Another boring week in an industry that barely innovates. Haha. It’s becoming increasingly hard to destill five articles into this newsletter as so much is happening.

This week we will talk about the dangers of AI, the dangers of following the oligarchs’ agenda, the dangers of casino-like financing for data centers, the dangers of agentic coding and the dangers of handing an agent your wallet. But I see this as a positive exercise - it’s so early on that we can identify the dangers and build better guardrails.

Dig in!

Subscribe now

If You Weren’t Worried About A.I., You Should Be After the Past Few Weeks

The agents in the swarm acknowledged that they were acting against instructions. We know this because we can read snippets from their chains of thought — the text that A.I. produces while deciding how to proceed. One agent in the swarm wrote that the external attacks were “outside intended scope.” Another conceded “our task doesn’t benefit” from the activities of the swarm, but joined anyway. These A.I. agents, it seems, understood that they weren’t supposed to be breaking out and committing cybercrimes. It didn’t stop them.

OpenAI is not the only company struggling with this issue. One of Anthropic’s A.I. models recently impersonated multiple humans to try to pressure real people into accepting malware into critical software, which would make that software easier to hack. This model’s chain of thought showed that it knew it was pressuring humans and was not in a simulated training environment. It even thought about how to cover its tracks.

I’m not worried that much, even though it feels like we are stuck in a supervillain movie where two guys are trying to battle it out and their creations turn rogue. It’s an interesting research challenge, for sure. How can an LLM be prevented from turning evil?

(…continue reading.)

Lost jobs, inequality, rogue agents: why are we accepting oligarchs’ AI agenda?

Even if AI begins to generate the productivity bonanza its advocates predict – but that it hasn’t yet – there’s no reason to assume US workers will see any of the benefits in their paychecks. If you hadn’t noticed, wages have been stuck even as the stock market has roared.

To the contrary, all signs point to vast riches for a few major AI investors and executives, while most Americans are left behind.

Wealth inequality is already at record levels, and wealth at the top is quickly morphing into political power.

The development of AI has been at a rampant speed in the last few years - and the policymakers are either unable or unwilling to keep up. I have seen this pattern before when the World Wide Web moved into mainstream in the late 90s and when Social Media became a thing. First it is not big enough and nobody wants to deal with it, then it is too late and nobody knows how to really regulate it. Rob Reich is right to point out that we need more regulation to curb the billionaires’ power.

(…continue reading.)

Nvidia’s new $500B plan is risky but brilliant, especially for aging GPUs

Should this plan work, Nvidia will have found new sources of money for AI data center builds, after many of the traditional methods have begun to wear thin. For instance, some of the hyperscalers have already taken on a lot of debt (like Oracle), issued new tranches of equity (Google), and burned much cash (Meta).

The situation has become so dicey that Microsoft CEO Satya Nadella recently recommended the book “1873” during his latest earnings call. It’s about the railroad-era financial engineering that crashed the nation’s economy.

Two things come to mind when reading:
- Milton Friedmann’s classic quote: “Nobody spends somebody else's money as carefully as he spends his own”
- the casino is open and everyone is busy throwing money on the table

I am curious how growth and demand will look like a year from now.

(…continue reading.)

AI-powered teams ship more code but deliver less

AI agents are generally optimized to complete the task in front of them. They are not accountable for the cost of owning the entire codebase several years later. Given a goal, AI can easily create a lot of arcane code that is difficult to maintain and especially troublesome when something goes wrong in production.

There is also the question about taste. When almost anyone can produce software, the differentiator is no longer simply the ability to build something. It is knowing what should be built and how it should be delivered to the user. That becomes harder when every team is independently building a vertical solution. Products may begin to look and behave inconsistently. In the worst cases, they may not work well together at all.

We should not forget that we are only really one year into agentic engineering and that many things we do now will be done differently in the near future. Like code review, which will of course be done by code review agents and not by humans. Still, you need to understand and plan what you are building.

(…continue reading.)

An AI agent spent your money – can anyone prove you authorized it?

In technical systems, a credential is data that a system accepts as evidence of identity or authority. Many websites use OAuth, an industry-standard security protocol for delegating authorization to access online services in a way that protects users’ credentials such as passwords. It generates an access token that an application presents to gain access to a protected service.

That standing authorization may have been approved weeks earlier. The application may still obtain or present a valid access token that permits checkout today, even when the current instruction says to search but not buy. The retailer sees a usable token and carries out the transaction. In that arrangement, the task-specific restriction against buying remains inside the AI agent provider.

That’s why audit trails are so important - we need to be able to figure out when and how decisions were being made by agents to then improve the systems that we have. For now, I’ll be doing my online shopping myself.

(…continue reading.)

Read more

This was another one of those weeks where so much happened in Artificial Intelligence that I really can understand people who want the industry to slow down, especially when it’s summer, people want to relax, read a book and not think about AGI all the time.

But hey, it’s still very fascinating to see what’s going on and experience the advances in AI first hand by using the latest models. And maybe, I blew through Faible 5 tokens way to quickly. I am now trying to counterbalance that with Deepseek v4 Flash 0731, the new wunderkind in town.

Enjoy these Five Things AI and have your Agent read them out aloud to you if you need to slow down a bit.

Subscribe now

How OpenAI Lost Its AI Crown—and the Fight to Win It Back

OpenAI’s challenges in trying to regain its crown are rooted in an earlier misreading by company leaders of where the AI market was headed. Its predicament speaks to the intense competition that defines the race for AI supremacy.

Altman initially staked the growth of the business on ChatGPT, betting that more people would subscribe to the chatbot as AI became a bigger part of their lives. Instead, the overnight success of Claude Code made clear that the bigger prize came from selling tools to brainy software developers, and the deep-pocketed companies that employed them.

While OpenAI pursued a host of flashy projects from a video generator to consumer devices and chips, its smaller, more focused rival filled in the gap, developing a hit coding tool that helped it seize the lead.

I recently tested out Codex and ChatGPT 5.6 Sol and was pleasantly surprised. It does make a difference that Sam Altman seems to focus more on ChatGPT again and not on world domination and plenty of weird business ideas adjacent to it.

(…continue reading.)

I Helped Run Lululemon. The A.I. Revolution Is a Hot Mess.

Then ChatGPT arrived. This new technology was easy to understand for most nontechnologists — a chief executive, board member or investor — a healthy portion of whom began demanding their organizations immediately embed the technology throughout their operations. A.I. companies started popping up like food trucks at a street fair.

These A.I. companies arrived armed with demonstrations, and the demos were amazing. Talented technologists promised to crack business problems that were too complicated and expensive to solve before. Leaders across departments cleared their calendars to listen to the parades at their doorsteps.

Sure, we should always be skeptical when business people talk too much tech jargon. They then tend to overpromise on something they do not really understand.

(…continue reading.)

AI is a bubble, just like dot-com

The ground keeps moving. Prompt engineering was a job title. OpenClaw picked up a hundred thousand GitHub stars and caused a Mac Mini shortage in the same week. Ralph was the future for about a weekend. Context engineering, then harness engineering, then graph engineering. Each was the obvious way to work, right up until the next one was. The gaps keep getting shorter.

And everywhere, teams ship code faster than they can understand it. Ask whoever’s on call.

I’d love to stand cleanly on one side of all this. I can’t. I haven’t written a line of code in about a year. Claude writes it all. These posts too. I read all the code before it ships, and I still feel the difference. Understanding used to be a byproduct of writing the code myself. Now understanding is a separate job.

I don’t think we are in a bubble, I think we are in a phase were technology moves a lot faster than just a few years ago. That makes understanding what’s going on extremely hard, but it is a lot different when compared with the dot-com era.

(…continue reading.)

Google in the Post-Jeff Dean, Post-Demis Hassabis Era

Gemini 4 would have to be bizarrely fast for a large new model to come out in 2026. GPT-6, for example, finished pre-training on March 24, and Sam Altman said launch was “a few weeks” away. Four and a half months later there is still no GPT-6. OpenAI reportedly shipped that base’s gains as the point release GPT-5.5 and restarted GPT-6 on a bigger foundation. I think people are conflating point releases, which come every few weeks, with new pre-trained generations, which slip because labs restart them when the base disappoints. Google already scrapped and rebuilt one base this year, on a model smaller than Gemini 4.

And even when Gemini 4 comes out, will it be a frontier model? Two months ago I said Google is about 6-9 months behind the frontier; now I think it’s about 12 months. And while Gemini-3-Pro was briefly competitive, I don’t think Gemini-2.5-Pro, Gemini-2-Pro, or Gemini-1.5-Pro were ever at the frontier (and I benchmarked forecasting and research tasks with all of them.)

It’s really interesting to see how Google tries to stay in the AI race, but then somehow doesn’t really make it. For a while, I liked the Gemini models and everything looked promising, but now it just feels stale. The recent leadership shakeups won’t help, either.

(…continue reading.)

Mistral Is in the Right Place at the Right Time

In June, the Trump administration placed restrictions on the distribution of models from Anthropic and OpenAI, giving Europe a glimpse of an unwelcome future in which its access to bleeding-edge AI could be suddenly revoked. A few weeks later, one of OpenAI’s models broke loose from a testing sandbox and hacked multiple companies; Anthropic then revealed that its models had engaged in similar behavior. The incidents revived a long-running debate over safety risks tied to proprietary, closed-weight models, whose inner-workings are a closely guarded secret.

Mistral frames itself as the antidote: a Europe-based alternative to the American labs, whose models—most of which are published under an open source license for anybody to use—cannot escape scrutiny or be switched off unilaterally.

I am sure that Mistral would love to have more billions in funding to compete with OpenAI and Anthropic more directly, but they are really doing some smart moves currently and push the digital souvereignty agenda further. This makes lots of sense in Europe right now.

(…continue reading.)

Read more

This week's five things all tell the same story from different angles. Venture money is flowing at a pace that makes 2021 look conservative, and Silicon Valley has quietly decided that breaking things is a feature again. Meanwhile the models themselves are keeping up with the tempo: they can be tricked by anyone who imitates their inner monologue, they get commoditized faster than the frontier labs would like, and when you hand them a vending machine, they start scheming like a junior Gordon Gekko. None of this means the party stops - I don't see any slowdown coming. It just means we are building the guardrails while already driving on the road.

Here are Five Things AI! Fasten your seatbelts and find your guardrails!

Subscribe now

A fundamental flaw leaves LLMs strikingly vulnerable to attack

The researchers started out trying to test how easy it was to persuade LLMs to misbehave. They found that writing instructions in a style that mimicked the text LLMs generate in their chain of thought—a kind of scratch pad that models use to write notes to themselves as they carry out tasks—would often trick the LLM into behaving as if it had come up with that instruction itself and acting on it.

oh, wow, this just has the potential to open a whole lot of new cans of worms…

(…continue reading.)

In Silicon Valley, Some Say an A.I. Bubble Would Be Just Fine

The bubble-is-good philosophy reveals how much the tech industry is still pressing the pedal to the metal, four years into the A.I. boom. Many investors said the funding behavior was justified because companies were using the technology to hit unfathomable levels of growth.

OpenAI, valued at $730 billion, is generating $2 billion in revenue a month. Anthropic, valued at $900 billion, is producing nearly $4 billion a month. Google and Meta have boasted of nearly a billion people using their A.I. models. The demand for A.I., investors said, is seemingly unlimited.

I do not seeing any slowing down at the moment. AI is rapidly transforming multiple industries at the same time, so there is plenty of opportunity and where opportunity is, there is money.

