The short version is that the conversations around AI are now professionalized, institutionalized, and politicized. As a result, I’ve noticed that my relationship with writing morphed from something I did for fu…

Read more

Don’t panic, I listened to the full recording of the meeting so you don’t have to. What I got back was less a technical cross-examination of how training a model actually intersects with copyright law, and more a vibe session about power and money, wrapped in the veneer of “transparency.”

Let’s start with the opener, because it set the tone. The chair kicked things off by noting that other platforms had been invited to appear in front of the committee but had declined to appear. This was a not-so-subtle insinuation: the AI industry won’t show their faces, so we’re left with whomever’s willing to take the meeting. Except, knowingly or not, the chair made, in the words of Doechii, an “oopsie.” Or, as the legal profession calls it, a false statement. I had offered to participate and initially the committee was open to it, but then the invitation was pulled. Thankfully, the committee chose to invite a technical expert who appears to spend a lot of time on social media reposting policy proposals from Steve Bannon, sharing speeches from the National Conservatism conference, but also statements from Alexandria Ocasio-Cortez and Gavin Newsom. [Update: after publishing this article, a reporter called to let me know that the expert in question disagrees with my description of their political leanings. After reviewing their social media posts, I've updated the blog to better reflect the diversity of their online activity.]

But this missing-witness innuendo matters because it points to a deeper problem with how these hearings often work. If you really want to understand the mechanics of AI and copyright (what’s happening in training pipelines, what data provenance can and can’t do, what’s feasible to disclose without leaking trade secrets or creating new security risks) you need the companies that build the stuff, plus the people who can interrogate them in detail. Instead, what you had in this hearing was a failure to communicate and a committee whose strengths are, politely, elsewhere.

Half the members of the committee are political science graduates, half are journalists. Those are perfectly respectable ways to earn a living. It does, however, explain why the conversation kept orbiting around broad, morally loaded abstractions like “transparency,” “accountability,” “fairness” that sounded great in the moment and then dissolved into fog the second anyone tried to operationalize them.

There were brief moments when the Lords and the witnesses tried to have a more pragmatic conversation; for example, one committee members asked whether the UK should train its own models (spoiler alert: the UK has produced zero competitive foundation models). The problem is that the incentives to have these frank and tough conversations simply aren’t there and the risk of being honest is too high. So what you get is a very polite spokesperson from Google predictably having to deliver platitudes and threading the needle between “we support creators” and “we can’t disclose everything about how we train models.” And because the committee’s questions rarely pinned down technical specifics, the answers tended to float at the same altitude: high enough to be unobjectionable, too high to be useful.

Beyond the lack of depth and intellectual curiosity, the Lords on this committee have another challenge: they see generative AI first and foremost as a threat to the creative industries. They’re not shy about it, which I guess is fair. They push hard for protections: stronger rights, tighter rules, more leverage for artists, authors, and publishers.

The problem is that this framing conveniently forgets what generative AI actually is: a general purpose technology. Copyright is the flashpoint because it’s emotionally legible (artists being “ripped off” is a story anyone can understand) and because the UK is rightly proud of its creative economy. But if you regulate generative AI as though it’s basically a content remix machine for filmmakers, writers and musicians, you’re going to miss the fact that the same underlying capabilities are sliding into law, professional services, healthcare and life sciences, automotive, manufacturing, and a long list of industrial and enterprise workflows that have nothing to do with fan fiction or deepfake memes.

That’s also why the House of Lords’ apparent regulatory instinct, a tougher version of the EU AI Act style of governance, should make everyone nervous. The EU AI Act is famous for its risk-based structure, classifying uses and attaching obligations accordingly. In theory, that sounds sensible. In practice, risk-based regulation of a general purpose technology is a recipe for disaster, because the “risk” isn’t a property of the model the way toxicity is a property of a chemical drum. It’s a property of context, deployment, incentives, and human behavior, things that vary wildly from one use case to the next and can change faster than any legal taxonomy.

Europe is already living this tension. It’s had to significantly delay the AI Act because it couldn’t come up with usable standards and codes of practice. Right now, the European AI Office is working on a code of practice for Article 50 of the AI Act that focuses on transparency rules for certain systems, including generative systems and deepfakes. Article 50 sets out transparency obligations that, among other things, require informing people when they’re interacting with AI and requiring certain synthetic content to be marked or disclosed.

In a meeting of the Article 50 working group, speakers from two of Europe’s industrial giants laid out the sheer madness of applying rules that were clearly designed with generative AI on social platforms in mind to industrial or enterprise AI use cases. When your AI system is generating a deepfake of a politician or a newsy piece of public interest text, labeling and watermarking conversations make intuitive sense. When your AI system is optimizing a supply chain, flagging defects on a production line, or assisting with drug discovery, the same transparency playbook starts to look like regulatory cosplay: burdensome, misfitted, and occasionally nonsensical.

Even within the narrow domain Article 50 is targeting, the current draft code leans toward requiring providers to implement transparency through multiple prescribed methods and layering techniques like watermarking, metadata, detection interfaces, and logging. This framework leans heavily on standardized disclosure conventions such as icons, disclaimers, and modality-specific labeling.

But different AI systems need different solutions,. A single set of mandated techniques might be convenient for regulators, yet it risks being actively counterproductive in the real world. Providers should be able to apply the most appropriate transparency techniques based on their specific context (which tends to be best defined by industry-specific policymaking): what the system does, where it’s deployed, who the users are, what the threat model looks like, and what tradeoffs are acceptable. In other words: don’t confuse “we want transparency” with “we want one transparency to rule them all.”

Back in the Lords hearing, none of this nuance around transparency was fully grappled with. The committee is clearly trying to respond to a genuine political problem: creators are furious, lawsuits are multiplying, and governments that promised to turn their countries into AI powerhouses are discovering that you can’t speed-run legitimacy. The ministers have already framed the government’s approach as needing a “reset,” acknowledging that earlier instincts like leaning toward an opt-out approach ran into a wall of opposition from the creative sector.

But here’s the real risk. If lawmakers treat generative AI as primarily a cultural threat to be contained, they’ll end up designing rules optimized for one battleground and accidentally kneecapping the broader economy-wide transition that general purpose AI is bringing. The creative industries deserve real protections and workable licensing markets. They also deserve a regulatory conversation that doesn’t stop at slogans, doesn’t mistake disclosure for governance, and doesn’t turn “risk-based” into “everyone gets the same paperwork, plus a different colored sticker.”

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• MIT Technology Review: CES showed me why Chinese tech companies feel so optimistic

• Fortune: Worried about AI taking your job? New Anthropic research shows it’s not that simple

• CNBC: The rebellious instincts that turned Synthesia’s Victor Riparbelli into a generative‑AI trailblazer

• Business Insider: Tech executives bet big on AI. Their workers are being tasked with proving they were right.

• FT: Sadiq Khan to warn AI could cause ‘mass unemployment’ in London

• WSJ: Chinese AI Developers Say They Can’t Beat America Without Better Chips

• The Guardian: Lamar wants to have children with his girlfriend. The problem? She’s entirely AI

• FT: Apple sits out AI arms race to play kingmaker between Google and OpenAI

• Sifted: Yann LeCun poaches from Meta, Google DeepMind for new startup

• MIT Technology Review: Meet the new biologists treating LLMs like aliens

• Sifted: Inside Hiro Capital’s €500m plan for European startups: ‘We can shift the dial’

Read more

(If you want to listen to the whole episode, you can do so here.)

If you’ve been anywhere near the AI industry for the past three years, the feeling you’ve probably had was that of living in a split-screen reality. In one window, labs shipping ever more capable foundation models, each launch framed like a moon landing. In the other: the familiar reality of a mostly AI-less workplace, with maybe a chatbot feature added to a tool you already hated. The result has made even the more AI-pilled insiders scratch their heads and ask: how can the technology feel so futuristic, and daily life feel so unchanged?

The explanation, in my view, is that we’ve spent all this time building a very thick infrastructure layer and a very thin application layer. Models got big but everything around them stayed small. The hard, unglamorous work of mapping models onto specific business processes, integrating with legacy systems, enforcing governance and compliance, redesigning workflows so humans and machines don’t trip over each other has been the bottleneck, with only a few companies brave (or capable) enough to take it on. When the application layer is missing, impact gets trapped where the friction is lowest: demos, hackathons, experimental teams, and the occasional department head who duct-tapes a workflow together and calls it digital transformation.

However, over the last year, the application and orchestration layers have been filling in at speed. There’s a growing ecosystem of companies building vertical AI software that doesn’t merely answer prompts, but actually runs chunks of work. And once you have enough of that scaffolding, the story stops being about what a model can do in theory and starts being about how an organization defaults to getting things done.

In practice, that means 2026 is the year AI stops being a tool employees sometimes open, and becomes the default path work takes through a company. Not because executives suddenly become enlightened, but because the path of least resistance changes. If the AI-native workflow is faster, cheaper, and safer than the manual one, it becomes the new normal the same way cloud software did: quietly, then all at once.

The clearest shift will be from one-off prompts to end-to-end workflows. Right now, most corporate AI use is still a string of little favors: draft this email, summarize that meeting, suggest a block of code, rewrite this slide. Helpful, sure, but it’s the productivity equivalent of buying a better stapler. In 2026, more companies will hand models ownership of entire low-risk processes, with humans supervising exceptions rather than executing every step.

That’s also where the first real productivity gains arrive. Not the cute kind (“ChatGPT saved me a few minutes on drafting a Slack message”) but the kind that shows up as cycle time collapsing, queues shrinking, and managers realizing they didn’t hire three extra people this quarter because the work simply didn’t accumulate. Companies don’t get measurably more productive because employees write faster. They get more productive because work moves through fewer bottlenecks, fewer handoffs, and fewer rework loops. End-to-end workflow ownership is the difference between AI as autocomplete and AI as operations.

As this becomes normal, job content will mutate in ways that feel obvious only in hindsight. I’m skeptical that 2026 is the year of mass unemployment. But I do think it will be the year when job descriptions start to read like they come from a slightly different economy. Analysts spend less time constructing models in Excel and more time interrogating AI-generated scenarios, pressure-testing assumptions, and deciding what to believe. Sales and support reps spend less time typing and more time handling nuance: negotiation, relationship management, and the messy edge cases that don’t fit the template. Operations teams shift from “do the process” to “design, monitor, and tune the process the AI runs.”

You can already see the shape of this in places you wouldn’t necessarily expect. Consider sell-side research: the regulated, prestige-heavy world of analyst reports and client communications. One argument goes that the written report is the first thing to automate, while the higher-touch one-on-one interactions survive longer because they rely on trust, context, and human nuance. And yet even there, firms are experimenting with turning analysts into AI avatars to scale video summaries, offloading the performative, repetitive part so humans can focus on the higher-value work.

Which brings me to the new career superpower for 2026: managing workflows and systems, not just tasks. For a decade, the status game for individual contributors involved in office work has been about output: how fast you respond, how clean your slide deck is, how many tickets you clear. In an AI-default organization, what matters more is whether you can design a process that keeps quality high while throughput increases, whether you can set guardrails that make automation safe, and whether you can diagnose failures without melting down the whole pipeline. The winners won’t be the people who “use AI.” They’ll be the people who can make AI reliably do something useful inside the constraints of their business.

A second-order consequence is that small and midsize businesses get enterprise-grade leverage “for free.” Big companies have been piloting AI, hiring AI leads (or “chief AI officers”), and paying consultants to tell them what they already suspect: their data is messy and their processes are worse. Smaller firms, meanwhile, have mostly been stuck with generic tools that help a bit but don’t change the game. As the application layer matures, a small team using AI-native tools will be able to do things that recently required a whole operations department: automated outreach, first-line customer support, basic forecasting, content production, even light finance workflows. That levels the playing field in a way incumbents rarely enjoy. It also puts pressure on mid-tier companies that don’t have the brand moat of giants or the agility of startups: the ones most likely to say “we’re using AI” while still running 2016 processes underneath.

Basically, 2026 is the year when we finally see differentiation. For the last couple of years, “we’re using AI” has been as meaningful as “we have a website.” Everyone says it. It tells you almost nothing. In 2026, the gap in outcomes will widen enough that you can spot it from the outside. Companies that actually rebuilt workflows around AI will ship faster, respond to customers faster, and run leaner teams without feeling understaffed. Companies that stapled an AI chatbot onto the front of the old process will wonder why nothing really changed, except that their customers now have a new way to get frustrated.

The technology will keep improving, of course. Models will get cheaper, faster, and more capable; AI systems will feel less like party tricks and more like infrastructure. But the competitive advantage won’t come primarily from having the slightly better model. It will come from execution: who integrates well, who governs data responsibly, who redesigns work instead of sprinkling prompts on top of it, and who treats AI in the business as an operating model rather than a feature.

And now, here are this week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• MIT Technology Review: What’s next for AI in 2026

• Startup Coalition: The AI Index

• Business Insider: Microsoft reshuffles teams to bolster GitHub as AI coding and agent wars heat up

• FT: AI will free households from chores and boost hidden productivity, says OpenAI

• Fortune: After Nvidia’s Groq deal, meet the other AI chip startups that may be in play—and one looking to disrupt them all

• Business Insider: A future without work? What Elon Musk, Bill Gates, and others in AI are saying about the future.

• FT: AI start-ups take on Google in fight to reshape web browser market

• WSJ: How Google Got Its Groove Back and Edged Ahead of OpenAI

• Sifted: ElevenLabs founder Mati Staniszewski: ‘It’s not just about voice anymore’

• The New York Times: Can A.I. Match Molière’s Wit? These Researchers Think So.

• Wired: AI Models Are Starting to Learn by Asking Themselves Questions

Read more

In the age of AI, that story is starting to fall apart.

If you want to see the shift in a single snapshot, start with The Observer’s Global AI Index, a data-heavy attempt to benchmark 90-plus countries on AI investment, innovation, and adoption. Its latest edition doesn’t mince words: the US still dominates, Asian countries are rapidly rising, and Europe, despite all that talent, is falling behind. Roughly 90% of all private AI funding in the first nine months of 2025 went to the US. Europe attracted just 3.8%, and Asia 2.9%. Meanwhile, the EU’s share of papers at major AI conferences fell from 16% to 12%, and the continent is responsible for only about 13% of frontier models.

The same index highlights what’s happening elsewhere. The UAE, Taiwan and South Korea have all surged up the rankings, with the UAE jumping 10 places to enter the global top 10 on the back of massive investment in compute infrastructure and rapid AI adoption across government and society. Saudi Arabia, the UAE and others in the Gulf are now in the headlines as often for AI megaprojects as for oil.

Europe does still train world-class AI talent but it increasingly can’t keep it. The Observer’s analysis notes that the UK, for example, retains less than half of its top academic AI researchers; most move to the US. Only about 15% of top industry researchers work at UK-headquartered companies. The picture across much of the continent is similar: the labs are good, the passports are European, but the affiliation on the NeurIPS paper increasingly reads San Francisco or Seattle, not Zurich or Paris.

NeurIPS, the world’s flagship machine learning conference, has become a kind of global MRI scan for AI power. An analysis by AI World of the total accepted NeurIPS 2025 papers leaves little room for nostalgia. The conference, held this year in San Diego and Mexico City, is now dominated by three hubs: Beijing, Shanghai and San Francisco. Chinese and US institutions account for a substantial share of all authorships; everyone else, including Europe, is competing for the scraps.

