2026-07-26 10:43:29
Hi,
Welcome to the latest Sunday briefing! I am off on holiday for a couple of weeks. The team will continue to tend to Exponential View while I’m gone, so you won’t miss a beat.
Azeem
On 1st September 1789, Samuel Slater set sail from England for New York. He’d learned that the Pennsylvania legislature had recently passed an act awarding £100 to a British textile worker who smuggled high-end machinery into the state. America was going to great lengths to acquire industrial know-how, by any means possible.
Britain had made it illegal to export textile machinery and technical drawings. The ban extended to prevent textile workers from emigrating. Slater had worked in mechanized textile mills and saw his chance.
He committed the entirety of Richard Arkwright’s system – the first factory method for spinning cotton – to memory. He disguised himself as a farm laborer and broke the laws of his nation as he carried himself across the Atlantic, valuable know-how secretly distilled into his brain.
By 1790, Slater was a partner in a cotton mill in Pawtucket, Rhode Island, built from the plans he memorized. President Jackson called him the “Father of American Manufactures.” He died in 1835 worth around a billion dollars in today’s terms.1
Today’s question is to what extent are Chinese AI labs distilling the outputs of American AI models – and is it really a problem? The easy answer is the hawkish one: hugely and yes, it is. But it’s not the only answer.
First of all, distillation is a decades-old machine learning technique in which a larger model can train a smaller, more efficient model. A lab running distillation internally is not an issue. But it is possible to distill a model from the outside (even without the provider’s permission). This is the accusation against Kimi and other Chinese labs.
Almeida, a four-year veteran of OpenAI, explains that effective distillation is much harder today than a few years ago, when AI models helpfully provided their reasoning traces. What Almeida calls “behavior parroting” is to learn from final answers, the least powerful approach but might still work well to help bootstrap another model.
There is substantial evidence that both Chinese and American researchers have trained models on the outputs of frontier systems. Stanford’s Alpaca project has admitted as much. Anthropic has alleged that DeepSeek, Moonshot and MiniMax have used more than 16 million Claude chats via 24,000 fake accounts. Michael Kratsios, Trump’s science chief, says he now has evidence of how Moonshot ran distillation attacks.
Anastasios Angelopoulos, the CEO of Arena, a benchmarking company, makes the case that Kimi K3 is exceeding the performance of some of the top US models, something distillation alone doesn’t allow, and he predicts that “American labs will start distilling Chinese intelligence.”
It isn’t clear that distillation is illegal yet. points out that “[t]here’s no legal precedent that model outputs are IP.” The US Copyright Office’s 2023 statement on AI confirms as much:
When an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship. As a result, that material is not protected by copyright.
Unlike the case of Samuel Slater, who knew he was breaking British law, the problem here is a case of the exponential gap: the technology has stepped ahead of the law.2 If labs have IP in outputs of their models, they will essentially have IP in every future economic activity of all their users.
It also weakens the labs’ own position – why should AI models be prevented from consuming Anthropic’s IP when Anthropic is allowed to consume yours and mine?
Bahrad Sokhansanj has some other sensible suggestions, summed up as follows. Address distillation but scope it narrowly, only focusing on the real harm done, with the right instruments. Is it about national security? Or something else? Regulate more broadly, and this will play into the labs’ desire to skew the regulatory field in their favor.
Again, there is precedent: in 1824, once he was settled as a prominent American industrialist, Samuel Slater lobbied for protectionist tariffs to stifle foreign competition. By then, he was also known as “Slater the Traitor” back in his hometown.
See also:
Satya Nadella was one of many tech leaders to force the case for open-weight models to quieten rumors that the American administration was considering limiting them. Jensen agrees:
When you and your team deploy AI agents, identity becomes your strategic infrastructure. An agent is an actor that reads your data, calls APIs, and does things on your behalf. You need to give it clear permissions and to audit its trails.
Okta’s AI Identity Readiness assessment will score the security of your AI agents and show exactly what you need to fix to go to production with peace of mind.
Run Okta’s 5-minute survey to evaluate your agents.
In 2015, Made in China set a target of 70% self-sufficiency in semiconductors within a decade. It was often mocked as fanciful and missed wildly. Data from Morgan Stanley now shows that domestic suppliers will have met about 41% of China’s AI chip demand in 2026, up from 20% in 2023. Beijing may hit its 70% threshold five years late.
