22 June 2025
News
Zuck goes beast mode
Last week Meta paid $15bn for 49% of Scale.ai, plus the CEO (while most of the clients will now leave). This week we hear that it’s trying to hire Nate Friedman (formerly GitHub CEO) and Daniel Gross (Ilya Sutskever’s co-founder at Safe Superintelligence) and buy out their AI venture fund. Sam Altman claimed Meta is making hiring offers worth $100m each to OpenAI staff. And it was reported that Meta has talked about buying Safe Superintelligence (which last raised at $32bn), Thinking Machines ($10bn), and Perplexity.
Even if all of those deals came off, that would still be less than the 10% of Facebook that Mark Zuckerberg paid for WhatsApp. Companies that are led and controlled by their founders have an arguably unique ability to make moves like this: generative AI is the future, Llama 4 was a disappointment, and Zuck is determined to get to a leading position, whatever it takes, because the alternative might be to get Blackberryed. (Remember, meanwhile, that Meta’s investment in AV and VR is now close to $100bn after more than a decade and still years from traction.) ALTMAN, FRIEDMAN, SSI, OTHER DEALS
Placing stories
Meanwhile, Bloomberg reports that Apple also looked at buying Perplexity - again, a giant tech company that needs to catch up with AI, but Apple doesn’t do big acquisitions and it’s hard to see what Perplexity has that would get Apple where it wants. Perplexity puzzles me - I know people who like the product, but it doesn’t have its own models, it struggles to rank in the top 200 in the App Store and Google Trends tells a similar story: the CEO is good at saying interesting things in public, but it’s hard to see that this has traction. So, when you see two M&A stories attached to its name in the same week, you wonder if the bankers are flying a kite. LINK
Another case where the leaks might be telling more than one story: the WSJ reports that OpenAI is so upset with its Microsoft relationship that it thought about making a competition complaint. Who placed that call to the WSJ, and why? And does Jony Ive have a WSJ subscription? LINK
Anthropic made another play for regulatory attention with a paper claiming that LLMs will ‘lie, steal and blackmail’ - which actually means that they asked an LLM what it would do in a situation where most people would do X, and the LLM returned the result ‘I would do X’. Yes, well done. All of these analyses have the same intellectual validity as writing ‘I’m alive!’ on a piece of paper, photocopying it, and then saying “Look! The photocopier says it’s alive!” LINK
The week in AI
Amazon’s CEO Andy Jassy got a lot of attention by stating the obvious: generative AI will change what kinds of white-collar jobs there are at Amazon. LINK
I don’t normally cover valuations, but it’s worth noting that Anysphere, which makes the super-hot coding assistant Cursor, is apparently being approached by investors offering a $20bn valuation. It raised $900m at a $10bn valuation earlier this month(!), and it’s one of the fastest-growing startups ever, with $500m ARR. LINK
Equally, Mira Murati’s pre-product AI lab ‘Thinking Machines’ raised $2bn at a $10bn valuation. LINK
Midjourney has launched a video model. LINK
OpenAI appears to have a US Department of Defence deal worth up to $200m. It feels like a long time since Silicon Valley companies had employee rebellions against defence work. LINK
Two noteworthy releases of curated training datasets this week: Harvard is publishing a collection of around 1m books, which is 394m pages and 242bn tokens, and Essential-Web is a structured dataset of the open web with 24tr tokens. For context, Llama 4 was trained on 30tr tokens. BOOKS, WEB
Bundling and unbundling TV advertising
As TV advertising goes smart, Disney and Amazon have formed a partnership to link their TV ad inventory, while in the UK the three biggest legacy networks (ITV, Channel 4 and Sky) have done the same. DISNEY, UK
Netflix streams linear
A fascinating story from Cannes: Netflix will distribute linear channels and on-demand content for TF1, the leading French legacy TV channel. This reminds me a tiny bit of Barnes & Noble outsourcing ‘online’ to Amazon. LINK
Re-regulating crypto