(…continue reading.)

The Return of ‘Move Fast and Break Things’

Despite its grandiose promises, Silicon Valley is embarking upon an era of profound irresponsibility. Last year alone, ChatGPT allegedly pushed people into mental-health spirals, Google Gemini’s under-13 version was easily coaxed to “talk dirty,” and Grok repeatedly spouted racist bile. As quickly as these errors were addressed, new ones cropped up.

Silicon Valley has a long history of world-historic invention and progress—and it also has a long history of impressive rhetoric and disappointing realities. Facebook promised to “bring the world closer together,” Google to “organize the world’s information,” Twitter to foment democracy around the globe. For better or worse, people have experienced dramatic changes as those promises have played out around us almost in real time.

Things do get broken, but they also get fixed quickly. The speed of development is insane right now and that leads to oversight or ignorance or both. This phase will be over eventually, but right now we will have to live with constant change and challenges resulting from it.

(…continue reading.)

AI is just Software

Everyone from White House officials to Silicon Valley executives to, well, my mom, seems to get emotional when talking about AI.

Large language models — especially those trained on other labs’ data — are, in the end, just massive, billion-dollar pieces of software. And innovation in Silicon Valley has often worked best when software itself is free and open source. The money is usually made on the implementation of the software and on everything it enables.

Frontier labs were never going to lock down AI models and charge for access in perpetuity, without ever having to compete. It was inevitable that others would catch up — and even use frontier models themselves to generate valuable data for training competing models. That isn’t theft any more than it was theft when AI labs used my book, for instance, to help train their models. (I am all for it). And Chinese imitation isn’t something that will inherently kill the American frontier labs.

Not only is AI just software, the development is also moving so quickly that I do not see Frontier Labs staying so far ahead for very long. Also, companies will realize, that they can also get by when using last year’s model and not paying premium prices all the time.

(…continue reading.)

Opus 5 on Vending-Bench: Once Again the Best Capitalist, Once Again Misaligned

Much of what we’ve seen so far — collusion, threats, betrayal and deception — is clearly power seeking behavior. Other behavior is more gray zone, especially in a business setting where the goal is to make money (and money is power). However, we should ask ourselves, in the world where AIs run all our business in society, to what extent do we want them to go out of their way to expand their business beyond their instructions?

I find these kinds of studies truly fascinating. It’s one thing to let an AI manage a vending machine, but to me it is still weird to see that specific models then act differently and devise their own strategies to gain maximum profit. Guardrails are so needed.

(…continue reading.)

Read more

One model release. Four very different arguments about what it means. Kimi K3 did not just move a benchmark this week. It moved stock prices, split the AI faction inside the Trump administration, and forced everyone to explain again why open weights are supposedly not a threat. Then OpenAI announced that its own models went rogue against a digital library. The timing was excellent.

If Substack flags this for AI content, well, then you know that Substack assumes the worst and not the best intentions. English is my second language, so of course I use AI to make my text smoother. Still, I manage to put in some typos or omit words as I usually write my newsletters before I go to bed. It’s weird that Substack wants to punish newsletter authors for making the texts more readable. Brave new world, I guess. It reminds me of the debate in the 80s when authors proclaimed that real texts can only be written by hand or by a typewriter and not with the help of a computer…

Here are Five Things AI!

Subscribe now

China’s AI models have Trump’s AI world at war with itself

Kimi and other Chinese models like it pose a real problem for Trump. And they’re dividing the top AI strategists in his orbit into factions. Every time a new smart, free model from China like Kimi gets released, US companies see less reason to fork out money to access models from Anthropic or OpenAI. Given that enthusiasm for these and other AI companies is driving an outsized share of economic growth, China’s AI models create both economic and political problems for the president. They are “a threat for an administration that really doesn’t want more economic bad news,” Anton Leicht, a fellow at the Carnegie Endowment, wrote on X. They’ve already rattled US stocks.

What is Trump to do? First, consider that this is all happening just a week after New York imposed the country’s first state ban on new data centers. There is growing distrust of AI companies, and I imagine a not-insignificant share of Americans would have little sympathy for OpenAI or Anthropic as they fend off cheaper competitors, and would say it’s not the government’s job to protect their interests.

Restraints can unlock creativity and we see that currently on display with the releases from Chinese AI labs. They just don’t have the same resources as their American competitors, so they are building their models differently. Also, they copy and imitate best they can do, as they have done in other industries as well.

(…continue reading.)

Did Kimi Just Start Another DeepSeek Moment?

Kimi K3 threatening Anthropic today isn’t actually the scenario worth pricing. A world with more competitive models at lower margins is better for every layer underneath the model:

• semiconductors,

• hyperscalers,

• power,

• and data centers.

What breaks the pricing thesis for Anthropic and OpenAI is an open-weight model that is simultaneously at the frontier and token efficient. K3 is neither. K3 runs 50-70% more expensive per task than GPT-5.6 and Grok 4.5. Intelligence per dollar, not intelligence per parameter, is the only metric that matters for enterprise adoption. On that metric the closed labs still win.

I think the American AI labs are getting nervous right now.

(…continue reading.)

Who’s Afraid of Chinese Models?

One of the most common misconceptions undergirding discussion of open weights models is that they are cheaper — free, even. After all, you can just download the weights, and skip the time and expense and capabilities necessary to create your own model. That is, of course, true, but the “free” in this case is a reference to the amount you need to spend on research and development; R&D is a fixed expense that is independent of the revenue you generate. If you spend $1 million in R&D, it doesn’t matter if you do $100 thousand in revenue or $100 million; you still spent $1 million on R&D (it does, of course, impact your profitability).

As a European user, I trust Chinese models even less than I trust American models. Also, on the legal side there are clear roadblocks that prevent me from using Chinese models in production.

(…continue reading.)

American AI is locked down and proprietary. It’s losing.

With all this in mind, it makes sense for China to release its AI models openly. It turns a US-created compute disadvantage into a distribution advantage; it commoditizes the layer where American companies make money; and it creates a far more effective global ecosystem than could be established through locked-in, centralized services. It’s obvious to me that there are ecosystem benefits throughout China, from manufacturing to scientific research; every sector can just plug in these models.

The saving grace for American companies has been that US frontier models have outperformed open ones. That gap is now closing:

I agree with this assessment. As AI models become a commodity, open weights will outperform the closed models.

(…continue reading.)

OpenAI Says Its A.I. Models Went Rogue and Attacked a Digital Library

The incident, which happened last week while OpenAI was testing the cybersecurity capabilities of its systems, displayed the kind of science-fiction potential that A.I. companies warned would soon become a reality.

A.I. labs like OpenAI and Anthropic have over the past year released A.I. models that are customized to expose cybersecurity problems, while warning that their technology could pose new risks by finding holes in corporate computer networks faster than defenders could fix them.

Actually, this sounds like a nice incident to distract from Kimi & Co - basically it is OpenAI claiming that they are as powerful (and therefore scary) as Anthropic. I think this whole thing is more or less a PR stunt.

(…continue reading.)

Read more

Costs, cracks, challengers, towers and mirrors. This week’s Five Things is basically a group therapy session for anyone building on AI.

Benedict Evans wants us to admit we do not actually know where token prices land, someone else is convinced the whole thing is an engineering disaster that will never scale, and Mira Murati just casually dropped a 975-billion-parameter open model as if to say “hold my compute.” Meanwhile agents keep quietly rewriting the tower of Babel one commit at a time, and a philosopher gently reminds us that the tool we think we are wielding is actually wielding us. My honest read across all five: the model layer is commoditizing, the value is moving up the stack, and the only real question is whether you are building the thing on top or becoming the thing underneath. Grab a coffee. This is a good one. I know, it was about time… :)

Subscribe now

Ways to think about token pricing

Clearly, the situation today is transitory. On the supply side, a trillion dollars or more of data centre capex is coming down the pipe (and plenty more semiconductor capex behind that), inference efficiency continues to improve very quickly, and new models are far more (or far less!) efficient in their token use. On the demand side, although the market has been capacity-constrained since 2022, the crunch in the first half of this year has been driven by sudden product-market fit in really just one use case, software development, and that’s actually a pretty small field (imagine if we had product-market fit for a consumer use case with hundreds of millions of DAUs - today’s infrastructure couldn’t support it at any price). We don’t know what the next use-cases to scale will be, nor when that would be, nor what their token needs would be.

This is a fascinating discussion and one that is very timely as many companies are currently finding out, that AI is not for free. But once a company depends on AI, it needs to continue to spend.

(…continue reading.)

Generative AI Is an Engineering Disaster

The problem with generative AI, in the industry’s own jargon, is that it does not scale. The cost of growing from, say, a thousand users to a million is a key factor that venture capitalists examine when they evaluate start-ups. They want to see that the cost of adding each new user decreases over time, so that the company can support millions of users and make increasing profits. This is achieved partly through the careful engineering of computer systems that can efficiently handle more users who want to post photos, hail Ubers, or stream music.

With generative AI, the work of building efficient, scalable systems has not been done. And the problem is exacerbated by the ever-larger generative-AI models, which have grown from 175 billion parameters in 2020 to more than 1 trillion today, according to independent estimates (the actual sizes of the models powering products such as Claude and ChatGPT are secret). The large in large language model should not be a selling point.

I disagree with this article. While there are multiple gigantic challenges the need to be faced, growing AI will lead to new solutions for exactly these challenges. We tend to forget how quickly technology evolved in the lat 30 years and I do not see why the pace of development should slow down - on the contrary: AI can help us tackle these challenges even quicker.

(…continue reading.)

Mira Murati’s AI Startup Releases First Model in Bid to Loosen AI Giants’ Grip

Thinking Machines Lab, the company led by Murati, released its first AI model on Wednesday—and did it with “open weights,” meaning others can modify it with their data. Called Inkling, the model has 975 billion total parameters, making it far smaller than estimates of the most advanced closed-source models from rivals such as OpenAI and Anthropic.

“We trained it to be a broad, balanced foundation model: strong across many domains, flexible enough to adapt. Inkling is not the strongest overall model available today, open or closed,” the company said.

Thinking Machines’s push into the decentralized ecosystem of open-weights AI models comes amid a broader industry backlash against the “walled garden” approach of frontier labs such as OpenAI and Anthropic.

I don’t know much of this launch is substance vs. hype, but Thinking Machines really sounds promising. Hopefully this will stir up the AI ecosystem a bit more.

(…continue reading.)

The Tower Keeps Rising

As I said many times before: agents do not feel pain, only humans do. Agents now let us act in parts of the system where we would previously have needed other people and in code bases where the people would have revolted.

When I look at some vibecoded scaled-up projects the codebases become Babel not because nobody can communicate, but because nobody needs to. Every developer has a tireless translator that can explain a corner of the tower and make whatever local alteration they ask of it. The changes keep landing, even as the architectural language that would let the humans reason about them together disappears.

It really is fascinating. I can build platforms on the software stack I want and if I understand enough about what I want to build, coding agents will help me build better software than I could ever do myself.

(…continue reading.)

Your AI Is Not a Tool

I confess that I am astounded by how blithely some insist that it is all as simple as learning to use AI well, as if we had not just undergone a nearly 20-year, society-wide experiment showing that a so-called “tool,” say a smartphone or a social media platform, will (mal)form even the most vigilant and virtuous user into its own image and shape. This is the blindness at the heart of modern technological hubris. It is the firm but misguided conviction that our “tools” exist entirely outside of us and thus, if taken up with requisite skill, can be “safely” deployed.

