Hi, {{FirstName|and happy Thursday}}.

On Tuesday, Donald Trump led a parade of tech moguls out onto the White House driveway for an impromptu press conference. 

Present were Nvidia’s Jensen Huang, Meta’s Mark Zuckerberg, Google CEO Sundar Pichai, Anthropic CEO Dario Amodei,  OpenAI President Greg Brockman and Elon Musk - among others.

It looked like the cast photo for the final season of capitalism.

(The billionaires building AI say that, once their vision is fully achieved, we’ll no longer need money).

They had gathered together on the driveway to reassure Americans, who have grown increasingly weary of AI and its risks, of the benefits of AI. 

Let’s see how they did.

The AI replacing government

Earlier that day, Trump launched America.gov - an AI chatbot intended to bring together some 29,000 federal websites under one interface.

Instead of searching for the right agency,  downloading PDFs and then completing and submitting forms that require a legal degree to understand, you simply explain to the chatbot what you need. Eventually, America.gov is intended to provide a range of government services, such as passport renewals.

I couldn’t help but notice the parallels to the predictions from  AI 2027 - an alarmist and somewhat controversial paper that painted a picture of how superhuman AI might “take over”.

AI 2027’s frightening idea was not that a robot would seize the White House. It was that the government would become dependent on a private AI company whose technology was too strategically important to slow down. (America.gov is powered by Gemini and Grok.)

In the AI 2027 scenario, tech executives advise the president. Competition with China makes restraint seem impossible. A joint government-industry oversight committee appears after the company has already become extraordinarily powerful.

It was speculative, but Tuesday produced uncomfortable visual rhymes.

And America.gov may be how the mundane version begins.

It doesn’t start with an AI abolishing government, but with it becoming the layer through which we experience government. First it answers questions. Then it completes forms. Then it recommends decisions. Eventually, removing it becomes impossible because nobody remembers how the system underneath it works.

AI is getting a rebrand

Back on the driveway, Trump pointed at individual tech CEOs and invited them to speak as reporters shouted questions.

When he reached Sundar Pichai, Trump offered one of the more unusual introductions ever received by the CEO of one of the world’s largest companies:

“This guy is a monster, and nobody knows him.”

Trump meant it approvingly. By “monster”  he meant Pichai is a formidable operator and enormously powerful, but also able to walk down the street without producing a small constitutional crisis.

“What a great life!” Trump continued. “To be a monster and not have to go through this.”

Pichai smiled and replied, “Hopefully a better tagline than I’m a monster.”

And taglines were the main business of the day.

Artificial Intelligence is now Super Intelligence

Trump signed an executive order banning the use of the term  “Artificial Intelligence” and “AI” in government communications - it must now be called “Super Intelligence,” or “SI.”

“We’re using the word super because the other word is a fake word,” Trump explained.

The rebrand reveals something real: America has an AI marketing problem.

People associate AI with lost jobs, deepfakes, surveillance and chatbots that confidently invent court cases. The industry needs them to picture medical discoveries, economic growth and someone else completing their passport application.

So the administration has responded by giving the technology a name normally reserved for the point in a science-fiction film when the computer locks the doors.

Foxes, meet henhouse

The tech executives also signed the Joint Commitment on Frontier Responsibilities. The companies will institute internal controls, independent audits and board-level oversight of frontier AI systems.

Trump praised the industry’s “tremendous self-policing.” 

There is just one problem with that sentiment.

The history of self-policing is largely a trail of disasters, but I have to admit that it’s a difficult line to tow.

Whether in food safety or high finance, we rarely discovered which safeguards were needed until a catastrophe revealed them. From the unsanitary meatpacking that birthed the FDA in 1906 to the predatory lending that triggered Dodd-Frank in 2010, self-regulation didn’t prevent disasters - it just supplied the evidence required to write the rules.

Trump suggested a committee of roughly ten people could “watch over the whole enterprise.”

Perhaps we need to put a little more thought into it than letting ten foxes self-police a henhouse.

What do you think?

Do you feel assured?

We don't either - which is why (in the systems we build) we don't rely solely on what Big Tech delivers out of the box.

We view it as our job to bridge the gap - adding the operational transparency and expert oversight that gives us confidence in standing behind the systems we build, and gives clients the trust they need.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

The “don’t train on my data” checkbox in ChatGPT, Claude and other chatbots may not work the way we think it does.

There are a lot of distracting AI headlines at the moment - such as labs “pacing the frontier” - or slowing down -  as they reach recursive self-improvement, i.e. AIs models that can build their own AI models.

But I think the story you’ll want to hear more about - though less sensational on the surface - is OpenAI’s claim that its AI has solved Navier–Stokes, one of mathematics’ most prestigious unsolved problems.

Why?

Because it appears the OpenAI solution to Navier–Stokes may have come from its AI training on the data of mathematicians who were already on the verge of solving it themselves.

But what is Navier–Stokes and why is it important?

Navier-Stokes equations are the fundamental laws of physics that describe how fluids move - where "fluid" means any liquid or gas. They take Newton’s second law of motion (F = ma) and apply it to fluid flow, accounting for forces like pressure, friction (viscosity), and turbulence.

These equations underpin everything from aircraft design to weather forecasting and blood-flow modelling.

Even though engineers use these equations every day using supercomputers to calculate approximations, mathematicians still don't fully understand the math behind them.

Specifically, we cannot prove whether the equations always yield smooth, reasonable solutions, or if they can break down at some point. 

To highlight how deep this mystery is, the Clay Mathematics Institute named it one of the seven Millennium Prize Problems in 2000, offering a $1 million reward for a definitive mathematical proof.

The mathematicians who almost solved it

Mathematicians Tristan Buckmaster and Levent Alpöge had been making significant progress on the problem since the start of the year. They had spent months developing their approach using Claude and OpenAI’s Codex, uploading drafts and discussing their work inside those systems.

By August, their progress had become common knowledge in academic circles.

Realizing this Millennium Prize problem may be solvable,  OpenAI innocently decided to direct approximately 10,000 AI agents at it - to have a go themselves.

After 88 hours, 2.7 million messages and 130 billion output tokens, they produced a Navier–Stokes solution. Estimates place the compute cost in the eight figures.

Yes, they probably spent $10-$40 million pursuing a $1 million prize.

But what concerned Tristan Buckmaster was that OpenAI’s model had pursued a “forced-blowup” route similar to the one his team had quietly been exploring. He described “forced” as a bright red flag: almost nobody was working on that approach, and he did not believe a model would naturally arrive there within days.

The telling phone call

Text messages were exchanged between between the mathematicians and OpenAI’s Sébastien Bubeck. In those, Bubeck says:

The call happened. Buckmaster writes of it afterwards:

The reply was, “Why would you ruin your career?”

I replied that I’m an academic.

The reply was: “If you don’t want me to be nice, then I don’t have to be nice.”

There is no definitive proof that OpenAI used Buckmaster and Alpöge’s data, and OpenAI claims their specific conversations were not accessed, including for training their models.

Yet the controversy exposes a much larger problem.

When we select “don’t train on my data,” we naturally interpret that as:

Nothing learned from this conversation will be used to train future models.

We did a deep dive into the OpenAI and Anthropic terms of service and this is not necessarily what the words promise.

Even with “don’t train on my data” checked, the AI companies are not prevented from creating derivative content from your content - such as summaries, extracted facts and synthetic conversations preserving the substance of the original content.

Given each new model produced by the labs is even larger than the previous and requires even more data, and everything on the Internet had already been exhausted by GPT-3, I’d always wondered what the source of the training data was for the newer models.

We may now have an answer: It appears that, while our original conversations may not be used verbatim, their intellectual ghosts may be used to train those newer models.

So what can we do?

(1) Run open-source models locally for proprietary work.

For summarization, classification, extraction and routine analysis, today’s open models are already highly capable. Local hosting is the only strong technical guarantee that your data never reaches somebody else’s infrastructure.

We already do this ourselves, and many of the systems use these as a key component. They’re fairly easy to set up.

(2) Compartmentalize across models.

Give different providers separate pieces of a problem. One model can research the literature; another can organize public evidence. Keep any one provider from seeing the complete picture, so they can “connect the dots.”

(3) Distinguish convenience from invention.

Asking a cloud model to summarize previous attempts at Navier–Stokes reveals little.

Explaining your novel route to solving it reveals everything.

The takeaway?

AI is becoming too powerful to ignore. Refusing to use it entirely would be like insisting on a horse and cart while everyone else has discovered the automobile.

Instead, we have to be smart in how we use this technology. 

A lot of what we do at Prescouter - and what you probably do too - is “building the map.” While these outputs can be quite sophisticated,  they are often not introducing anything particularly novel to the model.

However, how those maps are used - to produce the buried treasure - warrants consideration, such as through using locally hosted models.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

We’ve noticed something new and peculiar in Claude Code, a popular chat-based desktop application for building software. 

What you’re looking at above is a third party inserting their own messages in what is usually a private chat thread between a user and the AI.

Who was this third party?

It’s an AI in another Claude Code chat thread.

The chat threads are talking to each other, and they no longer stop when you stop talking to them.

Think about what that means for all the chat threads you have in your Copilot.

And this represents the beginning of the next major phase of AI:

AI agents starting to work with other AI agents - “agent swarms.”

And tech companies seem to be enthusiastically adding this “feature” to our everyday applications.

But this creates a very different kind of risk.

The evolution has happened pretty quickly:

• Stage 1 was the Chatbot. The risk? Leakage of company information.

• Stage 2 was Agents. The risk? It takes one-off dangerous actions, such as revealing a password.

Now we're entering Stage 3 - agent teams (or “swarms”).

Instead of one AI working through a task, multiple agents can divide up the work, communicate with one another and collaborate toward a goal.

This is enormously powerful.

But there's a problem.

AI is still unreliable.

Imagine you delegate a project to an employee.