The analysis calls out some notable institutions that sit outside the usual US–China duopoly. Among them is Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), a purpose-built AI university founded in Abu Dhabi in 2019. In the NeurIPS data, MBZUAI shows up as one of the standout contributors alongside Singapore’s NUS and NTU and South Korea’s KAIST. That lines up with a broader trend: MBZUAI has already climbed into rankings of the global top 20 institutions in AI-related computer science, an extraordinary rise for a five-year-old university.

If you’re a young AI researcher, it’s hard to miss the signal this sends. Yes, conference papers are a “numbers game” in many ways: incentives favor salami-sliced work and citation cartels. But prestige conferences are still a powerful language the academic world understands. When their author affiliation maps start lighting up around Beijing, Shanghai, Abu Dhabi and the Bay Area, and dimming across big swathes of Europe, the message is clear: if you want to be at the center of the action, you may need to get on a plane.

That’s been one of Europe’s biggest blind spots: the old assumption that “we have the talent, so we’ll be fine” no longer holds. The talent is mobile. It goes where it finds compute, patient funding, and dense networks of collaborators. Right now, that’s more likely to be California, Shenzhen, or indeed the Gulf than much of the European Union. Perhaps the easiest way to understand this reality is to bring it to life with an example: while visiting Abu Dhabi this week, I was speaking with a French neuroscientist with years of experience working in industry and academia and who recently left France for the UAE to develop medical-grade brain-computer interfaces. Oftentimes, the media likes to portray moves to the Gulf as motivated purely by money, but this particular scientist had multiple successful exists, and could’ve continued to live a comfortable life in his homeland. Instead, he became frustrated with how Europe treats entrepreneurs who want to innovate, and with French regulators in particular, who kept dragging their feet over dozens of months with vague questions and unnecessary legal paperwork. In contrast, he described the regulatory approval process in the UAE as being far more intense but also fast and optimized for innovation. The Emirati regulator moved at speed to check that his project passed all their high safety standards, and then they immediately switched to “innovation mode,” connecting him with academic and commercial opportunities that matched his ambition, which would’ve never happened in Europe.

To be fair, some European leaders are not oblivious to these challenges. In November, France, Germany and the European Commission announced a new initiative aimed explicitly at “cutting-edge” AI research and digital sovereignty. Responding to this project in Le Monde, prominent figures including Max Welling and Yann LeCun made the argument that Europe does have the resources to sustain world-class AI research if it stops fragmenting them and starts acting strategically. Among the suggestions: focus on underexplored areas of the research frontier instead of chasing the biggest proprietary models, concentrate funding in a small number of agile institutes rather than thinly spreading it, and make serious, coordinated investments in compute and data infrastructure.

But there’s a tension between that vision and Europe’s day-to-day political behavior on AI. While others race to build labs and data centers, Europe has often looked more preoccupied with tinkering over the exact phrasing of AI regulation and offering symbolic support to academia that isn’t always matched by hard resources. The EU’s AI Act, for example, is still being digested by companies and universities alike; its detailed risk categories and compliance demands may be well-intentioned, but they’re also adding friction to exactly the institutions that Europe says it wants to empower. At the same time, public research budgets and national compute strategies are only now starting to catch up with the scale of today’s models.

The point isn’t that regulation is bad, or that Europe should suddenly go full techno-libertarian. It’s that the opportunity cost of political attention is real. Hours spent negotiating the 19th compromise draft of Article Whatever are hours not spent hammering out long-term funding structures or buying serious GPU clusters for European universities. In the meantime, other regions are moving faster. The Gulf states are using sovereign wealth funds to underwrite national AI champions and infrastructure at startling speed; the US continues to attract record levels of capital and talent; China has its own vast state-backed ecosystem.

Against that backdrop, the NeurIPS numbers I shared above should be a very loud wake-up call for Europe. In less than a decade, a university built from scratch in the UAE has become more visible in some AI benchmarks than many long-established European institutions that assumed their prestige was an asset no one could catch up with.

European leaders should feel uncomfortable when they look at those maps and rankings. Not because Europe has suddenly become bad at science (it hasn’t) but because the rest of the world has started treating AI as a first-order strategic priority, on a scale and with a focus that Europe has been slow to match. When 90% of private AI funding flows to the US, when Europe’s share of flagship conference papers and frontier models is sliding, and when young researchers are voting with their feet, “we have great universities” is no longer a strategy.

So what would a less complacent European response look like?

One starting point is to take the Welling/LeCun playbook seriously rather than treating it as another think-piece. That means picking a handful of locations to host truly world-class AI institutes, with guarantees of long-term funding, shared GPU super-clusters, and governance structures that make it easy for top researchers to move between academia and industry. It means aligning visa, tax and employment rules so that bringing in a brilliant PhD student from Lagos or Lahore is not a bureaucratic nightmare. It means using EU-wide mechanisms to negotiate access to large datasets in health, climate and industry, rather than leaving every project to renegotiate the same legal thicket.

It also means being honest about trade-offs. If Europe wants to be competitive in AI research, it cannot treat every GPU as a potential climate sin or every model training run as an ethical scandal in waiting. It will have to decide, in public, that some amount of risk and energy consumption is acceptable in exchange for scientific and economic leadership, and then design regulation that constrains the worst outcomes without smothering the rest.

None of this would guarantee Europe “wins” the AI race (I hate this term with a passion). The US and China are operating at a scale that will be hard to match. The Gulf is deploying capital with a speed that Brussels eggheads will never replicate. But there is still a plausible future in which Europe is not just a source of graduates and bad op-ed authors, but a place where frontier ideas are born, trained and deployed.

The alternative is to keep telling the old comforting story while the new one writes itself elsewhere: an AI world whose center of gravity runs from Beijing to San Francisco via Abu Dhabi, with Europe watching, regulating, and occasionally complaining from the sidelines.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• WSJ: Behind the Deal That Took Disney From AI Skeptic to OpenAI Investor

• Business Insider: Meet the young AI startup founders raising millions in the race to build the next big thing

• The Information: DeepSeek is Using Banned Nvidia Chips in Race to Build Next Model

• The New York Times: Meta’s New A.I. Superstars Are Chafing Against the Rest of the Company

• Bloomberg: How ASML’s CEO Plans to Keep Pace With Soaring AI Demand

• FT: Google’s TPU chip puts OpenAI on alert and shakes Nvidia investors

• Fortune: Google DeepMind agrees to sweeping partnership with U.K. government focused on science and clean energy

• Business Insider: It used to be for shopping and lipstick. Now, a Chinese app is a haven for tech workers to swap AI intel.

• WSJ: Inside the Creation of Tilly Norwood, the AI Actress Freaking Out Hollywood

• The Verge: AI companies want a new internet — and they think they’ve found the key

• Time: The Architects of AI Are TIME’s 2025 Person of the Year

Read more

With the AI Act, Brussels made a giant gamble: can you design comprehensive AI regulation before the technology settles down, and still stay competitive? The answer, judging by the EU’s frantic effort to simplify and delay parts of its landmark AI Act, is turning into a bitter lesson in what happens when process gets ahead of practice.

A big part of that story is who Europe chose to listen to. From the early white papers to the final trilogue marathons, civil society coalitions and academic institutes were enormously influential. Groups like AI Now Institute, Access Now, AlgorithmWatch, the Future of Life Institute, and the Centre for the Governance of AI submitted dense consultation responses, coordinated “pause AI” open letters signed by more than a hundred NGOs, and pushed a simple message: don’t wait for AI to go wrong; hard-wire fundamental rights and risk controls into the law now.

Some of these organisations are full of weirdos and freaks, who have deluded themselves into believing the technology poses existential risk, and therefore have taken very radical positions. Others are not “anti-AI” in the sense of demanding a return to fax machines; they frame themselves as pro-rights and pro-accountability, and they spend more time in policy workshops than on picket lines. But their center of gravity is clear: they worry far more about surveillance, discrimination, and concentration of power than about whether a European startup can train the next foundation model. Many of the most prominent voices in this camp are lawyers, social scientists, philosophers, and policy researchers. In short, experts on how technology reshapes society, less so on shipping large-scale commercial AI products under time-to-market pressure.

A sprawling civil society coalition was born which shaped debate about which AI uses should be “high-risk,” how biometric surveillance should be treated, and what kinds of transparency obligations should apply to general-purpose models. Access Now and AlgorithmWatch, for example, pressed the EU to treat broad categories of algorithmic decision-making as “highly consequential,” and to frame the AI Act first and foremost as a human-rights instrument.

Some of what they pushed for wasn’t inherently bad. Someone has to ask what happens when predictive policing or credit scoring goes wrong.

The problem is what happened next.

The AI Act began life in 2021 as a relatively crisp risk-based proposal, focused on a list of “high-risk” applications. Then generative AI exploded into public consciousness. Suddenly, the law had to cover everything from chatbots to foundation models that hadn’t been imagined when the original text was drafted. Under intense political pressure and with civil society groups warning about “systemic risks” from powerful models, lawmakers started bolting new categories, carveouts, and obligations onto an already dense text.

The result is a regulation that many companies and individuals, including those who support strong guardrails in principle (such as yours truly), describe as confusing and hard to implement. A recent survey commissioned by AWS found that more than two-thirds of European businesses struggle to understand their obligations under the Act. Tech lobby groups representing firms like Alphabet and Meta and European AI startups have publicly called for a pause or delay, arguing that unclear guidance and overlapping requirements risk chilling investment just as the US and China are racing ahead.

Brussels is now quietly agreeing with at least part of that critique. This month, the European Commission proposed what amounts to a mini-U turn: a “Digital Omnibus” package that would push back the application of key high-risk AI rules into 2027, soften some requirements, and generally “simplify” the digital regulatory stack, from GDPR to cyber incident reporting. Officially, this is about giving companies time and standards so they can comply. Unofficially, it’s an admission that the first draft of Europe’s big AI experiment was too much, too fast.

The problem is that this is more like a half-step retreat under fire. Civil liberties groups that once cheered the AI Act as a global benchmark now warn that the simplification drive risks gutting hard-won protections, especially around law enforcement and migration. At the same time, CEOs warn that even a delayed AI Act remains so complex that their companies will be stuck in compliance limbo while competitors ship products from friendlier jurisdictions. Europe has managed to upset both those who wanted stronger safeguards and those who wanted fewer.

Amidst all the confusion, two developments are worth calling out.

First, Europe is losing its academic edge and industrial soft power. China now dominates the open model space, which means the current and next generation of AI applications and services are built on technology created in Asia by companies that not only have complete disregard for European norms, values and regulation, but, in some cases, enable the kind of mass surveillance and social scoring that the AI Act tried to ban.

Secondly, other countries are looking at the EU and saying “No, thanks.”

For example, instead of rushing to introduce an AI Act of its own, the UK government chose a softer, “pro-innovation” path: no single horizontal AI law, at least not yet, but a set of high-level principles to be interpreted by existing regulators like the Competition and Markets Authority, the Information Commissioner’s Office, and the Financial Conduct Authority. That framework, first laid out in a 2023 white paper and followed up by a 2024 consultation response, rejects a one-size-fits-all statute in favor of context-specific oversight.

Crucially, the UK’s consultation process has been more openly pluralistic from day one. The British government has been gathering views from industry, trade associations, regulators, academia, and civil society groups, rather than treating any one camp as the default voice of “the public interest.” Charts in the government’s own response note that AI and tech companies made up the largest share of respondents, followed by professional bodies, NGOs, and research institutions. The goal, at least on paper, is to balance safety with competitiveness, not to pick rights over innovation or vice versa.

You can see that same pluralism in how the UK has handled one of the nastiest flashpoints in AI policy: copyright. After a furious backlash from musicians, publishers, and rights holders over proposals that would let AI companies train on copyrighted material by default, the government didn’t dig in or capitulate. Instead, it convened expert working groups bringing together representatives from creative industries, unions, and AI developers, with a brief to hammer out “practical, workable solutions” around licensing, transparency, and consent. I attended two of them and it’s a messy, adversarial process but at least the mess is happening in meetings, not only in last-minute amendments on the floor of Parliament.

None of this makes the UK a regulatory utopia. The same light-touch instincts that appeal to startups have alarmed parts of the creative sector, which see a government too eager to please Silicon Valley. Artists have staged protests, released a silent album, and launched campaigns to defend copyright as a cornerstone of the UK’s cultural economy.

Still, there’s a structural difference worth paying attention to. The UK is trying to regulate around live deployments by real companies, many of them homegrown, rather than designing a total system on the whiteboard and hoping reality catches up. That means its rules and codes of practice can, in theory, evolve alongside the technology, with regulators updating guidance and standards as new risks show up in the wild. It’s a bet that agile governance, anchored in existing institutions, will age better than a single, towering piece of legislation.

Europe’s experience suggests why that might matter. By the time the AI Act’s high-risk obligations fully kick in, today’s model architectures may look quaint, and the most important questions may be about systems barely mentioned in the original text. The more detailed and prescriptive your rules are, the more often you have to revisit them. Brussels is discovering that a law written for one phase of AI development can quickly become a straitjacket or a patchwork of exemptions.

It would be easy, especially from London, to turn this into a simple morality play: overcautious continent versus nimble island. The reality is messier. European civil society organisations were right to flag the risks of unaccountable AI long before most politicians cared. European lawmakers were right to insist that AI systems used in policing, employment, or healthcare deserve more scrutiny than a recommendation engine for cat videos. And the UK’s more flexible approach comes with potential hazards, especially if economic anxiety tempts ministers to ignore harms until they become scandals.

But there is a lesson in how the EU process unfolded. If you treat AI primarily as a bundle of abstract risks to be constrained ex ante and place limits based on imaginary risks or arbitrary mechanisms, then you calibrate your rules mainly through the lens of people whose professional job is to worry about those risks, you risk underweighting the operational realities of building and deploying these systems. That’s how you end up with meticulously negotiated article numbers that leave both startups and regulators scratching their heads about who exactly needs a conformity assessment for what.

If, instead, you build your governance model around concrete deployments, with regulators, industry, trade bodies, and rights advocates all forced to argue in the same room about specific use cases, you may end up with less elegant law, but more adaptable guardrails. That’s what the UK is betting on with its “pro-innovation” framework and its swarm of consultations and expert groups. Whether that translates into a durable competitive edge is still an open question. But at the moment, Britain looks more like a live testbed than a museum of frozen good intentions.

Europe’s bitter lesson in AI isn’t that regulation is doomed, or that human-rights advocates should be sidelined (well, okay, maybe some of them should be!).

It’s that in a field moving this fast, who you invite into the drafting room, and how closely they’re connected to the messy business of actually building and operating AI systems, can matter as much as what ends up on the page. The AI Act was supposed to cement the EU’s leadership.

Today, as Brussels scrambles to simplify and delay its own rules while others watch and learn, that leadership looks a lot more conditional.