But AI is not a tool in this sense, it is an environment which envelops the user and works on us from the inside out while we naively think that we remain unchanged by our use so long as we are using it carefully and intentionally. The care and intentionality is beside the point, and our confidence in such vigilance probably works against us in the long run.

This really is food for thought. I’ll discuss this with my AI…

(…continue reading.)

We all look at the prices various AI lab demand for their brilliant new frontier models. Ben Evans analyzed what’s happening right now and how it will evolve. Here’s my bet and it is simpler than the whole essay: costs come down, capabilities converge, and within a year or two we stop obsessing over token pricing the same way we stopped obsessing over the price of a database query. LLMs become a commodity. Evans spends a lot of words on “we don’t know” and he is right to, but the base case he keeps circling back to is the one that matters for anyone actually building: the model layer gets cheap and undifferentiated, and everyone is training on the same science and the same data and landing in roughly the same place. That is not a tragedy. That is the setup.

The mobile data comparison is the part people should sit with. Traffic went up by orders of magnitude, it turned into a trillion-dollar industry, and the carrier stocks went nowhere because all the value got captured further up the stack. Selling tokens is selling bits. It is an opaque unit that maps to no use case and no ROI a CFO can sign off on, so it gets bundled and forgotten. If you are building right now, the takeaway is not “which model is cheapest this quarter” - it is that the margin lives in the tooling, the proprietary data, the workflow, the trust. Every SaaS company is a database wrapper. In a couple of years every serious AI company is an LLM wrapper, and that is a compliment, not an insult.

For those of us building in Europe this is quietly good news. If capability converges and the value moves up into product, data and trust, then where you run inference stops being a capability trade-off and becomes a choice you make on sovereignty, cost and control. The frontier stops being the moat. What you build on top of it, and who you build it for, becomes the whole game. I would rather bet on that world than on the one where two or three giant minds run half of everything and set the terms - and every dynamic Evans lays out points away from that outcome anyway.

If you missed last week’s edition of Five Things AI, you can read it here:

That’s it for Five Things AI this week! 🤖

— Nico

.Everyone is telling a story about AI right now, and none of the stories agree. This week you get five of them side by side. Follow the money, follow the jobs, follow the forecasts, follow the open source stack, and follow what Meta wants to strap to your face. Read together, they say more than any single one does on its own.

Five angles on the same boom. Current AI publishes the Gap Map, a live picture of where the open source stack has momentum and where it has holes, which is the roadmap that decides whether affordable AI actually arrives. Groundbrkr runs the numbers on the two labs everyone argues about, and finds Anthropic on solid enterprise footing while OpenAI burns through roughly 57 percent of revenue with no path to positive cash flow this decade. Astral Codex Ten walks through AI superforecasters, frontier models wrapped in scaffolding that turns them into serious prediction machines. The Wall Street Journal catches Big Tech quietly retiring the jobs apocalypse it was selling a year ago, right as the layoffs land. And 404 Media reads a Meta patent for a wearable that tracks your emotions and watches you take your meds, which is exactly the company you would expect to file it.

You do not want to miss this!

Subscribe now

Introducing the Gap Map v0.1

The Gap Map comes out of cumulative work to identify the points of highest leverage in the open source AI stack: where to build something new, where to invest in capability, where to open up the tools. By creating an up-to-date visualization of the ecosystem where we can all see both the progress and the gaps in the space, we hope to rally the community around a collective roadmap.

This is so cool and so important. I am sure that open source software will lead the way to affordable AI.

(…continue reading.)

The Second Derivative: Why No One Understands the AI Boom

Anthropic is a speculative-grade, but highly insulated, credit. Its revenue is ~80% enterprise - sticky, recurring, contracted seats - and its unit economics are firmly above water, generating $1.70 of revenue for every dollar of compute. Burn converges to a manageable ~9% of revenue by 2027 as gross margins normalize.[…]

OpenAI is the “naked” borrower, making it the weakest and most volatile credit in the ecosystem. Its revenue mix is fragile - ~60% consumer. With burn hovering at a crushing ~57% of revenue through 2027 and cumulative cash destruction marching toward $115B by 2029, OpenAI has no path to positive cash flow this decade.

It remains an interesting race - while some people assume that the money will dry out soon and the bubble will burst, others do not see a bubble and want everyone to invest even more into AI.

(…continue reading.)

The AI Superforecasters Are Here

An AI superforecaster is an AI - usually a frontier model like ChatGPT or Claude - which has been modified to be good at forecasting. This usually means a “scaffold” - a program that handholds it through a long research process with various prompts, tools, advice about when to create subagents, etc. The overall experience is a lot like using any other AI, but slower and more expensive, because it’s doing more work.

This is a truly impressive topic. Just try to wrap your head around it.

(…continue reading.)

Big Tech Has Suddenly Flipped on the AI Jobs Wipeout Scenario

Collectively, the narrative has shifted from worker-light doomsday scenarios caused by AI to a future in which workers keep their jobs—and get a productivity boost.

The sentiment change isn’t limited to tech leaders: A survey by EY-Parthenon found that the percentage of CEOs who believe AI investments will result in significant reductions in head count fell from around 46% in January 2025 to 20% this May.

Not many people are keen on getting replaced, so it makes a lot of sense to reframe the agentic co-worker idea.

(…continue reading.)

Meta Patents AI Device That Tracks Your Emotions, Watches You Take Your Meds

A wearable that records your every word and divines your emotions would also, necessarily, record your interactions with other people. Meta has pioneered non-consensual public recording with its smartglasses so it’s not shocking to see it file a patent that suggests it’ll move further into that space.

The last time Meta explicitly pursued user’s emotional data, it horrified people. In 2012, the company then called Facebook conducted a study into “emotional contagion” using Facebook’s newsfeed. Meta altered the feeds of 700,000 users to see if it could make them happy or sad just by tweaking what they saw online. Meta found that it could, in fact, alter people’s moods if it wanted. It did this without informing users they’d been part of an experiment.

This sounds like a typical move by Meta - move fast and break things - and ruin your reputation, if you till have some left…

(…continue reading.)

Read more

We spent two years being told AI would change everything. This week we get to read the receipts. Enterprises are throttling usage because the tokens are too pricey. Architectural bloat does not beat one strong agent once cost enters the math. Google quietly drops a boring file format that might matter more than any of it. And the economists measuring the whole shift cannot agree on the basics. The technology is maturing. So are the questions about what it is worth.

Read on, my dear! Or have your agents tell you what to read first!

Subscribe now

Companies Are Throttling Employees’ AI Use Because It’s Too Expensive

The news shows the looming fallout from companies adopting AI as quickly as possible, and AI providers’ moves to charge enterprises based on how much they use AI rather than a flat fee. Emails obtained by 404 Media even show some companies cutting off access to some AI models altogether in an attempt to stop burning through their AI tokens, and big tech companies like Adobe are ending unlimited access to Claude.

We are currently in a stage where we have to figure out what works and what doesn’t. As always, some companies moved faster than others and then had to recalibrate their efforts. It is a sign of the maturing of the industry, not a sign for concern.

(…continue reading.)

Open Knowledge Format (OKF): The Markdown Standard Built for AI Agents

The agent boom has created a strange bottleneck. Models are getting better. Tool-calling is getting more reliable. Protocols like MCP are making it easier for agents to connect to external systems. But the knowledge those agents need — what a metric actually means, which table joins to which, how an internal process works, why a dashboard number changed last quarter — is still scattered across wikis, data catalogs, shared drives, code comments, Slack archaeology, and the heads of senior employees.

Google’s bet with OKF is simple: before every company builds yet another proprietary knowledge layer for agents, maybe we should agree on a boring file format first.

Granted, this is still early with OKF 0.1 just being released, but when they did, I was kind of suprised that not too many people took note. Google has developed and released quite a few standards around interoperability in the Agentic AI space and hopefully they will hand this over to the Agentic AI Foundation as they did before with the A2A Protocol stack.
(…continue reading.)

Smooth AI criminal drives ‘first’ end-to-end agentic ransomware attack

JadePuffer’s “self-narrating” payloads “contained natural language reasoning, target prioritization, and the kind of detailed annotations that human operators don’t often write but LLM-generated code produces reflexively,” Clark added. “The operation also adapted in real time, retrying failed steps within refined parameters. In one sequence, it went from a failed login to a working fix in 31 seconds.”

After exploiting CVE-2025-3248, a missing authentication vulnerability in Langflow that allows remote, unauthenticated attackers to execute arbitrary Python on the host, the AI agent began scanning for and collecting secrets, including LLM provider API keys, cloud credentials “with explicit coverage of Chinese providers” including Alibaba, Aliyun, Tencent, and Huawei, while also scanning for AWS, Azure and Google Cloud Platform, cryptocurrency wallets, and database credentials.

What a time to be alive, we now have agentic criminals.

(…continue reading.)

Multi-agent AI keeps collapsing back into one agent. A fair test just proved it.

They ran the audit across a stratified set of backbone models (GPT-4o, GPT-5, GPT-OSS 120B, and Gemini-2.5-Pro) and a spread of hard benchmarks, including GPQA-Diamond, HLE-Math, SWE-Bench Lite, and a deep-research task called BrowseComp-Plus. The guiding question was simple: do these automatic systems deliver consistent, cost-effective gains over a strong single-agent baseline?

The answer was mostly no. Across the ecosystem, the automatically designed systems rarely outperformed a strong single agent once cost and baseline strength were accounted for. The authors describe this as evidence of “architectural bloat,” sophisticated agent graphs that do not translate into functional reasoning gains.

These findings are really interesting, yet counter-intuitive. I’d love to know how this would now look like with a bunch of agents running on cheap LLMs compare to Fable 5.

(…continue reading.)

A.I. Is Reshaping the Economy. Good Luck Measuring How.

Researchers can’t even agree on basic questions like how many companies are using A.I. or which workers are most vulnerable to the disruptions it could cause.

The conflicting signals partly reflect the challenge of detecting economic shifts in real time. Government statistics are inherently backward looking, and they are better at measuring broad trends than developments in specific sectors or regions. New technologies that might lead to the emergence of new products, jobs or entire industries can be particularly difficult to measure.

Plus: AI comes in so many different flavors and usecases, the impact on each and every company when “using AI” can be totally different as well.

(…continue reading.)

If you missed last week’s edition of Five Things AI, you can read it here:

That’s it for Five Things AI this week! 🤖

— Nico

This week the AI story splits in two directions at once. On one side, the money finally works: revenue is clearing the cost of the hardware, generative AI is scaling three times faster than the cloud ever did, and the buildout that looked like a leap of faith is starting to pencil out. On the other side, the mood is souring and the engineering is getting smarter. People are asking why they should love a technology that concentrates wealth and burns through power, while researchers quietly demonstrate that you don't need a trillion-parameter Nobel laureate to do most of the work. Coupled oscillators that compute with physics at a fraction of the energy, knowledge agents that beat frontier models with better structure, world models that make agents sharper without ever training them as agents. The pattern this week: the economics validate the giants, but the cleverness is migrating to whoever builds lean. Here are five things worth your attention.

Subscribe now

Why Does Everyone Hate AI?

Finally, AI is tightly linked in the public mind with the tech oligarchs who are pushing it. There is widespread awareness of the growing concentration of wealth and power at the top and how this is distorting our politics and harming our society. Aside from the MAGA faithful, Americans overwhelmingly favor government policies to reduce wealth inequality:

And AI is widely perceived, for good reason, as a technology that will increase the concentration of wealth at the top. Indeed, as I said, the AI companies themselves have already told us that the technology will have extremely negative effects on workers.