Compute-weighted, the 41% figure delivers less compute compared to Nvidia, but for the purposes of strategic autonomy, the quality gap is increasingly a non-issue.
The spark was Washington’s export restrictions. “If the U.S. hadn’t forced our country, our company and our industry into a corner, we would never have done something like this”, says Huawei’s deputy chairman. It has become an “all-out push” according to this excellent reporting.
A leaked conversation between DeepSeek’s boss, Liang WenFeng, and several investors supports this. Liang says:
What’s the gap with the U.S.? Only one thing: resources. We don’t have enough GPUs – our count is still small. […] Domestic chips now have a historic opportunity. Previously, adaptation was hindered by poor ecosystem… But that’s changing. NVIDIA CUDA’s moat is eroding rapidly.
Full transcript and context at ’s blog.
Peter McCrory, Anthropic’s Head of Economics, points out that the US labor market has shrugged at AI. Unemployment is at 4.2%, and Anthropic’s data finds no worsening unemployment even in the most exposed occupations.
AI augments rather than replaces, for now. Not a single profession has been 100% handed over to machines yet. Every job still needs human effort. It is changing how work gets done, and if workers get more productive, value shifts inside existing roles, and those who use the technology best stand to benefit.
The final hard-to-automate tasks, the “weak links”, as Professor Chad Jones calls them, are the things only a human can do. Companies will need people to get them done, and this protects employment, keeping a decent share of income in human paychecks
Here is another take. People still matter and will continue to matter. Europe creates far fewer successful innovative companies than the US. One reason I’ve often argued is the simple cost of changing the workforce. I think of startups as exercises in making mistakes and learning from them. Every additional cost to making a mistake means an opportunity to learn not taken. Yoram Wijngaarde finds a simple relationship (correlation is not causation) that shows that the more expensive it is to let go of staff, the lower the rate of unicorns per capita.
A new biography of Jean-Paul Marat, one of the leaders of the French Revolution, reviewed in the current LRB, is worth reading for anyone trying to make sense of today’s AI debate.
Stanford historian Keith Baker3 has a new biography of the journalist and politician. He argues that Marat hates mediation of any type, from Newtonian formulas to parliamentary assemblies and calm discussion- anything that stands between the people and the truth. He tried to write a daily pamphlet, shouted rather than argued and manufactured intimacy. “By making his journal ‘more interactive, more dynamic, more personal’, he fashioned an intimacy that allowed him to speak for the people.”
In amongst this, his paranoia did help him identify real corruption and institutional betrayal. Baker calls him the first modern populist.
High-frequency publishing, paranoia as analysis, a parasocial closeness and the constant insistency that any complexity is just conspiracy in disguise… Well, AI discourse is now selecting for exactly this Marat-like temperament.
While today’s keyboard warriors carry none of the physical violence of Marat’s Terror, there is a similar underlying logic against nuance and complexity. Doomers and accelerationists share patterns – purge rhetoric, aggressive polemics, and the framing of every whiff of nuance as corrupt. What is left are the extremes. Call to mind imminent economic disaster; catastrophic fraud; utopian abundance… or, simply, the transformation of the human condition.
What can get lost in these extremes is the reasoned position that admits and examines evidence. A position that balances probabilities and accepts answers might be complex, incomplete and – contingent.
See also:
💪🏼 Really stoked that AMD’s CEO Dr. Lisa Su opened her keynote with Exponential View data. You can get the same data here.
Become a member to receive our Sunday briefing every week in your inbox or the app.
Why AI-assisted bioweapons won’t kill us. via EV member
🏋🏼♀️ The share of UK businesses using AI has nearly tripled since 2023, but most firms are still dabbling.
Arsenal FC is building AI models for football (soccer).
Young people are more likely to gamble in financial markets when important life goals (like buying a house) feel out of reach.
Global air and sea surface temperatures are headed for a new record.
Good news from the Amazon: wildfires are at a record low this year and deforestation is at a 10-year low. h/t EV member
Vintage LLM 😎Training AI models only on pre-1931 texts help researchers study what AI can do without contamination from the modern web.
👀 Stripe is in talks to buy OpenRouter.
Intel is shipping the first chips with layers printed on ASML’s $380 million High-NA EUV machines.
Know thy maths. breaks down why the EU’s 46% electrification target by 2040 is mathematically unachievable. Relevant for anyone working in policy, really.
