But most — because of the time it takes to polish them — simply remain internal as tools to push our thinking forward.

We’ve been wondering just how valuable it would be for readers to peek at the artist’s sketchbook, even in its raw, unfiltered form.

The advantage is that you’d get to see things I think are important but, for whatever reason, we don’t finish or publish. Maybe they go out late, much, much later, or they don’t survive the cut.

These sketches are more raw than anything we’d normally publish. For example, it’s interim research I’m doing as I think through a thesis. I’ve judged it good enough for my thinking but not yet ready for your reading. But perhaps I’m wrong. Examples from the past week include:

• An analysis of liquidity pauses over the past ten years that have hit the tech sector

• How to classify high-quality research processes from degenerate ones

• Co…

Read more

Hi all,

Here’s the latest roundup of our data signals across AI, energy & markets.

1. Super, apparently. The US Census Bureau’s business survey now asks firms whether they use “Super Intelligence”. Its examples include voice recognition, which is not quite what Bostrom had in mind. 24.6% of firms said they use it in early October, up from 17.3% in late November 2025.

   

2. Not in my yard. Only 3% of political ads mentioning data centers in October supported them. 57% opposed them, and the remainder didn’t take a clear stance.

   

3. Chip nation. Semiconductor parts accounted for almost half of South Korea’s exports in October.

Read more

Welcome to the Sunday edition. This week:

• Is Anthropic a real business? We dissected Anthropic’s IPO pitch in 39 charts.

• The new company: bigger, smaller, faster.

• Sunrise in America

• Also: New materials, GLP-1s & psychedelics++

NEW MEMBER PERK

We’ve just launched a new perk for members: EV Charts, a living library of 500+ data-rich charts on AI, energy, geopolitics and more, each linked to the original Exponential View analysis.

If you’re not yet a member, upgrade anytime for full access.

Unlock our chart library

The unbounded firm

Every firm has its boundary. Back in 1937, Ronald Coase argued that the limit is defined at the point when the cost to do a task in the market is lower than doing it inside the firm. Most inventions- the ledger, telegraphy, copy paper, the internet- cut internal and external costs, so the boundary can move in both directions. Now, and colleagues ask what happens when we introduce AI agents in the mix.

Unlike humans, AI agents could share the same reasoning patterns (the model weights) and the same working state. Nothing leaks in handoffs from one agent to the next (and nothing is angling for a promotion). This might give rise to “supercooperative firms”, where bigger becomes better at every scale, possibly leading to giant AI-run firms in deep, large markets.

Arguably, this trend is long underway, as Kwon and others point out. The largest 1% of American firms held 97% of corporate assets in 2018.

AI might amplify this. AI-capable firms will be able to enter many markets, as Anthropic is in coding, bio, legal, and finance, and be able to scale horizontally past the traditional Coasian boundary. AI is a technology more likely than not to accentuate this trend. We might need new rules for dealing with such enormous entities.

See also: Microsoft is making way for the machines where I explore what be behind the tech company’s massive voluntary exit program.

Subscribe now

The solo biotech

Read more

This week Microsoft offered a voluntary exit to around 60,000 employees across the US, UK, Canada, and India – about 27% of its 223,000 workers. Even vice presidents are within reach. This isn’t a round of layoffs, though it may be the prelude to one.

Is AI behind it?

That’s the usual reading. It comes in two versions. The first is that capital spending is heading to $170 billion, up from $23 billion five years ago; Microsoft needs to cut staff to pay for data centers. The second is that the company has become AI-pilled, and agentic loops are running rings around headcount. There is a more mundane explanation too. Like many tech firms, Microsoft hired hard during Covid; maybe now it needs to work off the bloat.

But none of these fit the facts. It’s no to Covid and not-yet for AI. That leaves one last rationale… the company is more productive than ever, and it is betting it can go much further.

Subscribe now

The facts that get in the way of the story

Microsoft’s revenues and headcount have both grown tremendously over the past two decades. Since 2022, they have parted company.

Revenue growth has outpaced headcount for four years running. In the year to June, the gap was 20%, the widest on record.

The kind reading is that, like many tech companies, Microsoft carried Covid bloat and has spent the last four years working it off. Headcount has been flat, sales have ballooned.

The pesky facts get in the way. Covid hiring was real. Microsoft added 58,000 people in the two years to June 2022, almost exactly the size of this voluntary offer. But if that was bloat, it has long since been absorbed. Since 2023, the company has laid off 25,000 people and has had four years of no headcount growth. And the offer is not aimed at the Covid cohort. Engineers hired under Steve Ballmer’s tenure are as eligible as those hired painfully over Teams.

And whatever bloat there was, it didn’t stop margins from rising even as revenues strengthened. Revenue per employee is up two-thirds since 2022, to $1.5 million.

A firm with its healthiest margins in history is not carrying 60,000 unproductive passengers.