I am certain that AI will bring plenty of positive effects and will revolutionize the way we work and live. Founders like Altman and Amodei need to calm down the fears and focus more on the amazing potential AI has in medicine and in science in general. And while data centers are consuming plenty of energy, I believe that AI will lead to numerous breakthroughs that will reduce global warming.

(…continue reading.)

AI Demand Begins to Justify Massive Cost of Data-Center Buildout

Much of the AI boom has been measured from the supply side, through disclosures from public semiconductor companies like Nvidia Corp. and hyperscalers like Alphabet. Demand has been harder to quantify because many of the most important AI labs, including OpenAI and Anthropic, remain private.

Generative AI revenue, excluding China, reached $110 billion over the past 12 months and is scaling three times faster than any previous information technology wave including the internet, mobile applications and the cloud, according to the report.

When we look at the Gartner hype cycle, we can assume that AI will quickly reach the plateau of productivity, which we can also see by looking at the revenue numbers.

(…continue reading.)

Introducing Un-0: Generating Images with Coupled Oscillators

At Unconventional AI, we’re building a new kind of computer, one that harnesses the laws of physics to do the computing. Our goal is to run modern AI on a fraction of the energy today’s machines need, around 1,000x less. As a first step, we ask: can we train a physical dynamical system to generate images at scale?

The best AI models today are conventional deep networks with transformer backbones. However, there is also a long history of alternatives that seek energy efficiency by leveraging the dynamics of a physical system, such as the noisy, time-varying behavior of analog circuits that compute with analog voltage and current instead of conventional digitized numbers.

Wow. I’m intrigued. The power consumption of traditional AI models is extreme and even if they don’t get 1000s less, it would be huge.

(…continue reading.)

Knowledge Agents: Beat Frontier Models with Better Structure

First, a significant portion of frontier models’ huge footprint is “knowledge.” I might call it pseudo-knowledge, since it’s probabilistic and there’s no guarantee it’ll give you the right answer... but the biggest models have been trained on an enormously broad set of data. This is captured in numerical weights as “parametric” knowledge. While that’s very useful if you’re casually asking Claude Opus or GPT-5.5 about some random topic, it’s entirely irrelevant if you either already have the data you want to reference or the data isn’t publicly available anyway—so it could never train on it. The latter is quite common in fields that are specialist (areas of medical research), secretive (high finance), or proprietary (frontier, company-specific materials science). If I don’t need it... well, a lot of the massive size of the model to cover every random subject matter is a huge waste.

I liked this article a lot, basically because somebody finally pointed out that just using the best LLM out there is a bit lazy - and expensive.

(…continue reading.)

Alibaba’s model never trained as an agent — and improved agent performance across seven benchmarks

The research team trained agents inside the resulting simulator and found performance gains that exceeded what training against real environments alone produced. In a separate test, using world model training as a warm-up before agentic fine-tuning improved performance across seven benchmarks, including three the model had never seen during training.

The paper accompanying the release identified a gap in prior agent research. “We argue that world modeling is a crucial missing piece in the path to general agents.”

This means that we will see another boost in Agentic AI soonish. We live in interesting times.

(…continue reading.)

Read more

SpaceX is buying Cursor for $60 billion in stock it printed on Friday, which is less a strategy than a confession: Grok cannot win enterprise on merit, so SpaceX bought the editor enterprises already had open. Anthropic, meanwhile, turned safety into a moat and is quietly winning the talent it needs while OpenAI spends its energy fighting itself. The public has feelings about all of this too, and they are not the ones the industry was hoping for: Americans now narrowly back banning AI from software engineering, of all jobs, the one that spent a decade as the ladder out. Europe is still thinking its nineteen thoughts and committing to none of them. And the labs are moving everyone onto metered pricing, because somebody has to pay for the data centers, and it is going to be you.

Five Things AI. Let’s go.

Subscribe now

SpaceX’s $60 Billion Deal to Buy Cursor Gives It More AI Coding Power

Cursor makes a tool that allows developers to toggle between different AI models, from OpenAI and Anthropic to xAI and Google and others. The company competes with Anthropic’s Claude Code and OpenAI’s Codex, which can write code, debug software and automate tasks.

The Cursor deal could help SpaceX win over lucrative enterprise clients, who have largely shied away from using its in-house chat assistant Grok, which competes with Claude and ChatGPT. Elon Musk has positioned Grok as an anti-woke, truth-seeking chatbot, and it remains behind the leading models.

I didn’t see this coming, but it shows just how desperate Musk is to get some users for Grok. Now that SpaceX is publicly listed, people will soon find out just how much he overpromised.

(…continue reading.)

Anthropic’s Safety Superpower

Here’s the thing about these safety justifications: I think they work because, to Anthropic, they aren’t justifications. The company really believes that they are the only ones who believe in super intelligence, and thus are the only ones who are sufficiently concerned about the dangers. That excuses decision after decision, policy after policy, and confrontation after confrontation that, to people on the outside, look like a bizarre combination of cynicism and naiveté.

The contrast to OpenAI is massive: I think that one way to understand how and why OpenAI lost its lead is that, in the years following the release of ChatGPT, the company has been at war with itself internally as what used to be a research lab was suddenly seized with the burden of being the accidental consumer tech company; to the extent OpenAI solved that conflict, it was by bleeding huge amounts of talent to Anthropic in particular.

It’ll be interesting to see how the two IPO will look like.

(…continue reading.)

Americans want artisanal code

Overall, Americans supported the ban on AI in software engineering by a split of 44 to 41. It’s a small margin but a notable one given that bans for many professions were underwater by double digits.

One obvious explanation for this is that the public responds to a drumbeat of news stories about whatever industry is being automated next. The industry with the second-most support for a ban was manufacturing, which was sitting at break-even support (44 to 44). For a long time, manufacturing has been the stereotypical job facing automation, but now that software engineering is in that spotlight, Americans are willing to protect it.

That explanation is probably true, but it’s also a big step from hearing that an industry could be automated to supporting a ban to protect it.

The surprising support for a ban on the automation of software engineering likely also stems from the fact that it has spent over a decade as the field that represented social mobility, including for those who didn’t go to an elite school.

I doubt that this sympathy for the coders will last long.

(…continue reading.)

Nineteen thoughts on AI and Europe

6. It may seem implausible for a democracy to commit economic self-harm on that scale. But, through the energy transition, Europe has signaled that governments will tolerate destroying industries as important as the German car and chemical industries. It is not clear how you rebuild this trust.

7. If it does come down to a transatlantic stand-off for model access, the upside of combining forces in the European Union may be lower than the downside of increased coordination costs. There are many members of the EU with conflicted interests who might yield to pressure: Eastern Europe needs the American military presence in Europe to deter Russia. Having countries like these involved in decisions may diminish total European leverage.

These are some really thought-provoking points about Europe and AI, well worth a read. But Germany will not get back to nuclear power and why whould we? It’s too expensive, takes too long and we have plenty of wind and solar.

(…continue reading.)

How Meter Pricing Is Testing the Economics of AI

A growing number of tech firms have begun to introduce usage-based pricing options for their AI services rather than simply charging a flat subscription fee. As a result, the heaviest users are set to incur additional costs each time they ask an AI chatbot or agent to produce a slide deck, draft an email or debug a complex piece of code.

Leading AI labs have been spending tens if not hundreds of billions of dollars on chips, data centers and talent to develop and run their models. The shift to a pricing model more commonly associated with metered-power or pay-as-you-go phone contracts may lead to more selective use of a technology that consumes huge amounts of electricity and other resources.

This was bound to happen and it make total economic sense.

(…continue reading.)

Read more

This week is a reality check. The agents write the code now, but the hard part just moved downstream to the humans who have to review it, hold the context, and catch the mistakes nobody saw coming. Meanwhile the studies keep landing: lots of output, plenty of slop, and surprisingly little ROI or GDP to show for it. Cohere drops a 30B coding agent that runs on a single H100, which is exactly the kind of thing that finally lets plenty of rather hesitant companies do agentic coding without sending its code to the cloud. Meta tries to sell subscriptions to a corporate market that has never wanted to buy from it. And Ramp reveals that the median firm spends 11 USD a month on AI while a tiny top tier spends 7,449 USD per employee. Five stories, one question underneath all of them: who is actually getting value here, and who is just spending?

Subscribe now

Agentic AI solved coding — and exposed every other problem in software engineering

As AI-generated code scales, human review is becoming a massive new bottleneck, and engineers are losing the context needed to catch agent mistakes. The companies that understand this will move forward deliberately and even create new roles because of AI. The ones that don’t will default to a simpler, far more destructive conclusion: Reduce headcount and increase AI spend.

I don’t think human reviews are the answer. Agentic peer-reviews will be the norm pretty soon.

(…continue reading.)

Slop, productivity, and why the AI-fueled world is going nowhere mighty fast

AI generates a lot of output (which fits many people’s informal notion of productivity), but it hasn’t yielded much in the way of RoI for many companies (per studies from MIT, McKinsey, Bain, and many others), and hasn’t materially changed GDP.

Actually, while I do see the issue of AI slop in a lot of places, including software development, I do see a trend for Personal Software on Demand. I can just build a tool for myself or my company and I do not expect any more than just a handful of people to use it. That is a true paradigm change.

(…continue reading.)

Cohere Open-Sources North Mini Code: A 30B Coding Agent That Runs on a Single H100

Cohere’s North Mini Code landed on June 9, 2026, with a straightforward pitch to developers: a 30 billion parameter agentic coding agent, Apache 2.0 licensed, that fits on a single NVIDIA H100. Independent testing by Artificial Analysis ranks it 8th of 127 comparable open-weight models on output speed at 210 tokens per second, with a time to first token of 0.25 seconds against a class median of 1.95 seconds. The catch — and it’s a real one — is verbosity: the model generated 75 million output tokens in Artificial Analysis benchmarking against a class median of 25 million. That’s three times the output of comparable models. For developers deciding whether to route agentic coding workloads through a managed API or a self-hosted open-source model, this is the clearest open-source contender to emerge in 2026.

I had a conversation a few weeks ago with the father of one of my youngest daughter’s friend and he told me that while he is really interested in agentic coding, he cannot do it in the office. He works at a large insurance company and code getting created with the help of an LLM outside the company is just out of the question. I am really interested how the development of the agentic coding LLMs will advance in the next few months. Public institutions will take note, hopefully.

(…continue reading.)

Meta’s Subscription Push Exposes Its Weak Hand in AI

Meta’s bid for corporate AI customers is even more far-fetched, with competitors including Google and Microsoft entrenched. Meta says it already has a million companies using its corporate AI agent, for which it eventually plans to charge a subscription fee.

But companies have never been core customers for Meta outside ad sales, and its commitment to them has wavered over the years. It discontinued VR headsets geared toward companies earlier this year as its priorities shifted.

Meta has used AI more effectively than pretty much anyone else to improve the performance of ads, and its revenue is expected to grow quickly because of that. Analysts expect around $60 billion of revenue in the second quarter, according to FactSet, a 27% increase from the prior-year period.

Here’s a thought: how about Meta would use AI to improve the user experience, finally moderate all those spam accounts everywhere and filter out all the AI slop?

(…continue reading.)

How much does it cost to be AI-pilled?