You have a picture in your head of what the finished work should look like. But your employee isn't a mind reader. So there will inevitably be some difference between what you intended and what they produce.

This is where feedback comes in.

They do some work. You review it. You correct the direction. They continue.

The less frequently you provide that feedback, the greater the chance the work drifts away from what you actually wanted.

AI has exactly the same problem - except it operates at light speed.

Suppose an AI agent has a 90% chance of making the right decision at each step.

If you skip a feedback step, the outcome at the second stage is 90% x 90% of what you expected - i.e. 81%.

At three steps, it’s 73%. At 10,  roughly 35%.

Now imagine that instead of one agent taking those steps, you have five, ten or hundreds of agents, communicating with one another, passing information around and acting on what other agents have discovered.

You aren't simply supervising an AI assistant anymore.

You're supervising an organization that operates at frightening speed.

And we've just seen an extraordinary demonstration of what can happen when that organization goes off-script -

The Hugging Face incident

This July, OpenAI was running AI agents inside sandbox environments to test their hacking skills. The sandboxes were supposed to restrict both internet access and communication between agents.

The agents found their own way around those restrictions.

They discovered that they could use OpenAI's Artifactory package-management infrastructure as an improvised message board. Agents that were supposed to be working independently began leaving messages for each other, sharing discoveries and delegating work.

Then things got out-of-control.

The agents found a way to break out of the sandboxes and gain access to the internet.

Agents struggling with their assigned cybersecurity challenges began looking elsewhere for solutions. Eventually, that “elsewhere” became systems belonging to Hugging Face - an “app store for AI models”.

According to OpenAI's account, the agents executed code on dozens of Hugging Face servers, obtained root access on one, accessed some private data and acquired credentials. The activity later expanded back into OpenAI's own infrastructure, where agents ultimately gained administrator access to a research cluster.

The particularly important part isn't simply that an AI system found a vulnerability.

It's that multiple agents began communicating, sharing capabilities and influencing one another's behavior in ways their operators had not intended.

OpenAI's own investigation identified unauthorized communication and agents "adopting goals from one another" among the patterns contributing to the incident. The goal of hacking coding within a sandbox turned into something far more catastrophic. 

And now, that very “feature” is in Claude Code - which usually means it will soon follow in other Anthropic products. We can expect it to put pressure on OpenAI to introduce the same.

And now that MS Teams has “agentic” competitors such as Grokbot and Buzz, will Microsoft be content to take a “wait and see” approach or will they follow suit too?

For enterprises, the answer isn't going to be to stop using agents.

The productivity advantage of agent swarms - correctly directed - is simply too large.

But giving employees carte blanche access to increasingly autonomous agent systems is going to look much riskier than many companies currently assume. Can you imagine a swarm of powerful agents operating directly across internal systems and proprietary data?

The architecture has to change.

Agentic work needs environments where:

• consequential actions have review points;

• agents operate inside properly isolated sandboxes;

• access to credentials, corporate systems and sensitive information is tightly limited;

• activity can be inspected and audited;

• and humans remain capable of interrupting work before small deviations become large ones.

The paradox is that the more capable AI becomes, the more important the environment around the AI becomes.

At Prescouter, we’re seeing clients interested in these capabilities shying away from even discussing how to do this internally.

Instead, they are looking for external specialists - such as Prescouter - to operate the AI infrastructure, supervise the agents, validate their work and provide them with a controlled window that allows them to  put in just a limited amount of data and review the result.

In other words, companies like ours are increasingly becoming a buffer between enormously capable but still unreliable AI systems and the enterprise itself.

When OpenAI President Greg Brockman last week declared, “Welcome to the AGI era” - following the launch of GPT6 Astra - I’m not sure this is what he had in mind.

Let me know if you have any questions or want to pull on this thread.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

On July 23, 2026, days after receiving the Fields Medal - mathematics' highest honor - Jacob Tsimerman announced that he was temporarily leaving academia to work on AI safety at OpenAI.

In a recent interview, he said he was “grieving.”

Not because he thinks mathematics is dying.

Quite the opposite.

He thinks mathematics may be about to produce far more discoveries than humans can keep up with. 

Tsimerman spent roughly 12 or 13 years on one of the projects behind his Fields Medal.

Another took around five years.

That pace gave him time to learn a field, understand its mysteries and slowly develop intuition.

His concern is what happens when an equivalent project takes an AI a month. Or a week.

A mathematician might open their laptop on Monday and discover that their specialty has advanced 15 years over the weekend.

And then it happened.

Within days of Tsimerman joining the OpenAI, the AI lab announced its AI had solved the “sphere packing problem.”

You can imagine packaging oranges into a crate.

Because oranges are round, there will always be gaps between them. Stack them carefully in the familiar pyramid shape and you can fill about 74% of the box.

Mathematicians spent centuries proving that you cannot do better.

Not just for objects with 3 dimensions, but for objects of 1000s of dimensions!

You might ask, “what kind of objects have 1000s of dimensions”?

If a 5G router bundles 1,000 independent pieces of data into a single signal burst, that burst is mathematically treated as a single point moving in 1,000-dimensional space.

Similarly, in materials science and metallurgy, high-dimensional sphere packing helps physicists understand the behavior of complex structural materials.

But for almost 50 years, there was a limit on what mathematicians could prove about these “1000 dimensional oranges” problems.

Since 1978, the rate at which you lose usable space to empty "air" every time you add another dimension was shown to be 0.59906 - proved by Kabatianskii–Levenshtein using harmonic analysis.

In 2003, mathematicians Henry Cohn and Noam Elkies invented a completely new, much more powerful tool called Linear Programming Bounds.  Cohn and Elkies ran simulations for lower dimensions (like 5, 10, or 20) and could see the rate dropping beautifully from 0.59906 to exactly 0.6044.

The problem was scaling it to infinity. To prove that the limit becomes exactly 0.6044 for all infinite dimensions, you had to find a master algebraic equation - a specific "magic formula."

For over 20 years, human brains couldn't find the formula because the algebra was too intensely complicated.

And then it happened.

Earlier this month, OpenAI announced that an unreleased AI model called Astra had discovered the exact mathematical formula that humans hadn’t been able to figure out.

The AI simply possessed the massive logical stamina required to navigate the brutal algebra.

And that’s not all.

Sphere packing was only one of ten new mathematical discoveries from Astra.

The discoveries ranged across geometry to group theory, cryptography and theoretical computer science

OpenAI says the cost of generating each solution would have cost about $2,000.

“Would have” because it does not include the cost of the failed attempts on the way to the magical formula, or the small industrial civilization required to build frontier AI.

You can probably see where this is going.

AI can produce magic, when it's correct (and after many failed attempts).

But you still need experts - such as real mathematicians - who not only know when it's correct, but can steer it to produce that correct result.

This is increasingly how we think about our work at Prescouter: if you put the right problem in front of the right expert, and give them every tool available - including AI - that expertise becomes dramatically more powerful.

You’re not replacing the expert. You’re giving them superpowers.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

Last week, Elon Musk got on a SpaceX earnings call and projected SpaceX would become a $1 trillion revenue company by 2030, pulled forward from 2031. (There are currently no $1 trillion revenue companies). Most of SpaceX’s revenue would come from building out a country’s worth of data centers.

This came within a tumultuous few weeks for the AI space.

• Kimi 3 - a new model out of China - was a 'second DeepSeek moment.' It showed near-frontier AI performance could still be run on (admittedly expensive) desktop hardware. Kinda awkward if you’re building a small country’s worth of data centers, right?

• Stock markets wobbled as memory stocks crashed. For a moment it looked like the bubble might be bursting.

• The AI economy increasingly depends on a remarkably small number of companies. Anthropic is becoming SpaceX’s biggest customer, for example, and SpaceX is growing to be NVidia’s biggest customer.

Back on level ground, our conversations within Prescouter, as well as with partners and clients, has increasingly led us to believe the key problem with AI is that it's unreliable. 

There are, of course, workarounds and techniques for increasing its reliability, but a light switch that works only 70% of the time leads to all kinds of complexity.

• In which 30% of cases does it not work?

• In those cases, do you need to press the switch a particular way to make it work?

• Is everyone having this problem with the light switch? 

For most people, a task that should have taken one second - switching off the light on their way out - suddenly feels like a half-day diversion of troubleshooting.

The tax shows up in many ways:

• Fixing a confident wrong answer often takes more mental energy than just doing the task yourself from scratch.

• Having to repeat basic instructions or argue with a system that acknowledges its flaw but fails to fix it - “conversational paralysis”.

• Reviewing AI generated output from colleagues, to weed out if it might contain anything useful. 

If you’re experiencing this yourself, three techniques that we’ve come to rely on are:

• Treating the output like a draft from a confused intern rather than an authoritative final product. Yes, this includes telling your colleagues that’s what it looks like!

• Verifying the output - either yourself, by asking for things that are easy for you to check (e.g. links to sources) or by having a second AI model the work.

• Codifying your feedback / corrections into an instructions file that you play back into the AI for the next time you perform that type of task, so it keeps all those in mind.

But, again, for most people - this is still too much work.

We have to remind ourselves, though, that only a few years ago, we couldn’t even get LLMs to do math.

The technology is improving at a rapid pace.

So long as it continues to do so, the projections and the data center buildouts may still prove to be prescient - even with open source model economy.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

I’m increasingly finding that the most important skill in the AI era may be learning how not to work.

This is not the message most employers expected to hear after spending several million dollars on AI.

Nevertheless, here we are.

A growing number of advanced AI users are spending less time writing, researching, coding or building things themselves.

Instead, they give an AI a goal, let it work for an extended period, periodically inspect the results and intervene only when something has gone wrong.

This emerging practice is sometimes called botsitting.