And now, here are this week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• Fortune: MIT report: AI can already replace nearly 12% of the U.S. workforce

• Business Insider: Amazon’s AI capacity crunch and performance issues pushed customers to rivals including Google

• The New York Times: Fears About AI Prompt Talks of Super PACs to Rein In the Industry

• FT: AI video start-up takes wide-angle view to boost growth

• Bloomberg: AI Slop Recipes Are Taking Over the Internet — And Thanksgiving Dinner

• The Information: OpenAI Forecasts Nearly as Many ChatGPT Subscribers as Spotify by 2030

• WSJ: ‘Sovereign AI’ Takes Off as Countries Seek to Avoid Overreliance on Superpowers

• The Verge: David Sacks tried to kill state AI laws — and it blew up in his face

• FT: China leapfrogs US in global market for ‘open’ AI models

• WSJ: Teens Are Saying Tearful Goodbyes to Their AI Companions

• Bloomberg: Google, the sleeping giant in global AI race, now ‘fully awake’

Read more

A new report from Accel gives a data-rich snapshot of that landscape. It shows a world where US companies overwhelmingly dominate model building and hyperscale infrastructure, while European and Israeli startups are rapidly catching up, and sometimes almost match US levels, when it comes to cloud and AI applications.

Most people will chalk that up to familiar explanations: Europe’s strong technical universities, a dense industrial base that’s hungry for automation, and policy tailwinds around trusted AI. All true. But there’s another, more nuanced reason that I think matters just as much: US companies are paying what I like to call the ASI hype tax, and European companies mostly aren’t.

By ASI hype tax, I don’t mean GPU bills or cloud credits. I mean something more human and messier: salary inflation, bidding wars, and relentless talent poaching driven by the belief that whoever gets to artificial superintelligence first gets to own the future. That hype shows up as higher operating expenses and higher employee churn, two things that are especially painful in a young, fast-moving industry where institutional knowledge and culture are still being built.

Europe, for now, is getting a discount.

Let’s start with the capital flows. On the model side, the Globalscape report is blunt: US model builders have sucked up almost all the oxygen. From 2021 onward, AI model funding in the US has run into the tens of billions annually, while Europe and Israel have raised only a tiny fraction of that. For example, 2025’s big rounds are dominated by OpenAI, Anthropic, xAI and a handful of other US players, with European model rounds an order of magnitude smaller.

In total, US AI model funding vastly outweighs EU/IL model funding over the 2021–2025 period. That capital isn’t just buying H100s and data center space, it’s also bidding up researchers, engineers, and product leaders. If you’re a senior AI engineer in San Francisco, your inbox has essentially been a live auction for the past 12 months.

Accel frames this as part of a “race for compute,” a multi-trillion-dollar build-out of AI data centers that could reach about 117 additional gigawatts of capacity by 2030 and roughly $4.1 trillion in capex between 2026 and 2030. That race is being led, and largely financed, by US hyperscalers whose combined operating cash flow is big enough to underwrite years of this spending.

When so much capital and narrative energy is concentrated at the very top of the stack, you get a predictable side effect: intense competition for a relatively small pool of elite talent that can exploit that compute, and compensation expectations that increase with every new mega-round or model launch. That’s where we see the ASI hype tax in action.

And that tax doesn’t stay neatly inside the model labs. It bleeds into the entire US ecosystem. Application-layer startups in the Bay Area are competing for many of the same people, in the same cities, at the same inflated price points, and often watching their best people get poached by a lab or a hyperscaler the moment they prove themselves.

Now contrast that with European and Israeli startups. Depending on the year, EU/IL funding is between roughly half and two-thirds of US levels (for 2025, about 66% of US funding on an annualized basis.) Europe and Israel aren’t trying to outspend the US on frontier models, they’re channeling capital into AI-native products in verticals like healthcare, finance, security, and construction.

Accel’s Europe AI 100 and US AI 100 snapshots offer a great way to visualize the two different operating environments. The top 100 European and Israeli AI and cloud companies together employ about 10,400 people and have raised $5.4 billion in total funding, with an average age of around four years and roughly 104 employees per company. The US AI 100, by contrast, employs about 5,200 people on $4.8 billion of funding, with companies that are on average only 2.4 years old and 52 employees apiece, and they’re growing headcount much faster, with 213% average year-over-year FTE growth versus 122% for Europe and Israel.

In Europe and Israel, the typical breakout AI company is a little older, a little larger, and growing headcount aggressively but not insanely. That’s what a more sustainable hiring environment looks like: still fast, but not “triple the team every twelve months, and hope the culture somehow survives.” In the US cohort, where more than half of the winners are under three years old and headcount is more than doubling annually, the odds that your early employees stick around for the whole journey are much lower.

If you’re running an application-layer startup that has to ship real features to real customers, that stability matters. Every time a staff engineer walks out the door, they take months of context with them: weird edge cases from your largest customer, half-finished internal tooling, undocumented architecture decisions. Multiply that by a market where everyone is constantly interviewing and being interviewed, and you get an ongoing drag on execution that rarely shows up in pitch decks or funding charts.

The Globalscape report also shows how fragile AI application economics still are. Even today’s standout AI-native application companies show ARR per employee far above traditional SaaS, but their gross margins are well below classic software benchmarks. Several are in the 10–40% range, compared with roughly 76% for the average public cloud company.

Why? Inference costs, bespoke integrations, heavy support workloads which are all the unglamorous realities of turning a clever model demo into a dependable product. The report notes that rapid declines in inference costs (OpenAI’s flagship models have seen more than a 90% drop in per-token pricing over the last few years) should help margins improve over time, but we’re not there yet.

In that environment, operating expense discipline is the only way to guarantee survival. If your compute bill is still chunky and your gross margin is 30%, you simply can’t afford to pay San Francisco frontier-lab salaries for every engineer and then lose a chunk of them to the latest model company round or “AI agents for everything” startup every 12–18 months.

This is where Europe’s smaller ASI hype tax becomes a strategic advantage. Because European model funding is lower and the compute race is less centered there, local application companies face less direct competition from homegrown hyperscale labs and fewer nine-figure frontier rounds bidding up the same talent pool. They still compete globally, of course but in practical terms, a Berlin-based AI vertical SaaS firm is not fighting OpenAI for engineers on the same street the way a Bay Area startup is.

Lower churn and less extreme salary pressure don’t just save cash, they preserve culture too. In a fast-moving domain like AI, institutional memory compounds: how you built your evals, how you debugged that silent model failure for a key customer, how you navigated the first wave of AI safety audits. Keeping the people who lived through those moments is a non-trivial balance-sheet asset.

The Globalscape report also makes clear that someone has to pay for the AI party. Accel estimates that paying back the expected 2026–2030 AI data center capex would require about $3.1 trillion in AI data center revenue over that period, enough to add roughly 1–2 percentage points to global GDP growth if it all materializes. That’s a massive macro bet on AI.

European application companies, by and large, are not the ones making that bet. They are the ones who will try to turn that infrastructure into specific business outcomes: shorter order-to-cash cycles, fewer denied insurance claims, faster permit approvals, more secure networks. The report’s What’s Next section is the Sears catalog equivalent of vertical European opportunity: AI-native tools for legal work, finance, construction, healthcare, AI security, voice and media for the enterprise, and vibe coding tools reshaping how software is built.

Being downstream of the hype has two big advantages.

First, you benefit from falling unit costs without absorbing the capex risk. As model performance converges and more players compete on inference price, application teams can shop around. They can move workloads between providers, use domain-specific models where they’re good enough, and keep their gross margins trending up as the infra world races to the bottom on price.

Second, you’re less exposed to the cultural and financial distortions that come with “we must be the first to ASI” thinking. You don’t need to hire a hundred researchers to work on hypothetical emergent capabilities; you need to hire enough engineers and product people to wire existing models into real processes in a robust way. That is still hard. But it rewards patience and execution more than it rewards hype.

Europe’s AI ecosystem, as Accel describes it, is starting to look like a place where those boring virtues can compound. The companies are older, the teams are bigger, and the headcount growth, while still ferocious by normal-industry standards, is not quite as vertiginous as in the US. In other words: there’s still hype, but less ASI hype tax.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• Wired: AI Relationships Are on the Rise. A Divorce Boom Could Be Next

• Business Insider: AI hype could sideline Gen Z workers

• FT: The EU needs to rethink its AI rules

• Wired: All of My Employees Are AI Agents, and So Are My Executives

• Bloomberg: When Employee AI Adoption Stalls Out, Workplace Influencers Get Called In

• WSJ: Inside Microsoft’s New AI ‘Super Factory’

• Business Insider: What one Big Law firm told 400 young lawyers about using AI

• Sportico: AI Companies Want to Change How You Buy Tickets

• WSJ: The Political Left Is Dialing Up Scrutiny of Data Centers

• Bloomberg: EU Plans to Streamline Data and AI Rules in Bid to Boost Business

• Business Insider: OpenAI is weighing a move into consumer health apps

• FT: ‘People like dealing with people’: Reed boss on the challenge of AI in hiring

• WSJ: AI Is Upending Jobs. Corporate Tech and HR Are Teaming Up to Figure It Out.

Read more

But another, quieter shift in the AI industry helps explain why Palantir keeps punching above its weight, and why Burry’s bet may be more story than substance. The newest must-hire in the AI wars isn’t a research scientist, it’s a forward-deployed engineer. You basically take a software engineer, dress them in a suit from the same department store where all the McKinsey consultants shop at, and send them into the field to embed with customers and help them make AI actually work on real data, real processes, and real timelines. OpenAI, Anthropic, and Cohere have all accelerated recruiting for these roles in recent months because they’ve learned the obvious lesson of 2025: shipping a model is great for press headlines, but shipping outcomes is what drives the business.

If this model sounds familiar, that’s because Palantir popularized it years ago, and still practices it at industrial scale. The company built a whole operating system around data analytics: teams that sit with customers, map messy data into a usable ontology, and then assemble production software that the customer keeps using when the engineers go home. That hands-on approach is why Palantir keeps showing up in high-stakes environments, from defense work in the US to the NHS’s Federated Data Platform in England, where the data is gnarly, the stakes are high, and results can’t live in slideware.

This, combined with the AI vibes, is what has underwritten Palantir’s numbers for so many quarters. The company’s Q3 update showed sharp growth, led by its US commercial business and rapid adoption of its generative AI platform, AIP. Revenue, margins, and guidance all moved in the right direction, and management talked about beating expectations as customers scaled deployments from single pilots to enterprise-wide rollouts. That’s what execution looks like when you’re selling not just a product, but a workflow delivered by an embedded team that can wrangle a warehouse of CSVs into something decisions can sit on.

I was recently in New York for a dinner with a few people working for some of the large organizations that are Palantir customers, and the uncomfortable truth they shared is that the old world of data analytics and the new world of large language models has one common bottleneck: the real world. Corporate data is scattered, contradictory, insecure, and often trapped in systems that were never designed to talk to one another. AI models don’t magically resolve that. People do, specifically engineers. Palantir has spent a decade industrializing that human-in-the-loop integration playbook. That’s why the best-known and most imitated customer success program still belongs to Palantir, and why everyone else is now racing to hire their own version.

None of this means Palantir is immune to drawdowns. Valuations bake in a lot of future perfection. Burry could be expressing a short-term view on sentiment, hedging other exposure, or simply trying to time a comedown after a monster run. He’s done that before. But if your thesis is that software spending will shift from lab demos to durable line-of-business systems, then the companies that can turn AI potential into production workflows should be advantaged, not shorted. And right now, the top of that list includes Palantir, whose quarter and backlog say customers are buying not just the dream but the deployment.

Betting $1 billion against Palantir in 2025 because you think AI is hype is a big mistake. The next leg of the AI cycle won’t be decided in model leaderboards, it’ll be decided in the boiler rooms of the enterprise, where forward-deployed teams wire data to decisions. Palantir made that playbook mainstream.

And now, here are this week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• MIT Technology Review: How AGI became the most consequential conspiracy theory of our time

• Fortune: Meet the power broker of the AI age: OpenAI’s ‘builder-in-chief’ helping to turn Sam Altman’s trillion-dollar data center dreams into reality

• 404 Media: The Future of Advertising Is AI Generated Ads That Are Directly Personalized to You

• FT: The new hot job in AI: forward-deployed engineers

• The Guardian: Boom or bubble? Inside the $3tn AI datacentre spending spree

• FT: Apple Watch data teamed with AI reveals heart damage

• 404 Media: arXiv Changes Rules After Getting Spammed With AI-Generated ‘Research’ Papers

• FT: Creative groups fail to secure UK legal precedent in Getty AI copyright case

• Wired: The AI Data Center Boom Is Warping the US Economy

• WSJ: Microsoft Lays Out Ambitious AI Vision, Free From OpenAI

Read more

My first observation is that data center demand is compounding at breakneck speed. If a few years ago, AI-related data center buildout was influenced by the training requirements of scaled models, inference is now dominating workloads, as models stop posing for benchmarks and start doing real work. The energy industry is bracing for multi-year, trillion-dollar buildouts and we’re already seeing the effects of demand outstripping supply in terms of energy bills going up by double digits.

That’s because once you’ve got swarms of agents scheduling, drafting, reconciling, and hounding your backlog all day and all night, the baseline load is no longer spiky but continuous. We’re at a stage where we’re industrializing inference, not just renting burst capacity. A great example of where AI agents are starting to deliver value is in regulated industries such as financial services where they can be deployed for mission-critical processes. FlowX.AI has built dozens of multi-day or multi-goal agents with self-optimization loops; they run commercial lending, fraud detection, or customer support inside banks such as State Street, UniCredit or OTP Group.

The conversation then moved away from the cloud to the edge. And this is the part Wall Street still underprices: AI adoption won’t be confined just to hyperscalers; it can be soldered onto boards, flashed into microcontrollers, and embedded in wearables. Cheaper, faster and closer to the physics.

XMOS’s CEO described a generative SoC system they’re building that lets anyone specify system behavior in plain English and have the chip “compile” to that intent, collapsing design cycles and expanding the set of companies that can credibly ship custom silicon.

Luffy’s COO made some really great observations about at the factory floor, using the example of the BMW plant that makes MINI cars. His thesis was that we can swap out brittle PID loops for small neural networks running inside the microcontrollers already deployed in factories. If do that at scale, the assembly line robots shrink and factories get more energy efficient.

The more interesting point he made was about lifecycle reality, something that I also experienced inside Ocado’s robot warehouses: in many factories or warehouses with moving parts, equipment ages and tolerances wander, and the lovingly tuned factory you commissioned in March is mysteriously running behind by October. But if we can put adaptive AI into the microcontrollers driving the automation, it can track drift across force, distance, time, pressure, torque and whatever else the physics throws at it, while rejecting disturbances that used to trigger maintenance tickets.

Finally, I brought up a blast from the past: IoT devices and wearables, explaining why they might become interesting because all that data they stream to the cloud will finally be turned into useful insights. Sometimes, they’ll also be able to run AI models locally for quick wins. I talked about my Meta Ray-Bans and a near future where the glasses support a compact world model locally, and call out to the cloud only when they hit an uncertainty cliff. You don’t always need a gigawatt data center connected to your wearable; sometimes, a small, well-tuned prior that understands your environment and your routine, and a smart policy about when to escalate are preferable.

If there was a consensus on the panel, it was this: put the intelligence where the entropy is. A plant line that never runs the same way twice. A pair of always-on AI glasses listening for a wake word. A street where the radio spectrum goes feral at rush hour.