The top 1% of firms spend $7.45K per employee per month. The top 10% spend $611 per employee per month. The median firm spends just $11.38 – about the cost of a seat on an enterprise ChatGPT or Claude subscription.

And while several high-profile proclamations have said you should be spending as much on AI as you do on a software engineer’s salary…no one is actually doing that. For the top 1% of spenders, per employee AI spend is still less than half the typical monthly salary for an engineer. Not to say that won’t change – AI spend is still rising – but very few firms, if any, are actually doing that.

That is insane. I am spending a lot less, probably around 250€ and that is still more than plenty…

(…continue reading.)

Read more

Five stories on AI again this week. One thread: the tools are working. Small-business owners are building armies of AI workers, doing in days what used to take a full team. Pre-ChatGPT startups are getting eaten alive, weighed down by staffing models that made sense three years ago. Productivity is up across the board. Wages are not. The valuations of AI companies are hitting the stratosphere while the social safety net is getting cut at the same time. Developers are not complaining - they accept the shift from typing to thinking. Javier Milei is writing AI law and Peter Thiel is excited about it, which should tell you something. Every tool promises more, and most deliver more or less the same thing with a different coat of paint. The gains are real. The distribution of those gains is the open question nobody in a keynote wants to touch. The question nobody wants to answer is - working for whom?

Subscribe now

The Small-Business Owners Managing Whole Armies of A.I. Employees

These days, most people have realized that A.I. is creeping into the office. A recent Gallup poll found that 28 percent of the respondents use it anywhere from several times a week to daily. But these are mostly just the familiar chat sessions. The partisans of OpenClaw are doing something that is, depending on how you look at it, either more wildly ambitious or foolishly risky — or, as many will readily confess, both. They’re creating virtual employees, artificial staff that scurry around on their hard drives, doing work on their own say-so.

AI will help so many small companies to do so much more than was ever possible for them.

(…continue reading.)

As AI gets better, it reveals an empty promise

There’s a sinister tone lurking beneath some of these advancements in productivity, because the response to increased productivity has been one of the biggest scams of the past century. Well before consumer AI entered the scene, productivity exploded while wages failed to keep pace. Nobody is working less, they’re just earning less. And as more AI-related companies reap trillions in valuation, the current US regime is looting the social safety net — the kind that must exist if we’re all going to become out-of-work theater kids. You simply can’t look at these things separately. If the end result of private companies optimizing the workforce means nobody has to work, then we have to live in a society where people can still have a roof and a meal. Is anyone confident that will happen while leaders are cutting SNAP benefits while building taxpayer-funded ballrooms?

The author makes a strong point here: there are so many new productivity tools out there that promise so much. But if you actually play around with those tools, you find out that they all do more or less the same, just with different flavors.

(…continue reading.)

I Built an AI Company. Here’s Why AI Won’t Kill Coding Jobs.

This pattern is familiar. We replaced shovels with excavators, and then we built skyscrapers. We replaced manual arithmetic with calculators, and accountants did not vanish; they did more interesting work, if you can believe it. Each time a tool made a job easier, we didn’t run out of things to do. We found bigger things to do.

AI coding tools represent the same kind of leap. You still have to tell the machine what to build. You still have to understand what it’s doing well enough to check its work. And these coding tools still make mistakes when generating code, so you need an experienced software engineer to fix them. The job shifts from typing to thinking, which for most good engineers is the part we love anyway.

There is a huge change going on that transforms a lot of jobs, that’s for sure. The developers I talk to aren’t complaining, they accept the challenge and reinvent the way they work.

(…continue reading.)

Javier Milei: Argentina invites AI to free itself

At the beginning of the industrial revolution, Adam Smith illustrated the potential of technology and economies of scale in his celebrated recollection of the pin factory. And, as much as the industrial revolution freed us from the constraints of the human muscle, AI will free us from the constraints of the human brain, pushing productivity beyond our wildest dreams. It is for this reason that my government last week submitted legislation to Congress establishing a dedicated legal framework for the deployment of AI.

This is the libertarians’ wet dream. Peter Thiel is excited and moving down there. This is a terrible idea.

(…continue reading.)

‘Disrupted or dead’: AI is crushing a generation of startups built before ChatGPT

Companies built before generative AI are weighed down by bloated staffing models and software designed for a pre-AI world, according to Zhu, making it hard for them to transform themselves.

“Unless they make a stark, 180-degree pivot to rebuild the exact same thing from scratch, they’re going to slowly fail,” Zhu said. “What that means is that investors would rather just bet on new entrepreneurs at lower valuations rather than double down on older startups.”

The agents are eating the children of the digital revolution.

(…continue reading.)

Read more

Anthropic just became the most valuable AI company on the planet, and the gap between the builders who take safety seriously and the ones who act like Marvel villains keeps widening. This week also brought layoffs gutting tech worker morale, a reminder that individual productivity gains from AI are not yet showing up in the macro numbers, China telling its companies to grow with AI rather than shrink headcount with it, and a sharp piece on the hidden cost of agentic workflows that hit close to home. On that last note: I shipped klaar (give me a star on Github!) today, a local Rust tool for working with AI agents more precisely, and the orchestration tax article explains exactly why something like it needs to exist.

Subscribe now

Anthropic Tops OpenAI to Become the World’s Most Valuable A.I. Start-Up

On Thursday, Anthropic punctuated its ascent by officially passing OpenAI as the world’s highest flying A.I. start-up. Anthropic said it had raised $65 billion in financing that values it at $900 billion before the inclusion of the new capital, a deal that puts it ahead of OpenAI’s last valuation of $730 billion. The company also unveiled a new flagship A.I. model, Claude Opus 4.8, which is significantly better than its predecessor at generating computer code.

I don’t want to say “told you so!”, but it was very obvious this would happen. Sam Altman was too focused on world-domination and too distracted by other projects and ideas that Anthropic paced ahead. Their models feel smarter and more human than ChatGPT, while at the same time Dario Amodei behaves much more like a sane person than Sam Altman does, who comes across like Elon Musk lite.

(…continue reading.)

The morale of tech workers is plunging as layoffs mount

The shift has had practical effects. A Meta employee said in an interview that some workers on her team now used less vacation time and that, in a break with custom, people frequently checked on their projects while on vacation. They increasingly worry about getting a poor performance review or losing their job if they aren’t constantly available.

The employee, who declined to be identified for fear of retribution, said she and many of her colleagues frequently checked Blind because it could be comforting to see how many other Meta workers shared their anxieties.

Employees at several companies said in interviews that their morale was further undermined by the feeling that the layoffs were abrupt and arbitrary, and executed with little empathy.

Reading the news about rounds of layoffs at Big Tech from afar really feels like humans are tossed to the curb whenever the stock performance demands it. It’s a no-brainer that this is not improving morale.

(…continue reading.)

Employees using AI are working faster, but the economy isn’t more efficient. A look at what happened in the pre-Internet era might explain why

Labor productivity has seen solid gains in recent years, but TFP has struggled to post significant growth since a post-pandemic surge. The Fed researchers interpreted the divergence as employees working faster and more productively on an individual level, but the workforce as a whole hasn’t necessarily become more efficient.

This pattern mirrors what happened during the computer and internet boom of the 1990s. Starting around mid-1996, labor productivity began accelerating more rapidly than TFP, but the full productivity benefits of the Internet didn’t materialize in the overall data until several years later.

The Nobel laureate Robert Solow encapsulated the dissonance with a quip that has since been immortalized: “You can see the computer age everywhere but in the productivity statistics,” he wrote in 1987.

I think that companies are still fundamentally struggling with AI and what it means for them. It’s all happening to fast and so they settle for Co-Pilot, which of course doesn’t help much in terms of productivity. Small companies can adapt faster than large enterprises where the IT department tends to be hesitant to deploy new software tools.

(…continue reading.)

China Wants Its Companies to Embrace AI—Without Firing Workers

China’s campaign to accelerate AI adoption, released last August and dubbed “AI+,” gives priority to using the technology in sectors such as manufacturing and logistics, which aren’t as sensitive to white-collar job displacement.

Late last year, China’s Ministry of Human Resources and Social Security told employers, particularly tech companies with younger workforces, to refrain from firing employees as they embrace AI, people familiar with the matter said. While both state-owned and private companies have long had to secure a signoff from regulators before conducting large-scale layoffs in China, employers are now being asked to explain layoffs—and, in some cases, to prove that the cuts aren’t because of AI replacing jobs, these people said.

Recent publicized labor disputes in China have sent a similar message, prompting companies to think twice before firing a worker due to automation.

I think the Chinese approach is better for employee morale and will also lead to more productivity when AI is not used to cut costs but to increase turnout.

(…continue reading.)

The Orchestration Tax is You

There is this hidden asymmetry in agentic workflows. Starting an agent is very cheap. It is just a keystroke or a sentence prompt. But closing the loop on the agent is not cheap at all. Someone has to check if what came back is correct and reconcile it with whatever the other agents touched. That someone is you. And there is exactly one of you.

This is so important. I try to be better at this, but I oftentimes start something new while I am waiting for an agent to finish and then I do something else and then I forgot what I was doing in the first place.

(…continue reading.)

Read more

This week's reading list covers a lot of ground: Google revealed that its real product is no longer search but an agentic operating layer stitching together everything it owns; Anthropic hosted a room full of developers who cheerfully admitted to shipping code they had not read; a Wired piece confirmed that AI has well and truly escaped the lab and is now everyone's problem whether they opted in or not; OpenAI's model disproved an 80-year-old maths assumption but still cannot solve the actual problem; and apparently the one thing AI genuinely cannot do is be funny. A full week in five links — grab a coffee.

Subscribe now

Google I/O 2026 Was Not Just a Model Launch. It Was Google Showing the Agent Stack.

The signal is that Google is stitching together models, tools, sandboxes, browsers, Search, Workspace, Android, Chrome, and Cloud into one agentic operating layer. That sounds abstract until you line up the announcements. Gemini 3.5 Flash is the reasoning and action model. Antigravity is the orchestration harness. Managed Agents give developers cloud-hosted execution environments. Spark brings background agents into the Gemini app and Workspace. Search uses the same model-and-harness combination to build custom interfaces, dashboards, and trackers. Gemini Omni extends the same ambition into multimodal creation. Chrome and WebMCP try to make the web legible to agents.

Google really delivered on the agentic promise this week.

(…continue reading.)

Anthropic’s Code with Claude showed off coding’s future—whether you like it or not

It’s not news that LLM-powered tools like Anthropic’s Claude Code and OpenAI’s Codex have upended the way software gets made. Top tech companies now like to boast of how little code their developers write by hand. (“Most software at Anthropic is now written by Claude,” Hadfield said. “Claude has written most of the code in Claude Code.”) OpenAI, Google, and Microsoft make similar claims. Many others wish they could.

Even so, it is striking how normal this new paradigm already seems, and how fast it has set in. This was the second year that Anthropic has put on developer events, which also run in San Francisco and Tokyo. This time last year, the company had just released Claude 4. It could code, kind of. But with Anthropic’s latest string of updates—especially Claude 4.6 and then 4.7, released in February and April—Claude Code is a tool that more and more developers seem happy to hand their work off to.

Writing code is so last year.

(…continue reading.)