🐳 Orcas preparing food for their young? Amazing.
Thanks for reading!
His estate was worth $1 million. On a CPI basis, that is $40 million; on a relative income/wage equivalence, it is about $1 billion; on a share of the US economy, it is about $1.7 billion.
It is obviously unseemly to many people that the labs trained on other people’s outputs (like books and essays) en masse. But the courts haven’t yet decided that the training is a breach of copyright law. In Anthropic’s case, despite the settlement, they have decided it wasn’t.
Baker is a super historian whose work I have gotten to know over the past few years. His two sons run one of the world’s most successful (and least well-known) hedge funds.
2026-07-23 22:16:26
Kimi K3 has caused quite an uproar since its release last week. It’s the first time a Chinese model has taken the lead on the frontend Code Arena benchmark. And that’s three months since Moonshot AI’s previous impressive flagship model, Kimi K2.6, was released.
Following in Moonshot’s steps, Alibaba announced over the weekend that Qwen3.8 – a 2.4 trillion-parameter model – is coming soon, and unlike its last release, this one will be an open-weight model. No benchmarks or further details have been released as of yet.
Open models are now estimated to be 4-7 months behind the frontier in cyber capabilities, down from 6-10 months in 2025. And despite compute constraints, efficiency improvements mean these labs are doing more with less. Comparing the compute availability and model performance between US labs and Chinese labs, we estimated Chinese labs to be getting 4-7x more out of their compute.
and I spent some time with the Moonshot AI and Alibaba teams in China back in April and May, and we’ve had time to think about the economics of open-source models and how they affect the entire ecosystem.
Some have claimed that Kimi K3’s performance breaks the economic case for AI as it lowers the cost to complete various tasks at frontier standards. For instance, Microsoft engineers are reportedly testing whether Kimi K3 can be used within Copilot.
We don’t think this is the case, and in today’s post we’ll work through what might happen next.
In The State of the AI Economy report, we found that token usage is elastic across providers. This means that every drop in token price leads to a larger increase in token volume, more than offsetting the difference.
For every 10% price cut, token consumption rises 12-18%. A paper by Demirer et al, found a similar effect: a 10% price cut resulted in an 11% or so increase in volumes, which economists call an elasticity of -1.11.
The net effect is a rise in total token spend. But note that the effect is a weak one, not the cantering Jevons’ paradox sometimes presented. Reality might tilt the scales further in favor of more, not less, demand. Workflows are becoming more token-intensive as we rely on reasoning models and verification and approval loops. And the early evidence suggests that firms that adopt AI early tend to increase their relative spend alongside growing headcount. These effects might be short-term elasticities rather than ones that can be sustained for decades, but for now they indicate that falling prices increase volumes and, with that, revenue.
The model weights may be free, but the inference is not. Kimi K3 has 2.8 trillion parameters. The weights alone occupy 1.4 TB. It needs to be served on something like a 72-GPU NVIDIA GB200 NVL72 rack or equivalent. That’ll cost $3-4 million to buy and install. Operating it consumes about 120 kW continuously, over a million kWh per year, before you consider networking, storage, cooling, and humans. If you rented these in the open market, it would cost about $7 million a year.
But for infrastructure providers, the economics of hosting open-source models can be very attractive compared to serving closed-source models. A simple way to understand this is to think of the hyperscaler as needing to pay a license fee for a closed-source model but not for an open-source one1.
2026-07-20 22:05:20
Hi all,
Here’s our Monday roundup of data signals across AI, energy & markets.
Enjoy!
Kimi can code. Kimi-K3, released last week, overtakes Fable 5 and GPT-5.6 Sol to take first place on the frontend code arena benchmark.
Demand unlocked. DeepSeek nears $500 million in annualized revenue with gross margins as high as 70-80% on its V4 model.
Ask and receive. Tests of leading AI models on prompts based on real terrorist cases show that a third of responses would’ve provided useful help to attackers, and labeling the same prompts as “research” increased compliance from 17% to 42%.
2026-07-19 10:37:10
“You inspire me to think more exponentially. — Robin D., a paying subscriber
Moonshot AI’s new open model, Kimi K3, might be more of a shock to the US than the original DeepSeek model. As a model, it is very good.