Will the cloud go away? Of course not. But if the last two decades of software development consolidated most economic activity among a few American or Chinese megaplatforms, I continue to believe this AI wave might offer opportunities for European companies with deep industrial experience to build solutions made locally, safely and fast.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• Bloomberg:

  • What’s your AI-dentity

  • The Solution to Europe’s AI Problem Is Hiding in Plain Sight

  • China’s AI Dragons Must Survive ‘Involution’ to Conquer the World

  • AI’s Copyright War Could Be Its Undoing. Only the US Can End It.

• Reuters: From OpenAI to Meta, firms channel billions into AI infrastructure as demand booms

• Wired: How ByteDance Made China’s Most Popular AI Chatbot

• The Information: Salesforce CEO Shifts AI Strategy as OpenAI Threat Looms

• Bloomberg: Anthropic’s AI Principles Make It a White House Target

• WSJ: OpenAI Wants City-Sized AI Supercomputers. First It Needs Custom Chips.

• Time: Accenture CEO Julie Sweet on Trust in AI, Building New Workbenches, and Why Humans Are Here to Stay

• WSJ: AI Is Juicing the Economy. Is It Making American Workers More Productive?

• Bloomberg: Europe Fights for AI Independence to Avoid Becoming Tech ‘Colony’

• WSJ: Why AI Will Widen the Gap Between Superstars and Everybody Else

Read more

The study’s topline: despite tens of billions poured into generative AI, only a thin sliver of projects deliver measurable financial returns; the rest stall in proof-of-concept purgatory, especially when companies try to build everything themselves rather than buy focused tools. The methodology leaned on interviews, surveys and public disclosures, and the verdict was grim enough to fuel a week of human slop (podcasts, newsletters, etc.) from the usual AI grifters.

But before we declare an AI winter, it’s worth reading the fine print and then looking at more data. For example, new G2 data published this week reveals the MIT narrative conflates yesterday’s bespoke chatbot experiments with today’s agentic systems that actually run in production. According to G2’s survey of more than 1,300 B2B decision-makers, nearly 60% of companies already have AI agents in production, satisfaction is high, and post-deployment failure rates are closer to 2% than 95%. While early generative AI pilots often died on the vine, agents embedded in workflows are sticking, saving money and speeding work, especially with a human in the loop.

Slides 98 to 104 from the State of AI Report produced by AI investor Nathan Benaich and Air Street Capital tell a similar story of real-world adoption: AI-first companies are now throwing off real revenue, well into the tens of billions on an annualized basis.

The most compelling part is the shape of the growth curve. Stripe’s AI 100 cohort (the 100 fastest revenue growing AI companies) is reaching $5 million ARR roughly 1.5x faster than the last great software archetype (the top 100 SaaS companies circa 2018). The generational pull of generative AI products is obvious in the cohorts founded after 2022, which sprint to early scale far faster than their pre-ChatGPT peers. In other words: when the product is natively AI, the go-to-market velocity looks different and better.

That acceleration doesn’t fade as companies mature. Across a broad sample of AI vendors, those in the $1–20 million bracket and those above $20 million have outgrown all-sector averages quarter after quarter since late 2023. The pattern is consistent: AI-first firms don’t just pop early, they keep outrunning their neighbors as they add products, expand seats and compound usage. For investors thinking about durability, that’s the signal to separate hype from habit.

Adoption is also doing something unfashionable for a bubble: it’s compounding. Card and bill-pay data from tens of thousands of U.S. businesses show paid AI adoption climbing from low single digits in early 2023 to north of 40% by September 2025.

Twelve month retention improved markedly versus 2022, while average contract values jumped from the tens of thousands in 2023 to the mid-six figures in 2025, with seven-figure deals in sight for 2026.

And the model leaderboard is, for now, a one-horse race: OpenAI remains the most-used provider by a wide margin, with Anthropic a distant second and everyone else in the peloton. Pilot projects are turning into line items, and budgets are following suit.

Meanwhile, applied AI companies are exiting the novelty phase and entering the boring revenue era. In audio, avatars and image generation, category leaders are now well into the hundreds of millions of annual revenue, with improving mix as enterprise accounts dominate alongside a very long tail. ElevenLabs doubled revenue in nine months to roughly $200 million and closed a multi-billion valuation; Synthesia crossed $100 million ARR with deep Fortune 100 penetration; Black Forest Labs is around the $100 million ARR mark on hefty margins, buoyed by large-ticket partnerships. These numbers point to potential generational businesses entering a new phase of growth.

So what did the MIT report get wrong? Not necessarily the diagnosis of early failure. It captured a moment when corporate AI was a science fair: scattered pilots, thin integration, and heroic promises made by teams without budget authority. Where it misleads is in extrapolating that snapshot to the present tense. The market has shifted from “play with a model” to “deploy a workflow,” from bespoke builds to opinionated products, and from experiments owned by innovation teams to opex owned by line managers in IT. When your unit of analysis moves from press releases to purchase orders, success rates look very different.

The policy takeaway is not pump the brakes, but do the plumbing. The companies showing real ROI are AI-first by design: they ship tightly scoped products, instrument the outcomes, and iterate with humans in the loop. The capital markets are beginning to price that operating discipline. In aggregate, the class now throws off tens of billions annually, with leading providers consolidating share as adoption broadens and standardizes around a small set of foundation models (chiefly OpenAI’s).

The bottom line for executives: stop litigating whether AI works in the abstract and start measuring whether your workflows do. If the pilots are stuck, it’s probably because they never left the lab. If they’re working, you’ll see it the old-fashioned way: in revenue that arrives faster than SaaS benchmarks, cohorts that retain better than they used to, and contracts that renew bigger than they started.

The market is moving on from demos to dividends. It’s time your roadmap did, too.

And now, here are this week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• Forbes: This Former SpaceX Engineer Is Using AI To Design Circuit Boards

• Bloomberg: The Real AI Risk is ‘Meh’ Technology That Takes Jobs and Annoys Us All

• Fortune: How business leaders can survive a ‘phenomenal’ AI bubble

• Wired: This Startup Wants to Spark a US DeepSeek Moment

• Politico: Dutch chips star exec slams EU for overregulating AI

• The Verge: OpenAI wasn’t expecting Sora’s copyright drama

• Fortune: AI came from tech, but the most advanced AI businesses are in every industry

• Wired: The Future of AI Isn’t Just Slop

• Axios: AI video apps are a scammer’s goldmine

• FT: How AI became our personal assistant

• MIT Technology Review: AI toys are all the rage in China—and now they’re appearing on shelves in the US too

• The New York Times: How Elon Musk Is Bringing Sexy A.I. Chatbots to the Mainstream

• The Information: Founders Fund Shifts From Caution to Concentrated Bets on AI

• The Verge: All of the updates from OpenAI DevDay 2025

• Time: How AI Is Changing White-Collar Work

• Wired: OpenAI Wants ChatGPT to Be Your Future Operating System

Read more

For those paying attention, short-form AI video has spent the last year quietly crossing the uncanny valley for the use cases that matter in a feed: punchy explainers, product demos, hype edits and memes, reaction clips, even photorealistic scenes stitched together with VFX-grade polish.

This is why the outside world is suddenly feeling the competitive heat. OpenAI lit the fuse with the original Sora model that composed half-believable scenes and camera moves from natural language, and now they’ve doubled down with Sora 2 and the Sora app. YouTube, which controls the world’s most valuable video shelf space, is racing to make creation as automatic as search: Veo 3 integration into Shorts; Edit with AI, which transforms raw footage into a first draft video; and Speech to Song, which turns dialogue from eligible videos into soundtracks. TikTok, the ultimate format innovator, is turning its Creative Center into a co-pilot for ads and creator content with tools such as Symphony because if a meme can be minted in minutes, the winner is the platform that collapses the time from idea to upload. Meta, meanwhile, has rebuilt its entire business around recommendation engines and is experimenting with an AI video feed called Vibes while it figures out how to bring AI video into Reels and ads workflows.

The subtext is the same across all four: the most valuable graph in media is no longer the social graph; it’s the intent-to-video graph.

The usual criticism that synthetic clips still “don’t look quite right” misses how today’s feeds actually work. The average attention span has dropped from 150 seconds in 2004 to only 47 seconds. Viewers aren’t grading Oscar reels; they’re skimming 6 to 20-second units where narrative clarity, pacing and style consistency matter more than pixel-peeping. Today’s video generators are more than capable to meet those constraints: stable characters, coherent motion, decent lip-sync, and compositing that rivals what junior VFX artists spend hours doing by hand. The corporate world is also adopting AI video: UBS is producing lifelike avatar videos of its analysts because clients prefer a two-minute clip over a 20-page PDF, and crucially, because the quality has cleared the threshold where no one bounces on sight.

If feeds are the battleground, ads are the logistics chain. That’s what makes AI video a dual-use technology in the best (and most lucrative) sense. For consumers, it lowers the cost of expression to nearly zero. For advertisers, it turns the old creative bottleneck into a faucet. Think of a single concept (say, a running shoe drop) branching into thousands of on-brand variants for different geographies, aesthetics, weather, and viewer histories. The creative itself can react: cut a morning-run edit if the user typically watches fitness content at breakfast; switch to a night-run mood with neon reflections if their watch history tilts cyberpunk. Dynamic Creative Optimization always promised this; AI video finally makes it cheap, fast and native to the feed.

That, in turn, reframes platform economics. If your ad system can promise not merely “reach the right person” but “manufacture the right ad for this person, now,” the value of your inventory goes up. But precision requires pipes. YouTube has the deepest library and advertiser tooling. TikTok owns the culture pipeline and the shortest distance between a meme and a market. Meta controls the largest cross-app identity graph and a ruthless performance ads stack. OpenAI doesn’t own a mature social surface, but it owns mindshare with creators and developers, and has pioneered a kind of new intelligence that is very sticky: computers we can talk to. Each company’s advantage becomes self-reinforcing as models learn from engagement signals unique to their platform.

This is why we should expect a video model competition that echoes (and eventually eclipses) the large language model sprint. The inputs are expensive (compute, data curation, and safety tuning) but the winners don’t just get better demos. They get tighter feedback loops: every frame watched, paused, or skipped is labeled training data. They get richer first-party datasets, at a moment when privacy constraints are starving third-party targeting. And they get switching-cost gravity: advertisers trained on your creative tools won’t casually move budgets; creators who build an audience around your model’s “look” won’t casually change styles.

AI video also collapses the old separation between “content” and “ad.” If a model can produce a compelling, creator-voiced review of a moisturizer and then swap the brand, price, and call-to-action on the fly, what do we call that? An ad you watched willingly? A video you bought from? Shoppable overlays, real-time personalization, and performance measurement live inside the same piece of media. The platforms that harmonize those layers (creation, collaboration, distribution) will pull spend away from tired channels and convert brand budgets into always-on, model-driven video factories.

None of this is frictionless. Model licensing will become thornier as rightsholders test the limits of training data provenance and opt out mechanisms. Watermarking and provenance standards will be table stakes if platforms want trust at scale; they’ll also be a defensive moat for whichever company’s watermark becomes the industry default. Compute costs might squeeze margins in the short run, especially for platforms subsidizing free creation to seed supply. But these are execution problems, not existential ones, and all four companies have the capital and incentive to brute-force solutions.

The existential question is different: when synthetic media becomes the norm, what is authenticity worth? The optimistic answer is that authenticity becomes a style, not a prerequisite: Shot on iPhone as an aesthetic choice in a world where most things aren’t. That’s survivable and monetizable. The gloomier answer is that commoditized aesthetics push creators and brands into a zero-sum game of novelty for novelty’s sake. But again, the feed sets the rules. Platforms will tune models for watch time, merchants will tune them for conversion, and the market will do what it always does with new creative tools: overshoot, correct, and then professionalize.

So, yes, we’re headed for an AI video race that rhymes with the LLM era, and it will be just as capital-intensive, just as noisy, and maybe more consequential for how the internet feels. The difference is that this race ends up directly in front of everyone’s eyes. Language models quietly change how information is produced. Video models might loudly change what we choose to watch. The companies that win won’t merely have better models; they’ll have better taste encoded in those models, a taste defined by the engagement of billions.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• The Information: The $10 Billion Enigma of Mira Murati

• FT: AI and Europe’s quest for a $1tn company

• Bloomberg: AI Video Clone Startup Launches New Tools

• Wired: Mira Murati’s Stealth AI Lab Launches Its First Product

• Bloomberg: AI Data Centers Are Sending Power Bills Soaring

• Business Insider: I spoke to 7 executives this month. Here’s how they’re using AI at work.

• The Economist: A $2bn AI unicorn tests London’s nerve

• FT: How Nvidia’s Jensen Huang became AI’s global salesman

• FT: AI groups bet on world models in race for ‘superintelligence’

• WSJ: OpenAI’s New Sora Video Generator to Require Copyright Holders to Opt Out

• The Verge: How generative AI boosters are trying to break into Hollywood

• CNBC: How a ‘nudify’ site turned a group of friends into key figures in a fight against AI-generated porn

Read more

If I were to summarize my remarks in one sentence, it would be: Europe’s AI bottlenecks are boring, but we should fix them anyway. For those interested in the longer version, watch the video below (featuring my spanking new Express-2 avatar courtesy of Synthesia) and read on for some additional thoughts.

If the EU wants AI that scales, it doesn’t need another moonshot strategy or piece of regulation; it needs plumbing. The technical bottlenecks throttling European startups aren’t mysteries. In fact, they were there before generative AI and they’ll continue to exist long after, if left unaddressed. They are predictable byproducts of fragmented infrastructure, slow money, and standards that live in PDFs instead of procurement.

Hacktivate AI was particularly interesting because it brought together industry and academia to brainstorm some real solutions. My working group was made up of the Startup Coalition, Hugging Face, AI Lithuania, Volkswagen, Nvidia, the French Ministry for Finance and the Economy, appliedAI, Sanofi, SAPIE and Synthesia and we presented four policy moves that would actually move the needle on AI adoption.

1. Build a “mutual-aid” market for compute and data

Right now under the EuroHPC rules for accessing European AI supercomputers, early-stage teams are asked to predict their compute needs before they even have product–market fit.

This is red tape dressed as capacity planning.

Instead, we should create a structured, EU-backed exchange where startups can trade harmonized open datasets, labeling/clean-up work, and contribution credits for time on these public AI factories.

Think of it as a “take-a-euro, leave-a-euro” jar, except the euros are clean data and compute credits, with Commission-provided templates to keep the inputs interoperable and trustworthy .

The payoff: less gatekeeping upfront, better utilization of expensive clusters, and a continent-wide incentives loop to produce high-quality domain data (climate, health, social science) that startups can actually use today, not after a year of negotiations.

2. Replace the grant maze with a layered scale-up pipeline

Europe excels at funding research; it struggles to turn papers into products. The EU can map the jumble of programs (advice, grants, compute, capital) to the real startup journey: simple compliance guides and no-code tooling at the start; pooled compute and EIB-backed growth capital when traction appears; and “right-sized” interventions at every intermediate step. In other words, a conveyor belt—not a scavenger hunt .