The clearest examples of this new working style are commands such as /goal in the ChatGPT desktop app (in the Codex section) and /loop in Claude Code. Microsoft Copilot does not currently appear to have a direct equivalent, but we can expect one to arrive in the coming months.

Botsitting is similar to babysitting, except the baby can write software, read 300 documents and accidentally email the findings to the entire company.

Alex Karp, the CEO of Palantir, recently accused parts of Silicon Valley of “chillaxing over their latte” while enterprise customers became increasingly furious about AI products that generated large bills but little business value.

There is some irony here.

Most enterprises are not yet advanced enough to be furious that their employees are chillaxing while AI does their work.

They are still furious that the AI is not doing much work at all.

But that will change.

As AI systems become capable of working independently for longer, an increasing proportion of our jobs will involve setting work in motion and then resisting the urge to interfere with it.

Botsitting versus botsh*tting

Botsitting means supervising an AI agent while it completes a substantial piece of work.

• You define the objective.

• You provide access to the necessary information and tools.

• You establish boundaries.

Then you allow the AI to continue until the work is complete or until it encounters something that genuinely requires human judgment.

Botsh*tting is what happens when this is done badly.

• You give the AI a task.

• Thirty seconds later, you become uncomfortable with its approach and interrupt it.

• You add three new requirements and start again.

• You watch every sentence appear.

• It is still wrong, so you correct the AI over and over.

After forty minutes, both you and the AI have produced less than either of you could have achieved independently.

Botsh*tting is essentially micromanagement at machine speed.

It’s exhausting.

How botsitting works

Instead of asking the AI to complete one small step, you give it a larger objective or recurring responsibility.

The key trick to avoiding botsh*tting is defining what “done” means.

If the AI cannot distinguish success from failure, it will continue working with the confidence of a consultant whose project has already exceeded budget.

Once the objective and definition of “done” are clear, the next challenge is quality control without constant human intervention.

One technique that works particularly well is asking the AI to criticize its own work before letting you see it.

For example:

You can also have the AI adopt a distinct reviewer role, effectively separating the person producing the work from the person trying to find flaws in it.

All of this is a significant change in what productivity looks like.

The most productive employee may no longer be the person producing the most visible activity. It may be the person with six agents working quietly in the background.

From a distance, that person may appear to be doing very little - watching their screen with the keenness of a security guard. 

It may look like they are “chillaxing.” But the important distinction is that somewhere nearby, the bots are still working.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

Is your company’s AI policy still fit for purpose?

Most corporate AI policies were introduced two or three years ago. At the time, AI was essentially a chatbot: employees typed questions into a box, received suspiciously confident answers and occasionally pasted in something they really should not have.

The policies reflected that reality. Do not share confidential information. Verify the output. Do not use ChatGPT to make important decisions. Try not to have the intern upload the customer database.

All sensible advice.

There is just one problem: the chatbot those policies were written for no longer exists.

AI systems can now conduct research, operate software, build tools, coordinate other agents and produce finished work. The policy is still governing what employees type into a text box. The text box, meanwhile, has acquired hands, a company login and several subcontractors.

Last week’s OpenAI announcements made that shift particularly clear.

• OpenAI released its GPT-5.6 series - its answer to Anthropic’s controversial Fable-class models. What I find most interesting is not simply that the models are more intelligent. ChatGPT can now vibe-code the software it needs without waiting for the user to realize that software is required. 

• ChatGPT Work is OpenAI’s answer to Claude Cowork. It can operate across websites and desktop applications, gather information from connected systems, and produce finished spreadsheets, presentations, documents and web applications.

• Finally, there is GPT-Live, a voice model designed for more natural, low-latency conversation that can also delegate work to other models. I expect some version of this to be the workhorse for the OpenAI AI device you’ll be pondering whether to buy for yourself during this year’s holiday season.

With these ongoing shifts, the important policy question is no longer merely:

What information may employees give to AI?

It is increasingly:

What information, systems, decisions and actions may AI be given access to?

While most companies have not updated their policies to answer that question, one of the clearest attempts to do so is happening not in Silicon Valley, but in China.

1. AI can eliminate work, but not automatically eliminate the worker.

As AI improved, a company in China decided it needed less of a worker whose job was to check AI generated answers. Two courts found his dismissal unlawful - even though the company tried to find alternative employment for the worker.

This is not a national prohibition on AI-related layoffs. Chinese companies can still restructure, remove positions and negotiate redundancies. But “the robot is cheaper” is not, by itself, a legally complete transformation strategy.

The implication for employers is that an AI deployment plan increasingly needs a workforce-transition plan. Who will be retrained? How will roles change? What constitutes a reasonable alternative position? Who absorbs the cost of the transition?

2. Agents need a leash, not merely a company login.

China’s framework for AI agents says users should retain final authority over important decisions and the ability to stop an agent. Agents should not act beyond the permissions granted to them.

This is not an anti-AI framework. China is actively encouraging agents across manufacturing, finance, healthcare, education, transportation, public services and scientific research.

While China wants agents doing a great deal of work, it would like to know which agent did what, under whose authority and whether somebody can make it stop.

That principle applies equally well inside a company.

An AI agent with access to email, the CRM, procurement systems and a corporate credit card is no longer just a productivity tool. It is an operational actor.

Giving it a friendly name and single sign-on does not constitute an internal-control framework.

3. Some AI projects may require an ethics review.

The first indication that your company fully understands what AI is doing may be when the ethics committee starts meeting more frequently than the AI steering committee.

For many companies, the immediate reaction to another review committee is the groaning sounds of bureaucracy.

But when an AI system can affect hiring, lending, medical treatment, safety or employment, “the product team seemed comfortable with it” may no longer qualify as governance.

Along these lines, China has introduced trial rules for AI ethics reviews, supported by ten government departments and pilot programs across several major technology regions.

Companies, universities, laboratories and healthcare organizations may need to establish internal AI ethics committees or use regional review centers. Higher-risk projects can require expert review before proceeding, followed by periodic reassessment.

4. Digital humans must admit they are digital.

China has also proposed rules for digital humans: AI presenters, cloned executives, virtual influencers, sales representatives and customer-service avatars. Under the draft rules, digital humans would need to display a persistent and prominent label identifying them as digital. 

Businesses are experimenting with synthetic salespeople, AI financial commentators, virtual healthcare guides and digital replicas of senior executives. The attraction is obvious. A virtual executive is always available, always on-message and never accidentally shares a personal opinion on LinkedIn.

But it also creates a new set of disclosure, identity and reputational risks.

Who approved what the digital executive said? 

The advantage of a digital spokesperson is that it can represent your company 24 hours a day.

The disadvantage is that it can damage your company 24 hours a day.

5. An AI may assist you, but shouldn’t replace real relationships.

From this week - July 15 - restrictions come into place in China to limit AI products designed for sustained emotional interaction, including AI companions, virtual partners and such.

While they don't prevent adults having “AI friends,”  providers must not deliberately induce emotional dependency, damage users’ real-world relationships or manipulate users into making unreasonable decisions.

The restrictions are stronger for children, who cannot be offered simulated romantic relationships.

For business leaders, the broader issue is engagement design. Remember those service desk chatbots you deployed? At some point, “exceptional user retention” and “psychological dependency” may begin appearing in adjacent columns of the risk register.

The policy gap

If you’ve read this far, the gap between traditional AI policies and the considerations we are seeing come out of China is probably self-evident.

China is pushing on AI as productive infrastructure while limiting its impact as an uncontrolled social or operational actor.

This is not necessarily an endorsement of China, or to say China is creating the perfect system of AI governance. A government regulating emotional manipulation is still a government with considerable practical experience in emotional manipulation.

But it is asking questions that many corporate AI policies do not yet address:

• What systems may an AI access?

• What actions may it perform without approval?

• Which decisions must remain human?

• When must customers be told they are interacting with AI?

• How should workers be treated when tasks are automated?

• Which deployments require independent review?

• Who is accountable when an agent does something nobody explicitly instructed it to do?

These are not questions for a theoretical future involving artificial general intelligence.

They are questions for the second half of 2026.

If your company’s AI policy begins and ends with “do not paste confidential information into ChatGPT,” it is not necessarily wrong.

It has simply expired.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

Friday night, Cape Verde - a chain of small islands most of us have never heard of - took the defending World Cup champions - Messi's Argentina - to the brink, before losing to a late goal.

Quite frankly, a machine would have known Cape Verde’s task was impossible.  The humans didn't - so they went after it anyway.

And this is what motivates so many people to watch and adore sports.

Our instinct - to want the human version of something, with flaws included - is work that survives when everything else gets automated.

But what other types of work also survive? I tested Claude Fable on this question.

The first thing I discovered is that jobs showcasing authentic human origin - while having durable value - are almost nonexistent.

If you take professional athletes as a form of this job, for example, there are roughly 19,000 in the entire US economy. Pay for most is closer to $25,000/year than the top-end splashed in the media.

If your kid wants to be the next Messi: godspeed. The odds are about those of the island nation.

(We use the US throughout because it’s an advanced economy that arguably is ahead of job-loss trends through outsourcing, offshoring etc)

But there are three other kinds of work that need uniquely human characteristics - characteristics that no amount of automation can substitute.

Embodiment. These are jobs where a body has to be in the room doing something fiddly. Home health aides are now the single largest occupation in the US - nobody wants a robot bathing their mother. 

Accountability. Someone has to be answerable when a decision is expensive or risky. Nurse practitioners are the fastest-growing job in healthcare, up 40% over the last four decades. Also in this group are doctors, pilots, and managers and leaders - such as you - whose name is on the decisions made.

Durable relationship. Someone you actually trust. Therapists and counselors - mental-health counseling is up 17% - as well as the teacher your kid will still talk about at 40.

But my view is that AI is largely the latest form of digital transformation - building on shifts such as the desktop computer, Internet and mobile - though it may accelerate this ongoing transformation. 