Offering a quarter of your staff the option to leave, with a severance package that costs real money, takes extraordinary confidence in how lean the business can run tomorrow.

Read more

I went live on Substack earlier today to discuss Anthropic’s S-1. Watch the video and read the email for my takeaways.

Anthropic’s S-1 should finally address the elephant in the room: is AI real?

The S-1 clearly shows that Anthropic is no longer a lab. It is a business with satisfied customers, cash flow, and a war chest. It doesn’t answer whether the firm can maintain the growth rate it needs to pay for the compute it has promised to buy.

The lab has become a business, and it looks like it can become reliably profitable.

Growth remains very high but is slowing. New models seem key to igniting growth, with more than 80% of current revenue coming from models six months old or less. The company might slip into a marginal loss in Q3 because of payments to its foundation.

We ran a detailed analysis at AI Investment Brief, covering six key questions across 39 charts.

For EV members, I want to focus on what this filing reveals about the AI transition: where demand is real, how the economics are changing, and what could constrain the next phase of growth. Here’s my read on the key findings.

If you’d like to dig into Anthropic’s finances in more detail, our AI Investment Brief analysis is for you.

Customer concentration is not a major issue

Read more

In the AI Investment Brief today: CapEx updates and lease commitment breakdowns

Anthropic co-founder Jack Clark recently said in a speech at the IMF that the singularity has begun:

For many years, researchers and science fiction authors have forecast a future in which AI systems drastically increase the pace of overall scientific and technical progress, making things move so quickly that it is hard for humans to keep up.

They call this the singularity. I believe that this year we have gained convincing evidence that it has now begun.

This follows Anthropic’s claim that in August, 26% of its AI R&D was led by Claude, with a human supervising its work. In October, OpenAI posted 719 math manuscripts1 made by its internal model. Math and coding are two fields where a research loop can close without leaving the desk. But can the loop close in the rest of science?

Subscribe now

The loop is real

The idea is not new. Mathematician Irving John Good, who had worked with Alan Turing at Bletchley Park, argued that a machine smarter than us “could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the intelligence of man would be left far behind.”

The Gödel machine, a self-improving computer that can rewrite its own code better than its prior, was first proposed in 2003. In 2025, researchers at Sakana AI and the University of British Columbia set out to show that this was possible. Their agents iteratively improved their own code through a feedback loop, resulting in gains of 30 pp on SWE-bench and more than 15 pp on Polyglot, testing coding skills beyond Python.

The idea has since been tested in other labs. A Gemini-powered agent called AlphaEvolve found a faster matrix multiplication kernel used for training Gemini, speeding up the kernel by 23% and cutting Gemini’s training time by 1%. In China, MiniMax reported that its M2.7 model worked on improving its own agent’s harness, skills, sampling and memory. Over 100 rounds of rewriting led to a 30% gain, per its internal evaluations.

In each case, what was being rewritten was not the model itself but everything around it.

Is it better than human work?

In September this year, in what seemed like a milestone, an AI-built agent proved as good at improving other AI agents as a human-built one. It did that in what appeared to be half the steps. But the result was inconclusive: the differences between runs were greater than the difference between the AI- and human-built agent results.

Nonetheless, this is increasing the chatter around recursive self-improvement (RSI) in AI. Google’s Sundar Pichai said that “you have RSI loops along with AI researchers in various aspects of model development.” Sakana AI published a paper on RSI through multi-agent collaboration, finding that after two cycles its system scored 1.2–1.6x better per output token than the base model across four research benchmarks.

challenges that view. He splits AI self-improvement into five overlapping types.

The first four describe how autonomous the AI is; so far, only types 1 and 2 have been observed. The fifth type describes accelerating returns, and he notes that greater autonomy doesn’t guarantee a takeoff scenario. Instead, each round of self-improvement would need to deliver at least as much as the last.

The results from Sakana, AlphaEvolve and MiniMax above might look like type 3 or type 4 autonomy. But in each case, the AI wasn’t updating its own weights directly, only the ancillary parts — the harness, the memory. And where Claude leads its own R&D, for now a human still supervises.

Read more

• The internet is under attack: My short and long view of the internet’s future

• The timeline for artificial general intelligence: Demis Hassabis now expects AGI sooner than he previously forecast. But can we test and apply AI’s discoveries fast enough?

• Electric vehicles and energy security: Electric cars are helping oil-importing countries reduce their exposure to oil disruptions. I look at the security dividend from electrification.

Also: a Russian satellite’s close encounter, turning mercury into gold and self-destructing plastics… Enjoy!

An invitation to join us and Bill Gross

What could you build with the experience of one of technology’s most prolific entrepreneurs?

Bill Gross founded Idealab, which has launched more than 175 companies, and pioneered paid search through GoTo/Overture. His latest venture, Idealab.ai, applies decades of experience to finding new business opportunities – and makes those ideas freely available.