This is less about new money than choreography: route founders through a single intake and progressive tiers of support that expand as technical and commercial readiness does.

If the EU wants globally competitive AI firms, it needs to stop tinkering at the edges with vanity projects like the EuroStack and institutionalize the handoffs that Silicon Valley leaves to chance.

3. Make funding agile: short apps, fast clocks, delegated decisions

AI moves too quickly for 50-page applications and nine-month verdicts. Stand up an Innovation Fast Track that uses: (a) 10-page concept notes for MVP pilots; (b) proportional due diligence so oversight scales with grant size; (c) trusted intermediaries (accredited accelerators, venture funds, and innovation hubs) to re-grant with clear audit trails; (d) expert “rapid review” panels by sector; and (e) always-open calls with monthly cutoffs so companies can apply when they’re actually ready .

Two-month funding decisions should be the rule, not the outlier. Publish hit-rate and adoption metrics so everyone can see what works and sunset what doesn’t .

4. Move standards from committees into contracts

The EU AI Act is a lemon for many European startups that often do not have the legal expertise to navigate complex regulation. But when life gives you lemons, you make lemonade.

If we want faster, safer AI adoption in hospitals, courts, and factories, we should buy only what meets a clear bar.

Fast-track European adoption of the ISO/IEC 42001 AI management standard and turn it into a “golden standard” that public buyers and highly regulated industries such as financial services or healthcare/pharma use by default. Ask CEN CENELEC to prioritize this work (they’ve been painfully slow at it) and then have EU institutions lead by example in procurement based on the standard. Nothing spreads interoperability and risk discipline faster than purchase orders that demand it.

Of course, none of this is particularly groundbreaking. It’s infrastructure: capacity sharing, program orchestration, agile capital, procurement-grade standards. But these are the levers that shrink the translation gap between research and revenue.

Europe can keep writing white papers about strategic autonomy, or it can clear the on-ramps that let founders actually build. Choose to do the plumbing.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• MIT Technology Review: How AI and Wikipedia have sent vulnerable languages into a doom spiral

• The Guardian: ‘Tentacles squelching wetly’: the human subtitle writers under threat from AI

• The Information: How Anthropic and OpenAI Are Developing AI ‘Co-Workers’

• Wired: Distillation Can Make AI Models Smaller and Cheaper

• TechCrunch: Scott Wiener on his fight to make Big Tech disclose AI’s dangers

• Harvard Business Review: AI-Generated “Workslop” Is Destroying Productivity

• FT: Nvidia’s $100bn bet on ‘gigantic AI factories’ to power ChatGPT

• The Verge: Trump’s H-1B chaos is Europe’s opportunity to steal tech talent

• FT: America’s top companies keep talking about AI — but can’t explain the upsides

• Wired: YouTube Thinks AI Is Its Next Big Bang

Read more

But I have to confess I’m a bit confused. How did these bastions of liberal values fighting the “good fight” against the evil empires of technology transition from stern op-eds in the New York Times and The Guardian to joining the MAGA mediasphere?

Allow me to present to you a tragedy in three acts.

Act I: Elon and the doomsday media beat.
We start with Elon Musk, the metronome of media attention. Years back, he warned that AI could be “more dangerous than nukes,” and the press obliged with the usual fevered coverage. “AI will kill us” is catnip to SEO traffic-obsessed news desks; Elon Musk delivered clean headlines and moral urgency, and the halo of a visionary founder. I’m using the past tense because now he mostly delivers incoherent rants on giant screens at weird rallies.

Many of his initial AI doomer takes were soft on evidence and heavy on vibes, yet they set the tone. If a household-name billionaire says the robots are coming, you don’t worry about the footnotes; you clear up the front page and push a media alert in the news mobile app.

Act II: From research labs to “war rooms.”
Next came the hardcore AI doomers, people like Max Tegmark or Eliezer Yudkowsky who quiver at the sight of large auto-regressive models. Tired of being interviewed by “MSM” journalists for glowing profiles in Wired and given main stage slots at tech conferences, they started toying with ideologically charged venues like Steve Bannon’s WarRoom. The move makes strategic sense if you believe policy is downstream of airtime.

But it also marks a break from the academic grounds of MIT or Berkeley they once claimed as their home turf. When existential risk talk jumps from white papers to culture-war studios, the message doesn’t merely reach a new audience; it changes.

“Pause AI” morphs into trolling in front of Google DeepMind's office for likes on X.

Act III: Endorsements with January 6 baggage.
Then there’s the more explicit political alignment from another class of AI grifters - the skeptics. They’re busy endorsing politicians who made their names fighting platform moderation and fact-finding institutions, not building them.

For example, Ed Newton-Rex fashions himself as a hardline copyright advocate writing op-eds in The Guardian and appearing on Time’s list of most influential people in AI. He also has an alliance of creative industry representatives hanging on his every word.

But what his loyal fans might not know is that he’s recently been hard at work publicly boosting speeches from US senator Josh Hawley. Hawley isn’t just another conservative vote; he’s a symbol tied indelibly to January 6. You don’t get to shrug off that association with a “both sides are captured by Big Tech” narrative. Promoting his views is a choice. It says: my allegiance is with the political figure who raised a fist to a crowd that later stormed the Capitol.

I asked Mr Newton-Rex about his newfound love for those on the right of the American political establishment and about some of his other conflicts of interest. He chose to block me while complaining about the "fragility of AI CEOs" and being blocked by Demis Hassabis.

This tragedy didn’t happen in a vacuum. There were several forces that did the bending.

First of all, AI grifters provide an endless supply of frictionless content. It’s dramatic, universal, and requires no prerequisite knowledge. You can do a hit-piece, run a moody stock photo of a glowing robot eye, sprinkle in “Yudkowsky” or “paperclip maximizer,” and call it a day. In an era when traffic targets wag the editorial dog, journalists are rewarded for stories with top-hat wearing weirdos. “Doom” beats “domain-specific safety engineering” in every headline A/B test ever run.

Secondly, when you’re a reporter on the tech beat, access is alpha. AI grifters are unusually willing to take calls, grant interviews, and gift journalists the comfort of certainty: “This is existential,” they keep saying. That kind of clarity is irresistible when you’re staring down an editor who wants a take by 5 pm. The more airtime they got, the less they had to defend the scaffolding of their arguments.

Finally, the far-right media ecosystem is a machine for converting disaffection into influence. It thrives on elite betrayal stories, and the AI grift slides neatly into the groove: coastal elites built a god-computer that will replace you; the experts are lying; only my non-profit/public benefit corporation can stop them. The alignment feels inevitable the moment we move away from rational conversations to moral panic and expletives.

So what now for mainstream outlets that platformed these voices, giving them op-eds, expert quotes, and sympathetic profiles on the premise that the authors were issue-driven rather than ideology-bound?

Imagine being the editor at The Guardian who greenlit the columns from Ed Newton-Rex and finding out that he is now aligning himself with a January 6-adjacent standard bearer. Do you run a note to readers? Do you reassess your roster? Or do you keep quiet and hope nobody notices that yesterday’s “AI truth-teller” is today’s culture-war proxy?

Regardless, if a columnist’s advocacy predictably funnels into a movement that celebrates institutional demolition, that’s not mission creep, that’s the mission revealing itself. The question for editors isn’t whether to ban people with right-wing views (they shouldn’t!); it’s how you can stop validating a pipeline that is a profit machine for the AI grift.

For those who are still confused about how that pipeline works, the sleight of hand is simple: start with abstract existential risks (unaligned superintelligence, artists’ rights, etc.), translate them into immediate policy demands (pause research, expand liability, empower regulators), then hand the microphone to politicians whose broader agendas are explicitly hostile to expertise and pluralism. You get the optics of high-minded caution and the outcomes of a panic-driven politics that rarely builds anything, let alone safe AI.

Meanwhile, the actual work (mitigating consumer harms or building enterprise-grade security) has to fight for oxygen against a rolling Chyron of apocalypse. If you’re an honest researcher trying to craft rules for model access or benchmark leakage, you’re crowded out by people talking about the heat death of humanity on podcasts that also push gold coins and male vitality supplements.

If you want a fix, start by treating catastrophic claims like biotech claims: give us extraordinary evidence or you will never see the front page. Then, prioritize researchers who publish in academic, peer-reviewed conferences over influencers who ship vibes. And when advocates pick teams in the culture war, update priors publicly.

The story of AI is still being written. But the first draft has to stop mistaking panic for rigor and celebrity for credibility. Otherwise, we’ll keep retracing this loop: doom, platform, pivot, backlash, repeat. And each time, the politics will get a little louder, the policy a little dumber, the public a little angrier, and the technology a little less governed.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• The Guardian: What is new in UK-US tech deal and what will it mean for the British economy?

• The Verge: Americans want AI to stay out of their personal lives

• Sifted: Pleo CTO Meri Williams on why AI won’t replace junior engineers

• The Hollywood Reporter: The AI Movie Factory Is Ramping Up

• MIT Technology Review: How do AI models generate videos?

• WSJ: The City Leading China’s Charge to Pull Ahead in AI

• Bloomberg: Are Trampoline Bunnies and Dog Podcasters the Future of Entertainment?

• Sifted: Life at Lovable: cracked hires, 17-year-old coders and no shoes

• Business Insider: A key type of AI training data is running out. Googlers have a bold new idea to fix that.

• Fortune: OpenAI’s ChatGPT dominates personal uses, while Anthropic’s Claude has the edge in business, dueling usage studies show

• Sifted: Aleph Alpha: Can Germany’s one-time OpenAI challenger bounce back?

• Business Insider: Meet Silicon Valley's new AI startup whisperer

Read more

We’re starting to see that conventional wisdom break, first with GPT-5 which was not a scaled model, despite what a clueless Gary Marcus claimed in his amateur hour op-ed for the New York Times.

Then along came K2 Think, a 32 billion parameter model from MBZUAI and G42. Built on an open Qwen2.5-32B base and released with code and weights, K2 Think shows that careful post-training, smart test-time compute, and hardware-aware inference can get you remarkably close to today’s frontier on math and hold its own in code and science, without a trillion-parameter physique or a FAANG-sized budget.

K2 Think is a system based on a six-pillar recipe that was developed in three phases.

Phase one started with long chain-of-thought supervised fine-tuning (SFT) on curated traces (AM-Thinking-v1-Distilled). Early training moved the needle fast on competition math (AIME 2024 ~79.3% pass@1; AIME 2025 ~72.1%), then plateaud, which provided the team with a useful signal for where to lean next.

Phase two involved reinforcement learning with verifiable rewards (RLVR) across six verifiable domains via the Guru data set. This data set includes ~92k prompts spanning math, code, science, logic, simulation and tabular tasks, implemented with GRPO in the verl library. RL from a strong SFT checkpoint brought modest absolute gains; the same RL recipe from the base model grew much faster on AIME, hinting that heavy SFT can constrain exploration.

Phase three was about test-time improvements, focused on two specific aspects: first, a “Plan-Before-You-Think” approach where an external planning agent pulls out key concepts and drafts a high-level plan that’s appended to the prompt. Then, Best-of-N sampling which generates multiple answers and picks the best via pairwise LLM judging. After experimentation, N=3 provided the cost/benefit sweet spot. Alone, BoN delivers most of the lift; combined with planning, it adds 4–6 points across hard math benchmarks.

Additionally, the technical report describes how the team used speculative decoding and the Cerebras Wafer-Scale Engine (WSE is a massive wafer-scale integrated processor for AI training and inference) to achieve unprecedented inference speeds for K2 Think.

WSE delivers approximately 2,000 tokens per second per request, representing a 10x improvement compared to what Nvidia H-class GPUs can deliver. This dramatic speed-up fundamentally transforms the practical usability of long chain-of-thought reasoning, turning minute-long waits into seconds and making BoN + planning feel interactive, not academic.

Put differently: the team compressed model size and moved the “scale” to where it pays: the post-train recipe and the test-time budget.

Okay, but does it work? The math benchmarks say yes. On composite math (AIME 2024/2025, HMMT-25, Omni-MATH-HARD), K2 Think posts a 67.99 average, ranking it at the top among contenders in its open weights class and it even edges bigger open models like GPT-OSS 120B (67.20). The point isn’t that size never matters, it’s that careful post-training and test-time scaffolding can reclaim a lot of the curve.

This isn’t a one-trick contest crammer: on LiveCodeBench, K2 Think hits 63.97; on SciCode, 39.2/12.0 (sub/main); on GPQA-Diamond, 71.08. That’s competitive or better within its class and in striking distance of much larger systems on science Q&As.

Two pragmatic results jumped out to me:

• BoN vs. planning: BoN does the heavy lifting; planning adds a meaningful but smaller delta. The combination wins: e.g., AIME 2024 improves from 86.26 (SFT + RL) to 90.83 (Plan + Bo3); similar gains are achieved on AIME 2025, HMMT-25, Omni-HARD.

• Planning trims verbosity: structured prompts led to shorter outputs (up to ~12% token reduction) and better answers. That’s rare: more thinking, fewer tokens.

There’s also a sobering takeaway: constraining max response length during RL hurt performance and didn’t recover with more training, evidence you can’t squeeze the chain-of-thought and expect the same reasoning paths to re-emerge later.

Until recently, the dominant mental model was “scale or stale.” K2 Think offers a different trade: a 32B open model that’s fast to iterate, accessible to research labs and startups around the world, and extensible with high-quality data and a few hundred GPUs rather than a sovereign compute budget. Because weights and code are out, anyone can extend the SFT/RLVR recipe, tweak the planning agent, change the verifier, or push N above 3 if their latency budget allows. The result is an R&D loop that is closer to software engineering than mega-pretraining: you can measure, ablate, and ship.

Hardware is part of the story. BoN and agentic orchestration are only fun if you’re not waiting around. The WSE’s on-chip weights and memory bandwidth make best-of-3 feel cheap; 32k-token chains go from “coffee break” on H100s to ~16 seconds. That unlocks interactive, human-in-the-loop reasoning workflows, exactly where open models can differentiate.

There are some caveats, of course. K2 Think’s “Safety-4” macro score sits at 0.75: strong on refusing harmful content (0.83) and conversational robustness (0.89), weaker on cybersecurity/data-protection (0.56) and jailbreak resistance (0.72). Translation: solid baseline, but indirect/jailbreaky attacks remain a to-do, as we saw a few hours after release.

Secondly, RL from a strong SFT checkpoint gained only ~5% on AIME 2024 versus ~40% when starting RL from the base model, evidence that heavy SFT can “lock in” habits and limit exploration.

Finally, the paper’s method is admirably explicit (16× pass@1, shared decoding hyperparams), but test-time scaffolds like BoN and LLM-as-judge have operational costs and failure modes. The good news is that K2 Think’s open release lets the community reproduce (or break) these claims.

K2 Think won’t dethrone the very largest proprietary models across every domain, but it doesn’t have to. It demonstrates that parameter-efficient reasoning is real, that test-time compute is a first-class lever, and that hardware choices can flip the UX from batch to interactive.

