13 July 2025

News

AI hiring wars

The AI hiring wars continue. Meta is continuing to buy top execs and researchers with unprecedented pay offers, picking up Apple’s head of models for what Bloomberg says is ‘tens of millions of dollars a year’. Down the road, OpenAI’s $3bn deal to buy Wiresurf (AI coding tools) fell apart, and instead Google hired the CEO and a bunch of senior employees while taking a licence on its technology for $2bn. The Information reports that OpenAI expects to spend $6bn on stock comp and another $1.5bn on cash comp this year as it tries to hang onto the people Meta hasn’t hired yet. See this week’s column. APPLE, WINDSURF, OPENAI COMP

AI browser wars

Perplexity released its own web browser (‘Comet’), and Reuters reports that OpenAI plans to do the same. 

If you want to become the starting point of people’s internet experience and computing experience, and to replace Google as the gateway to the internet’s information, then the browser is the key point of aggregation and experience. Given that so many people’s day-to-day computing experience now happens entirely in the web (at least on the desktop), the browser is also effectively the OS. And if you want to build a personalised LLM that understands someone, then you want to see everything they do online. So you make a browser. But, of course, that doesn’t help you on mobile, where most people’s real computing experience moved a long time ago. You can see my Amazon orders on the desktop, but not on mobile, and you won’t see my Instagram or my TikTok. OPENAI, PERPLEXITY

AI Infrastructure wars 

Surprising no one, OpenAI is building a physical infrastructure team, as it continues to diversify away from Microsoft. LINK

Coreweave, a cloud GPU company/Nvidia proxy/financial instrument (opinions vary) merged with Core Scientific, another crypto GPU company that backed into LLMs, for $9bn in stock. LINK

And Nvidia became the first company with a $4tr market cap. LINK

Meta glasses

Meta thinks it can get out ahead on AI-powered glasses, and this week it bought 3% of EssilorLuxottica for $3.6bn. It could have got at least half-a-dozen AI researchers for that money. LINK

Apple does F1

Apple’s Brad Pitt Formula One movie seems to be a hit, and now it’s trying to buy the streaming rights for F1 races in the USA. I still don’t understand what strategic benefit Apple gets from being in Hollywood, nor what it adds. LINK

US fintech fees

JPMorgan startled the US fintech scene, declaring that from now on it will charge for API access to customer data—for example, using Plaid to put your payments into a wealth management service. Up until now, this had been free. LINK

Elon’s Nazi bots?

After a new software update, Grok, the LLM from Elon Musk’s xAI, started giving explicitly Nazi responses to questions, praising Hitler and attacking Jews, and referring to itself as ‘MechaHitler’. The posts were deleted. This seems at least partly due to new instructions to the model: “The response should not shy away from making claims which are politically incorrect, as long as they are well substantiated”. As Elon never seems to learn, content moderation is harder than rocket science. LINK

Meanwhile, the CEO of Twitter, Linda Yaccarino, has stepped down after two years in which she never appeared to have any actual authority or control over what was happening at the company. Her job was supposed to be running ad sales, which is hard when your boss tells advertisers to “go f*ck yourselves” and the product, well... see above. LINK

Ideas

The media’s ‘traffic apocalypse’. LINK

Walmart’s internal AI apps platform. LINK

This week’s interesting AI paper - the ‘chain of thoughts’ that reasoning models generate to show how they reach their conclusions might not actually be how they’re working things out at all. LINK

Analysing the performance and cost of xAI’s Grok model. LINK

With everyone in publishing preoccupied by the replacement of SEO with LLMs, it’s symbolic that Demand Media, the original SEO content farm, was finally scrapped this week. LINK

Outside interests

A narco-sub remote-controlled over Starlink. Paging William Gibson. LINK

Jane Birkin’s original Hermès ‘Birkin’ bag sold at auction for $10.1m. LINK

On movie silencers. LINK

Data

MIDIA’s global music forecasts. LINK

Updated and pretty comprehensive study on how much people in the US are using LLMs at work so far. Around 45% say they’re using it at all, of which a third claim to use it every weekday - so 13-14% DAU. LINK

Vercel surveyed AI developers. LINK

Column

AI and acquihires

An acquihire used to mean that you bought a small company that hadn’t really worked, for single or low double-digit millions, so that you could pick up the entire team. But now it means you pay a couple of billion dollars for a minority stake, or even just to ‘licence’ the tech, and the key members of the team abandon the company to work for you, leaving a shell with some startled investors, confused customers, and a bunch of employees holding suddenly-worthless stock options. 

This is, obviously, an end-run around competition regulators - you didn’t buy control of the company, only a minority, and in the case of Wiresurf Google apparently hasn’t bought any equity at all. Regulators are already unhappy about this (while I’m not a lawyer, I wonder if it might be easier for them to intervene regardless in principles-based jurisdictions like the EU or UK than in the more courts-bound USA). It’s also an end-run around conventional start-up governance and in particular of liquidation preferences. Distorting capital structures and governance structures like this tends to mean prosecutions when the dust settles (check your email retention policies!). And what happens to those employees and their options if their company got bought at a $4bn valuation but didn’t actually get bought at all? Do they get made good? Maybe, maybe not. What happens to that social contract? I also wonder about morale for employees at Meta and Google that are just as good and have their names on just as many papers and didn’t get a buyout offer like this. And if you got a $100m 4-year deal, how hard do you work after the 1st $50m?

All that aside, in the last platform shift Meta bought products and user bases in Instagram and WhatsApp: this time people are buying the team alone. There’s a contradiction between the pretty widespread perception that models seem to be commodities with open source the growing trend and the fact that several giant companies that already have model-building teams are willing to spend billions of dollars to get more. 

Of course, a technology can be both a commodity and very hard - think of flat-panel screens, say. Meta’s spending spree follows Llama 4 being a huge disappointment. I think we could also compare this to web search: Google has advantages of scale and network effects, but it also has a 25-year accumulation of institutional knowledge - of people and teams that know what all the problems are and how to solve them. LLM research has been around for a lot less time, but there has been an equivalent explosion in the complexity and sophistication of building a SOTA LLM, and people who know how all of that fits together remain rare. And when the Big Four plan to spend over $300bn on capex for AI this year, what’s another $100m for an SVP? 

But a long time ago there was a number sticker in Silicon Valley that said ‘Please God, just one more bubble!’. All important new technologies tend to lead to a bubble and a crash. With the crash, we generally uncover fraud on the margin that got swept up in the rush to get the deals done, and we can see the waste that came as well. But we also get railways or the internet.