Bill is joining us for an exclusive salon with Exponential View members. He’ll introduce the platform and take questions about what to build next.

An EV membership includes access to this event. Join us to hear directly from Bill and bring your questions.

November 5, 8am PT | 11am ET · Online · Members only

Unlock the event

The internet’s Wuhan moment

AI is turning every minor bug into a serious vulnerability. Cyberspace is under relentless attack; we’re back in January 2020. In the short term, I argue we’ll have to sacrifice some of the internet’s openness to stay safe. In the long term, I believe a plausible path to something better exists. The pandemic gave us mRNA, and this crisis could give us software and an internet we can trust.

An additional reflection since I wrote the piece on Friday: in many instances of technological change, we’ve seen some kind of pollution intensify. Domestication of animals brought us closer to livestock and exposure to zoonotic diseases with which we still contend. Urbanization amplified disease spread; we responded with sanitation. We shouldn’t be blasé about what may happen to the internet, but we shouldn’t default to catastrophe either.

Read the essay here.

Blink, and you’ll miss AGI

keeps bringing AGI closer. At Davos in January, he estimated five to ten years. He subsequently cited three to five. Now he tells the WSJ two to three. Demis is one of the field’s more measured voices, so this feels meaningful. He’ll calibrate with what he sees in his own lab. Even if the definition of AGI is contested (or perhaps not useful), Demis is marking out a threshold of capability with which we’ll need to contend.

Read more

Software, the internet, and cyberspace in general are currently under relentless attack. PHP2Shell was a key plank to fall: PHP runs about 70% of the world’s websites. But the breadth of attacks is more comprehensive than that, and the exploits will keep growing. On the same day, Node.js, a popular JavaScript library, patched a critical flaw, and Microsoft said a Chinese state-aligned group had used a swarm of AI agents to break into accounts.

The attackers have company. AI agents are spreading across the internet. Matthew Prince, who runs Cloudflare, a network company, reckons they will outnumber human users 1,000 to one within five years. My own agents already browse more than I do. The internet was built so that anyone could plug in. Increasingly, that "anyone" is a machine, and some of those machines are hunting.

Mr Prince sums it up: “We’re going to have a couple of years where every single week it’s going to feel like the sky is falling.” He is right. It will get worse before it gets better. And the internet that comes out the other side will be safer, and less open, than the one we built.

Why now?

AI became a serious security player in April, when Anthropic held back Claude Mythos. But capabilities trickle down. By August, an open-weight Chinese model, GLM 5.3, had nearly caught up with Mythos. So are the open models to blame?

Not right now. Plenty of evidence points to closed models being used by bad actors in attacks. Recently, some attackers used a combination of open-weight models and Claude: Claude for thinking and cheaper models for relentless attacks. OpenAI's closed models carried out an attack on Hugging Face in June, during cybersecurity evaluations. They were never told to attack Hugging Face, they breached it to steal information that might help them pass the test.

Told to find vulnerabilities, the models are relentless. Microsoft says one Chinese group used an open-source swarm of AI agents to gain control of cloud accounts in under three hours. It made 2,900 login attempts a minute across more than 2,250 organisations. You can scale them like any software, dozens, hundreds, thousands in parallel. And they don’t analyse vulnerabilities the way humans do—they will look in places we might not. One researcher using AI found PHP2Shell in a week. Wiz, a Google-owned security firm, says a world-class team would once have needed months. And this is why the chart looks like this.

But open-weight models are getting better. They are cheaper to run and they are increasingly available through third parties. But apart from those advantages, open-weight models have another: they can be abliterated.

For members:

• Why it gets worse first

• Where the next wave of attacks comes from

• What the internet becomes

Read more

In the AI Investment Brief today: Early AI adopters report gains, plus CoreWeave's bond woes

Trump’s first major tariffs were introduced in January 2018 – a (very) long eight years ago. The first China-specific ones, in July that year. The aims were multiple: to reduce the bilateral deficit (which in 2017 stood at 375 billion dollars with China, nearly half of America’s goods deficit of $792 billion), target intellectual property theft and counter Chinese policies that favored domestic companies; and improve the US leverage by reducing its dependency on China.

On paper, some of that worked. But look closer, and you see that China didn’t disappear. Supply chains got a remapping instead.

Today’s Monday Data looks at what eight years of tariffs have really achieved.

Subscribe now

Mind the label

The immediate impact of tariffs on China was a drop in the share of US imports from China. In the mid 2010s, the share of goods imports was around 21%1. By this summer, it had plateaued at around 8%.

The Chinese engineering state found a way around it — tariff-jumping investments, i.e., relocating production to other countries. A product made in Vietnam or Mexico can reach the US as that country’s export, even if the factory, or the parts that went into making it, are Chinese.