I’m wary about benchmarks because benchmarks aren’t the real world. But the emerging consensus appears to be better than Claude Opus 4.8 and, in some cases, on par with Claude’s Fable and OpenAI’s GPT 5.6. The real question, of course, is on which dimensions does K3 beat the frontier labs, and for which workloads are those dimensions important?
Price-wise, it is expensive for an open-weight model. According to Artificial Analysis, it’s about the same price as GPT 5.6 Sol but about 24x more expensive than DeepSeek V4 Pro. Indeed, on a per-token basis, it is only half the price of OpenAI’s GPT 5.6 Sol, far from the usual price advantages of Chinese models.
For the AI economy as a whole, for companies around the world, for governments that aren’t rich, this is probably a net positive. The inference margins that OpenAI and Anthropic enjoy are significant, and they can maintain them because they have the very best models. But it’s pressure, not displacement. Enterprises don’t buy on price alone. They value security, support and possibly the fancy professional services on offer. And the harnesses OpenAI and Anthropic have built remain a differentiator.
As we argued in the State of the AI Economy, token demand is elastic. Falling prices drive demand, and that demand drives infrastructure usage. Hyperscalers and neoclouds will serve these higher-end open models, further fueling demand for compute and everything around it. This pushes more of the revenue pool towards the compute layer and away from the model layer margin. This strengthens rather than weakens the infrastructure payback case—and, of course, the chip and memory suppliers that sit below them. A tempering note: cheaper intelligence still waits for monthly management meetings and a slow-moving approval process.
Your AI pilot worked, but now comes the hard part – deploying AI agents across a company that wasn’t designed for autonomous software.
To deploy your agents safely, you need to know what AI can access, what it can act on, and where accountability sits.
Okta’s AI Identity Readiness assessment will score the security of your AI agents and show exactly what you need to fix to go to production with peace of mind.
Run the 5-minute assessment to benchmark your organization.
Lazard’s latest energy cost report shows that the levelized cost of solar photovoltaic electricity has risen in the US on a year-on-year basis. Back in 2021, this was $38 per MWh; in 2026, it’s $69. The price of gas generation has also risen from $60 to $90.
You’d be right to point out that we’ve long argued that because solar panels – one of the key cost elements for solar power generation – are on such strong learning curves, the price will keep trending down.
To make sense of this, I looked at the evolution of solar electricity costs in thirteen markets between 2020 and 2025 using data from IRENA, the International Renewable Energy Agency. The headline story is that, yes, PV modules remain on an aggressive learning curve, with unit costs dropping as production increases. Overall systems costs continue to trend downward, but the levelized cost for delivering electricity has risen slightly since 2023.
Member-only
This post includes bonus content for Exponential View members, in which we explain in more detail why this is happening and what to do about it. Upgrade to unlock.
See also:
In June, solar supplied a record 25% of the EU’s electricity, more than nuclear, gas, wind and hydro.
Households in three Australian states are getting three hours of free power every day during peak solar hours.
Britain’s first experiments with copyright began in the decades after William Caxton introduced printing in the 1470s. The Crown licensed a single company to police what went to print, and in return, its booksellers had an exclusive right to copy the texts with no end date. This license manufactured scarcity for over a century.
In 1695, Parliament refused to renew the Licensing Act, which enabled the booksellers’ monopoly. That 15-year interregnum brought an explosion of ideas: London gained 70 political periodicals (from one) and print culture spread to the provinces and American colonies. Then, in 1710, Parliament enacted the Statute of Anne: the world’s first copyright law. It gave exclusive rights for a fixed term, just 14 years, renewable once, and then the work went into the public domain. The law’s title was “An Act for the Encouragement of Learning.” Enough scarcity to incentivize creation and freedom after that so that knowledge would compound.
Twenty-first-century economics agrees. Joel Mokyr took the 2025 Nobel Prize for showing how useful knowledge becomes self-generating. Scientific understanding enables new technologies. The problems encountered in applying those technologies stimulate further science. And a society open to new ideas allows each advance to become the foundation for the next. Knowledge does not simply compound; it helps produce more knowledge
In a world of AI, this compounding will be doubly true. And will set up greater tension between content industries, who, like Britain’s booksellers in the 17th century, will want to protect their old business models, and the potential to drive open-source AI models and a raft of complementary startups. Brian Williamson argues that the EU’s Anne-style settlement- its text-and-data mining exception is key to the EU staying at the AI frontier.1 It allows AI to learn from lawfully accessible material:
Machines as well as humans should be free to learn; what matters in terms of protecting creators is whether outputs, not inputs, duplicate existing work.