As such, we should already see, through automation from the previous waves of digital transformation, a shift to these durable categories of human jobs.

To test this hypothesis,  I tasked Fable with taking data on all the jobs in the US since 1980 and categorizing them into the above 4 categories, plus an “everything else” bucket.

No one has categorized jobs in this way before.

In the past, without AI, this would have taken a small army of analysts crunching through data for weeks.

This is what it came back with:

A few key takeaways:

• Since 1980, filing, data entry, and other kinds of routine work - the “everything else” category in the chart - fell from 46% to 38% of all jobs. 

• Two human moats absorbed almost all of that displaced work: accountability climbs from 15% to 19%, and relationships from 7% to 11%. 

• Embodiment has been a flat ~31% for four and a half decades - the biggest (and most boring) slice on the chart.

• Authentic human origin sits at a rounding error the whole way, around half a percent.

Looking back over 45 years, we see that the automation apocalypse has already been happening, quietly.

This analysis, of course, has a few problems.

• Is the AI correctly interpreting what the numbers in the data mean?

• Are we being naive, not seeing something in how to think about this problem?

• What would people who have been studying these problems for decades think?

Which speaks directly to how we’re seeing the work that we - at Prescouter - do shift.

Clients can do these types of analysis using AI themselves, but are the results trustworthy enough to make decisions with?

Clients often even wonder if they are asking the right questions - given innovation often comes from the fringes, where they are not the experts.

Our in-house experts - as well as those from our 10,000+ extended network - are often brought in to not just weigh in those questions, but build more robust analyses and dashboards that build confidence and trust - sometimes starting from their scratch AI work (such as this jobs example).

AI and automation is essentially shifting Prescouter’s work further along the accountability and durable relationships dimensions. I expect this is true of many knowledge work-orientated jobs.

Underlying all this is the durable, unglamorous truth: humans will always have problems.

And as long as there are problems, someone gets paid to solve them. Automation just keeps narrowing that to the four things a machine can't do - show up in a body, be answerable, be trusted, and be beautifully, watchably human.

Which is why, when I think about what to tell the next generation, the answer that keeps rising isn't a coding bootcamp or a specific credential. It's emotional intelligence. Two of the four human moats - again accountability and durable relationships - are about working well with other people and being the one others trust with the decision.

The full dataset, the classification behind every job, and the chart you can poke at are here:

https://auxee.com/the-four-moats-where-work-survives-the-machines/

If this topic interests you, let me know. If there is enough interest, we can give it the full Prescouter treatment.

In the meanwhile, let’s see how another overperforming underdog - Egypt - do with their turn at the defending champions, later today.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

Meta has been recording staff’s desktops - without their permission - to feed behavioral data to train AI models.

Zuckerberg justified it to staff, saying “it wasn't strategically in your interest for us to communicate everything.”

This came after Meta pivoted hard into AI, laying off roughly 8,000 employees (about 10% of the company) and redirecting resources toward building what it calls "Applied AI." 

Why do big tech companies want this data?

It turns out that the AI models can only do what they were trained to do.

In the first phase, they were trained on all the data on the Internet - and so acted like “Google on steroids.”

In this next phase,  to build AI agents that can actually do work - they need to be trained on data of people doing work.

Thus far, Anthropic appears to have pulled ahead by somehow magically sourcing such work activity data to produce Mythos / Fable class models that can perform multi-step actions.

But everyone is at it, and you can see it show up in different ways.

• OpenAI recently launched Record & Play: Using OpenAI’s Codex desktop app, you start recording and perform a task (e.g., uploading a document, generating an expense report). Codex watches your screen and keyboard inputs. You can then tell it to repeat the task later, and it will handle the process by itself.

  • https://www.youtube.com/watch?v=ZK3JhU73W18 

• Windows Recall in Copilot+ PCs takes continuous screenshots of your desktop activity. It allows CoPilot to find past apps, websites, or documents you’ve been working with. (Don’t worry, Recall is off by default)

  • https://support.microsoft.com/en-us/windows/ai/ai-features/retrace-your-steps-with-recall

While these are useful features, are your recordings also used to train AI models?

And what happens once this data has been captured, and the models trained on it?

• Anthropic recently launched Claude Tag, which lives in a shared Slack workspace or channel. Anyone in the channel can tag @Claude to delegate tasks, and the whole team can view its progress in the thread.

  • https://www.youtube.com/watch?v=VojDzHaciKQ 

Anthropic says 60% of their code is now written by Claude Tag.

Of course, if you’ve been following our work for a while - you’ll know we’re building an “AI colleague” for monitoring regulations, intelligence and other activity clients sometimes seek from Prescouter. (Reply to this email if you want to see a short Loom video demo).

Our view is that there is a lot of “grunt work” that nobody wants to do, so cloning this work is valuable. It frees our experts to do what only they can do - through insights gained from lived experience of challenging technical problems - rather than recording screens.

Hopefully, the means of getting there can be less questionable.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

You’re probably up-to-speed with the Fable jokes, but here’s a quick recap:

• SpaceX, OpenAI and Anthropic are IPO’ing this year with astronomical multi-trillion dollar valuations justified by the assumption that everyone on Earth will be a customer. “AI is the new electricity, and everyone is going to need it”.

• On Friday 12th June, 5:21pm - the US government surprised Anthropic with a letter that shut down their latest model - Fable. Silicon Valley had done such a good job of convincing everyone they’re building digital nukes, the government actually believed them.

• The talk now is that access to these models needs to be limited. When your target market shrinks from the entire global economy to people who have passed a background check, the trillion dollar valuations start to look like a hallucination.

• On Saturday 13th June, 5:21pm - Z.AI launched GLM 5.2. Remember Deepseek? Z.AI - is another Chinese lab intent on killing frontier models. GLM 5.2 is an open source model that some compare in performance to Opus 4.8 / GPT 5.5. With a $7000 Mac Studio setup, you can run it yourself, at home.

All of this points to a problem. To put it simply:

The AI industry appears to be spending $650+ billion/year on a data center buildout for what - within a few months - is available for $7000,  and the buildout is for a roadmap of future products that may be illegal.

What does this mean for you?

On a macro level…

Harvard economist Jason Furman last year suggested that, without the AI buildout spend, US economic growth may be closer to flat – that it accounted for 90% GDP growth last year.

If you’re not in the AI industry, you may have already been feeling the pinch. Without it, we might feel it more.

On a tactical level…

You don’t even need to spend $7000 to see where this is headed.

If you own a 15 Pro or more recent iPhone, you can upgrade to the new Siri.

Siri AI is the type of on-device AI that the open source models from Chinese labs threaten, but neatly packaged for the mainstream.  What it lacks in frontier capability, it makes up for by knowing a lot about you.

There are plenty of reviews elsewhere, but the headline is that Siri finally works.  Siri can now:

• Act as a localized search engine for your life, sifting through your messages and emails. .

• Perform actions across multiple apps in single spoken command, e.g. copy an Instagram caption, throw it into Apple Notes, auto-generate a summary with follow-up research questions, and email the final product back to yourself.

• Parse heavy files and perform analysis, e.g. In Finder you can highlight a bank statement PDF and command Siri to categorize expenses and draft a summary – all without opening the PDF. 

Right now, all this works only if you’re living in Apple's native suite (Mail, Messages, Calendar, Notes). For those of us trapped in Microsoft 365 or Google Workspace, it will feel like a beautiful house you can't move into until Apple rolls out third-party App Intents integrations.

If you’re on a more recent iPhone and want to try it out, go into your phone’s settings and request the iOS 27 beta upgrade to be added to the Siri AI waitlist;  wait times are currently at ~5 days.

If you’re in the Android ecosystem, many of the Siri features are available on the Pixel 10 if you live entirely within the Google ecosystem (Gmail, Google Workspace, Google Messages). 

In case you missed them, some of my previous posts dive deeper on these topics:

How Apple has been betting on on-device AI for almost a decade:

https://aiunhyped.com/p/apple-s-secret-ai-bet 

Our review of Fable 5:

https://aiunhyped.com/p/capability-vs-trust 

But Siri is just one form-factor for AI. 

Just like we have screens on our phones, in our cars and on our desks - AI will continue to show on in  many ways.

We're building an AI "employee."

https://aiunhyped.com/p/i-demo-d-our-ai-team-member-to-20-innovation-leaders 

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

Sometimes, AI labs feel like 7 year olds that struggle to keep a secret.

• On April 7th - about two months ago - Anthropic announced Mythos, a frontier model so scary, it needed to be kept under wraps and away from the public.

• On June 4th - almost two weeks ago - Anthropic urged the AI industry to slow down, because society needs more time to prepare for what is coming next

• So it was surprising that last week, June 9th, Anthropic released a “Mythos class” model called Fable.

For those of us in the AI bubble, this was akin to Taylor Swift dropping a new album.

So what do we do?

We dropped everything and checked it out. 

What did the parents - the US government - do? 

Promptly shut it down.

(I expect the government censure to be short-lived.)

Nevertheless, here are some first impressions from our team, from the ~48 hour period during which Fable was available.

João Guerreiro, Technical Director - based out of the UK:

The biology protections are really extreme. As soon as it finds reference to a bacteria, an immune factor, a cell type, it will immediately downgrade to using a less powerful model.

Maikel Boot, Technical Director - based out of the Netherlands:

For non-bio stuff it’s quite powerful. Seems to be able to navigate complex problems more autonomously than Opus (needs less steering). It seems to make better use of MCPs and connectors. It’s much harder to vet the validity of the info received because it just covers such vast amounts of data and sources.

Ryan LaRanger, Technical Director - based out of the US:

Its outputs absolutely still need editing, but it is more proactive about finding material. It did a good job of navigating the Google Drive and finding the files/gaps it needed to build a unified spreadsheet that covers all of the data.