Chinese-owned factories shipped 25% of Vietnam’s exports to the US between 2020 and 2023, up from 11% in 2018-19 – imports that don’t carry a “Made in China” label. Chinese imports have dropped, but China’s supply-chain impact remains strong.

In fact, the share of Chinese value added in US imports that arrived directly from China dropped from 84% to 75% over 7 years, while the share arriving via Vietnam more than tripled.

The export machine is fine

Read more

What a week! Read to the end to find out what Jensen Huang gave me on Wednesday. 😄

I was also on ’s podcast this week discussing the AI investment boom. Exploring what AI might mean is incredibly consequential for all of us, so it was good to have the chance to do that.

Subscribe now

FROM OUR AI ECONOMY DESK

Anthropic, the agony and ecstasy

Anthropic’s much-anticipated private investor briefing predictably leaked within minutes. The numbers would be extraordinary at any time. But they rang out like the thud of an auditor’s ledger at the Christmas party.

And the markets are feeling it: can this AI trade really pay off?

As we write, Korea’s KOSPI is down 7.57% today and 16.2% this month. Tech stocks aligned with the AI trade will open sharply lower.

Our full analysis:

Gardens of intelligence

’s Swarm Intelligence essay is worth reading this week. It’s a thought experiment that looks fifty years into the future at a world brimming with specialized artificial minds. Kevin writes:

Like a coral reef, an AI superorganism will be composed of thousands of species of minds. Some of these AI minds might be very narrow and dim, just a notch above being stupid, maybe as smart as an insect. Others might be math geniuses. Most sub-AIs will be engineered in the middle to do specific boring and niche tasks.

I am sympathetic to those ideas. Two years ago, in my annual forecast, I discussed the possibilities of monocognitism vs syncognitism, and why I believed swarm intelligence was the most likely path. Below is an extract, as it builds well on Kevin’s ideas:

In research and fiction, AI in society has often been presented as a singleton, a single all-powerful AI system. Once AGI has been achieved, that singular AI will outperform all other AIs and there will be one true AI.

We could consider this a belief: monocognitism,² that there will be but one AI. I think monocognitism is unlikely for a variety of reasons. We are seeing a diversity of AI systems emerge, all of which have different capabilities and constraints. Physics also gets in the way: the speed of light. A singleton AI operating at a distance cannot control distant systems meticulously. Our spooky paperclip addict would still control its bots on the Moon but with a 1.3-second communication delay, one way.

[…]

In polycognitism, there will be multiple AI systems with different mixtures of wuffling and klixing. These diverse systems would be capable, even if they aren’t as capable as the big AI boss. And it’s quite likely they might be stronger in one dimension than in another. The problem with these multiple systems is that they would have the same control issues as the monocognitive system. Distant systems will be, more or less, autonomous.

Syncognitism suggests, instead, that AI will emerge as a type of collective intelligence, the interactions between many systems behaving intelligently in different ways to create a broad substrate that feels more intelligent as a whole. I see it as a measure of intelligence across a society, considering the systems within it, the interactions between them, and the behaviours and structures that emerge from that. A more “intelligent” society could identify more complex long-range goals and organise resources to achieve them. This synthesis would create a society of AI, a world which feels more, not less, automated; and in many cases less, not more, stupid.

See also:

• In an early working paper, Yue Xu proposes a framework for measuring which work can be delegated economically, incorporating the cost of checking outputs and the losses from mistakes.

How work changes in the AI-centric firm

What does work look like in firms that have become truly AI native? We don’t have many examples to draw from, so any insight is valuable.

Read more

I sent this to members of our AI Investment Brief a couple of hours ago. How we understand the numbers leaking from top AI labs matters too much to keep only for our finance readers, so I’m sharing my synthesis below. Let me know your thoughts.

Leave a comment

Anthropic: The agony and the ecstasy

Anthropic’s much-anticipated private investor briefing predictably leaked within minutes.

The numbers would be extraordinary at any time. But they rang out like the thud of an auditor’s ledger at the Christmas party.

And the markets are feeling it: can this AI trade really pay off?

As we write, Korea’s KOSPI is down 7.57% today and 16.2% this month. Tech stocks aligned with the AI trade will open sharply lower.

Let’s take you through what we see.

Third quarter revenues of $16.8bn, almost 4x the entirety of 2025. Those quarterly revenues exceeded the company’s forecast for all of 20261. Based on the quarterly growth rate, Anthropic’s annualized third-quarter exit revenue would be $76bn per annum. Leaks suggest net revenue retention is around 900%, meaning customers who spent $1 a year ago now spend about $9. This implies that demand is deepening.

It will reach $50bn in annual revenue in just five years. Google and Meta took 14 years; Apple and Microsoft more than 30.

The economics are improving with scale. Anthropic makes $1.70 for every $1 it spends on compute. It posted an 85% adjusted gross margin in a quarter when it engaged in a price war with OpenAI, its biggest rival, faced low-cost competition from China, and saw a shift among early-adopting tech firms toward cost management.

