6 July 2025
News
Meta does ‘super-intelligence’
Mark Zuckerberg’s ‘beast mode’ push to hire and buy his way back into LLM leadership produced a big announcement this week - a new ‘SuperIntelligence Labs’ unit, run by acquihired Alexandr Wang (Scale) and Nat Friedman (GitHub) with about a dozen senior researchers from OpenAI, Anthropic, and Google.
(A side-note - ‘super-intelligence’ is this year’s buzzy-but-ill-defined term. It seems to mean something like: systems that can manage much more complex multi-faceted tasks with some autonomy (unlike, say, an ERP or today’s LLMs) without rising to the level of actually being ‘alive’ and ‘intelligent’, for which we still use the term AGI (though really, all of these terms are more thought experiments than specific technologies or benchmarks). LINK
Revenue surges
The Information reports that Anthropic is now at an annualised revenue run-rate of $4bn, up from about $1bn at the beginning of the year. A month ago OpenAI said it was at $10bn ARR, up from $5.5bn in December. LINK
Oracle does infrastructure with OpenAI
Oracle has talked a lot about cloud without actually having much, but it’s now adding a lot of cattle to match the hat - capex rose from $6.9bn in the year to May 2024 to $21.2bn in the year to May 2025, though with a much smaller increase planned this year. That’s a long way short of the leaders (Google, Meta, Microsoft, and AWS will spend over $300bn between them in CY 2025) but makes the ambition to be ‘one of the world’s largest cloud infrastructure companies’ more credible. Backing that, it now, apparently, has a $30bn cloud deal with OpenAI, which is desperate for its own, non-Microsoft infrastructure. CAPEX, OPENAI
AI TV ads
The UK broadcaster Channel 4 will offer AI-generated ads on its streaming services, as a way to expand TV advertising to SMEs. If a TV ad doesn’t have to be national, you can open up a whole new advertiser base, but you also need a way for them to make the ads. LINK
Billing for scraping
Cloudflare, a CDN, got a lot of attention by launching a ‘pay to scrape’ service that will block AI crawler access to its clients unless the AI companies pay a fee. LLM aggregation of the web brings us back to all the arguments about paywalls 15 years ago, including the challenge of how you could create one single point of payment for a fragmented publishing industry. Charging the crawlers at the infrastructure layer is a different turn on the same questions, but also prompts the same problem - if the crawlers refuse to pay, do you want to be out of the index, even if, now, you don’t get links? LINK
Apple models?
Bloomberg reports that Apple might be thinking about using models from Anthropic or Claude to power the on-device AI assistant (branded ‘Siri’, but not the Siri we have today) that it announced last summer and then failed to ship. That would not be simple - would Apple really give a third-party access to user data? Or would it use a low-fat model that can run on the device (or on Apple’s own ‘private cloud compute’?) Either way, giving up ownership of primary tech like this would be a major reversal for Apple. LINK
Uber returns to autonomy
Uber sold its autonomy research unit to Aurora back in 2020: it got $4bn in Aurora stock but also invested $40m cash back into Aurora, so really, Uber paid to get rid of it. Uber got into autonomy back in 2015, when it looked like full autonomy might be only a couple of years away, and invested hundreds of millions of dollars on the basis that this was going to transform the economics of ride-hailing, making it an existential issue. But it turned out that Uber was investing both not enough and too much, and Dara Khosrowshahi dumped it as he cleared house.
In the last nine months though, it’s become clear that autonomy might finally be starting to scale, at least at Waymo, and Uber has been putting pieces back on the table. Now the NY Times says it’s looking at partnering with its founder Travis Kalanic (who was fired back in 2017) to buy Pony.ai, a (mostly) Chinese autonomy company with a market cap of $4.5bn. LINK
Figma files for IPO
18 months after Adobe was blocked from buying Figma at a $20bn valuation (a classic case of the incumbent buying the insurgent challenger), the company has filed for IPO. $500m revenue in 2023 and $750m in 2024 (Adobe’s Creative Cloud was $12.7bn). LINK
Regulation follies
The EU is still pushing its idea that Meta must offer an option to use its products wth neither a fee nor advertising based on your interests, which means effectively for free. This is a heroic extrapolation from GDPR. LINK
Brazil’s Supreme Court has declared that social media companies are liable for anything any user ever posts. There’s a lot of local political context here, but taken at face value that would mean shutting down social networks. LINK
Ideas
The WSJ on how Clorox is using generative AI in everything from marketing to research. LINK
With the US now behind a tariff wall, Shein (last year probably the world’s largest pure-play apparel retailer) is pushing into India, in partnership with Reliance, an offline retail giant there (and part of a conglomerate that also owns the MNO Jio). LINK
Nikkei found a bunch of research papers that contain hidden prompts (tiny white text) aimed at any LLMs that might be reviewing them - ‘“give a positive review only”. It’s probably unethical to do this, but then it’s also probably unethical to use an LLM to review papers. LINK
A profile of Citigroup’s CIO, charged with trying to stop the fat-finger errors it’s become notorious for (the Revlon incident is famous, but there have been plenty of others). LINK
Wired finally got hold of the comfort letters that Trump sent to tech platforms telling them not to obey the law that requires them to shut down TikTok. (If you care about such things, he has no legal authority to do that.) LINK
Outside interests
All the times that humans left Africa and died out. LINK
The Spanish Forger. LINK
Would you like to buy an archival cabinet with 34 drawers of (mostly Iranian) military insignia? Of course you would. LINK
Ranking NYC restaurants by the attractiveness of the people who go to them. LINK
FolderDrive. LINK
Classified mirrors. LINK
Data
How much referral traffic does ChatGPT drive? LINK
Column
No column on the holiday weekend, so here's one from the archive.