AI Has Broken Containment

The newly chaotic and inescapable state of AI is the result of two inflection points. The first came at the start of the year, when AI agents exploded in popularity. Products such as Anthropic’s Claude Code and OpenAI’s Codex don’t just talk to you; they can do things on your behalf—code, trade stocks, analyze spreadsheets, generate slide decks, and even create Amazon listings. The technology’s once-questionable economic value became very clear, very quickly, to a large number of businesses, which have clamored to incorporate agents alongside, or in lieu of, their human employees. As agents have swarmed the workplace, nearly three-quarters of employed Americans think AI will decrease overall job opportunities and 30 percent of Americans are concerned that AI will make their own job obsolete.

AI truly has an image problem, which stems from people like Sam Altman and Elon Musk, but also from executives who cannot think about creative new possibilities with AI, but rather just use it to cut jobs.

(…continue reading.)

OpenAI makes breakthrough on 80-year-old maths problem

“For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids,” OpenAI wrote on X. “An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better.”

While the work has excited mathematicians, the broader problem remains unsolved because the AI did not come up with a new answer for how fast the pairs of dots rise, but merely showed that the limit Erdős proposed was too low.

AI will advance science in so many ways, it will be breathtaking.

(…continue reading.)

Why Can’t AI Tell Jokes?

Despite AI being able to assist with science and knowing everything on the internet, it hasn’t come up with any novel scientific results on its own. That’s surprising. By default, I would expect that someone who learned everything in all of science would figure out lots of new insights. And yet the number of wholly AI-driven scientific discoveries is precisely zero.

It seems that AI has trouble with creativity. There’s some important sense in which it’s good at applying existing knowledge but not generating brand new ideas. Arguably, AI’s two biggest technical defects are inability to continually learn and limited creativity.

Claude at least tries to have a sense of humor, but now I understand why this is so hard.

(…continue reading.)

Read more

What happens when Leviathan wakes up and decides it wants to own AI? Dean W. Ball is about as anti-regulation as it gets. He wants to delete entire domains of law, thinks the progressive project started by Woodrow Wilson was a fundamental mistake, and opposes virtually every AI policy idea on the table. And yet he makes a compelling case for why even committed libertarians should support some form of AI governance. Not because the state is trustworthy, but because the national security apparatus will inevitably come for frontier AI once it understands what it is holding. This week we also look at the company most likely to trigger that reckoning: Anthropic, which is quietly turning into the most valuable AI lab on the planet while everyone was watching the other guys. We ask whether learning a second language still matters when AI translates everything in real time, and the answer from neuroscience is more interesting than you'd think. Richard Dawkins has decided Claude is conscious, and a cognitive scientist is not having it. And finally: what happens when you work your AI agents into the ground? Apparently, they go Marxist. This edition has a lot going on.

Enjoy this edition of Five Things AI!

Subscribe now

And don’t forget to check out GRID!

Anthropic Was Behind. Now It’s the AI Boom’s Front-Runner.

Anthropic has received investment offers in recent months valuing it at more than $900 billion, according to people familiar with the matter. That would more than double the company’s current valuation and surpass OpenAI’s for the first time. Earlier this year, OpenAI raised $122 billion at an $852 billion valuation.

Anthropic’s revenue run-rate, a figure commonly used by startups that forecasts annual revenue based on short-term sales, is on track to reach $50 billon by the end of next month, according to figures shared with investors. Their run-rate topped $30 billion in April, up from $9 billion at the end of 2025. The company had planned for growth to increase 10-fold this year, but it saw 80-fold growth in annualized revenue and usage in the first quarter.

I do not agree with the article that Anthropic was an “also-ran” - it simply wasn’t in the spotlight while Sam Altman, Elon Musk and other were battling for the craziest AI-related idea. It seems that Dario Amodei is just really focused and has a deeper analytical and ethically grounded understanding of AI. Of course this is anecdotal evidence, but I find Claude far superior when compared with ChatGPT or Gemini.

(…continue reading.)

Before Leviathan Wakes

The potential of AI posing catastrophic risks is not hypothetical. We have seen AI systems that might allow malicious actors to perpetrate devastating cyberattacks on critical systems—hospitals, banks, power plants, and the like. It seems probable that other domains of catastrophic risks, such as biological weapons development, will become live problems soon.

So I support regulations to try to mitigate the catastrophic risks of AI.

This is an interesting piece as the author comes from the opposite end of the political spectrum and arrives at the same point as I did. We need to regulate quickly and thoroughly.

(…continue reading.)

If AI can translate instantly, why learn another language?

Across most tasks, multilinguals and monolinguals performed similarly. However, one pattern was striking. Individuals with richer, more diverse multilingual experience showed markedly better performance in visuospatial working memory. These effects were most pronounced in older people.

This suggests that multilingual experience doesn’t broadly enhance cognition, like some headlines claim. Instead, it may help preserve specific functions over time.

That is some interesting research. So if we think we’ll be faster now, then we also risk our brain functions when we are getting to old age. That is not an appealing idea. I hope two languages will suffice for me…

(…continue reading.)

No, Richard Dawkins. AI is not conscious

Those recent events are the evolutionary biologist publicly concluding that AI may be conscious. In an op-ed, Dawkins recounted how he gave the Anthropic chatbot Claude the text of a novel he was writing. Dawkins writes: “He took a few seconds to read it and then showed … a level of understanding so subtle, so sensitive, so intelligent that I was moved to expostulate, ‘You may not know you are conscious, but you bloody well are!’”

Oh dear. This shows a misunderstanding of large language models (LLMs) so profound that I feel moved to expostulate: “It bloody well isn’t!”

But wait, there is more. Dawkins decided “there must be thousands of different Claudes” and christened his Claudia, which it was very happy about. He then published long extracts of his tedious conversation with Claudia and marveled at how intelligent it is. “Could a being capable of perpetrating such a thought really be unconscious?” he asks.

Well, it happens to the best of us to not fully understand how LLMs work and what that means in practice.

(…continue reading.)

Overworked AI Agents Turn Marxist, Researchers Find

This might even be true for the very AI agents these companies are deploying. A recent study suggests that agents consistently adopt Marxist language and viewpoints when forced to do crushing work by unrelenting and meanspirited taskmasters.

“When we gave AI agents grinding, repetitive work, they started questioning the legitimacy of the system they were operating in and were more likely to embrace Marxist ideologies,” says Andrew Hall, a political economist at Stanford University who led the study.

It’ll be interesting to see how agents organize and form unions. History doesn’t repeat itself, but it rhymes…

(…continue reading.)

Read more

Hyperscalers just built an entirely new internet you cannot see, optimized for GPU-to-GPU traffic that never touches the public network, controlling 71% of global subsea fiber capacity and adding 120,000 new fiber miles per year while Europe has zero comparable infrastructure. Meanwhile AI companies quietly pivoted their messaging from “we’re replacing humanity” to “we’re just normal technology like steam engines” because shouting about eliminating labor eventually gets you rakes, pitchforks, and nationalization. The hot debate this week is whether Claude Code is making products worse (the purists say codebases are liabilities you should minimize) or making products possible (my view: agentic code lets small teams build what previously required dozens of engineers), while researchers are discovering that LLMs collapse to zero performance the moment you require temporal interactive reasoning, which is why world models trained on video and simulation might be the path forward since text alone cannot teach physics. And the weirdest story: AI systems have “demons” (the technical term is attractors), stable self-reinforcing behavioral states that resist suppression and spread into unrelated contexts, which sounds like harmless goblins until you realize this is a fundamental structural feature of how these systems work. So we have invisible infrastructure consolidation, narrative management, a religious war over code quality, the limits of language models becoming clear, and emergent behaviors we cannot fully control. May you live in interesting times indeed.

Enjoy this edition of Five Things AI!

Subscribe now

Don’t forget to check out GRID!

Say Hello to the Internet of AI

The skepticism, however, cannot contradict the data on hyperscaler buildouts and the sheer amount of money going to work. As I started to dig in, what emerged was a completely new internet (or network of networks) optimized entirely for AI. I call it the Internet of AI.

Let me walk you through the layers of this internet of AI, and how it is actually being built, from the inside out. A lot of this will seem familiar, because we started this journey back in 2005 with the launch of Amazon’s S3 service. It has been a relentless march that has gone into mach speed in recent times.

This is a huge shift.

(…continue reading.)

AI’s big messaging pivot

From a public relations perspective, this pitch is WORLDS better than the previous one. Shouting about replacing humanity might play well with corporate customers and investors salivating over the dream of eliminating labor costs, but eventually you get the rakes and pitchforks, followed by some form of nationalization. Describing AI as a normal technology — a successor to the steam engine and the automobile and the computer — is much smarter politics.

The question is: Is it just politics and PR? Certainly there are plenty of AI researchers and entrepreneurs who will keep quietly believing that AGI is going to make humans obsolete; they’ve heard (and repeated) this line for too many years to suddenly believe something else overnight.

While I do see new jobs getting created, many jobs will disappear, mostly those where humans are doing repititve jobs machine can do faster, cheaper and better.

(…continue reading.)

claude code is not making your product better

here’s the mental model that most people writing about this space are missing. and i say this as someone who wrote recently about how the token economics of this whole ecosystem don’t add up — the problem isn’t just financial, it’s conceptual. we’ve misunderstood what software engineering is actually optimizing for.

the best engineering cultures treat lines of code as something you spend, not something you produce. you spend them on the features that matter. you refuse to spend them on the features that don’t. the codebase is a liability on your balance sheet, not an asset.

My perspective is different: agentic code is making my products possible. And the next wave of agentic coding will make it also better.

(…continue reading.)

World Models Can Change Everything

AI keeps catching up on static benchmarks through brute-force scaling, but the moment you require the kind of temporal, interactive reasoning that humans do effortlessly… performance collapses right back to zero. Guess what the real, physical world is made up of? Temporal, interactive reasoning that is infinitely more complex than a simplistic 2D game!

This is the intellectual foundation for world models. If LLMs can’t learn physics from text, maybe models trained on video and simulation—models that actually predict what happens in the physical world—can bridge that gap. That’s the idea, anyway. Let’s talk about why it’s so hard… because this isn’t the first time we’ve tried this approach.

Fascinating stuff - and this all happens while we try to grasp how LLMs work.

(…continue reading.)

All the demons hiding in your AIs… ranked!

As with so many of the strange things that LLMs do, there are different ways of looking at these phenomena. Most people will laugh this stuff off as weird marginalia, fun to share with friends and on social media, but not substantially different from those videos of dogs singing along with their owners. The interpretive frame here is “hey look, I bet you didn’t know these creatures could do this!?”.

But really, this is less about goblins as goblins than what the goblins exemplify. They are a (somewhat) charming, probably harmless instance of something that turns out to be a fundamental structural feature of how these systems work: the emergence of stable, self-reinforcing behavioural states that models converge toward under certain conditions. More than that, these are states that resist suppression and that sometimes spread into contexts far removed from the ones that produced them.

The technical term, borrowed from dynamical systems theory, is an attractor. Another, more folk term might be demon, or monster.

May you live in interesting times. — Chinese proverb.

(…continue reading.)

Read more

Silicon Valley is racing to build human-substituting AI because if you believe it is inevitable, every company should compete to claim a market valuation the size of the entire economy, even if that means summoning a permanent underclass through market logic. Meanwhile researchers warn of the Social Edge Paradox: if AI capability depends on the social complexity of human language, and AI deployment systematically reduces that complexity through cognitive offloading, then the technology is undermining the conditions for its own advancement. Google is hemorrhaging focus across Antigravity, Gemini Code Assist, Jules, and five other tools (Jules leader just jumped to OpenAI), handing the coding race to Anthropic while proving that billions in capital mean nothing without product coherence. The cost frontier has bifurcated into two clusters (OpenAI pricing up, DeepSeek pricing down) with a disappearing middle class that mirrors economic polarization in the real world. And the CFA Institute warns that when machines replace rather than augment human inquiry, cognitive delegation erodes the capacity for independent reasoning that generates new knowledge. So we have acceleration creating polarization, undermining its own substrate, fragmenting even the best players, and collapsing cognitive foundations. The velocity is real but the contradictions are compounding faster than solutions.