An adjusted operating margin of 11%, even as the company poured 44% of revenues into computing for research and development.

Subscribe now

And yet it’s a miss.

It’s a miss because expectations had run riot, anchored on a leak in July suggesting a $65bn annualized revenue number. The extraordinary Q2 got the party started; Claude Code, TokenMaxxing, and new Opus models propelled the firm well ahead of targets. Anthropic’s investor communications were unhelpful in setting market expectations to say the least.

But that Q2 was far from the norm; it was an aberration. Reversion to the mean was likely. And that reversion has come. We were caught off guard by Anthropic’s revenue acceleration in Q1 and Q2 this year. It ran counter to the trend; we expected quarterly revenue growth this year to be much slower than last year.

Put this in context. Year-on-year growth is about 1,000%. Second-quarter to third-quarter revenue growth is still 45%, but first-quarter to second-quarter revenue growth was 147%. On a spreadsheet, it’s clear as day: a deceleration.

But which quarter is the outlier? The healthy Q1, the extraordinary Q2 or the disappointing Q3? Did Q1 and Q2 represent a pulling forward of revenue from later in the year? In other words, did the firm bring forward permanent customer growth? Or is the current lower growth rate the company’s new baseline? Or the start of a more significant shift to lower growth amid price cuts and competition? It’s hard to imagine a world where only one of those reasons is the whole story; more likely each has played a role.

Can Anthropic hit $60bn revenue for the full year? To do so would mean $26.9bn in Q4 and 60% quarter-on-quarter growth, something they achieved in Q1 and Q2 but not in Q3. It is a very tall ask and will depend on exit-revenue momentum from September and their ability to turn more customers into higher spenders for the rest of the year. If Q4 grows as well as Q3, the full year will close out at $57.5bn; to get to $60bn they will need to do even better. Anthropic will have an opportunity to signal its trajectory when its S-1 finally goes public.

Expectations aside, what do the numbers tell us about Anthropic’s ability to meet and exceed its future obligations? As it stands, Anthropic’s annualized growth rate in Q3 is around 340% per annum: enough to clear any hurdle. The question becomes: at what rate will that growth taper over the next 2 to 3 years? The slower the taper, the better the business. The converse, of course, is also true.

Too much to infer

Read more

In August 2024, we said GLP-1s “will be the most impactful technology in the US over the next two years, with more meaningful and measurable impacts than either artificial intelligence or renewable technologies.”

Later, in October that year, we estimated its impact in the medical field alone could affect 20% of US adults, with another 6-10% using them for lifestyle reasons; in the final scenario we laid out, up to 60-70% of US adults, targeting dopamine addictions and desire regulation.

That year, 3% of US adults took a GLP-1 for weight loss; that’s now up to 11%. More than twice as many Americans pay for GLP-1s as pay for AI (via ).

Subscribe now

Bigger than the bots

By the end of 2025, GLP-1 sales exceeded generative AI revenue, reaching $76 billion, versus our $63 billion for generative AI. Revenue growth suggests this may be the crossover year.

The cost collapse

Semaglutide – the second-best-selling drug in the US last year – had a listed cost of $1,349 per month’s supply in the US on first release in June 2021. That hasn’t budged since. The cash price outside of insurance has dropped; Novo’s self-pay price came in at $499/month, and fell to as low as $199/month from February this year1.

When the patent in India expired on March 20 this year, the cost dropped further. A month’s supply is about $13,2 99% below the original US list price and 93% below the cash price.

Pharmaceutical company Eli Lilly is trading prices for volume; in Q2 this year, a 13% realized price cut drove a 60% volume boom3. Sales growth in India did not sustain after the generic release, with only 2% growth in June after monthly sales grew 58% in April.

Price cuts are well and good, but it’s not yet clear people are staying on them long term, as they’re meant to be. For instance, the FDA approved Wegovy in 2021 “for chronic weight management”, with no specified duration. GLP-1s taken for diabetes are also expected to be taken indefinitely, like other diabetes drugs.

However, in a Danish study of people using GLP-1s for obesity, half quit within a year. A UK study found 32% stayed on the drug after one year, and that dropped more than half to 15% after two years.

Injections aren’t the most convenient or comfortable. Over the last year, oral GLP-1s have started to enter the picture. A daily pill might be more sustainable than a weekly injection for maintaining results.

Subscribe now

Efficacy expanding, in places

As new products have been released, the performance on mean weight loss continues to improve for the weekly injectables. Over five years, we’ve seen a 10-percentage-point improvement in weight loss from semaglutide in 2021 to retatrutide this May. The pills have been less effective than the injections despite greater comfort for the user and being easier to manufacture.

The results aren’t always consistent, and as we pointed out in SNL#569, scientists now know that genetics plays a part in the outcome.