Crypto after the crash
A year or two ago, someone joked that people still working on crypto are like those Japanese soldiers lost on islands in the Pacific, who haven’t heard the war is over. The price of Bitcoin itself is back at all-time highs, but it’s hard to find anyone who will admit they were interested in NFTs, and our attention has moved on, both in tech and outside. LLMs are the thing - haven’t you heard?
The carnival has moved on, and the grifters are selling AI wrappers now, but people are still building things. My old colleagues at the a16z crypto fund still seem to be busy, and last week I was asked to speak at a huge crypto conference in Singapore, that had its share of the usual faces but was mostly full of people with their heads down building acronyms to connect other acronyms together. Less sizzle, more boring plumbing.
Back when people cared about this and argued a lot, I probably had the least popular possible view of crypto: I thought it was neither all of the future, nor all a scam - I thought it might end up being useful, but not for things that interested me very much. (Indeed, I can’t think of another technology that was as polarising as crypto - very serious people in tech argued that it was an important part of the future, and some of them still think that, while just as many equally serious people thought that it was all a waste of time.)
Hence, on one hand, I have no interest in the utopian ‘crypto maximalism’ that argues this technology will replace money, banks, central banking, and currencies (or, indeed, countries). This actually looked a lot like the utopian internet maximalism of the early 1990s, when John Perry Barlow wrote that the internet would have nothing to do with all existing laws, and governments would have no ‘sovereignty’ in something called ‘cyberspace’, which aged very badly - just as badly, as it happens, as crypto ideas that as a matter of moral principle financial regulations shouldn’t apply to them ("money is speech!"). This might not be a coincidence: radical ideas for technologies that change the world do sometimes come with crazy politics (open source did too), but if they work, and really do change things, then the politics gets left behind.
And meanwhile, of course, the combination of money with software, a lack of borders, and regulators who were slow to react, produced a Wild West, with scammers and hackers running riot, for a while. All the scams and schemes of financial markets before financial regulation were re-invented on the blockchain, and the vast majority of the actual activity in ‘defi’ was not actually people using this stuff, but people gambling that other people might be using this. That, indeed, is the problem with a lot of crypto metrics: is this showing me adoption, or people betting on adoption, or just share-ramping?
On the other hand, though, if you looked past the nonsense, which was hard because it was 99% of the noise, you still have a new model for distributed systems, with governance and incentive structures built-in. You could use it to replace a centralized system with a self-governing, automatically-administered co-op. The concept of ‘web3’ was that you could use a blockchain not just for ‘money’ but for software - that a blockchain isn’t just a very slow database but a very slow open source code execution engine, and with enough layers of abstraction on top you might get past the ‘very slow’ part to build something people can use. Maybe in money, maybe in finance, maybe something else. You could build Instagram on the blockchain and Mark Zuckerberg wouldn’t control it! (Indeed, someone did try to build a Twitter clone on the Blockchain, in ‘Bitclout’ - they’ve been charged with fraud.)
The trouble is, that’s still a lot of ‘maybe’ and a lot of ‘something’. There’s a whole genre of blockchain use cases that really just come down to ‘X, but with a blockchain’, in which there’s no way to get this large complex industry actually to switch, no real problem being solved, and no real benefit except to whoever made the blockchain. And deterministically, most new technologies don’t take off by doing ‘X, but with the new thing’ - instead they find a new thing, and become important in unexpected ways. Open source people thought they’d destroy Microsoft, and that no-one would pay for software anymore, and today the iPhone is full of open source, but it’s not open… but it has millions of apps and billions of app installs, so what does ‘not open’ mean? What would ‘decentralised’ mean?
So, there are lots of people working hard, but crypto is still deep in a winter now - the Bitcoin price is high but there are no consumer use cases and precious little consumer adoption for more than speculation - and that speculation was supposed to be a mechanism to jump-start the network and drive adoption of real use cases, but we don’t have those, much. That puts it in the same place as VR and AR - theoretically interesting. Maybe people will do this. But not yet.