One of the things that drove parts of Silicon Valley to vote Trump was Biden’s attempt to treat crypto as nothing more than a branch of the securities industry, prosecuting anyone trying to turn it into software. Now the policy flips, with a bill creating a formal regulatory regime for so-called ‘stablecoins’ (tokens that are pegged to a currency and don’t fluctuate on their own). Coincidentally, Trump has enriched himself by issuing all sorts of crypto assets of his own since the election, which is probably something that the writers of the US Constitution had thoughts about. LINK
No ads, no games, no gimmicks
WhatsApp famously had a sign on the wall saying that back when Meta bought it, but now Mark Zuckerberg, perhaps as part of the ‘beast mode’ seen above, will add advertising. The inevitable hand-wringing interests me less than how this will work - WhatsApp has far less data about what you’re really interested in than Facebook or Instagram. LINK
Tesla ‘robotaxis’
Waymo is doing hundreds of thousands of full-self-driving robotaxis trips every week, but with a very expensive LIDAR/RADAR/camera sensor stack, while Tesla’s dream has always been to make this work with nothing but cheap cameras plus the brute force of raw data from all the cars it’s already sold. Deterministically, that has always sounded great, but as a simple matter of fact, it hasn’t actually worked yet: Tesla has been promising without delivering for a decade now. Now it will launch a very limited full self-driving (as opposed to ‘Full Self Driving’) taxi service in Austin - with a handful of vehicles in a geofenced area, using current Model Y vehicles, each with a human safety driver, much as Uber and Waymo were doing this back in the 2010s. LINK
Ideas
Andrej Karpathy on how AI will change software development. LINK
The Great PowerPoint Panic of 2003- worth reading as people start panicking that generative AI will somehow rot our brains. LINK
Analysing Apple’s latest on-device LLMs. LINK
Krebs breaks down a malicious content/malware distribution network. Worth reading for a glimpse of how the Russians spend their time. ILINK
A McKinsey piece on a second wave of corporate genAI deployment: after giving everyone Copilot (with limited success), companies look at ways to automate very specific workflows with agents instead. Very jargon-heavy but worth reading between the lines. LINK
Google on how to analyse marketing effectiveness. LINK
Outside interests
Christies has a sale of ‘legendary trunks’. LINK
RIP to Fred Smith, founder of FedEx. LINK
Data
Coatue published a comprehensive deck on the state of venture markets and the valuations of corporate AI strategies. LINK
Bloomberg’s latest EV outlook. LINK
The 2025 Reuters Institute Digital News report. LINK
Bloomberg’s credit card panel reports that Temu’s US sales are down by over a quarter - exactly as expected, given how it was hit by tariffs and the end of the de minimus rule, and that it reacted by cutting ad spend. LINK
Column
I've been on too many planes to write a column this week, so here's one from the archive. Normal service resumes next week).
AI and the death of links
In the beginning, we called it ‘the World Wide Web’ because it was a web of links. Pages linked to pages and sites to sites, and you could surf or browse from one end of the web to the other by following the links. People made pages with lists of their favourite links, and ‘web rings’ and blog rolls. This was a flat, two-dimensional structure.
But though a web page itself is free, permissionless, scalable global publishing, how do you find things? When there are billions of links, you can’t follow all of them. How would you find some specific thing that you wanted, and how would you find something that you didn’t know you wanted? So we tried ‘portals’, and Yahoo tried building a hand-curated hierarchical directory. In hindsight, this sounds like the Borges story about an empire that tries to create a 1:1 scale map of itself - ‘imagine a website that listed every website that there was’ - and that didn’t scale either, with Yahoo giving up at over 2.5m listings.
Then Google made search work, by treating that web of links as an information system rather than as navigation, and as a vast mechanical Turk. Google made search work at scale, and then Facebook made recommendation work at scale as another mechanical Turk (though who it worked for is a matter of opinion), so now you might discover things you liked that you didn’t know about. This was centralisation, and a star network instead of a mesh, but it was still links. Google and Meta absorbed little bits of publishers around the margins (restaurant reviews, maps), but Google and Facebook don’t distribute content - they send you to content.