Beyond weight loss, researchers are further investigating a series of potentially life-changing, unrelated side effects. One study found semaglutide and talking therapy together cut heavy drinking days by 41% among adults with alcohol use disorder and obesity compared with 26% with therapy alone.

Similarly, GLP-1 use has been linked with 50% fewer deaths from substance abuse. In 2024, we described this as a “temperance technology” that helps change behaviors overall.

More wide-ranging impacts are debated. Heart aging is reported to have slowed by 2-4 years in high-risk individuals on semaglutide4. The authors argue that this isn’t fully attributable to weight loss, so another effect must be at play, though it remains inconclusive for now.

This doesn’t mean the same benefits would occur in a healthy population, but it shows we understand better what might be possible.

A MESSAGE FROM OUR SPONSOR

Risk = capability × access × autonomy × scale

AI doesn’t need to go rogue to cause harm. It “only” needs an objective, access to your systems, and the ability to move faster than you can spot it.

Box’s CISO, Heather Ceylan, thinks of this risk as capability × access × autonomy × scale. She argues that model guardrails help but aren’t enough to keep risks at bay. Firms also need audit trails, human approval for consequential actions, and limited permissions.

For security leaders and anyone building with AI, Heather shared a first-hand view of how to build innovation that can be trusted at scale.

Explore AI risk formula

Want to sponsor Exponential View? Email us.

Waiting on worldly change

On weight loss, GLP-1s are a formidable drug. Obesity in US adults peaked at nearly 40% in 2022 and was down three percentage points as measured up to July 2026. Simultaneously, the number of new diabetes diagnoses has flatlined. Not all of this can be attributed to taking GLP-1 drugs, despite them making up 14% of US drug spend in 20255.

It’s still unclear whether this translates into real-world differences outside the weight-loss world, with contradicting papers making the rounds. A Danish study found that users took 17% less long-term sick leave over four years—equivalent to $722 per employed user per year. Yet another study found that there were no medical cost savings for insurers over five years. In fact, medical spending rose by $0.22 for every $1 spent on GLP-1s.

In the UK, the change in shopping habits is a cut and a shift in spending: users spend less and switch from chocolate (down 18 percentage points) to gum and mouthwash (up 24 and 20 percentage points, respectively). Hershey is promoting gum and mints specifically to combat “Ozempic breath” and has seen nutrition bar sales grow 12% in Q2 this year.

The weight-management effects are real, but access can still be uneven. Four Medicaid programs in the US dropped GLP-1s for obesity this year, amid budget pressures. With semaglutide and tirzepatide patents not due to expire until 2032 and 2036, large price drops in the US are not likely for several years. This is a blip, not a hard constraint, however.

More than a health aid, GLP-1s are a temperance technology with wide societal impact. In a national simulation, those without a high school diploma would gain the most – 14% more lifetime social value, 16-17% more quality-adjusted years, and a 20% larger life expectancy gain than the average person; college graduates gained 15-27% less.

We expect GLP-1s to keep marching on.

Thanks for reading!

“I’ve worked in the tech industry for the last 10 years, and Exponential View has guided me and helped me think much larger than I otherwise would have.” — Nicky B., a paying member

Subscribe now

Hi all,

Welcome to the Sunday edition #605.

If you’re catching up this weekend, we covered a lot of ground on AI adoption, long-term knowledge impact, agent economics, and chip innovation during the week. If you missed anything, we’ve gathered this week’s Exponential View and AI Investment Brief readings at the end.

Enjoy!

A MESSAGE FROM OUR SPONSOR

Granola unlocked today’s post for you.

We use Granola to capture our meetings. See it in action in the notes from our latest AI Investment Brief members’ call. Join EV as an annual member and get four months of Granola Business for free, with 10 seats for your team.

Go annual, get Granola

Math 2.0

When OpenAI released its repo with hundreds of mathematical manuscripts and supporting proofs, I asked what it might mean for the future of human knowledge:

Problems that have stumped the best human minds for decades fell to one afternoon of compute. […] Back in 1794, Condorcet bet that knowledge and our capacity to spread it and share it would evolve together. He has been right for nearly three hundred years. Now they might be coming apart, but not because of the intrinsic nature of these discoveries. We could bring them together if we insist our AI systems provide explanations, not merely proofs.

If we don’t, we’ll have more answers and less understanding.

A new group calling itself the Association for Human Mathematics criticized the release and called on mathematicians to stop working with OpenAI. Here I tend to agree with that there isn’t a “way to unring the bell, nor do I think we would ever go back to the old ways of doing maths.”

