17 August 2025
News
OpenAI fumbles?
As I noted last week, ChatGPT 5 continues the steady, incremental model improvement we’ve seen in the last 2-3 years, rather than being a step change. However, because it’s called ‘5’ rather than, say, 4.6, and because of the way Sam Altman and other people at OpenAI had talked about it, and casually tossed around (essentially meaningless) terms like ’super-intelligence’, there’s now something of a backlash because the model isn’t a big further jump. It’s about time for the hype cycle to turn, or at any rate people are trying.
Slightly more interesting: the most significant part of GPT5 is the ‘router’ that auto-selects which model to send your query based on its complexity, and as part of that OpenAI got rid of the model picker entirely and retired the 4o model. But it turned out that a small but very loud group of early adopters preferred the tone of voice of 4o, so OpenAI brought it back. This is a classic step in the maturing of a space: as you build out the product, you simplify and abstract, and remove access to the nuts and bolts at the lower levels, but your earliest users like twiddling the nuts and bolts themselves. So you either keep that complexity and then have to maintain it (Windows, Android) or hide it (the web, Mac, iOS).
Finally, Sam Altman did a PR dinner with a group of SF tech journalists, with many maybe-casual remarks, such as that the new head of apps, Instacart’s Fidji Simo, will be launching multiple consumer apps outside of ChatGPT. That gets to the core of all LLM debates: is this one universal UI and one universal product, or an enabling layer and API for many other things? Fewer and fewer people really believe the former. SLOWDOWN, OLD MODELS, DINNER
Chips to China
Chinese open-model labs are doing well, but they’d be doing better with access to Nvidia’s latest and best compute. DeepSeek’s latest model is apparently delayed by attempts to use Huawei’s chips instead, and there’s a lot of smuggling to fill the gap: the US has been hiding location trackers in shipments and two people were just arrested in California.
Trump’s position, on this as so much else, has been spinning like a weather vane, but the latest idea is that Nvidia (and AMB) will pay the US government 20% of revenue from sales in China - for Nvidia this applies to the H20 product that was designed to comply with a previous set of export restrictions. This belies the point of export controls - if you don’t want China to have the chips, the price is irrelevant. DEEPSEEK, TRACKERS, ARRESTS, FIFTEEN PERCENT
Bailing out Intel?
After Trump (briefly) claimed that Intel’s new turnaround CFO is too Chinese, apparently the White House is now considering buying a stake in Intel, presumably to support the investment it needs to get its next-generation process to market and stay in the game. As I’ve written before, you don’t have to be Trump to think it would be a major strategic problem for the USA not to have its own cutting-edge chip manufacturing capability. LINK
Perplexity pumps
Perplexity claimed to be bidding $35bn for Chrome, which Google may be obliged to sell as part of one of the ongoing antitrust cases it’s involved in. This was another of the publicity stunts that this company favours (so transparent that literally everyone saw through it - giving an ‘exclusive’ to both the WSJ and Bloomberg didn’t help), but it prompts two thoughts. First, we should remember that while antitrust cases are very boring and rumble on for ever, Chrome really might be subject to a forced sale, with its billions of eyeballs but with no revenue of its own, and that no-one really knows what that would look like. And second, no-one really knows what the UX of LLMs will be (note OpenAI above talking about spinning off apps), but memory, user data and distribution all seem important. LINK
Ideas
China’s top 10 open source AI labs. LINK
Bloomberg and the FT both have stories on how financial markets (especially credit markets) are putting together financing for AI datacenter contraction over the next few years. Meta just raised $29bn for its project in Louisiana, and if there isn’t a big contraction (a huge IF), then this is something like $3tr by 2029. BLOOMBERG, FT
A claim that the ad-tech company AppLovin (market cap: $150bn) side-loads ad-filled games onto Android phones without user permission. (Even if not true, it could be: the openness of Android versus iOS is a trade-off.) LINK
India’s AI sovereignty plans. LINK
Perplexity is doing telco distribution deals in India, and with Deutsche Telekom. INDIA, (MAYBE), TELEKOM
A16Z on how LLMs might change e-commerce search and discovery. LINK
An analysis of the chatbot assistants that have been deployed by major retailer websites. LINK
NY Times - 21 ways that people use generative AI at work. LINK
Havas launched a tool to help brands track what LLMs say about them. Expect far more of this. LINK
On the other hand, plenty of people talk about strategies to influence LLMs’ brand ‘opinions’, but there’s no science behind it yet. LINK
Outside interests
This house in Astoria is a labour of love and commitment. LINK
A first look at Sothebys’ refurbishment of Marcel Breuer’s old Whitney, one of the most beautiful buildings in New York. LINK
This Russian oligarch’s super-yacht, impounded by the USA, is now up for auction if you have $300m or so to spend, plus running costs. The interiors are huge, ugly, and uncomfortable, but you do get two pianos, a glass elevator, and a helipad (only one, though— real super-yachts have two, so your friends can visit). LINK
Data
Threads now has 400m MAUs, roughly the same as Twitter. LINK
UK traffic to the biggest adult sites fell by a third, other than a half, in the last couple of weeks after the UK brought in age verification requirements. LINK
Character AI has about 20m MAUs for its LLM-based virtual friends. LINK
The UK’s Press Gazette surveyed over 500 publishers and found that Google referral traffic is indeed roughly flat, as Google claimed last week, despite the deployment of ‘AI Overviews’. LINK
The amount of money now being spent on US datacenter construction (excluding compute) is now twice as large as retail and about to overtake office space. LINK
Column
Personal context and some second steps for AI
LLMs are fuzzy reasoning engines that grow out of a fuzzy encyclopaedia. They have a meaningful portion of the world's information, locked up in there somewhere, or at least they've seen that information in the training process, but it can't be consistently or precisely retrieved.