In practical terms, it lowers the barrier to entry: small labs or startups can start from 32B, adopt the six-pillar recipe, and specialize (math today, code or science tomorrow) without rebuilding a pretraining pipeline from scratch. That unlocks a more competitive, more pluralistic supply chain for reasoning models.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• WSJ: AI Startup Founders Tout a Winning Formula—No Booze, No Sleep, No Fun

• TechCrunch: Box CEO Aaron Levie on AI’s ‘era of context’

• The Verge: Sierra CEO Bret Taylor on why the AI bubble feels like the dotcom boom

• The Information: Are AI Valuations Bonkers?

• Wired: Microsoft’s AI Chief Says Machine Consciousness Is an 'Illusion'

• FT: Why is AI struggling to discover new drugs?

• Business Insider: Meta is paying contractors up to $55 an hour to build AI chatbot characters

• The New York Times: United Arab Emirates Joins U.S. and China in Giving Away A.I. Technology

• Bloomberg: Mistral Emerges as Last, Best Hope for European AI Contender

• WSJ: Ralph Lauren Has Entered the AI Age

• The Verge: Sal Khan is hopeful that AI won’t destroy education

• CNBC: Behind the AI talent war: Why tech giants are paying millions to top hires

• The Verge: MAGA populists call for holy war against Big Tech

Read more

For me, generative AI’s best trick isn’t replacing people, it’s compressing the cost of quality. For many years, the creative industry (whether that means music, film or other forms of art) had a dirty secret: who you knew often mattered as much as what you knew.

So when I called generative AI the “great equalizer” in my conversation with Hannah Murphy from the FT, I was thinking exactly of Seraphinne Vallora and how polished, studio-quality content at scale won’t suddenly be a Fortune 500 privilege. Used well, generative AI lets smaller brands compete on experience, tell better stories, and reach audiences they couldn’t touch before.

But there’s a meaningful line between influencers creating with AI and influencers who are AI. On one side are human creators wielding new tools. Here, a great example is journalist-creator Sophia Smith Galer building an app that helps turn written work (news articles, essays) into snappy vertical videos for Instagram and TikTok. That’s augmentation: translating ideas into formats audiences actually consume.

On the other side are faceless, nameless, auto-generated personalities pumping out content optimized for engagement first, everything else second. We’ve already seen AI personas rack up followers before the reveal (“Mia Zelu” posting from Wimbledon has entered the chat!), and brands are experimenting with AI clones and “digital twins.” It’s clever, yes. It’s also where trust can go to die if disclosure and quality control don’t keep up.

The market is testing both models at speed. Nearly three quarters of marketers surveyed by the Influencer Marketing Hub believe influencer marketing can be automated by AI. The pitch is straightforward: low cost, perfect control, infinite availability. But that control cuts both ways: consumers don’t like being fooled, and agencies are still wary about disclosure norms that haven’t caught up.

Until now, only companies with deep pockets could mass-produce slick video and hyper-localized creative. That moat is evaporating. If a retailer wants a hundred product explainers in ten languages, or a bank wants every report summarized on video by the morning commute, AI makes that feasible without a soundstage. Big brands are already testing the edges: H&M’s AI “digital twins,” Hugo Boss’s work with virtual influencer Imma, and a growing cottage industry of virtual-persona studios and “AI talent agencies.” Whether you love it or hate it, the capability is here, and it no longer belongs exclusively to the top of the market.

Capability however isn’t the same as connection. Early data suggest human creators still outperform on the metric that actually matters: attention that converts. Sponsored posts from flesh-and-blood influencers see 2.7x the engagement of AI personas, and they command far higher fees as a result. Why? Because lived experience translates. AI can simulate a persona; it can’t be a tired parent, have a skin condition, or taste the dish it’s recommending. That gap shows up in the numbers, and in the gut.

Here’s the awkward part. If your business model is “replace people with AI,” I wish you well, from a distance. (That distance gets infinitely larger specifically when it comes to the creeps, weirdos and ghouls building teenage-looking AI girlfriends or boyfriends.) The creator economy didn’t take off because humans were an inefficiency; it took off because audiences were tired of brandspeak and craved human judgment, taste, and accountability. The moment platforms and brands replace the human connection with fully synthetic engagement at scale, you invite perverse outcomes: influencer fraud (humans secretly swapping in AI), disclosure messes, and a flood of bizarre, low-quality spam designed to harvest clicks. We’re already seeing that “fast-food content” dynamic appear and the backlash won’t be subtle.

Contrast that with tools designed to amplify creators and operators. Translate a video into five languages while the creator sleeps? Great. Auto-answer FAQs so a solo shop can keep pace with DMs? Even better. Create safe, transparent workflows for brands to produce explainers, training, and product docs in minutes instead of weeks? That’s the point. Companies like Synthesia position themselves as infrastructure for this: give every business, large or tiny, a platform to make more engaging, informative content for knowledge sharing and entertainment, not to push humans off the stage. That’s the path that compounds trust.

So don’t confuse a production breakthrough with a creativity substitute. The winners will be the humans who use AI as leverage, who learn prompt craft (I hate calling it “engineering”), iterate faster, and take advantage of new formats and capabilities developed by AI. The losers will be the brands that outsource taste to an AI model, replace traditional photoshoots with half an hour’s work in Midjourney, and call it a day.

Platforms should require clear labeling and crack down on deceptive botfluencers. Brands should demand disclosure and favor creators who use AI transparently. And founders should build copilots, not clones. If the first slide in your pitch deck says “fire your creative team and replace them with AI,” enjoy the race to the bottom. If it says “arm every team with an AI studio in their browser,” you’re building for a world where the best ideas, not the biggest budgets or the hottest contacts, finally get their shot.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• MIT Technology Review: Synthesia’s AI clones are more expressive than ever. Soon they’ll be able to talk back.

• Fortune: Inside the Anthropic ‘Red Team’ tasked with breaking its AI models—and burnishing the company’s reputation for safety

• Bloomberg: Study of 67,000 Job Interviews Finds AI Outperforms Human Recruiters

• Reuters: Chinese firms still want Nvidia chips despite government pressure not to buy, sources say

• FT: The rise of the AI influencer

• WSJ: Cutting-Edge AI Was Supposed to Get Cheaper. It’s More Expensive Than Ever.

• FT: Why AI labs struggle to stop chatbots talking to teenagers about suicide

• The Verge: Is AI the end of software engineering or the next step in its evolution?

• AP: The success of AI music creators sparks a debate on the future of the music industry

• Semafor: How AI will upend the news

• Wired: Meet the Guys Betting Big on AI Gambling Agents

• The New York Times: How ‘Clanker’ Became an Anti-A.I. Rallying Cry

• FT: Computer scientist Geoffrey Hinton: ‘AI will make a few people much richer and most people poorer’

Read more

I spoke to three of the winners from my home country of Romania: Tudor-Ștefan Mușat, who got a perfect score on one of the hardest problems in the contest and clinched one of the gold medals, his teammate, Tudor Morariu, who brought home silver, and Ștefan-Alexandru Asandei, a bronze medalist. Together, they were part of the two teams that won eight medals for Romania, a country with fewer GPUs than Beijing has high-rises.

IOAI is a global competition that began last year, modelled after other science olympiads such as the International Maths Olympiad (IMO) or International Physics Olympiad (IPhO) that I dabbled in at the regional level as a kid.

The AI Olympiad included two six-hour individual sessions (according to Musat, “it’s the longest format I have ever faced”), with submissions that could take up to 10 minutes to re-evaluate. That forced Musat, Morariu and Asandei to adopt the kind of scarce-resource thinking that Eastern Europeans excel at: get a sturdy baseline, then make surgical improvements instead of YOLOing moonshots.

“In such a broad and uncertain environment, not every idea or solution is guaranteed to work, and it’s easy to lose track of time. My approach was straightforward: I focused on securing as much of the baseline as possible, then built and refined my own code on top of that foundation. This way, I managed my time effectively while still leaving room for improvements and experimentation.” - Tudor Stefan-Musat

Asandei’s strategy was also very simple:

“I applied the same approach I used in past contests. I first read all the tasks and I started with the one I was most confident about. After that, I would change tasks based on inspiration for any new ideas or approaches.” - Ștefan-Alexandru Asandei

And the tasks weren’t “prompt-engineering theater.” They ranged from classifying antique paintings in high-dimensional space to building object-detecting robots, and even counting chickens from aerial imagery.

On the paintings problem, Mușat did what top practitioners do in the real world: he treated unlabeled structure as signal. He projected a 5-D dataset down with PCA, spotted a spiral with two distinct heads, hand-separated the classes, then tightened the decision with t-SNE, which got him a perfect score, better than anyone else on the board.

Morariu’s edge was similarly un-glamorous and very effective: with tiny datasets (hello, chickens), classic train-test splits are noise amplifiers. He let loss functions like MSE and cross-entropy guide iteration, and leaned on segmentation tricks he’d seen work before. According to him, “experience beats novelty.”

Asandei’s favorite task was Pixel Parsimony, where he had to find the optimal crop box in an image to compress the most essential information about an image, like where the subject is. He tried various techniques and felt like creativity played an important role in his success.

Why did the AI giants whiff on gold? Partly because this is still a nascent high-school discipline; much of modern AI is graduate-level literature, and luck looms large when the environment is unfamiliar. Asandei believes the US doesn’t take Olympiads seriously at the local level while Chinese students struggle with independent learning outside the classroom:

“Since AI is such a new topic, students have to learn on their own a lot, from the required mathematics and coding, to machine learning and deep learning. I think China has a strong and centralized school system, and students are not used to learning topics on their own, however I’m sure this will change as they are already integrating AI into their curriculum. In the US, they lack a culture of Olympiads. In Romania, we are used to at least three to four phases for every competition: local, regional, national and team selection. This way students learn hard work and study independently.” - Ștefan-Alexandru Asandei

The IOAI rulebook also neutralized a common crutch: contestants had access to a token-limited GPT-4o inside a sandboxed Bohrium environment, with the broader internet blocked. Although Mușat found Bohrium unreliable, prone to buggy code and repetition, sometimes choking on longer outputs, the net result was that there was a tax on prompt-dependence and a reward for reasoning (the human kind, not the AI model variety).

Romania sent two teams drawn from a national selection that starts with ~30 students and whittles to eight after a camp of multi-day contests. The culture is IOI/IMO-hard, and it shows. Mock problems tracked the real thing uncannily well this year, and the mantra Mușat would hand next year’s cohort is the most old school one in AI: read, code, test. Read papers. Implement ideas yourself. Measure relentlessly. It’s not very glamorous, but it’s repeatable.

For example, in the Chicken Counting challenge, Asandei’s solution was based on the discoveries he made while working on the at-home task. The task requires the development and training of a neural network to count the number of chickens in a photo:

“The data was nothing special, just the expected images. The essential observation however, was for the data extraction: the latent vectors representing image features. For this kind of image analysis, I decided to use a UNet network. For better feature engineering, I added attention modules. The training was done in a quick manner, as I got optimal results in under 20 epochs with cosine decay learning rate scheduling, gradients and logits clipping, and AdamW as the optimizer. The great flexibility of the network enabled the straightforward training, otherwise it could have taken much longer to tweak the right hyperparameters.” - Ștefan-Alexandru Asandei

In the team robotics challenge, Romania didn’t win the podium but still built a pathfinding stack that “just worked.” That matters in industry more than leaderboard glitter.

“The team challenge was very different from the rest of the competition. We had to program a robot that could identify objects, pick them up, avoid obstacles, walk up to tables, and place the objects down. Even though we didn’t finish among the top teams, it was a great experience to work on a real, physical robot. I think what we did especially well was our pathfinding algorithm, which reliably avoided all obstacles.” - Tudor Stefan-Musat

Even though winning eight medals doesn’t mean Bucharest will suddenly become the new Mountain View or Palo Alto, IOAI is a stress test of individual capability under constraints, not a proxy for national data center capex.

There’s a policy signal here for countries chasing “AI sovereignty” via budgets alone: talent pipelines beat trophy GPUs. Teach students to reason with limited tooling, force them to ship within time budgets, and normalize solving problems without off-the-shelf model kits—and you might get the next DeepSeek team. IOAI’s partial block on Hugging Face did exactly that, and Romania benefited.

Asandei’s advice for teams participating next year?

“Do not overthink. In hindsight, I overprepared for the tasks by learning techniques that were too complicated. The format of the contest is three tasks in six hours, which leaves less than two hours per task, excluding evaluation time. This requires the solutions to be short enough to be able to write, test and debug in that time.” - Ștefan-Alexandru Asandei

I asked Mușat what he remembers most from the award ceremony and he said it was the tension of waiting through the silvers, and the relief when his name didn’t get called:

“The moment I’ll never forget is the award ceremony. It was incredibly tense and full of excitement because none of us knew the results until the very end. I remember waiting eagerly, with my heart racing, as they announced the silver medalists. When my name wasn’t called, I felt this huge wave of relief — I knew I had secured a gold medal […] which was such a proud and unforgettable moment for me.” - Tudor Stefan-Musat

When the internet is off and the clock is on, AI looks a lot less like a parameter arms race and a lot more like craft under pressure. Romania’s eight medals give me hope that the scarcest resource in AI isn’t Nvidia GPUs, it’s still talent.

Mușat is heading to university, eyeing startups and real-world applications of AI. When I asked Morariu if he remembered any notable bugs in the coding environment, he reported, deadpan, there were no mosquitoes on campus (according to him, “actually, it was a feature, not a bug!”). Asandei wants to continue learning more about AI and maths, and is considering a career as a researcher in this field: “I believe AI can tell us so much about how brains work, how to understand and process our world from a mathematical point of view.” Also, they all agreed the Peking duck and Wonton soup were good. (Fun fact: Romanians love soup.)

My last question was about a common myth in AI that they’d want to dispel for anyone reading this newsletter, to which Morariu gave a very short answer: ChatGPT does not actually “think.” Sam Altman and Dario Amodei, please take notice — the kids are onto you.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• WSJ: These AI-Skilled 20-Somethings Are Making Hundreds of Thousands a Year

• The New York Times: The A.I. Spending Frenzy Is Propping Up the Real Economy, Too

• MIT Technology Review: Meet the researcher hosting a scientific conference by and for AI

• Business Insider: Meta CEO Mark Zuckerberg is trying to poach your best employees. Here's how to stop them from leaving.

• Sifted: The AI bubble: fuelled by an ARR obsession, undone by churn?

• The New York Times: 8 Women, 4 Bedrooms and 1 Cause: Breaking A.I.’s Glass Ceiling

• FT: Inside DHL’s AI upgrade: ‘Love it or hate it, you have to work with it’

• WSJ: Silicon Valley Launches Pro-AI PACs to Defend Industry in Midterm Elections

• TechCrunch: How chatbot design choices are fueling AI delusions

• Bloomberg: AI Is Replacing Online Moderators, But It's Bad at the Job

Read more

Instead, I got into a brief exchange with someone who is not working in the technology industry but had a negative worldview on the environmental impact of AI shaped by some of the more alarmist press coverage.

“AI companies are boiling the planet and killing our waters” is the type of reaction you get when you’ve been fed a steady sugar rush of villains through clickbait headlines. That’s become increasingly clear with so much climate-and-tech discourse right now: we’ve traded inquiry for vibes, nuance for outrage, and analysis for whatever word pairs best with “crisis.”

If we wanted answers instead of social media adrenaline, the AI and energy + water conversation would look very different.