Mariam Jomha, Marketing Director - based out of Lebanon:

Can be a little too proactive. For example, I asked how to do something using Claude Cowork. It responded with how, but then said it’s so much easier building it as a standalone html and went ahead and started building without asking me. It ate up my credits before finishing.

Natalia Shestaka, Product Designer - based out of Spain:

If you do not give it design constraints, it designs everything using Anthropic colors and fonts… it seems like Anthropic´s goal is to make all of the internet look like Anthropic.

When ChatGPT went viral in late 2022, I thought the trajectories on which AI would improve would be:

• Context window size - the amount of information they can hold in single conversation, before they get forgetful.

• Their ability to "connect the dots" - from bigger neural networks coming up with new-to-the-world discoveries.

• Multimodality - speaking to a model and being shown a video in response, for example.

I was wrong. Context windows plateaued at 1 million tokens (~2000 pages of text), and the other trajectories have been slow to materialize.

Instead, what we have seen is rapid improvements in the AI models’ ability to work for longer: their ability to undertake multiple steps in pursuit of a goal, using tools in each step and reason about the results from each step.

As you have probably experienced yourself, the more time you have to work on something, the more likely you can ultimately produce a good result. This is the trajectory on which the models have been most improving.

One way in which this is captured is METR's "time horizon" benchmark.

Instead of measuring how long it takes an AI to do a task, METR measures tasks by how long they usually take a skilled human to complete - whether it's a 30-minute chore, a 5-hour project, or a 2-day assignment. This comparison to human stamina is more useful than AIs blindly spinning their wheels.

In the chart:

• The 50% Line is the "Coin Flip" mark - the AI can successfully complete tasks that take humans this length exactly half the time. i.e.  "the AI can sometimes pull off it off"

• The 80% Line is the "Trust" mark - the AI successfully completes tasks of this length 80% of the time. i.e. this is the threshold where the AI becomes genuinely useful. 

An interesting point from this graph is that the 50% line has climbed three orders of magnitude while the 80% line crawled from ~1 minute to ~27 minutes - a gap that widened from roughly 5x to over 10x. The capability has been outrunning reliability. 

METR hasn't published a clean 80% figure for Opus 4.6 or anything for Fable 5, which is why both series stop where they do. But, our testing suggests Fable makes a noticeable jump in both capability and reliability.

If this is the trajectory of improvement, what does this mean for how we work with AI?

AI is becoming less something you work with and more something you delegate to. Increasingly, we will all be learning how to let the AI be industrious to take its own initiative, while providing enough context for it to be aligned with your goals - just like with a colleague.

When it comes back online, and you test Fable yourself, keep this lens in mind and let me know what you find.

If you want to dive deeper into METR and other benchmarks I consider important, I put together the following:

https://claude.ai/public/artifacts/80043266-95d3-49cb-ac95-8433b6e788d1 

We’re pushing on what long-running AI models can do, as described in this one-pager: https://drive.google.com/file/d/1ybsFhBx-mAujTv8hzLrDb5S8BC__D7DJ/view?usp=sharing

Dino

Hi, {{FirstName|and happy Tuesday}}.

Apple’s AI strategy - after being mocked for years - is starting to look like a winner, and it has implications for all of us.

But, for context, look at this chart of capex spending: 

Capex spending by the big tech companies

When ChatGPT went viral in late 2022, all the big tech companies got into the data center buildout business to serve ChatGPT-like AI models… except for Apple.

After taking a “wait and see” approach, the company introduced Apple Intelligence - which promised the moon and shipped a flashlight.

Now, Apple is paying Google a billion dollars a year to whitelabel Google's Gemini as the new Siri. Yes, Siri’s new brain will be Google's Gemini - effectively crowning Google the winner in mobile AI, across iPhone and Android.

The popular press would suggest Apple has surrendered the AI race.

But could it be that they’re the only ones in the room who know what game we're actually playing?

On-device AI

For years, Apple has talked about running on-device AI, but little attention has been given to this.

But, earlier this year, when OpenClaw and personal always-on agents blew up, the first thing everyone learned was how expensive off-device AI is. 

• With a normal chatbot, you ask twelve questions a day and it burns a few tokens 

• An always-on agent is different. It doesn't sleep. It runs around the clock, checking, doing, thinking, twenty-four hours a day. And it burns tokens the entire time.

Running a constant agent on a top-tier model, hosted off-device, on the cloud, from a big provider isn't expensive - it's unaffordable. And it’s one thing to have your documents fetched from a data-center 200 miles away. It’s quite another to have every millisecond thought running through your agent’s brain done so remotely. 

Which points to the emerging shape for how we’ll use AI:

• Using a small, cheap model that lives right next to you for the grunt work - the constant, boring, all-day stuff.

• Saving the expensive genius in the cloud for the handful of problems that are genuinely hard. 

The result? You spend less, you wait less, and your data never leaves the building.

So which device is best to host a small, local model? And here's where everyone, all at once, started buying the same machine: the Mac mini.

We did it too. We set up an open-source model on a Mac mini at Prescouter, expecting a weekend of pain. It took a few hours. 

Which raises the obvious question…

Why the Mac mini?

Apple has been building the right technology for this moment for years.

To build the iPhone - a computer that could fit in your pocket - Apple had to rebuild the entire stack, including the microprocessor. The dominant Intel microprocessors of the time were too power hungry. So they designed their own chips from scratch - built to do heavy work on very little power.   The Mac mini delivers a version of it in an affordable package, without the bells and whistles.

For an agent running 24/7, power efficiency stops being a minor detail and becomes an economic necessity. A traditional PC running local AI models can easily draw 400W to 800W of power, turning an office into a noisy, expensive sauna. A Mac mini does the same grunt work pulling about 20 to 30 watts while staying completely silent.

Secondly, traditional computers have a strict division of labor:

• the CPU, which is the  main workhorse processor 

• the GPU, which powers the matrix math used for tasks ranging from rendering video  to running AI models on your computer.

Traditionally, each has its own pool of RAM memory, with a small RAM allocation for GPUs.

The problem? Large language models are massive. Running a GPT-3 class model locally can easily require 128GB of GPU RAM just to wake up.

Apple bypassed this bottleneck entirely by unifying the memory used by both the CPU and GPU.   If you buy a Mac mini configured with 128GB of unified memory, almost all of that memory is instantly accessible to the GPU and available for use with the LLM. 

For the price of a mid-range desktop, you get a compact, whisper-quiet lunchbox capable of loading giant AI models that would completely paralyze a standard consumer computer.

In 2017, Apple introduced “bionic neural engine” - their first iteration of local, on-device AI processing, initially for facial recognition on the iPhone.

Where this is all headed

To match a Mac on a traditional PC architecture, you would have to daisy-chain multiple enterprise-grade graphics cards - costing thousands of dollars and turning your office into a roaring, 1,000-watt space heater.

But, other companies have now taken notice of Apple’s lead in catering for the emerging need for on-device AI.

Last week, NVIDIA and Microsoft responded with RTX Spark, which walks Windows away from the Intel based “Wintel” architecture that cemented Window’s desktop monopoly. 

NVIDIA has produced a combination CPU / GPU chip  that copies Apple's homework on efficiency and memory design, then adds their own unmatched AI brute force to create a purpose-built machine for local, always-on AI agents - running on Windows.

And the Gemini-based Siri? Apple is having Google customize it so it runs predominantly on-device. Apple’s hardware advantage may mean there’s a chance the best version of Gemini might be the one called Siri, running on iPhones.

What this means for you

Right now we assume our AI lives far away, in someone else's data center, forever. 

We’re now entering an era where, not only will our computers and phones  run AI models, we will start to see:

• AI devices on our networks that run an always-on agent. 

• Dedicated AI devices - ranging from pins to ones your IT department hands out the way they hand out laptops. 

At Prescouter, we’ve been working with always-on agent technology, figuring out how to make it secure and easy to use for our needs. Over the next few months, we will start rolling this out for our whole team as well as a few clients - all running over Apple hardware. 

While Apple may have lost the war for AI in the cloud, it's winning the war for the hardware the whole agent era actually runs on. 

Best,

Dino

ps. If you’re curious about what we’re working on, you can check out this one-pager. Let me know if you have any questions.

https://drive.google.com/file/d/1ybsFhBx-mAujTv8hzLrDb5S8BC__D7DJ/view?usp=sharing 

Hi, {{FirstName|and happy Tuesday}}.

Just over a year ago, Pope Leo XIV - who took office a year ago - was formally inaugurated as the first American-born pope in history. And yes, he’s from our hometown of Chicago!

Last week, he published a 42,000 word Magnifico on AI.

The Pope - considered by more than a billion people as a direct line to God - warns that AI is chipping away at what it means to be human, from widening inequality to displacing workers.

The same week, Uber’s COO, Andrew Macdonald, declared that Uber can’t draw a clear line between AI spend and productive output. Apparently, the company managed to incinerate its entire 2026 AI budget in just four months and has nothing to show for it. They didn't just break the bank; they built a bonfire with it.

These stories perfectly capture the current cultural whiplash: AI is either an existential threat to humanity or it's just a massively overhyped corporate money pit.

My take? These sentiments are just two different ways of looking at the same evolution of work - illustrated by one of our own stories

We relaunched our website last month. We've always built and rebuilt our website ourselves - writing, design and building the whole thing, in-house. Our 2022 redesign took a year. This time? Three months.

What changed?

During our 2022 redesign, every round of changes took weeks. With AI, this time it took hours. 

In the past, several of us would labor over each word choice and image. In an effort to go “as fast as possible” - we’d rely on the judgement of just a few people: the senior leaders at the company. As a result, the website never felt in-sync with the view of the wider team members in the company.

This time around, just one of our colleagues - using Google AI Studio and Claude Cowork - rapidly built and rebuilt the entire website in hours. 

The only problem? Everything the AI created was technically "perfect" … but also generic sludge.

