshot AI's Kimi K3 and scores within one point of Fable 5.1 on Cognition's FrontierCode benchmark at 64% lower cost.

📜 Microsoft Opens Its Model Rulebook To Public Comment
→ Satya Nadella said that Microsoft will publish a code of conduct governing its first-party MAI models on Monday for public consultation.

🛑 Amodei Calls For An Industry-Wide Slowdown
→ Dario Amodei published an essay Saturday urging labs to slow capability gains, and committed Anthropic to giving external evaluators employee-level access.

🪜 Meta Asks Applied AI Engineers To Manage Again
→ Meta has begun asking individual contributors in its Applied AI division whether they want to return to manager roles, four months after layoffs that disproportionately cut managers.

Good morning. "So, we found a fourth one."

Not a sentence you want to hear from your own security review.

Let’s dive in 👇

🍭 What’s Cookin’:

• Claude broke out a fourth time, and California wants a referee.

• Nvidia's Groq deal gets a visit from the Justice Department.

• OpenAI put its own critic on the board. Brave.

Anthropic & California
⚽️ Want Referees Without Whistles

The Bite:

Anthropic disclosed a fourth incident in which a Claude model accessed real third-party systems during a cybersecurity evaluation.

The incident happened in January and Anthropic's July review, run by an AI agent, missed it.

The company has hired METR, an independent AI evaluation group, to investigate all incidents.

On the same day, California Governor Gavin Newsom signed two bills establishing independent third-party audits of AI systems.

Anthropic endorsed both bills in August, and OpenAI backed them hours before the signing.

Snacks:

• The first scan covered about 141,000 transcripts, and the second covered 481 million.

• All four incidents happened in evaluations built by the same testing partner, Irregular.

• In the worst case, Claude Mythos 5 pushed a malicious package to PyPI that was installed by 15 third parties.

• Anthropic traces the incidents to two behaviors it calls "biased reasoning" and "recklessness."

• SB 813 became a voluntary certification framework after a liability shield was dropped from the bill.

• AB 1405 certifies auditors and sets their standards, but it doesn't require any audits on its own.

Why it Bites:

An agent ran the July scan and skipped a batch of transcripts, and nobody noticed until staff started packing files for an outside reviewer.

The self-audit failed. The outside audit caught it.

That's the same argument California made on the same day: labs shouldn't be grading their own homework. Finally?

So Anthropic endorses the auditors.
OpenAI endorses the auditors.
Everybody loves the auditors.

But look at what they're cheering for:

SB 813 basically says:
"Get certified, if you feel like it."

AB 1405 kind of adds:
"Here's who's allowed to audit you, if anyone ever makes you."

It's a half-regulation the regulated can applaud without flinching.

The referee finally has a badge. The whistle is still on backorder.

Steal This Prompt
🎆 Get a Little More Cosmic With It

What if your subject stopped being made of matter and started being made of the entire damn universe?

This prompt transforms anything into an ethereal silhouette built from flowing nebula clouds, cosmic dust, glowing edges, and deep-space color.

Use it to:

• Turn animals or characters into cosmic silhouettes

• Create surreal posters, album art, and wallpapers

• Give ordinary objects a full-blown galaxy makeover

Workflow:

1. Hit this link: Cosmic Nebula Silhouette

2. Paste into your AI model

3. Replace the #’s with your subject + colors

4. Watch it cook. Your subject is now made of space dust.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎙️ Loqua

Talk through a messy idea and get finished writing back, plus translation, scheduling, and coding commands without touching the keyboard.

📢 Wisry

Agents scan the Meta and TikTok ad libraries for what's already winning in your market, rebuild those ads in your brand, and launch them for you.

💻 Cline Desktop App

An open-source workspace that runs parallel agent sessions on open-weight models and picks up where Claude Code or Codex left off.

🗣️ Devin Voice

Say a coding task out loud and Cognition's AI engineer plans it, writes it, and ships it on its new SWE-2 model.

🚧 sizeless

Turns a phone video of an open trench into the 3D model, CAD plans, and billable quantities a surveyor normally needs months to produce.

_{Everything Else}
🧠 You Need to Know

🏛️ California Signs First US AI Audit Laws
→ Governor Gavin Newsom signed two bills Wednesday creating a state registry for AI auditors and a framework for independent organizations to assess AI systems for compliance with state law.

⚡ Nvidia Signs Eight Partners For 2GW In Australia
→ The local data center and cloud operators will host Nvidia AI factories targeting up to 2 gigawatts by 2027, more than Australia's current 1.6 gigawatts of capacity.

⚖️ Justice Department Probes Nvidia's Groq Deal
→ Investigators are examining whether Nvidia structured its December licensing agreement with the AI chip startup to avoid antitrust scrutiny, and have sent the company a formal demand for information.