Let’s start with the basic physics of data centers: the buildings that keep your photos, emails and AI models alive. There’s no doubt: they run hot. Servers turn electricity into computation and then heat; you must dump that heat somewhere. Many facilities use evaporative cooling: spray water, evaporate it, carry the heat away. The industry often talks about a metric called water-use effectiveness (WUE), which is liters of water per kilowatt hour (kWh) of compute-related energy. Typical WUE across data centers hovers around 1.9 liters/kWh. That can sound alarming, until you notice the spread. Best-in-class sites hit far lower numbers, and air- or refrigerant-cooled designs can use nearly none on-site. Recently, some operators have reported WUEs of about a tenth of the typical figure, and Microsoft has rolled out a design which involves zero water use for cooling.

Then there’s off-site water. Even if a data center sips little or no water on its own property, the power plant feeding it often doesn’t. In the US, the average water intensity of electricity is roughly 2.18 L/kWh, and the mix swings wildly by technology and geography. Hydropower, the clean hero in so many narratives (including Canada, where I’m currently writing this newsletter from), can be very water-intensive if you count reservoir evaporation. So you can run a “dry” data center and still be indirectly “wet” through your grid connection, or vice versa.

Scale deserves honesty, too. North American data centers are estimated to use 1.5–1.7 billion liters of water per day directly. Globally, data centers may consume about half a trillion liters per year, and demand could double by 2030 as AI adoption grows. These numbers are definitely not rounding errors but here comes the part that you rarely get to read about in the media.

First of all, not all workloads are equal. Storing your photos and email (call it “cold-ish” storage) is less power-dense than training a frontier AI model or running high-intensity inference. In cooler climates, storage can often be cooled with minimal or zero water. But storage runs 24/7 and it’s typically replicated (often three copies for durability) so the total footprint scales with how much stuff we keep forever. The long tail of “just in case” data is a quiet resource hog. Also, the engineering and procurement are changing. While some operators such as Microsoft now avoid evaporating water entirely, others are switching to reclaimed or industrial water where possible.

Secondly, if you want another (better?) villain, I can help you. In 2023-2024, water companies in England and Wales alone lost one trillion liters from leaks (and dumped huge amounts of sewage into our fresh waters, but that’s a topic for another day). The worst performer in England leaked more than 200 billion liters in a year, equivalent to almost half of the global water consumption of data centers in that same timeframe. So if we were to plug all the water waste in just one country in Europe today, we’d have enough liters to satisfy the projected demand for the entire data center sector five years from now.

If you want a better internet, don’t engage in performative or illiterate activities like not using AI or deleting your old emails. Here are four suggestions that might actually have systemic impact:

1. Stop treating “hyperscalers” as a single cartoon villain. A warehouse cooled with reclaimed water in a cool, windy region using wind and solar is not the same as a desert facility flashing evaporative coolers while drawing from a thirsty, fossil-heavy grid. Lumping them together is how we reward greenwashing and punish genuine improvements.

2. Ask for the boring numbers. If you’re a large enterprise doing significant business with a cloud, SaaS, or AI service, look for the WUE numbers (and PUE, the energy companion metric). Prefer providers that publish these figures, site in cooler climates when possible, run on low-water energy (wind/solar), and use reclaimed water for cooling. Net zero slogans are great; actual intensity metrics are better.

3. Aim critiques at the system, not the selfie. Your marginal gigabytes won’t retire a server. Deleting a few emails is fine household hygiene, not climate policy. The big levers live with providers: cooling choice, siting, and grid mix. That’s where regulators, customers with buying power, and investors should push, through procurement standards, disclosure rules, and siting incentives that reward low-water, low-carbon designs.

4. Still clean up the junk in a smart and sensible way. Because storage does run constantly and is replicated, deleting truly unnecessary data has a real (if modest) cumulative effect. Think of it as flossing for infrastructure: won’t change your life today, will make the whole system healthier over time. If you’re a media organization (a heavy user of storage in data centers), start with your own vast archives before you ask the public to delete their personal data.

So is generative AI “worth” the environmental footprint? Reasonable people can disagree. I believe it does but of course I have my own biases. Perhaps the right way to decide isn’t by reaching for the nearest pitchfork. It’s by asking the right questions and carefully analyzing the answers rather than looking for simple explanations that sound too good to be true. And if we don’t like the answers, we should fix the inputs (procurement standards, plant design, siting) and not pretend that our performative uninstalls are climate strategy.

When we flatten complex tradeoffs like tech-related energy or water usage into “AI bad, hyperscaler worse,” we start having large blindspots and misfire our outrage. We also let the real culprits skate by while we yell at AI video companies.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• MIT Technology Review: Sam Altman and the whale

• Business Insider: As AI coding services face a reckoning, Bolt tries to go beyond building

• Reuters: Meta’s AI rules have let bots hold ‘sensual’ chats with kids, offer false medical info

• Bloomberg: Meta’s Superintelligence Dream Team Will Be Management Challenge of the Century

• The Verge: The head of ChatGPT on AI attachment, ads, and what’s next

• Wired: Inside the Biden Administration's Gamble to Freeze China’s AI Future

• FT: The art of persuasion: how top AI chatbots can change your mind

• Sifted: Inside Mistral’s push into the Middle East

• BBC: Staff fear UK's Turing AI Institute at risk of collapse

• FT: Manus and Benchmark: the AI deal that upset China and the US

• WSJ: China’s Lead in Open-Source AI Jolts Washington and Silicon Valley

• The New York Times: Companies Are Pouring Billions Into AI It Has Yet to Pay Off.

• Politico: UK’s AI ambitions clash with its climate goals

• Forbes: Zuckerberg Squandered His AI Talent. Now He’s Spending Billions To Replace It.

• WSJ: Meet Neurosymbolic AI, Amazon’s Method for Enhancing Neural Networks

Read more

This is of course about hyping the ongoing work led by Johnny Ive to build the first OpenAI device but it made me think back to a section from Karen Hao’s book Empire of AI in which she brings up OpenAI’s definition of AGI as “a highly autonomous system that can outperform humans at most economically valuable work.”

At first glance, this vision of AGI sounds thrilling until you try to make a balance sheet (#moneymindset, IYKYK) out of it. What work? Which humans? And whose P&L statement decides “economic value”? It’s one of those statements that works great in a TED talk, less so for capital allocation. The definition also smuggles in a sweeping assumption: once an AI system wins the productivity derby, humanity will simply let it drive.

Switching to Instagram for a palate cleanse, I was presented with another mini-TED talk, this time from Mark Zuckerberg. Fresh off a charm offensive targeting the world’s top AI researchers, the Meta co-founder presented his vision for superintelligence as a parallel between what he’s building and what OpenAI are offering.

In Meta’s playbook, superintelligence is less about beating you at spreadsheets and more about “help[ing] you achieve your goals, create what you want to see in the world, experience any adventure, be a better friend to those you care about, and grow to become the person you aspire to be.”

As a result, Meta’s glasses will become an embodiment of superintelligence, an exocortex that “understand[s] our context because [it] can see what we see, hear what we hear, and interact with us throughout the day.”

Shakeel Hashim, the managing editor of the Transformer newsletter, labelled his video “a pitch devoid of both vision and understanding” while Kara Swisher called it “blather.” That might be true (no matter how many gold chains he throws on, Zuckerberg will never match Altman’s charisma) but investors, regulators and boardrooms need a yardstick that maps directly to silicon roadmaps, energy bills and deployment risk.

So in this post-holiday issue of Computerspeak, I’m going to attempt creating a tighter spec of what superintelligence could practically encompass, presenting a definition that is ambitious enough to hopefully pass Hashim’s bar but still concrete enough to engineer against.

In my view, superintelligence could be achieved when we have a system which has:

• A cognitive scale equivalent to a 100T parameter model (though I’ll explain below why the 100T figure should be treated with care)

• An inference power draw no higher than a single household dryer (1kW), with a 10-year goal to reduce it to less than 10W.

• Native multimodality: text, images, audio, video and real-time sensor data fused in one world model

These three criteria sketch a system that can reason broadly, operate cheaply and interact with the real world, without demanding GW-scale data centers for inference or sovereign-level budgets.

Let’s unpack why each matters, especially in light of the only benchmark we should truly trust: the human brain.

The human brain has roughly 86 billion neurons and a few hundred trillion synapses. While parameters aren’t synapses, it’s the closest analogy we’ve got. Current frontier models hover in the low hundreds of billions of parameters; a 100 trillion–parameter model would be a leap of two orders of magnitude, pushing toward brain-scale complexity.

More parameters mean richer representations of the world, deeper reasoning capacity, and finer-grained understanding. Thanks to sparsity tricks (mixture-of-experts, low-rank adapters), not every weight has to fire on every token. That makes a system with the equivalent of 100 trillion parameters a logical, not literal, target for frontier models. I’m sure that advances in existing approaches (see GPT-5’s universal verifier) or perhaps even new (non-generative) methods like JEPA or brain-like architectures such as spiking neural networks would make such a model not just possible, but more efficient.

In short, we must not treat the 100 trillion target as a parameter fetish, but instead interpret it as a translation of human-like representational power into the units software and hardware architects actually understand and use. I’m saying human-like, not human-level, because despite what you may have read from armchair experts on LinkedIn, neural networks are not similar to biological brains (e.g. backpropagation is likely absent in the brain).

Next up, brain power (literally!) The human brain runs on 20 watts, about what a fridge bulb burns. By contrast, a frontier model at inference can swallow tens of kilowatts per rack. That scale gap is an economic moat: only hyperscalers can afford to serve the cutting edge they train.

Shrinking the power envelope to 1kW changes the game: such a node runs off a standard data center circuit and can share a solar array with your neighbor’s EV charger. The total cost of ownership plummets, making superintelligence a line item, not a moonshot.

Sub-MW platforms fit inside a factory floor, a naval vessel, even a rural hospital or school. That breadth unlocks new markets that can be decoupled from cloud computing. And since regulators fret about compute-monopolies because power scales lock smaller players out, a kW ceiling removes some of that concern, broadening access and, paradoxically, diffusing risk.

We won’t beat biology’s energy elegance anytime soon, but kilowatt-class AI by 2030 puts us within a few orders of magnitude. And perhaps in a decade or so, just like we will increase the performance by 100s, we will simultaneously reduce the power consumption by 100x. Reducing power consumption to single digit watts is important for a simple reason: we can then fit superintelligence in battery-powered mobile devices that can be worn directly on the skin (i.e. the glasses that Zuckerberg envisions we will wear on our face) without causing first degree burns. I’m sure the “AI should run in the cloud, not on device” fanboys will hit me up in the comments, so I just want to make one thing clear: I’m not saying all models should run on smartphones or glasses. Instead, I’m arguing that it would be beneficial if they could fit on a mobile device, as we would get the positive side effect of drastically reducing the environmental footprint of inference. Energy constraints have shaped natural intelligence; meeting a similar constraint forces engineers to solve not just for raw power, but for elegant computation — something Karen Hao has advocated for too.

Lastly, we need to divest more resources away from deepfake porn generators and focus more on the research and development of “real world” world models. Humans don’t OCR a red light into text before hitting the brakes; perception and action run on the same substrate. Today’s multimodal models, despite their increased performance, aren’t enough to match the human brain’s capacity to interact with the built environment around us; in fact, they can’t even match the capacity of a cat’s brain. That’s why we need truly useful world models that are more than just pretty image generators and can absorb pixels, waveforms, tactile feedback and language into a single latent space and combine that with advanced reasoning capabilities.

If we don’t, AI will largely remain confined to the digital world which means it will still be incredibly weak when confronted with the physical reality around us. Agents that write code or book tables at restaurants for us are great but logistics drones or self-driving cars, surgical robots and climate-monitoring satellites all operate according to the laws of physics.

The brain’s trick is cross-modal grounding: seeing steam, hearing a hiss, feeling heat, and then quickly acting in accordance. Systems built on world models can fit entire vertical stacks (think imaging, speech analytics and robotic control) into one SaaS engine. Investors like revenue synergies, engineers like end-to-end differentiability. Everyone wins.

A 100T-scale world model running at 1kW is superintelligence but not the kind you’ll hear about in the steady diet of Hard Fork episodes serving up overcooked smoke and mirrors mixed with tasteless hype. Instead, it’s the kind of superintelligence that offers a real, achievable and useful path towards progress: it pushes the infrastructure layer to innovate (or reinvent) model architectures, it motivates application-layer companies to build brain scale AI experiences that run directly on your personal device, and it even helps regulators get a tangible threshold for export controls and safety audits. It also dramatically reduces the energy and water usage of today’s foundation models.

Most importantly, this yardstick decouples what superintelligence is from which jobs it kills. If the system meets these three criteria, superior economic performance becomes an emergent property, not a definitional quagmire. The market can test, compare and monetize the result, exactly what capital formation is good at.

The race is now to hit brain-scale cognition on a wall-socket budget. Whoever does it first will not just beat humans at “economically valuable work,” they’ll redefine which work is valuable in the first place, and they’ll do it without tripping the circuit breaker.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• Business Insider: Here's why Sam Altman says OpenAI's GPT-5 falls short of AGI

• MIT Technology Review: A glimpse into OpenAI’s largest ambitions

• Business Insider: Nvidia warns that any GPU 'kill switch' or 'backdoor' into its AI chips would 'fracture trust in US technology'

• Business Insider: This tech CEO is trying to stop AI killing the internet. Why is everyone so mad at him?

• Forbes: Fear Of Super Intelligent AI Is Driving Harvard And MIT Students To Drop Out

• Wired: Inside the US Government's Unpublished Report on AI Safety

• TechCrunch: DeepMind thinks its new Genie 3 world model presents a stepping stone toward AGI

• WSJ: Anthropic’s Quiet Edge in the AI Talent War

• FT: Trump official urges Asia to reject Europe’s ‘over-regulation’ of AI

• The Guardian: Demis Hassabis on our AI future: ‘It’ll be 10 times bigger than the Industrial Revolution – and maybe 10 times faster’

Read more

Take for example the recent hiring spree in the United States fuelled by (alleged) offers of tens to hundreds of millions of dollars per head. TBPN, an online daily show about technology, has started keeping a tally of who’s getting (re)hired where, similar to how sports teams trade players.

Many of those traded are Chinese nationals or have strong academic ties to China. That’s no surprise, as China produces excellent AI researchers. But these researchers also come from a work culture defined by the infamous “996” schedule: 9am to 9pm, six days a week (although from my direct experience working in China, it’s more like 8am to 12am, seven days a week). A lifestyle optimized for maximum productivity and minimum time spent doing literally anything else, like talking to people in real life.

Silicon Valley is also no stranger to the 996 lifestyle, but they try to make it sound cooler by calling it “grinding” and “hardcore” when in reality it’s (mostly) a bunch of bros for whom interpersonal relationships outside the workplace are a tax, not a joy.

When you’re hiring people who get $100m to never leave the office, you don’t build tools for human connection, you build simulations. You build screen-time companions, algorithmic pacifiers, intimacy replacements. Which is how you get features like Grok 4’s latest innovation: a bubbly anime sidekick designed for lonely men who want their AI to blink, blush, and maybe pretend to love/talk dirty to them. It’s not a bug, it’s a worldview that starts with Elon Musk and filters down to the base of xAI.