And they have a reasoning engine, or they look like a reasoning engine, or they might become one, eventually. How much each of those matters depends on the use-case - most enterprise LLM use cases would work fine if you could have the ‘reasoning’ without the ‘encyclopaedia’ - what some people call a ‘world model’.
But besides those two aspects, then there's a third idea, which Apple and then Google have called personal context, which is a model that might actually know everything about you. An LLM is trained on all the data that the lab could get as of a cut-off date some time in the past (2024 for ChatGPT5), but that’s the same model for everyone, and it doesn’t have anything about you in particular. Then, a model’s productisation can include ‘memory’, but that only covers what you do in the product, not anything else. You can use the ‘context window’ to pipe more data in, or connect ChatGPT to your Gmail. Part of the current enthusiasm for somehow adding an LLM to a web browser is about distribution, much like Internet Explorer toolbars 20 years ago, but much more of it is about memory. ChatGPT and Gemini can remember what you asked, but what if they could see what you browsed? What if that could work on mobile, and all your apps, not just a desktop browser?
The challenge for any tech product or company in really getting to know you in this sense is a lot like the old story of the blind men feeling an elephant. One feels the trunk and says it's a snake, another feels the legs and thinks it's a tree, and so on. Instagram and TikTok have some sense of a taste graph, Google has some search and intent, and Amazon knows what you bought, but none of them really have that connected together (after 20 years of trying). Google’s Android and Apple’s iOS have a view from another angle - they see some of what you do in lots of apps, and have location and attention, but they can’t really look into other companies’ apps (or choose not to - Chinese apps on Chinese Android are very happy to scrape their competitors), and even if they could, they won’t know the graph behind it - Android might know you saw that image in Instagram but it doesn’t know the graph that led Meta to show it to you.
Then, of course, there’s ChatGPT - Sam Altman clearly wants to turn OpenAI into an old-fashioned trillion-dollar platform company, and ChatGPT might be an ‘everything’ app, but it will still be a partial view - it still won’t know what TikToks you watched, and a Shopify ChatGPT integration won’t give OpenAI your Amazon purchase history. Indeed, by itself, it’s unlikely to do any better than Google or Meta’s attempts to pull shopping into their services.
Meanwhile, none of them really know what things are - they know that you looked at this picture or watched that video and then watched others, and what other people watched, but not what's in those things. This is why people joke about Amazon saying 'this person bought a toilet seat - let's show them ten more toilet seats!" - Amazon knows the SKU but doesn't know what the SKU is.
All of this is a problem statement. What happens if you could plug each of those graphs together, into something that knows what toilet seats are?
This summer, I spoke to a bunch of senior people in retail, who said, in essence, that everyone has had a dozen big AI presentations now. They've had Google/Microsoft, Bain/BCG/McKinsey, WPP and Accenture, and maybe Stripe and Shopify. They've got search working, so you can ask “what would I take on a picnic?” - a query that would never have worked before LLMs. They have review summarisation and a SKU re-tagging project, and their marketing people have strong opinions about the stuff that Google and Meta have launched to optimise advertising. But now they wonder, “okay, is that it?” What's the step two? And it seems to me that an LLM that knows who you are, and what you care about, and what you're interested in, and what these things ARE, not just what their SKUs look like, would be a big part of the answer. But then again, no one can see the whole elephant, and no one wants to let anyone else see it first.