For Europe, woefully behind in semiconductors, compute infrastructure and foundation models, yielding to copyright lobbies would further weaken its relative position in AI.
AI impact on jobs: Junior roles are being “seniorized,” and employers are more open to humanities graduates.
Russian soldiers’ average survival time after reaching the front is 20-30 minutes.
💸 Prediction markets are starting to bet on AI compute costs.
estimates that China will have a Mythos-like model in February 2027. His full analysis is worth reading.
Claude’s personality changes depending on the model and the language you use with it.
💪🏼 There’s now a promising candidate for the first vaccine to prevent pancreatic cancer.
Apple is testing PrismML’s tech to run big AI models directly on iPhones.
Good short essay by : “We tend to conflate power-seeking AI and superintelligent AI.”
😷 How Palantir embedded itself in the UK state, an investigation: “Despite having no real history of working with health data, Palantir began positioning itself as the go-to expert and Global Counsel started hiring Westminster insiders who had contacts in healthcare.”
Thanks for reading!
Caveat: It is an independent report, but it’s paid for by Google. However, I think the argument is salient enough to present to you.
2026-07-13 19:46:30
Hi all,
Here’s our short Monday roundup of signals to kick off your week.
Another scaling law? ByteDance researchers found that newer AI models learn on the job1 about twice as fast as models from just three months earlier.
Execs talk down job cuts. The share of CEOs expecting significant headcount cuts from AI fell from 46% in January 2025 to just 20% in May 2026.2
AI’s audience splits. ChatGPT’s user share fell below 50% for the first time in March.3
2026-07-12 18:32:10
“With so much hype around the tech, your no-nonsense unbiased assesment is essential.” — RB, a paying member
Roberto Serrano, a professor at Brown, suspected his economics class was relying on ChatGPT to do their exams. He made the final paper a closed-book exam, and scores for 56 of his 59 students collapsed (by as much as 100%). Kudos to the two students who seem to work unaided.
This is a problem of incentives. The students appear to value the high score rather than mastering the subject. Perhaps because the value of a degree is increasingly less about intellectual excellence and more about job-market signalling.
The percentage of Americans who read for pleasure on any given day fell to 16% by 2023, down from 28% in 2004. Literacy, not a natural human state, is a learned skill that needs practice. Americans aren’t practising – postliteracy loves short-form video, after all. These data are from before the rise of generative AI.
Reading in America is a different problem. The data, even pre-AI, is terrible and doesn’t support the notion that post-literacy arrived after ChatGPT. Rather, the trend has been ongoing for at least two decades.
When access to intelligence is uncapped, AI could divide people based on our willingness to think and engage with what is difficult. David Brooks makes the case :
What really matters, therefore, is not brainpower but the willingness to run the mental marathons that produce high-quality results. […] The crucial task before us is to cultivate people’s desire to seek out cognitive complexity. How do we train people to see their life as a hero’s journey in which they take on difficult missions that they may fail at and that will certainly involve pain and suffering?
Reading long-form, constructed arguments that have been closely fought through by an author forces the reader to engage with the material far more deeply than a stream of summaries does. (These may give the illusion of thinking, but that isn’t the case.) I’ve been working on my second book over the past few months, and yes, R Mini Arnold, my agent, has been an extraordinarily helpful research associate. The latest AI models, Fable and GPT 5.6, can be prompted to produce outstanding (almost) research, but you really need to know what to ask and how to ask for it. In my case, that’s meant building up my mental map the old-fashioned way. Which means I need to sit quietly, read original material, consider it critically and handwrite my notes.
Elsewhere:
LinkedIn is awash with AI-generated posts; Substack is less so. I have built a Chrome extension that hides AI-generated content. It makes X more manageable to browse.
The University of Chicago Law School is piloting device-free first-year core classes and requires students to learn to use AI effectively.
“Fable is better than me at my job, but Fable alone would be a mediocre investor,” says one VC.
No, GPU demand is not softening. Silicon Data’s one-year H100 contract index bottomed near $1.70/hour last October and has since rebounded ~38% to $2.35. Spot prices are up 10% this year.
The SpaceX S1 offers some clues about how future demand might shape up as it discloses the infra deals the firm has signed: a three-year tenor with unusually permissive 90-day cancellation terms.