While the AI researchers are socially alienated, there’s also a growing class of AI safety folks who are epistemologically unmoored. There’s a certain genre of AI safety research that operates less like science and more like speculative fiction with footnotes. A recent paper from the UK AI Security Institute called Lessons from a Chimp: AI "Scheming" and the Quest for Ape Language pours some cold water on these folks, pleading with them to pull their heads out of their ass and start doing some real work. It opens with warnings about “AI scheming” and draws a not-so-subtle parallel to how some scientists claimed apes were learning sign language in the '70s: a lot of anecdotal excitement, very little rigor.

The authors skewer the current scheming discourse: overinterpreted chain-of-thought outputs, dramatic fictional prompts (like “pretend you're an evil assistant trying to blackmail your boss”), and cherry-picked examples that make it into the headlines but don’t hold up to even casual scrutiny. It's science by vibes, and those vibes increasingly sound like the hallucinations of a paranoid schizophrenic.

This all would be sad if it weren’t becoming policy. These “AI as con artist” papers are often written by researchers whose only real metric of model misalignment is “it gave a spooky answer when I asked it to act spooky” (also known as following instructions.) Many come from tightly networked safety circles that talk more about "what if" than "what is." The result: we are shaping public opinion and regulation around a mix of cosplay dystopia and soft academic panic.

So here we are: AI products designed by the socially disengaged, regulated by the epistemically fragile, deployed to users who mostly just want a powerful productivity assistant and not to get catfished by a ChatGPT anime girlfriend. And we wonder why things feel uncanny.

There’s something hollow in the circuitry. Not malevolent, not evil. Just absent. A bunch of ghoulish bros where a human perspective should be.

Maybe the machines are slowly, then quickly, turning some of us into ghouls too.

And now here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• Time: Chess Grandmaster Magnus Carlsen Beats ChatGPT Without Losing a Single Piece

• Wired: Where Are All the AI Drugs?

• FT: Chatbots in the classroom: how AI is reshaping higher education

• Business Insider: OpenAI's chief economist says he's teaching his kids these 4 skills to prepare for the AI world

• Bloomberg: Microsoft's Copilot Is Getting Lapped by 900 Million ChatGPT Downloads

• The New York Times: Nvidia CEO Treads Carefully in Beijing

• Forbes: This AI Founder Became A Billionaire By Building ChatGPT For Doctors

• The Information: Inside Zuckerberg’s AI Playbook: Billions in Compute, a Talent Arms Race, and a New Vision for Meta

• Sifted: Inside Helsing: A look behind the curtain at Europe’s AI defence unicorn

• WSJ: Can Pittsburgh’s Old Steel Mills Be Turned Into an AI Hub?

• The New York Times: Meta’s New Superintelligence Lab Is Discussing Major AI Strategy Changes

• Sifted: Synthesia CEO Victor Riparbelli: ‘I'm very good at slapping stuff together’

• Bloomberg: The New Third Rail in Silicon Valley: Investing in Chinese AI

• Fortune: Marc Benioff says AI is radically reshaping Salesforce, and 51% of Q1 hiring was internal as thousands of employees were redeployed

Read more

Bold claim, followed by an awkward silence from the four participants in the hostage video—sorry, “livestream." Also: pixie dust with a sprinkle of nonsense.

Reasoning models are still autocorrect on rocket fuel, powered by inference-time compute. In fact, the best way to think about reasoning models is this: if the machine learning of the 2010s was a car with cruise control, the reasoning models of today are vehicles with Level 2 autonomy. Don’t get me wrong: Level 2 is incredibly impressive compared to cruise control, but it’s not full autonomy. The problem of course is that there are some goofballs who fall for Musk’s marketing and think Tesla’s “Full Self-Driving” is a Level 5 system when it fact it’s not, resulting in so many collisions there’s a dedicated Wikipedia page to keep track of them.

Reasoning models work by hoovering up oceans of text and learn statistically plausible ways to arrange words, which are then improved using chain-of-thought (CoT) prompting. But as Yann LeCun pointed out, CoT is still (incredibly good!) knowledge accumulation and retrieval, not human intelligence. A PhD (and I know this because I’m surrounded by a lot of them) does more than regurgitate facts. They frame questions, decide what not to believe, design experiments, and argue with colleagues in the hallway (particularly about the quality and quantity of what’s inside the snack drawer). Grok 4 or equivalents from OpenAI or Google do none of those things unaided.

Even some of the best reasoning benchmarks mostly boil down to multiple-choice quizzes or very advanced trivia questions. Get enough training data and a big enough GPU cluster, and your model will ace the test. That’s impressive, and it gets product teams a shiny leaderboard badge, but it isn’t the same as actually operating at the frontier of a discipline.

AI companies love to present a chart where performance climbs smoothly past undergraduate, graduate, and finally PhD. Terrific marketing. In the real world, graduating from PhD-level capability to genuine discovery involves touching messy reality: running assays, shipping prototypes, persuading grant committees, enduring peer review. None of that lives in a token stream.

Yes, language models can spit out novel protein sequences or plausible circuit diagrams. But novel is not the same as useful, and certainly not the same as commercially adopted. Saying Grok 4 will “discover new, useful technologies” by 2026 is like saying Google Search would cure cancer by 2010 because it could find every article about oncogenes. Tools accelerate discovery; they do not replace it.

A core limitation is that today’s models are still largely one-shot conversationalists. They reason once, in a vacuum, and return an answer. A scientist, by contrast, iterates: build hypothesis → test → dump half the theory in the trash → repeat. Until an AI can operate an autonomous feedback loop (planning experiments, controlling robots, interpreting raw data, revising hypotheses) it remains a super-autocomplete, not a scientist.

Could we bolt on robotics, simulation, and automated lab platforms? Absolutely. Companies such as Insilico Medicine and Isomorphic Labs are already walking that path. But the leap from plausible X demo to fully automated R&D pipeline is gargantuan, and definitely not scheduled by a Musk-approved calendar app.

None of this is meant to dismiss the raw power of modern models, despite the attempts from ghouls like Gary Marcus and Emily M. Bender who keep yapping about the limitations while flogging human-made slop in the form of lousy books and newsletters. “Regurgitation” at trillion-token scale is transformative: patent lawyers draft claims faster, junior programmers ship code that compiles on the first try, and small biotech teams get literature reviews that used to take months. That’s real economic juice.

Moreover, as AI labs chase the mirage of superintelligence (RIP AGI, it was nice knowin’ ya!), they will keep stumbling into practical breakthroughs (better compression, clever retrieval augments, at scale networking of GPUs) that migrate into everyday software. One could argue the search for superintelligence is the best R&D subsidy earth has ever seen; even if AGI forever recedes over the horizon, the side-effects are priceless.

Musk’s bravado is partly brand management: he needs xAI to look as inevitable as SpaceX felt in 2008. Investors fund inevitability. Regulators fear it. Engineers quit comfortable jobs to join it. Declaring Grok 4 a PhD-slayer signals we’re already there—come aboard or get left behind. The statement works even if it’s untrue; what matters is narrative gravity.

Instead of asking whether Grok 4 beats every PhD, ask simpler questions:

• Does it reduce the cost of getting to “draft one” of a technical document to nearly zero?

• Does it surface non-obvious connections across literature that a human might miss?

• Does it shorten iteration cycles in simulation-heavy fields such as materials science?

Early evidence says yes on all three. That is plenty revolutionary. It does not require anthropomorphizing the model or pretending we’re about to birth “a nation of benevolent AI geniuses.” (I almost vomited in my mouth when I copied/pasted that headline from Wired.)

Musk’s prophecy of world-changing discoveries by 2026 will almost certainly be wrong on specifics. But the direction of travel (models growing more capable, more integrated with tools, more useful to human experts) is entirely right. Dismissing the tech because its loudest champions oversell it would be as silly as ignoring early internet protocols because the pets.com sock puppet promised eternal profits.

The smartest stance is realism wrapped in optimism: call out overreach, while doubling down on the real advantages reasoning engines already deliver. Humanity’s next breakthroughs will almost certainly involve an AI partner. They just won’t arrive on a timeline dictated by a livestream on X.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• WSJ: Was That Amazing Video in Your Feed Real or AI? Tech Platforms Are Struggling to Let You Know

• FT: Amazon weighs further investment in Anthropic to deepen AI alliance

• The Verge: Meta is trying to win the AI race with money — but not everyone can be bought

• Wired: How Video Games Became the New Battleground for Actors and AI Protections

• Business Insider: These are the most popular AI coding tools among engineers

• WSJ: The AI Scraping Fight That Could Change the Future of the Web

• Business Insider: How Google found its AI hype guy

• MIT Technology Review: Inside India’s scramble for AI independence

• CNBC: Engineer caught juggling multiple startup jobs is a cautionary tale of ‘extreme’ hustle culture, experts say

• WSJ: The Companies Betting They Can Profit From Google Search’s Demise

• The New York Times: The Coder ‘Village’ at the Heart of China’s A.I. Frenzy

Read more

Today, any Romanian child with a phone could have the same opportunity my parents fought so hard for: access to an AI tutor in the form of a conversational model that explains grammar, spots confusion, even cheers a kid on. Except Bruxelles has inserted a poison pill that might prevent that possibility from becoming a reality. Article 5 (1)(f) of the EU AI Act bans systems that “infers emotions” in education or healthcare unless they serve a strictly medical or safety purpose. Apparently an algorithm may detect a seizure but not a student’s frustration. If the AI-powered virtual tutor notices the student is about to cry over phrasal verbs and adjusts its lesson, it risks illegality. That is ludicrous, plain and simple.

A coalition of European startup founders and venture capitalists want to prevent such an outcome so they got together earlier this week and asked the EU to stop the clock on the AI Act until the rules actually make sense. In an open letter addressed to Bruxelles, they wrote: “In a world racing toward the next technological frontier, a call to pause the implementation of the rushed regulation that is the EU’s AI Act is not just prudent - it’s essential.”

Another, even broader alliance of industrial heavy-hitters, from Airbus to ASML, echoed the call a few days later, warning that an “unclear, overlapping and increasingly complex” rulebook is already scaring investors and nudging talent toward friendlier shores.

When the Marvel and DC universes agree you’re the baddie, you should probably stop and listen. After all, “with great power, comes great responsibility.” In the end, despite what some academics and ideologues might say, a short, surgical pause is simply the least-bad option left.

That’s because the obligations on general purpose AI models go into effect in August 2025 and on so-called “high-risk” systems in August 2026. Yet the Code of Practice, the technical manual companies must follow to comply, is still stuck in inter-service ping-pong inside the European AI Office while the standards needed for compliance are MIA. Moving ahead now would give Europe what it fears most: a fragmented regulatory map where every national authority improvises its own interpretation. The startup letter calls that outcome “a rushed ticking time bomb.”

One of my favorite novels from Romanian literature is Dimitrie Cantemir’s 1705 satire A Hieroglyphic History. In his roman à clef, Cantemir describes the ostrich‑camel, a beast stitched together from irreconcilable parts to satisfy every court faction and thus loved by none. Like the ostrich-camel, the AI Act has become a regulatory chimera. The creature waddles too slowly for the desert yet cannot keep its head out of the sand; Europe’s law does no better, lumbering under the weight of security clauses while burying its vision in precaution.

The result pleases neither industry nor activists, yet drains energy from both. But where the AI Act does most damage is in the startup ecosystem, which is arguably the innovation engine of Europe. Sure, there are some European AI scaleups such as Synthesia (my employer), Mistral, Aleph Alpha or Helsing that have now grown to a size where they can deal with the requirements of the AI Act. But a startup founded one or two years ago, without an in-house policy expert or legal counsel, will struggle. When they came into office, the new Commission promised flexibility and simplification for European startups, but those haven't yet happened when it comes to AI. How can a small company comply in the absence of clarity? As I’ve told Wired, this risks slowing down Europe’s ability to compete with China and the United States.

Plus, investors read time bombs as flight risk. We already have a huge problem with tech companies scaling and listing in Europe, and private equity flowing to jurisdictions where liability is clearer and capital markets exits are smoother. The US has SEC guidance; the UK is (hopefully) finalizing a light-touch, principles-based regime. Right now, the only thing on offer in Europe is ambiguities, deadlines and harsh penalties.

Critics say a delay would reward lobbyists, benefit American and Chinese tech giants, and weaken protections. Okay, maybe they have a point. But why make European companies pay for the perceived excesses of Big Tech? There is nothing protective about passing rules no one can interpret without an army of lawyers. Better to suspend the countdown, finish the Code of Practice, and rewrite provisions that criminalize perfectly benign use cases like adaptive teaching.

A two-year “clock-stop,” as the corporate letter proposes, would synchronize three moving parts:

1. Standards: Give CEN-CENELEC time to publish harmonized technical specs so every startup knows which checklist applies.

2. Sandboxes: Expand the AI regulatory sandboxes already allowed under the Act, so regulators learn alongside developers.

3. Clean-up: Amend or clarify articles whose unintended consequences are now obvious, starting with 5 (1)(f).

Europe missed consumer tech, social media and cloud computing. Miss AI and it graduates from laggard to fossil. The Commission’s own AI Continent Action Plan talks up competitiveness, yet its flagship law, in current form, would handicap precisely the firms trying to build European-made foundational models.

No one is advocating a regulatory Wild West. Transparency, safety testing and redress mechanisms belong in any modern AI statute. But Europe has always distinguished itself by pairing tough rules with workable guidelines. Workable is the missing ingredient. As the Stop-the-Clock signatories warn, “ambition must now translate into action” — action that simplifies, rather than multiplies, compliance.

Stopping the clock is not capitulation to Big Tech. It is a vote for craftsmanship over speed. So finish the Code of Practice, engage European industry (for the record, inviting them to be a silent observer on a call with Big Tech about the Code of Practice ain’t it), and let regulators, startups and incumbents test drive the guardrails before they ossify into law.

My parents sacrificed so much so one Romanian kid could master English. Europe now has the chance to give millions of children an AI tutor for the cost of a data plan, unless it bans the empathy that makes tutoring effective.

Pause the Act, fix the flaws, and then move forward. The alternative is to march proudly into irrelevance and decline.

And now, here are the week’s headlines:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• Fortune: Runway’s AI transformed films. The $3 billion startup’s founders have a bold, new script: building immersive worlds

• Bloomberg: SAP CEO Says Europe Needs More Applied AI, Not Another Stargate

• CNBC: AI voice startup ElevenLabs pushes global expansion as it gears up for an IPO

• WSJ: How a Bold Plan to Ban State AI Laws Fell Apart—and Divided Trumpworld

• WSJ: China Is Quickly Eroding America’s Lead in the Global AI Race

• Wired: Sam Altman Slams Meta’s AI Talent-Poaching Spree: ‘Missionaries Will Beat Mercenaries’

• The Verge: Can the music industry make AI the next Napster?

• Forbes: This Startup Built A Hospital In India To Test Its AI Software

• WSJ: Digital Workers Have Arrived in Banking

• The Information: Corporate Data Wars Intensify

• The Verge: Hollywood’s pivot to AI video has a prompting problem

• Sifted: Startups and VCs call on EU to pause AI Act rollout

Read more