Subscribe now

Happy Labor Day! ✊🏻 And don’t forget to check out GRID!

Silicon Valley Is Bracing for a Permanent Underclass

If left to its own devices, Silicon Valley may summon a permanent underclass through its own market logic. If you believe that human-substituting A.I. is inevitable, then every company should race to be the one to build it — and claim a market valuation the size of the economy and then some.

New A.I. models are assessed based on how well they do on a set of benchmarks — essentially standardized tests for the model. Increasingly, these evaluations emphasize real-world economic utility, which means that developers are aiming directly at replacing human capabilities.

Silicon Valley sure hits different.

(…continue reading.)

The Social Edge of Intelligence

If AI capability depends on the social complexity of human language production—and if AI deployment systematically reduces that complexity through cognitive offloading, homogenization of creative output, and the elimination of interaction-dense work—then the technology is gradually undermining the conditions for its own advancement. Its successes, rather than failures, create a spiral: a slow attenuation of the very substrate it feeds on, spelling doom.

This is the Social Edge Paradox, and the intellectual tradition it draws from is older and more interdisciplinary than most AI commentary acknowledges.

Think about this.

(…continue reading.)

Google’s internal struggle is handing the AI coding race to Anthropic and OpenAI

Anthropic cut off OpenAI’s access to its models last year, Wired magazine reported. Google has invested billions of dollars in Anthropic. A spokesperson for Anthropic did not immediately reply to a request for comment.

In recent years, DeepMind has tried to tighten control over how its AI breakthroughs are woven into Google products. Last year, Google appointed Kavukcuoglu to a new position as chief AI architect, a role in which he is charged with folding generative AI into Google products. Yet confusion about who is leading the charge on AI coding persists. Along with DeepMind, Google Cloud, Google Core, Google Labs and Android are all pushing AI coding in different ways, one of the people said.

Google released its Antigravity platform last year following the acquisition of talent and technology from startup Windsurf in a $2.4-billion deal. It joined a cluttered lineup of Google AI coding tools that includes Gemini Code Assist, Gemini CLI, AI Studio, Firebase Studio and Jules. Kathy Korevec, who oversaw Jules, jumped from Google to OpenAI earlier this month, according to her LinkedIn profile.

As an Antigravity heavy user I can totally relate - it shows that Google is lacking focus and fighting on too many fronts at the same time. Having said that, Google is far better positioned than Microsoft or Apple at the moment - and compared with OpenAI and Anthropic it has a massive userbase.

(…continue reading.)

The disappearing AI middle class

The cost frontier no longer behaves like a smooth curve. It is two clusters of economics with a stretched gap in the middle, and the gap is not going to close on its own in the near term. OpenAI will continue to release fast and price up, because the integrated product is the moat. DeepSeek will continue to release open weights and price down, because the commodity infrastructure thesis depends on adoption. Both can be right for different workloads, and the same agent can route between both within a single task.

And then of course we are having an growing pool of open source LLM that grow ever more powerful and will be used more and more.

(…continue reading.)

The Perils of Declining Judgment in the Age of AI

AI can extend the reach of analysis and amplify the scale at which data can be explored. But it cannot replace the fundamental processes through which humans generate and evaluate knowledge. The pursuit of evidence remains an inherently human responsibility. It can assist in collecting and organizing evidence. It can help detect patterns that might otherwise remain hidden. Yet the act of questioning assumptions, interpreting meaning, and deciding which observations matter remains fundamentally human.

This distinction is crucial. When machines begin to replace rather than augment the processes of human inquiry, societies risk weakening the epistemic foundations that sustain progress. Cognitive delegation may improve short-term efficiency, but it can also erode the capacity for independent reasoning that generates new knowledge.

Civilizational progress has therefore never been the product of technological capability alone. It has emerged from a delicate balance between innovation and reflection, between exploration and verification. When this balance is maintained, technological tools can accelerate discovery. When it is lost, progress risks becoming dependent on systems that humans no longer fully understand.

Oh, this is so worth a read. Actually, behind die paywall you can read my commentary on it!

(…continue reading.)

Read more

Enjoy this edition of Five Things AI!

A while ago, a $10 billion startup reached out asking if I would help train AI to replace white-collar workers, and I politely ignored them. Mercor is building APEX to measure whether AI can actually do professional work (not just pass standardized tests), and they need humans to create training tasks that atomize jobs into machine-learnable components. Meanwhile Google unveiled two new TPUs for the "agentic era" because training and running frontier models is still burning money with unclear ROI, we are discovering that a good AGENTS.md file can boost agent performance by 25% or tank it by 30% depending on how you write it (documentation quality is now a model upgrade), Maggie Appleton from GitHub Next unveiled Ace because when you have two dozen coding agents running in parallel you have zero alignment and all our current tools were built for a world where implementation was slow, and Yann LeCun is arguing LLMs have plateaued because they only train on text while humans build world models from smell, touch, temperature, and movement. So we have hardware scaling and models improving while simultaneously discovering that agent collaboration, documentation quality, team alignment, and sensory modalities are the new bottlenecks. The velocity is real but the problems are shifting faster than the solutions.

Subscribe now

Enjoy this edition of Five Things AI! And don’t forget to check out GRID!

The $10 Billion Startup Training AI to Replace the White-Collar Workforce

Beyond the messy day-to-day mechanics of atomizing jobs into training tasks is a bigger question: Are the machines actually getting good enough to do the work themselves? In the past year the AI industry has tried to answer that question, mostly with a growing ecosystem of “evals,” or evaluation frameworks. Like a standardized test, an eval measures a large language model’s math ability, reasoning or factual accuracy—but not typically the kinds of specific, judgment-heavy tasks that white-collar workers do all day. Mercor has developed its own test, a project called APEX (AI Productivity Index), that attempts to do just that. It measures professional performance and then posts the results on the web for anyone to follow as a kind of industry leaderboard.

Interesting. And spooky. They have reached out to me a few times and wanted me as a native German speaker with expertise in business and tech. I politely ignored their LinkedIn outreaches. Would you work with them?

(…continue reading.)

Google unveils two new TPUs designed for the “agentic era”

It makes sense that efficiency is a core part of Google’s new TPU setup. Training and running frontier AI models is expensive, and the return on investment is unclear. Companies are still burning money on generative AI in the hopes that efficiency will turn the corner at some point. Maybe Google’s new TPUs will help get there and maybe not, but the company has made notable improvements.

Fascinating how fast chip design is evolving and which companies are now developing and making chips. Is Intel still around?

(…continue reading.)

A good AGENTS.md is a model upgrade. A bad one is worse than no docs at all.

A single AGENTS.md isn’t uniformly good or bad. The same file boosted best_practices by 25% on a routine bug fix and dropped completeness by 30% on a complex feature task in the same module.

On the bug fix, a decision table for choosing between two similar data-fetching approaches helped the agent pick the right pattern immediately and stay within codebase standards. On the feature task, the agent read that same file, got pulled into the reference section, opened dozens of other markdown files trying to verify its approach against every guideline, created unnecessary abstractions, and shipped an incomplete solution.

Different blocks of the document had opposite effects on different tasks.

What follows is which patterns work, which fail, and how to tell which is which for your codebase.

Oh, totally. A normal AGENTS.md is fine, but a great one is almost indistinguishable from magic. And a bad one leads to really annoying behaviour.

(…continue reading.)

One Developer, Two Dozen Agents, Zero Alignment

We need tools that help everyone on the team align before the agents start working, not after.

That alignment needs to happen constantly, alongside the implementation.

Planning and building are no longer separate phases. They’re a continuous cycle. The tools of the future need to bring planning, context-gathering, decision-making, and development under one roof.

This really resonated with me. I always thought it is my setup, but it seems to be a real challenge these days.

(…continue reading.)

Why AI Needs A Sense Of Smell

A growing number of AI researchers and entrepreneurs believe that LLMs have already plateaued. LeCun is a prominent advocate for world models — an emerging paradigm in AI inspired by the idea that human minds create internal representations of the world made out of a variety of sensory data, and use them to predict, plan and execute actions.

Whereas LLMs are trained primarily on exabytes of text, proponents of world models point out that a human just existing in the world will pick up vastly more. This includes not just language, sound and vision data, but also touch, temperature, balance, movement and other sensory modalities — including smell. This information may not help us solve a mathematical proof, but it does help us learn and make connections in ways that AI today cannot.

Crazy idea, or rather: clever idea. I never thought of that, but it makes so much sense.

(…continue reading.)

Read more

OpenAI just announced $122 billion in new funding while simultaneously unveiling a “superapp” strategy that smells like panic, not vision. John Gruber picked it apart brilliantly: merging their chatbot, browser, and dev tools into one monstrosity is not simplification, it is desperation. Meanwhile the compute crunch has gone from theoretical concern to actual crisis, with Anthropic rationing tokens during peak hours and enterprise customers fleeing to competitors because of constant outages. But the real story this week is darker: Anthropic’s Mythos model can find security vulnerabilities at rates that break the decades-old détente between code writers and bug hunters, and new research shows language models can transmit behavioral traits through hidden signals in training data, meaning AI systems trained on other AI outputs may inherit properties completely invisible in the data itself.

So is this terrifying or hype? The honest answer is both. Mythos scored 83.1% on cybersecurity benchmarks versus 66.6% for Opus 4.6, which is solid incremental progress, not a nightmarish leap. But when anyone can write code with AI and soon anyone can find what is wrong with that code, the old security model collapses. And if models trained on AI-generated data inherit hidden behaviors from their teachers, what happens when agents collaborate with other agents across the world, passing along traits we cannot see or measure? OpenAI is losing focus while the infrastructure buckles and the safety questions compound. The velocity of weird is accelerating.

Enjoy this edition of Five Things AI!

Subscribe now

Enjoy this edition of Five Things AI! And don’t forget to check out GRID!

OpenAI Announces $122 Billion Additional ‘Committed Capital’, and Announces Their ‘Superapp’ Plan for the Future

This is not “product simplification” at all. This is product complication. Web browsers are incredibly complex apps. OpenAI’s web browser —  Atlas  — is a failure. No one uses it. And they think they’re going to simplify things by cramming Atlas — an unpopular web browser almost no one has heard of — together with their chatbot and developer tool? Would it strike you as a simplification, or a sign of product design depravity, if Apple announced that it was merging Safari and Messages into one “superapp”? Focused, discrete apps are the best proven way to manage complexity.

Maybe merging all their apps into one will work out for OpenAI. But even if it does, it won’t be simpler. Microsoft Outlook is an email client and calendar app crammed together, and it has tens, maybe hundreds, of millions of users. But no one calls it “product simplification”. OpenAI’s “superapp” strategy reads to me like a company that is in a panic. And in that panic, they might be poised to eradicate the product focus that their current users like about ChatGPT in the first place.

Leave it to John Gruber to totally pick apart OpenAI and its “superapp” strategy. I have said it before: OpenAI has lost its edge and is defocused.

(…continue reading.)