The next step, I think, was not apps (which, conceptually if not technically, are just websites encapsulated), but the walled garden model of YouTube, Instagram or now TikTok, and Twitter, LinkedIn and Threads. The publishing, the search & discovery, the consumption, the metrics and the revenue model are wrapped up into one sealed system. There are links in, but not many links out, and the ‘links’ within the system aren’t a web anymore, but controlled and optimised by the product. These are systems, not protocols, but like Google or Facebook, they’re mechanical Turks. YouTube or Instagram don’t recommend something to you because they understand you and understand videos or pictures: they analyse what millions of people do, and LinkedIn knows that I know that person because I said so.
The theory of ‘Web3’ a few years ago (before that term became just another synonym for crypto) was that you could build something like Twitter or Instagram on a blockchain, as a decentralised system where all the participants would have votes and everyone could see the rules, as code running in real-time. It would be ‘open source’ not just in the sense that you could download the code but in the sense that you could see it as it runs, and it would work something like a co-operative, with votes, rules and economics determined by that code. Now there wouldn’t be one person or one company in charge. This overlaps a little with the idea behind Mastodon and behind the ‘fediverse’ (which is also the second coming of RSS) - can you combine the higher-level protocols of Instagram or LinkedIn with the permissionless decentralisation of the web? A little reflection suggests that this creates as many problems as it solves (if it solves any), but meanwhile, like all open source projects, it’s more appealing to developers than users, and either way it hasn’t happened yet.
However, even within the walled garden, whether it’s Instagram or Mastodon, and regardless of who controls it, the atomic units of content remain decentralised and permissionless, and they are all created and consumed, one at a time, by people.
LLMs suggest something else. Generative search shows me an answer, not a link, that’s synthesised from everything on the web - it synthesises all of those atomic units into a new one. Generative content, as Meta is already discussing, means that my Instagram or TikTok feed doesn’t just show the thing I might like, but creates things I might like. Does this matter to me? It depends on why I want to see it. If I take a photo of my fridge and ask what to cook, should I get a perfect match to a recipe, somewhere on the web, that already exists, or should something that synthesises those recipes, or a new recipe based on all of them?
As plenty of people have pointed out, this is a discontinuity from the underlying or at least implicit social contract of all the phases of the internet I’ve suggested above. For any content creator, that contract was that people can discover your content on a platform, and the platform makes money, but to engage with it, people come to you (links), or the platform pays you (YouTube, TikTok), or both. This exchange has often been fraught, with people endlessly complaining about how the search algorithm works or the payout ratio, but there is still always an exchange. If, now, my content is just the raw material for you to synthesise something else, that doesn’t connect to me or pay me at all, then why am I participating?
One answer, as a former sell-side analyst, is that we no longer have 20 people putting out a quarterly results note on Verizon, all saying that “the results were in line with expectations,” partly because Bloomberg just generates that automatically. If I can take a photo of five ingredients and my phone suggests a recipe instead of linking to you, how valuable was your recipe? We have a million tomato soup recipes, becaue the web means we can, but how many do we need? If my work can be synthesised into generic and predictable answer, what was it worth? If your business was to say something that people tend to search for or link to, and get ranked or linked, and make some money from ads… well, no arbitrage business lasts forever. As Tim O’Reilly says, data isn’t oil, data is sand (and now LLMs will generate a lot more sand).
Conversely, if I’m a marketer, what’s the SEO for LLMs? If people ask Perplexity what life insurance to buy, how do I persuade Perplexity that I should be considered? Who’s reading the ads and what persuades them? And if the SEO/SEM/link model has room for 10 interchangeable companies in your space, what happens if an LLM gives everyone the same two answers?
In other words, neither blocking the LLM crawlers, nor demanding payment (even with a blockchain), nor priming them with your own answers, seem like solutions unless you have something unique and timely that they actually need from you in particular. (To that point, we saw in the last few years just how little news is really worth to Google and Meta - people don’t want it that much - and it’s hard to see why it would be worth more to OpenAI).
On the other hand, if you have something different and unique, then you might rise above all of that permission-less publishing, and be easier to find. If the generic and obvious has less value, what does that mean for the unusual and different? The creative challenge for systems based on aggregating data from human activity (and of course LLMs are also mechanical Turks) is that creativity is variance and variance looks like an error, so theoretically, the more that you know what’s differentbut good, the better you’ll be.