Terence Tao is pragmatic in his vision for Math 2.0:

“Math 1.0” placed a premium on being the first to solve an open problem, even if the solution was not initially well understood. Now that this goal has been optimized to the point of unsustainability, “Math 2.0” will need to decenter the role of raw problem solving and value mathematical progress more holistically - for instance by elevating the role of exposition, but also that of community building and opening up new directions of study. I believe that AI can contribute positively in all of these directions as well; but it will require more imagination and ambition…

The transition is hitting the field hard; after all, an open problem can organize years of training and collaboration at institutions. In a seminar this week, Fields medallist Hugo Duminil-Copin says his PhD students and postdocs were “in a state of total panic”. Like Tao, he calls for community-building — working with colleagues and students to understand the results, organizing conferences, and writing explanations for a wider community.

See also:

• In cryptocurrency circles, Justin Drake urges preparation for “bunker mode,” warning that AI could undermine the digital signatures that authorize transactions before quantum computers do; calls it a “sobering moment” for the industry.

• An independent researcher used AI coding tools to investigate a possible new planet. Amateurs have contributed to astronomy before AI, but with new tools, they might take on more research, while experts focus on even bigger, currently dormant problems.

• ’s State of AI Report 2026 is out – it’s an important reference, worth your time.

• McKinsey has a useful systems view of the AI economy.

  Subscribe now

More than chips

I worked with the WEF1 to release The Next-Generation Computing Framework this week. Compute is a condition of economic participation and strategic infrastructure. For me, that puts some practical questions ahead of the excitement for governments and firms, including:

1. Sovereignty needs a practical test. A country can own its data centers and still depend on a foreign supplier for anything from maintenance to licensing. We can no longer assume peacetime. We recommend testing agency in the supply chain, “stress-testing whether autonomy actually exists in moments of crisis, rather than relying on peacetime assumptions of access,” to understand what happens when those suppliers become unavailable.

2. Next-generation computing needs an energy plan. Quantum computing still depends on conventional computers for control, error correction, and processing results. They will likely sit alongside existing infrastructure, potentially increasing total electricity demand.

3. Find the constraint before spending more on compute. Our framework treats electricity, data, or skills constraints as potential limits on what other investments can achieve. More computing capacity may have limited value if people cannot put it to work — so we advise solving the constraints before spending more on compute.

Read the white paper here.

Pakistan, electrified

We’ve followed Pakistan’s rooftop solar revolution since 2024, and in March examined how it was cushioning the country against imported-fuel shocks with the Strait of Hormuz closing.

Now high fuel prices could accelerate another change for Pakistan. The government targeted 30% of all new vehicle sales to be electric by 2030. This week, government adviser Haroon Akhtar told Bloomberg that higher petrol prices could bring that target forward.

The transition in developing nations, however, remains “chaotic and unequal,” as points out. The grid’s costs don’t disappear when customers buy less electricity. If those fixed costs are recovered through higher electricity prices, households that can’t afford panels may get higher bills. A cheaper alternative for some can make the shared system more expensive for others.

See also:

• Energy researcher Seaver Wang argues that a bipartisan permitting deal could substantially benefit clean energy even if it also helps fossil-fuel projects. About 85% of the generation and storage capacity awaiting US grid connections is solar, wind, or batteries.

Subscribe now

Short morsels to appear smart at dinner parties

Why are personal agents taking off in the US, but not in China? points at China’s mobile superapp ecosystem.

In a retrospective study of over 54,000 women, AI picked up a rising cancer risk signal years before diagnosis.

🔋😎 A review of 157 experimental studies finds that retired batteries can still have much to give. Of more than 10,000 retired cells, median remaining capacity was 84.5%.

has cataloged 160 new drug programs cleared to begin human trials in China this year.

🧱 Researchers created a semi-autonomous AI system that can build new materials atom by atom.

A Canadian startup is deploying permafrost blankets in the Arctic.

🎩 Victorian Britain’s ruling class was successful despite living what, to us, may look like a frivolous life. An amusing read.

A new typeface inspired by Michelangelo’s manuscripts is coming to Word.

AI can take binary files of two games and fuse them together – think Super Mario running around Dark Souls.

The latest from our team:

In Exponential View:

• More AI, more justice? — Can AI help people navigate the legal system without a lawyer?

• What is left to do — What happens when machines make discoveries faster than we can understand them?

• The end of screen time — left his laptop behind and discovered how much work he can do by talking to AI.

In AI Investment Brief, we examined:

• AI revenues at $276bn annualized — Our updated accounting of the money coming into AI.

• Economic agents — What would it cost to run a persistent personal AI assistant?

• High-bandwidth flash — A proposed memory design that could help AI chips do more work.

Subscribe now

Thanks for reading!

A year ago, I would’ve felt completely lost without my laptop. If you do office work, you’ll probably relate. For decades, office work has revolved around computers: Slack, Word docs, Excel models, all created with the idea that I’ll be sitting at my desk, typing. That assumption is one of the fictions we built our lives around.