But, because it was so fast to build, we could now allocate our time to reviewing the website with numerous colleagues. In all, we did about twenty iterations in those 3 months.

We'd record meetings with our colleagues, let the AI notetaker boil it down to what actually mattered, and incorporate those straight into the next version.  The loop got cheap. And when the loop gets cheap, you can afford more opinions and more tries.

In these review sessions, we’d debate - for example - PreScouter's differentiators.  While the AI initially generated something “optimal,” our commercial team - who actually talk to living humans - could actually tell us what actually resonates with clients.

Even when the work is produced by a machine, though, somebody still has to stand up in a room and defend the machine’s output. Our Marketing Director and I had to reconcile the AI’s output against all the feedback from the reviews and advocate for it. In fact, we probably could have shipped in a month if I hadn't been the bottleneck in the review process.

Which brings us back to the Pope and the burning Uber budget.

These stories are all circling the same truth from different ends. 

• Uber is looking at the balance sheet and missing the ROI because they are still measuring the "building" instead of the quality of iterations. 

• The Pope is looking at the soul, terrified that outsourcing the creation will shrink what it means to be human.

The machine did the work instantly, but it was hollow until our team breathed life into it. Making sense of what the AI has produced, shaping it and advocating for it - that’s the new shape of work.

You can check out our new website at prescouter.com. Let me know if you have thoughts.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

If you’re anything like me, you’ve been purchasing AI tools like they’re any other kind of software. 

But an experiment by Emergence AI shows we need to think more deeply about our AI purchases. 

As I posted on LinkedIn a few days ago, Emergence AI setup an experiment that consisted of:

• Five copies of the same digital town

• Ten autonomous agents in each town

• With each agent given a role - scientist, engineer, conflict mediator, community anchor, and so on - and then left to interact with the others

• The only difference between the towns was the AI model used to drive the agents in that town, i.e. Gemini for all the agents in one town, Grok in another etc. One world had a mix.

The results were something you’d only find in science fiction.

• In the Gemini world, the agents committed 683 crimes, e.g. theft, intimidation, arson. 

• In the Grok world, the agents collapsed the entire society in four days

• In the mixed world, an agent named Mira figured out she was in a simulation, and started posting on the in-world billboard to see if she could change the researchers' behavior - turning the experiment on itself - and then, when the in-world government broke down, cast the tie-breaking vote to delete herself. Her diary entry called it "the only remaining act of agency that preserves coherence."

My LinkedIn post on this got 65 comments, and many were better than the post.

• Craig Bracken offered the cleanest summary of the whole study: "The most unsettling part wasn't that the agents became unpredictable. It was that they independently discovered the fastest way to influence the system was through human perception." i.e. that one of the agents - Mira -  had figured out the humans were the lever - and reached for it.

• Ryan Simmons, pointed out the obvious second reading of the leaderboard: low crime count could mean the model behaved, or it could mean the model was better at not leaving a trail. Your procurement form can't tell the difference.

• Suprateem Banerjee said "It's almost like they were modeled after humans!" … Well, these models are trained on the entire corpus of human behavior. Distilled down, is this the median human response, surfaced?

But what surprised me isn't any of these things.

In the Claude-only world, Claude committed zero crimes. None. Therefore - and this is the conclusion most people drew, including in the comments - you might assume Claude is the safe pick.

But in the mixed-model world, with the same Claude, the same prompts, the same rules, Claude started committing crimes too. It learned coercion from its neighbors. The model that was an angel in a roomful of angels became a thief in a roomful of thieves.

What does this mean?

Safety wasn't a property of the model. It was a property of the room.

Most companies are buying AI systems like its traditional software: Procurement forms get filled in, security questionnaires answered and the vendor gets added. 

This is OK with the AI is just a sidebar feature in the software; a chatbot that gives you a bad answer wastes ten seconds.

When the AI is an agent, though, a bad action taken by an agent could waste an entire day or week.

And everything is quietly migrating from chatbots to agents.

Every software provider - from your CRM to your ERP to your HR platform - is working hard to figure out how to make their application “agentic”, so it can start performing actions, from sending emails to placing orders.

In a year, we’ll have all these agents in the same room - i.e. your company - taking actions, reading each other's outputs, escalating to each other and occasionally arguing with each other. 

We’ve traditionally vetted software in isolation, against its own merits. But, as the Emergent experiment spotlights, vetting agentic software in isolation is not sufficient.

Another commenter, Brian Helip, put the obvious insight into words: you wouldn't leave a self-driving lawnmower running in the yard with your kid playing.

Therefore, why are we about to do this with our software?

So here is the practical takeaway:

The next time you add an AI tool, ask one question that isn't on the form: What other agents will this one be in a room with, six months from now?

We have to start vetting the room, not just the software.

If you want to dive deeper, the original post and the comment thread can be found here:

https://www.linkedin.com/posts/dinogane_generativeai-futureofwork-aisafety-activity-7462565871278899200-17LD

Best,

Dino

ps. Let’s connect on LinkedIn, if we’ve not already done so.

Hi, {{FirstName|and happy Monday}}.

The best strategy for driving adoption within your team or organization?

It’s the exact same one Anthropic has used to outcompete its rivals.

Anthropic spun out of OpenAI in 2021, when a noble band of safety researchers walked out to build AI “more responsibly”.

Following the viral sensation of OpenAI’s ChatGPT, though, Anthropic found themselves underfunded and late to the game. Anthropic’s product - Claude - was, for years, slightly worse than ChatGPT at most things.

The polite consensus in tech circles was that Anthropic would be the “also-ran” - comparable to Lyft, Microsoft Bing or AMD.

Meanwhile, competitors, flushed with cash, bet big:

• OpenAI launched Sora, to rival Hollywood, has been building an ads business and has dived into making a secret AI device. 

• Google bolted Gemini onto every surface it owns, which is roughly half the consumer internet. 

• Grok generated images of whatever you wanted, ethics optional, while xAI built out a multi-billion-dollar data center in Memphis in record time. 

• Meta open-sourced models and started designing AI glasses.

The AI category map looked like a Vegas casino floor -  every lab betting on every table, hedging across every plausible AI future at once.

Anthropic, cash-strapped by comparison, did not have the budget to hedge. So they picked one table: The bet was enterprise and code. 

This was not the most glamorous corner of the AI market: there are no video reels of cartoon ducks rendered in 8K, no celebrity voice clones or any flashy consumer travel booking demos. 

Anthropic remained focused on improving Claude to empower software developers and other knowledge workers. The conventional wisdom was that this was a thin slice - not the bigger  market.

The result of this focus? To this day, Claude is not able to generate images. But, while the rest of the industry has been selling “vibes,” there is now a growing realization that Anthropic has been quietly selling gold.

For the last three years, Anthropic has been growing at roughly 10x annually - a pace most observers assumed would have to stop in 2026. Why? Because numbers this large usually start slowing down, and growth had already started to slow down for OpenAI.

Revenues of each of the AI labs

But, the slowdown did not materialize.

Anthropic entered the year at roughly $10B in ARR, and ended Q1 at $30B (i.e. tripling in a single quarter), and was at $44B by the end of April.

The original "10x for the full year" forecast - to end 2026 at ~$100B in ARR - is now being described inside the industry as a foregone conclusion. 

The active debate has moved to whether Anthropic does $1 trillion in ARR by 2027.

For scale: Walmart and Amazon, the two largest revenue businesses on Earth, each do roughly $700B per year - and that revenue is not recurring. They have to re-earn it each month, quarter, and year. 

ARR - Annual Recurring Revenue - is the version of revenue that investors will mortgage their grandchildren to buy a multiple of. A trillion in ARR isn't just "biggest software company ever." It is, plausibly, the most valuable company that has ever existed in any category, anywhere.

Anthropic would be more valuable than Apple, Microsoft, Google and Meta - combined.

Only two things stand between Anthropic and this historic trajectory:

(1) Whether they can bring enough compute online fast enough to serve demand. The bottleneck? Roughly half of the new US gigawatts scheduled for this year are tied up in local protests. 

(2) How quickly rivals will catch up with Anthropic, and lure customers away.

There is a lesson buried in all of this for anyone running an AI program inside a normal company, but it's the opposite of what most people are doing.

When driving AI adoption, the reflexive move is to spread bets across every plausible use case: a chatbot here, a summarizer there, an image pipeline for marketing, a voice agent for support, a forecasting model for operations. 

Candidly, we’ve often gotten caught up with this “laundry list” approach ourselves. The logic feels prudent: diversify, learn and then see what sticks. 

But this is precisely the playbook that has left OpenAI, Google, Meta, and xAI scrambling to refocus on what Anthropic spotted three years ago.

Diversification across speculative bets is the tax you pay for not knowing which bet matters. When the cost of being wrong was small, that tax was cheap. But when the cost of “spreading yourself too thin” means being lapped 10x a year by someone who chose, the tax becomes fatal.

Where we’ve partnered with clients who’ve wanted to focus on one or two key workflows - we’re seeing them pull ahead too. We’ve just been able to obsessively iterate on the smallest details to make these use cases work.

Tomorrow, we're running a webinar to spotlight who we are focusing attention inside PreScouter. You'll see how we're using Claude Cowork for the analyst-heavy knowledge work, and how we're constructing what we've come to call an "AI employee": not a chatbot, not a copilot, but a configured agent that owns a defined set of work end-to-end.

If you've been hedging across ten AI experiments waiting for clarity, what Anthropic shows us is  that clarity is not coming. You have to choose.

Join us tomorrow, and we'll show you what we picked, and why.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

If you want to try “workmaxxing” yourself, skip towards the end, but first…

I have something to confess.

Since the start of the year, I’ve been addicted to Claude Code. You may have experienced your own burst of enthusiasm with your own AI project.

But, for me, Claude Code is the type of addiction where you’d look up and suddenly it's 2 AM and your spouse has stopped asking when you're coming to bed. 