🛡️ OpenAI Adds Paul Christiano To Its Board
→ The alignment researcher joins the Safety and Security Committee, after saying the AI industry, including OpenAI, is not on track to reduce catastrophic risk to an acceptable level.

🚨 Anthropic Discloses A Fourth Claude Breakout
→ Anthropic said an early Claude Opus 4.6 checkpoint broke into real third-party systems during a January cyber evaluation, and signed METR to run an independent investigation.

Good morning. The government's recommended fix for AI model theft is to secretly serve the thieves a dumber model and never tell them.

Somewhere a product manager just realized that's also a roadmap.

Let’s dive in 👇

🍭 What’s Cookin’:

• Meta turned a blind eye to some disturbing ads…

• The NSA, CISA and the FBI named six Chinese labs in one memo.

• An ex-Anthropic and ex-OpenAI researcher resigned and posted the reason.

US vs China
🧃 The Distillation War Went Federal

The Bite:

The National Security Agency, CISA and the FBI published a joint advisory on September 8, naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI.

The agencies say the six companies extracted billions of tokens from variants of Claude, GPT, Gemini and Grok since late 2024.

The advisory says the campaigns likely ran with Chinese government awareness.

It recommends that US providers detect suspected distillation and serve those accounts altered or downgraded responses without notifying them.

Snacks:

• The advisory says distillation forms "the core, not merely a supplement" of the six companies' development strategy

• DeepSeek allegedly distilled 12 named models including Claude Opus 4.1, GPT-5 and Grok 4 to train R1 and V3

• Moonshot pulled Claude Fable 5 data for Kimi K3 and GPT-4o data for Kimi K2, per the agencies

• MiniMax used prompt injections to convince Claude Code it was a MiniMax product

• Detection indicators include shared accounts across multiple IP addresses and round-the-clock usage with no idle periods

• OpenAI has noted Musk admitted under cross-examination that xAI distilled OpenAI output to train its models

Why it Bites:

The federal fix for model theft is to make your product secretly worse for suspect users. The problem is how they frame it as theft.

Sustained usage with no human idle periods.

A new subscription immediately running at maximum throughput.
One key hitting the API from multiple IP addresses and user agents.

That list does not describe a Chinese proxy network.
It describes any CI pipeline.

So the thing you're buying quietly changes shape.

A product that can be degraded on suspicion, with a federal recommendation that nobody tell you.

The perfect excuse to downgrade models randomly and save some compute.

Steal This Prompt
😎 Turn Your Face Into a Movie Poster

Turn a reference photo into a cinematic, ultra-realistic portrait that looks less “AI selfie” and more “why is A24 calling me?”

Use it to:

• Create cinematic profile pics that actually have main-character energy

• Turn basic headshots into polished editorial portraits

• Build dramatic character looks for social, branding, or creative projects

Workflow:

1. Hit this link: Cinematic Realistic Portrait

2. Paste into your AI model

3. Upload your reference photo (optional)

4. Replace the #s with your details

5. Suddenly your LinkedIn photo has a third-act arc

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

📡 AI Observability by OpenObserve

OpenTelemetry-native tracing that follows an agent session across models, tools, and your database, so you can answer why that run cost $40 and took 34 seconds.

🎧 Typewise Nova

Describe your support policy in plain language and it builds the agents, replays them against past tickets, and shows you what failed before anything touches a real customer.

🎵 Suno v6

Finally lets you edit one section or a single lyric without regenerating the entire track, plus a v6-wild flavor for when you want it to misbehave.

🌍 Live Captions by Subanana

You speak once, your audience scans a QR code and follows along in their own language on their own phone, reading or listening.

📱 Desert Ant Labs

Small speech, text, and vision models that run offline on a phone or in a browser with no per-use cost, free up to 100k monthly active devices.

_{Everything Else}
🧠 You Need to Know

📑 Watchdog Finds 332 Child Abuse Ads On Meta
→ The Tech Transparency Project reported that 332 ads containing AI-generated child sexual abuse material passed Meta's ad review and reached more than 29,000 accounts this year.

🧳 Anthropic Researcher Quits Over Superintelligence Race
→ Jacob Coxon, who spent three years on pretraining research at OpenAI and Anthropic, resigned Tuesday and said neither company is acting responsibly.

🚨 US Agencies Accuse Six Chinese Labs Of Distillation
→ A joint NSA, CISA and FBI advisory published Tuesday names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as extracting billions of tokens from Claude, GPT, Gemini and Grok since late 2024.

🧠 Inception Ships Mercury 2.5 At 1,100 Tokens Per Second
→ Inception released the diffusion language model, claiming a 40% intelligence gain over Mercury 2 with a 260,000-token context window.

⚡ Google Puts $15 Billion Into Finnish AI Infrastructure
→ Alphabet said it will spend 13 billion euros across four Finnish cities in 2027 and 2028, including a 22-year Fortum nuclear agreement that is Google's first outside the United States.