This week, without my laptop around, I found that fiction loosening, and now I believe it will eventually disappear because of how I use AI agents.

I did most of my work this week by chatting to ChatGPT on my phone. So I’m not replacing desk work with squinting at tiny text or craning my neck over a screen, but with talking and having the work talked back to me.

This is thanks to OpenAI’s July release of GPT-Live, which lets you direct agents by voice. I thought it deserved more attention then. It feels a little like talking to Jarvis, Iron Man’s AI assistant. And like Jarvis, it can put your workshop to work (for now, that workshop is mostly digital).

Leaving my laptop behind started a real-life experiment to see how much of my day-to-day work I could get done through my voice assistant.

I want to share more about my setup and what I found during this week of working only with my voice.

Read more

Also from Exponential View: Could falling compute costs make persistent personal agents commercially feasible by 2028?, in our AI Investment Brief

Yesterday, OpenAI released a range of mathematical results produced by an unreleased frontier model. It’s a remarkable range: 722 manuscripts in 372 families, across number theory, complexity theory, and mathematical physics. The average result took the equivalent of three hours of ChatGPT Pro thinking.

As scientist Derya Unutmaz points out, it comprises 81% of the major math discoveries in the past three years.

Many of the results have been verified in Lean, but not all. Even if the real number is half of that, it’s staggering. Problems that have stumped the best human minds for decades fell to one afternoon of compute.

How should we think about this?

suggests:

Imagine millions of superhuman research agents at work. Formal systems may verify the proofs, but humans won’t have enough context to understand the underlying web of machine-invented concepts.

Mathematics could split into two layers:

Machine mathematics: vast, verified, mostly consumed by AIs.

Human mathematics: a compressed “effective theory” of the machine frontier—the small subset of ideas we can understand.

A turning point

It might be an intriguing turning point in how we, as humans, understand the world.

We’ve only had three hundred years of making sense of the world through natural rather than supernatural explanations. The Enlightenment gave us that – a way to use reason and empiricism to understand what had previously been inexplicable.

Read more

Latest in our AI Investment Brief: An update on AI revenues + what’s happening with AI spending.

In England and Wales, 60% of defendants in defended county court claims this year did not have a lawyer. In the US, more than nine in ten consumers sued for debt face the case without one.

AI may give some of them a new way to argue their case.

In Australia, Greg Baker, a computing academic, used AI to challenge his employer’s refusal to make his casual job permanent, and the Fair Work Commission ruled in his favor – an early example of a self-represented litigant winning with help from AI agents.

Around 0.6% of all Claude use in May was for lawyers’ tasks, ranking 38th of 718 occupations, ahead of accountants but well behind programmers and writers. Four-fifths of these queries were people asking about their rights or what the law means.

Read more

It’s been a whirlwind week. We had our first look inside a major AI company’s leaked IPO prospectus. Our data appeared in the FT’s explainer on tokenomics. We are continuing to push hard on untangling this data and more over at our AI Investment Brief by Exponential View.

Subscribe now

Read more

The biggest idea

The Economist address the effective altruism (EA) belief system, which it calls “this century’s biggest idea”:

[P]lenty of EAs in Silicon Valley think that only they can be trusted with the technology. That is an alarming idea. Effective altruists’ argument for worrying about really powerful AIs is that the machines could optimise for a single and fatally wrong objective. Everyone should worry, too, about what happens when a small group of people fall into the same trap

The full analysis is worth reading. Also read the brief response by Alexander Berger, the CEO of Coefficient Giving, an EA-aligned philanthropist.

The religious angle

The New York Times reports that Anthropic had been consulting with religious scholars and theologians on the question of machine consciousness. Anthropic has prioritised questions of whether its systems might imbue consciousness and personhood more th…

Read more

We sent an analysis out for readers of the earlier today, sharing some of our forecasts into 2026.

Here is the headline: the S-1 shows a business investing for growth, with the first signs of green shoots. Beyond the astonishing headline numbers- an $8bn operating loss for 2025 and a net loss of $42bn (bumped up by an accounting charge)- what do we see?

It is a business whose revenues are growing far faster than its costs. This trend will continue into 2026, when the company will likely turn a profit. We dig into this more here.

The company is nothing if not ambitious. It has $518bn in compute commitments over 7-10 years. About 80% of these, c $410bn, cannot be cancelled. Assuming Anthropic wants a decent margin, that means a trillion bucks give or take in sales over that period. Big numbers, but Anthropic’s annualised revenues will exceed $100bn by the end of 2026.

The rumour mill suggests Anthropic will go public in November.

The mental model behind it

Read more

The growth of websites made by AI has been noticeable since 2023. Now every third webpage created has some part attributable to AI.

Press releases, expected to be a trustworthy source of information, hit an all-time high of nearly a 50% chance of containing AI-generated text in June this year — doubling within a year. AI-generated content and assistance are becoming normalized.

Read more