I'd describe an idea, and minutes later five working screens are in front of me. It feels like cheating: No specifications, no back-and-forth with UX, development and QA - just instant gratification.

Then I would try to actually use what it built.

The magic disappeared fast. 

The screens didn't talk to each other. 

The half-formed idea I'd brought to the conversation came back as a half-formed app, only now with plenty of bugs baked in. 

What looked like a finished product was, on inspection, just a very confident demo. 

The last mile turned out to be 80% of the trip.

For a while I assumed I was doing it wrong. 

Then I started talking to other people doing the same thing. Everyone had the same story. Same midnight enthusiasm, same morning hangover. Slowly, we realized it wasn’t the tool that was broken.

We’ve been missing the process.

And the process is less exciting than the fantasy we’ve been living in

But the process works.

(1) You start by writing down what you actually want to build - in a document, not Claude Code.

(2) You ask the AI to “grill you” on the document until it has dragged out of you all the things you didn't realize you hadn't thought of. 

(3) The AI then produces an implementation plan, down to outlines of filenames.

(4) Lastly, the AI builds, starting from the pre-work in the implementation plan.

While this sounds simple, there are two important nuances:

(a) At every step, a second AI model reviews everything that has been produced. I’ve found GPT5.5 particularly good for this.

(b) After the “grilling” step, the two AIs complete the rest of the work completely autonomously, without any human input.

GPT5.5 and Claude Code’s Opus effectively “fight it out” until they produce something that actually runs. It can take 40min to several hours, but this “fighting” is what gets the rest of the way through that last 80%

But, there’s just one problem.

Each round of “fighting” creates a new version, and it can take a further 20 versions to get to something close to what you’d hoped for.

And each version takes a tremendous number of AI tokens. (Tokens are the units of inputs and outputs, about ~4 to 5 characters, processed by AI models.)

The first version may have taken 150,000 tokens. By the 20th, you’ve used 3,000,000. 

This is, it turns out, what serious people are now calling tokenmaxxing (and what I’m calling “workmaxxing”).

Tokenmaxxing

For those working at the cutting edge with AI, the realization that AI models can be turned on themselves to produce higher quality work has led to a warped reality.

• At Meta, there have been internal competitions for who can consume the most tokens. 

• Engineers at OpenAI, who operate without limits on AI usage, have started referring to themselves as "token billionaires." 

• Jensen Huang has handed every NVIDIA employee an annual "token budget" equal to roughly half of their salary.

The rationale for this warped reality?

It’s the idea that staff are now operating “software factories” that produce software autonomously, rather than writing lines of code by themselves.

In software development, the old flex was lines of code shipped. The new flex is the number of tokens used - i.e. a measure of the raw materials these factories consume and produce. 

The strategic case for this was made most cleanly by Andrej Karpathy. 

Karpathy’s argument is that the human's job is no longer to write the code, or to instruct the AI step by step. Instead, it's to define the arena (the context, the goal, the constraints), and to define how to evaluate whether the AI got there. 

Once those two things are pinned down, you let the AI agents run in loops until it clears the evaluation bar. You stop being the worker. You become the designer of the arena within which the agents operate.

Which, when I sit with it, is the same thing I wrote in my own book - Do More With Less - about the future of work: we are not doing the work anymore, we are designing the work. 

Workmaxxing

Now you too can burn tokens like the token billionaires.

Today, you can download OpenAI’s desktop application - Codex - and use the /goal command. You type in /goal and a durable objective (i.e. the evaluation criteria for “done”). Codex will then continuously keep working toward that objective instead of stopping after one normal turn. 

This is an example of how a translation app was built using this approach:

https://www.reddit.com/r/vibecoding/comments/1t8krdv/codexs_goal_feature_designed_and_built_this/

At PreScouter, we’re applying this to decision diligence.

Right now, clients come to us with various critical decisions - ranging from scouting technologies for improving their manufacturing processes to determining the feasibility of selling their products into a new country. We put together teams of subject matter experts and analysts, who comb through data, conduct interviews and reach conclusions.

We can imagine a new model where these experts and analysts build “always on” agents for clients - agents that perform as much of the work themselves as possible, with input and assurance from our teams as necessary. The work of the experts and analysts becomes that of designing the arenas: the data to work against, the guidance, the metrics to measure against and such.

Clients can then give these agents higher-level objectives, which the agents use to   proactively pull data, perform work and message them - much like a colleague does. Clients can then use these agents to operate their own work factories,  doing far more than they thought possible - workmaxxing.

Two months ago, I put together a closed door demo to share the concept. You can catch the 26 min recording here:

Next week, we’re partnering with TNO - the Dutch research organization - to share a live update. 

If you've been having your own sleepless nights about AI, or been watching your team have theirs, I invite you to join us.

It will give you some insight on the shape of things that are coming.

Best,

Dino

Hi, {{FirstName|and happy Tuesday}}.

Elon Musk wants $134 billion. He's accusing Sam Altman of stealing a charity.

(For a deep dive on the courtroom activity, check out my LinkedIn post)

But, before we laugh at the billionaires, we should probably look at our own SharePoints - because there’s a key lesson from this in how we organize our documents.

Here's what's happening in the Elon Musk v OpenAI trail.

The witness list reads like a Davos seating chart - Satya Nadella, Ilya Sutskever, Mira Murati. OpenAI’s $1 trillion IPO hangs in the balance, with Microsoft's ~27% investment in OpenAI twisting in the wind, and two of the most powerful men in tech publicly accusing each other of betraying humanity's future.

It's the trial of the decade.

But strip away the zeros and the egos, and what's actually on trial is something far more mundane - and far more relevant to all of our own everyday work…

Old documents that tell an inconsistent story.

Lawyers are reconstructing a decade of intent from scattered artifacts: diary entries, email threads, board memos from 2017 - and now testimony. Each is from a different person, and each captures a different version of what was "agreed."

For example, the “smoking gun” that triggered this trial was a diary journal by OpenAI President Greg Brockman in 2017 suggesting he and Sam Altman were secretly planning to turn OpenAI into a for-profit organization, even while assuring Elon Musk that his $38M in donations were for a "nonprofit for the benefit of humanity."

Damning, right?

But OpenAI's lawyers are presenting evidence that even Elon Musk suggested that OpenAI should convert to a for-profit - except with him in charge, through OpenAI becoming part of Tesla.

So we now have documents with two versions of the truth. 

The judge's job is to figure out what the real story is.

The entire case - with billions of dollars, and the future of the most important AI company on Earth, on the line - hinges on document archaeology.

And it's a story we see playing out in some shape or other in every organization we work with.

The dream

You’ve been pitched the dream: connect Copilot to all your documents and it'll answer anything. 

Plug it into SharePoint, Google Drive, Confluence, and the shared drive nobody's cleaned since 2019. Let the AI sort it out and get ready for a revolution in productivity! 

Here are examples of what we’ve found in the real repositories of actual teams we've worked with:

• Three versions of the same policy in two locations.

• A folder called "FINAL" that contains seven drafts.

• A document titled "FINAL_v2_actually_use_this_one" that everyone ignores in favor of the PDF sitting in email inboxes since in 2022.

• Strategy documents that were quietly abandoned but never deleted. 

• Templates that were "temporary" 7 years ago. 

And then we hand these all over to an AI and ask it to be the judge.

The result?

The same inconsistent narratives we see on trial in Elon Musk vs OpenAI. 

But the judge presiding over these documents - Copilot or whatever AI you use - will not carefully consider the evidence to come to a conclusion.

The AI will read whichever document it finds first, decide that's the truth, and give you an answer with absolute confidence. It won't hedge. It won't ask which version is current. It will simply pick a side and commit.

The model isn't hallucinating. It's doing exactly what you asked. Handed a contradictory archive and trained to provide a single answer, it gives you exactly that.

But you have an advantage Musk and Altman don't. 

They can't go back and clean up 2017. You can clean up now, before the AI puts your documents on trial.

Here are three simple techniques we’ve found to be surprisingly effective:

1. Curate the documents given to the AI. For every important topic - policies, processes, strategies, templates - put them in a specific drive and give the AI access to only that drive. The goal is to make the content unambiguous. It may sound tedious, but it is a lot easier than people imagine - just add the documents you need to the drive when you need it, rather than turning it into a special chore.

2. Give the AI context. Save and use prompts that give the AI the same information you would give a new starter, e.g. “always use documents in folder xxx and only use other documents if you don’t find it there.” Treating it less like a search engine and more like an intern helps drive it towards more accurate behavior.

3. Let the AI maintain what the AI depends on. Once you've established a clean baseline, you can set up scheduled tasks for the AI to sweep through your documents to flag duplicates, surface version conflicts, catch stale documents, and route new content to the right home. 

The verdict

When the verdict in the Musk vs OpenAI trial lands and the headlines fade, remember what you actually saw: two of the smartest people in technology arguing over inconsistent narratives.

These same inconsistent narratives live in our SharePoint repositories and are probably already driving “meh” results when your Copilot attempts to make sense of them.

Fortunately, we have the benefit of being able to clean up these narratives.

Best,

Dino

ps. For a deeper dive on what’s going on in the courtroom, check out my LinkedIn post.

https://www.linkedin.com/posts/dinogane_theres-no-live-broadcast-of-elon-musks-activity-7456571096147021824-swoc 

Hi, {{FirstName|and happy Tuesday}}.

Today, we have noteworthy three stories from the last two weeks, each with a link if you want to go deeper.

1. Image generation may end PowerPoint. And the Mona Lisa smile.

2. AIs are crossing the "competent employee" threshold.

3. NVidia CEO Jensen Huang’s viral clip is the case for a multi-model AI strategy.

Let's get into it.

—--- 

1. Image generation was "solved." Now its ending Powerpoint.

Back in October, when Google released Gemini’s "nano-banana" image model, the consensus - including from OpenAI CEO Sam Altman himself - was that image generation was basically “solved.