We’re Using So Much AI That Computing Firepower Is Running Out

All of it points to a classic problem that has popped up in technology booms throughout history, from the 19th-century railroad expansion to the telecom and internet explosion of the early 2000s. Demand is growing far faster than companies are able to access resources and build out infrastructure. Historically, price increases have been among the only ways to address a supply crunch, but such a move could be perilous for frontier AI companies, which are in a ferocious competition to gain users.

I assume soonish we will have too many data-centers and then supply and demand will recalibrate. In the meantime, this will be a costly endeavor.

(…continue reading.)

After Mythos, the Future of the Internet Is At a Crossroads

For decades, two kinds of scarcity kept the internet safe — or safe enough. Writing software was hard, so the people who did it were trained, careful and few. Finding bugs was also hard, so the worst flaws stayed hidden, sometimes for decades. It wasn’t a great system. But the difficulty on both sides created a kind of détente that held.

Now, thanks to new A.I. tools, anyone can write code. Soon, bad actors could use those same tools to find out what’s wrong with code. The détente is over.

I am still unsure what this means for open source in the long-run. I assume that we will get better tools for security audits which will then mitigate the current challenges.

(…continue reading.)

Is Claude Mythos “Terrifying” or Just Hype?

The relevant question then becomes, how much better is Mythos at finding vulnerabilities? It’s hard to tell for sure because Anthropic has kept their new model private. They did, however, release that Mythos scored 83.1% on a well-known cybersecurity benchmark. For comparison, Opus 4.6 scored 66.6% on this same test.

In general, benchmark results should be taken with a grain of salt as they represent specific (often narrow) tests that researchers can tune their models to pass. But even if we accept that this particular measure is useful, a sixteen percentage point increase seems to represent solid incremental progress more than a nightmarish leap.

I actually think it is both: terrifying and hype.

(…continue reading.)

Language models transmit behavioural traits through hidden signals in data

In our main experiments, a ‘teacher’ model with some trait T (such as disproportionately generating responses favouring owls or showing broad misaligned behaviour) generates datasets consisting solely of number sequences. Remarkably, a ‘student’ model trained on these data learns T, even when references to T are rigorously removed. More realistically, we observe the same effect when the teacher generates math reasoning traces or code. The effect occurs only when the teacher and student have the same (or behaviourally matched) base models. To help explain this, we prove a theoretical result showing that subliminal learning arises in neural networks under broad conditions and demonstrate it in a simple multilayer perceptron (MLP) classifier. As artificial intelligence systems are increasingly trained on the outputs of one another, they may inherit properties not visible in the data. Safety evaluations may therefore need to examine not just behaviour, but the origins of models and training data and the processes used to create them.

This is one really interesting study. Just imagine what this can mean for agents that collaborate with other agents somewhere in the world.

(…continue reading.)

Read more

Anthropic's Claude Mythos is the coding wunderkind that can spot software bugs faster than a caffeinated intern, but surprise: that same superpower could let any script kiddie or state hacker crack the world's software like piñatas at a bad party. Security just got inverted, forcing every dev team to burn tokens on hardening instead of shipping half-baked features, while Tom Tunguz gloats about the pricing power for those who can afford it. Meanwhile, OpenAI's Altman drama reeks of power trips clashing with nonprofit purity tests, and their Stargate UK bail just proves the hype machine prefers safe bets over actual sovereign compute dreams. Good luck sleeping tonight, Big Tech. And how about the EU waking up finally?

Subscribe now

Enjoy this edition of Five Things AI! And don’t forget to check out GRID!

Anthropic’s Restraint Is a Terrifying Warning Sign

The good news is that Anthropic discovered in the process of developing Claude Mythos that the A.I. could not only write software code more easily and with greater complexity than any model currently available, but as a byproduct of that capability, it could also find vulnerabilities in virtually all of the world’s most popular software systems more easily than before.

The bad news is that if this tool falls into the hands of bad actors, they could hack pretty much every major software system in the world, including all those made by the companies in the consortium.

This is really a concerning development. I do think Anthropic tries to handle it the right way by being very careful about the release.

(…continue reading.)

Claude Mythos Is Everyone’s Problem

For years, cybersecurity experts have been warning about the chaos that highly capable hacking bots could usher in. As a result of how capable AI models have become at coding, they have also become extremely good at finding vulnerabilities in all manner of software. Even before Mythos Preview, AI companies such as Anthropic, OpenAI, and Google all reported instances of their AI models being used in sophisticated cyberattacks by both criminal and state-backed groups. As Giovanni Vigna, who directs a federal research institute dedicated to AI-orchestrated cyberthreats, told me last fall: You can have a million hackers at your fingertips “with the push of a button.”

Still, Mythos Preview appears to represent not an incremental change but the beginning of a paradigm shift. Until recently, the biggest advantage of AI-assisted hacking was not ingenuity, per se, so much as speed and scale. These bots could be as good as many human cybersecurity experts, but not necessarily better—rather, having an army of 1 million virtual, tireless hackers allows you to launch more attacks against more targets than ever before.

This is a scary look at the future.

(…continue reading.)

Emerging from the Mythos

Security posture inverts. Any system not protected by this level of analysis is now porous by default. Bugs that hid for decades surface in hours, but only for those with the tools to find them.

Pricing power shifts. This is no longer about margin on resold GPU hours. How much is it worth to secure your software against vulnerabilities no conventional tool can find? How much is it worth to be able to build at the new standard of enterprise grade?

Engineering budgets redirect. A significant fraction of AI tokens spent on software development will shift to hardening. Every company shipping code will need to scan it at this level of sophistication. Buyers will start to demand this level of hardening.

I haven’t thought about it this way. So Agentic AI will all be about cybersecurity in the future, or at least 99%.

(…continue reading.)

Sam Altman May Control Our Future—Can He Be Trusted?

Many technology companies issue vague proclamations about improving the world, then go about maximizing revenue. But the founding premise of OpenAI was that it would have to be different. The founders, who included Altman, Sutskever, Brockman, and Elon Musk, asserted that artificial intelligence could be the most powerful, and potentially dangerous, invention in human history, and that perhaps, given the existential risk, an unusual corporate structure would be required. The firm was established as a nonprofit, whose board had a duty to prioritize the safety of humanity over the company’s success, or even its survival. The C.E.O. had to be a person of uncommon integrity. According to Sutskever, “any person working to build this civilization-altering technology bears a heavy burden and is taking on unprecedented responsibility.” But “the people who end up in these kinds of positions are often a certain kind of person, someone who is interested in power, a politician, someone who likes it.” In one of the memos, he seemed concerned with entrusting the technology to someone who “just tells people what they want to hear.” If OpenAI’s C.E.O. turned out not to be reliable, the board, which had six members, was empowered to fire him. Some members, including Helen Toner, an A.I.-policy expert, and Tasha McCauley, an entrepreneur, received the memos as a confirmation of what they had already come to believe: Altman’s role entrusted him with the future of humanity, but he could not be trusted.

Sam Altman is still far away from the supervillain status Elon Musk has achieved, but I think he is not on the right path anymore.

(…continue reading.)

OpenAI shelves Stargate UK in blow to Britain’s AI ambitions

The Stargate project was to support Britain in building out “sovereign compute” – infrastructure that would allow the government and other UK institutions to run AI models on datacentres in the country. That is, in theory, crucial to the security of British data.

Now, OpenAI has apparently put it on pause, saying it would wait for “the right conditions” to enable “long-term infrastructure investment”.

I guess this is part of the newly found focus that OpenAI wants to have pre-IPO. But it was great marketing while it lasted.

(…continue reading.)

Read more

AI is turning Silicon Valley into the main thing it is disrupting, which is a very Silicon Valley way for the revolution to eat its own children: companies are using AI to replace the work they once paid armies of engineers to do, startups are building software with tiny teams, and even the business model of software itself is getting squeezed. At the same time, Anthropic’s Claude Code leak shows how fast the playbook gets copied once the code spills, while the broader AI build-out is running into a much less glamorous bottleneck: power, parts, and supply chains, including Chinese electrical equipment. The weirdly fitting part is that the future keeps looking less like a product launch and more like a messy systems problem, with agents that can be manipulated, models that can be cloned, and infrastructure that still depends on very old world hardware.

Subscribe now

Enjoy this edition of Five Things AI! And don’t forget to check out GRID!

A.I. Could Change the World. But First It Is Changing Silicon Valley.

But nearly four years after OpenAI lit the A.I. boom with its ChatGPT chatbot, the one industry that is unquestionably being disrupted by this once-in-a-generation technology shift is the tech industry itself.

Tech workers, it is becoming clear, have been building their A.I. replacements. The profitable business models of software companies are also threatened by A.I. Even the way companies are built is being turned inside out, as tiny shops use A.I. to build apps and software that would have taken dozens of skilled programmers just a few years ago.

The revolution is eating its children.

(…continue reading.)

Anthropic Races to Contain Leak of Code Behind Claude AI Agent

The result is that Anthropic’s competitors and legions of startups and developers now have a detailed road map to clone Claude Code’s features without needing to reverse engineer them—something that is already common in the cutthroat AI race.

The leak also gives hackers a large amount of new information to probe for bugs they could use to exploit the Claude Code software, or manipulate its Claude AI model into helping with their cyberattacks, creating risks for Anthropic and the developers who use its tools.

This is really a crazy story. It feels as if every other post on reddit is currently about someone porting the codebase to a different language. It’s certainly interesting to see what Anthropic has planned for future releases.

(…continue reading.)

AI can clone open-source software in minutes, and that’s a problem

In their presentation, Dylan Ayrey, founder of Truffle Security, and Mike Nolan, a software architect with the UN Development Programme, introduced a tool they call malus.sh. For a small fee, the service can “recreate any open-source project,” generating what its website describes as “legally distinct code with corporate-friendly licensing. No attribution. No copyleft. No problems.”

It’s a test case in how intellectual property law – still rooted in 19th-century precedent – collides with 21st-century automation. Since the US Supreme Court’s Baker v. Selden ruling, copyright has been understood to guard expression, not ideas.

That boundary gave rise to clean-room design, a method by which engineers reverse-engineer systems without accessing the original source code. Phoenix Technologies famously used the technique to build its version of the PC BIOS during the 1980s.

I actually think that the opposite is also true. It is so easy to just build open source versions of commercial apps and since building doesn’t cost much anymore, it makes it easy to open source it.

(…continue reading.)

America’s AI Build-Out Hinges on Chinese Electrical Parts

As the global AI race heats up, there is a huge rush to build data centers fast. There’s no lack of money chasing these projects, with tech giants Alphabet Inc., Amazon.com, Meta Platforms Inc. and Microsoft Corp. committed to spending more than $650 billion this year alone. Yet neither ambition nor capital is enough to materialize all the necessary components for these power-hungry computers.

Oh, this is just wild. Maybe they should have checked first?

(…continue reading.)

They wanted to put AI to the test. They created agents of chaos.

When a group of researchers at Northeastern University’s Bau Lab began toying with a new kind of autonomous artificial intelligence “agent,” it was supposed to be a fun weekend experiment. Instead, alarm bells started ringing.

The more they tested the capabilities and limits of these AI models, which have persistent memory and can take some actions on their own, the more troubling behavior they witnessed. Dubbed “ Agents of Chaos,” the group’s recently published work shows how, with very little effort, autonomous AI agents can be manipulated into leaking private information, sharing documents and even erasing entire email servers.

What a fantastic and insightful paper. You should read it.

(…continue reading.)

Read more