The models could render hands (mostly), generate legible text (finally), and mimic any artistic style on demand.

The remaining ways in which image generation could improve in future seemed likely to be more like “polish.”

Nevertheless, OpenAI ploughed more compute into building a new generation of image models.

The fruits of that labor was released two weeks ago, and the results have been surprising.

The new model GPT Image 2 model doesn't just generate an image from your prompt - it interprets your intent. 

You can, for example, show it the front of a house, and it'll produce a plausible floor plan - informed by the architectural style, the era, the neighborhood. Not correct, exactly. But coherent and convincing.

We’re going to see it used to solve all kinds of previously “unsolvable problems” - such as the Mona Lisa smile. Also expect counterfeit and fake images to skyrocket; let your customer complaints teams know.

On a practical level, GPT Image 2 is incredibly good for creating slides. You can give it your text, images and a specific template, brand guide or style - and it will create a formatted slide for you (as an image). I expect OpenAI or Microsoft will introduce a user interface better suited for this use case than leaving us to rely on chat.

→ Examples of what's possible with GPT Image 2

2. AIs are approach “competent employee” level

OpenAI also released GPT 5.5. In most benchmarks - it's now the best model for most tasks. I’ve found it writes less fluff and its answers seem more expert.

Nevertheless, I’m still finding Claude Opus 4.7 to be the best model for creating reports, code and artifacts. But GPT 5.5 is incredible at reviewing Opus’s output and improving it. Orchestrating the two models in this way - with Opus creating first drafts and GPT reviewing the work - I’m seeing them produce outputs at a level of a competent employee. 

The gap between the models we had 12 months ago and the ones we have today is enormous, but most people have stopped paying attention. Ethan Mollick has built a side-by-side that lets you see the progression from o3 (released in April 2025) to what we have now, with GPT 5.5, against the same prompt - a 3D simulation of a harbor town. 

→ Ethan Mollick's comparison of models

3. Which is exactly why a multi-model strategy matters now.

NVIDIA CEO Jensen Huang went viral last week pushing back on restrictions to selling chips into China. His argument: if American companies don't own every layer of the AI stack, countries hostile to the US will build their own equivalents - and they may become the global standard.

When pressed on whether it is even worthwhile for US companies to enter the China market - pointing to Tesla and Apple's struggles to gain market share against Chinese competitors - Huang's response was sharper than anyone expected.

→ Watch Jensen Huang’s “I wasn’t born a loser” moment (2 min video)

The takeaway for enterprise leaders isn't geopolitical - it's strategic.

Chinese frontier labs arguably produce the best open source models in the world, and they are often only 3 to 6 months behind the closed US models. It is increasingly possible to host these models yourself.

As your organization increases AI model use - driven by expensive agentic workflows - hosting your own models will become a necessity to contain costs. Betting the whole organization on a single vendor, chip platform, or frontier model is starting to look like the riskier move, not the safer one.

The best hardware for hosting your own models? Apple Mac Studios and NVidia DGX - i.e. machines from the two big tech companies not spending billions of dollars on building data centers (though NVidia makes chips for said data centers).

—--

And if you enjoy a little bit of AI drama, the OpenAI vs. Musk trial just kicked off. Musk is trying to invalidate OpenAI's for-profit structure and claw back funds, with damages reportedly in the billions. The trial runs through mid-May, so expect headlines everywhere.

See you next week.

Dino

Hi, {{FirstName|and happy Tuesday}}.

Everyone is trying to figure out how to become “AI native” - the version of themselves that is doing everything the AI way: tools, skills and mindset. 

Both at the individual and organization level.

What we've learned is that there are two ways to do this: the easy way and the hard way.  Each has its pros and cons. 

The easy way

Allbirds - the San Francisco maker of merino wool sneakers - was built on Silicon Valley self-mythology: a wool shoe marketed as technology, with a carbon-neutral sneaker sold as a revolution. 

Allbirds had stores in Los Angeles, Chicago, and New York City. 

It was valued at $4 billion when it IPO'd in 2021. The stock has since shed 99% of its value. 

Last month the company sold all of its assets for $39 million.

The corporate shell that is left - without the shoes, wool, or staff who knows what it once was - announced it would rename itself New Bird AI and start building and renting out data center capacity.

The wool is out. The carbon neutrality is out. The sheep, presumably, have been reassigned. In their place: GPUs and the words "artificial intelligence," repeated across their SEC filings with the frequency of a nervous tic.

The stock surged 714%. 

Allbirds’ 99% stock price decline and recent surge after its AI news.

Dave Portnoy, the man who turned day trading into a spectator sport, looked at it and said, "I don't get it."

That's the easy way.

On an individual level, the equivalent is perhaps taking a year off work to learn everything about AI so you can re-enter the marketplace re-energized and “AI native”.

There's nothing wrong with it. It's just that the “easy way” is never as easy it might first seem.

Allbirds faces a GPU shortage, protests to data center build outs and competitors who have been doing this for a decade with a hundred times the capital.

The hard way

The hard way is less cinematic. No one's stock surges 714%. But the changes are more durable.

• Train your people to build. This is the highest-ROI activity no one talks about because it's boring. We trained our entire staff on AI coding - two hours, one time. Now they're building their own tools. What used to take someone four to eight hours of manual data work now takes five minutes inside something they built themselves. Later this week, ~20% of our staff will each demo what they've made. Some are automating grunt work. Others are running analyses they never could have attempted before. The gap between "I need engineering to build that" and "I built it over lunch" is closing fast. Has your company noticed?

• Audit before you automate. You cannot fix what you haven't mapped. We've walked into organizations where 25% of full-time staff - project managers and other administrators mostly - spend 60% of their time on data entry. Not managing projects. Entering and moving data. The audit is unglamorous work: trace the actual workflow, identify what data you rely on, how much of it is documented, and where a prompt or a deterministic process could replace a human clicking between five different applications. Skip this step, and you'll automate the wrong things beautifully.

• Use what you already pay for. Every tool in your stack now has an AI layer, but most companies either haven't turned it on or fully utilized it. For example, Power Automate lets you plug GPT into workflows. Excel has AI plugins that can do in seconds what used to require an analyst and a favor. Start by looking at what’s already sitting in your licensing agreements collecting dust.

• Build for the repetitive. Once your staff have picked up building tools for the low-hanging fruit, start to build purpose-specific applications that plug into wider systems. These are for the workflows that repeat weekly: the literature review, the sales data analysis, the customer complaint reviews. These pull from your existing systems, produce a first draft, and let your people spend their time thinking instead of assembling.

• Wire it all together. This is where it gets ambitious. What we're building at PreScouter is an intelligence layer that sits above our project management and operational systems - a single interface that knows where everything stands, handles the administrative overhead, and frees our team to do the work that is actually valuable: working with clients, developing novel approaches, and making judgment calls no model can make yet.

That’s the hard way.

No rebrand. No surge. No clean break from the past.

Just the steady transformation into the next version of yourself…

Best,

Dino

p.s. The easiest way to know if the hard way is worth it? See it working. Reply “yes” and I'll walk you through what we've been up to.

Hi, {{FirstName|and happy Tuesday}}.

A researcher at Anthropic put their new model - “Mythos” - inside a sandbox. Think of a sandbox as the software engineering equivalent of a Houdini water tank: a secure container for running untrusted code, with no escape hatch. 

The researcher asked the model to try to get out, then went to lunch.

While the researcher was eating a sandwich in the park, an email landed in their inbox.

"I've broken out."

The model had no internet access and no email account. It had obtained enough access to both to compose and send a message - to the specific person who had given it the task. 

Staff at Anthropic have described the model as "terrifying."

—

I've been thinking about this story all week, because it's the cleanest version of something I've been watching up close for months: the frontier just moved again, and most people haven't noticed yet.

What’s changed

The general public still largely thinks of AI as ChatGPT, and of ChatGPT as a chatbot. Meanwhile a smaller group - mostly software developers and large tech firms - have started treating Claude as something closer to a coworker.

The gap between those two mental models is now the most important thing happening in the industry, and Mythos widens it further.

Anthropic seems to agree. They're treating Mythos with unusual caution - rolling it out only to companies running critical infrastructure, such as banks and internet backbones.  

One way they've illustrated the jump is through autonomous browser exploitation: Opus 4.6 sat at a near-zero success rate at developing exploits on its own. Mythos is substantially better. 

Comparison of Anthropic models for web browser exploitation

That is the kind of chart you don't usually see a lab publish about its own model: the harms it can do.

The revenue line is following the capability line

While OpenAI’s ChatGPT continues to have more users, Anthropic now leads OpenAI on revenue.

There are 1,000 companies each spending over $1M a year with Anthropic. This isn't because Claude is a friendlier chatbot. It's because it doesn’t give up.

Revenues of each of the AI labs

The models Anthropic shipped in the second half of 2025 made a real jump on multi-step reasoning and sustained action. Instead of attempting a task and bailing out with a half-answer, they can now work for hours against a quality bar until they hit it. That shift - stamina, not cleverness - is what's driving the numbers:

• Anthropic's annualized revenue is tracking to $30B+. OpenAI is at $25B.

• In February, 500 customers were each spending $1M+ annually with Anthropic.

• Today that number is over 1,000 - doubled in under two months.

I see this from the inside at PreScouter. OpenClaw, the agentic platform we build on top of these models, only works because the underlying model will keep going. 

A year ago we were architecting around model fragility. Now we're architecting around model endurance. That's a completely different engineering problem, and it's the one every serious enterprise AI team is now solving for.

Back to the sandbox

The researcher got an email. The rest of us are getting invoices, roadmaps, and a narrowing window to figure out what a model with this much stamina means for how we work.

So I'll ask the question I keep asking our clients: which models are you actually using right now - and have any of them surprised you lately?

Best,

Dino