27 April 2025
News
Humanoid robots in China
In an echo of the famous 2004 DARPA Challenge that kicked off the autonomous car boom, 21 humanoid robots were entered in a half-marathon in Beijing this week. Their engineers were allowed to pick them up if they fell over and swap batteries. This whole space has become very hot, since humanoid robots now kind of work, due to a combination of better motors and AI-powered bipedal balance. It’s rather less clear what they’re useful for, other than stairs: where is a bipedal robot more useful than an autonomous forklift? LINK
Google gives up on killing cookies
If Google didn’t own Chrome, how would it manage privacy? We don’t really know what privacy means in an LLM era, but this week Google finally gave up on ‘Privacy Sandbox’, its attempt to kill third-party cookies and use the browser to do anonymous targeting instead. People have been arguing about this for ages, since it probably would be good for privacy, but would also have given more control of the adtech stack to Google. Now it’s gone, and cookie banners will be eternal. GOOGLE, ANALYSIS
iPhones sold in the USA will be made in India?
The FT reports that Apple is ramping up iPhone production in India to supply all US sales. Last year it made about 20m there, less than 10% of the total, and to cover the US that would have increased to more like 60m over the next few years. Of course, this really means that components from Japan, Taiwan, South Korea, and China itself are shipped to India and assembled using Chinese equipment by Taiwanese and Chinese contractors. This isn’t exactly ‘bringing manufacturing back to the USA’. Meanwhile, The Information reports that earlier this year Chinese officials tried to stop the export of manufacturing equipment involved in this to India. INDIA, EQUIPMENT
The week in AI
Google updated its music AI sandbox. This is still more a tech demo than a replacement for ProTools or Logic, but it also points to how how much grunt-work in the creative industries will be automated away (just as has happened half a dozen or a dozen times before). LINK
Perplexity updated its iOS app to do some basic assistant features like drafting an email or adding an appointment - all of which the old, legacy Siri could do a decade ago. But hey, this time it’s ‘AI’. LINK
Meanwhile, Motorola (i.e. Lenovo) is embedding assistants from Meta, Google, Microsoft, and Perplexity. As a consumer, which one should you choose? LINK
And, it emerged that Google had briefly tried to block this before realising that this was exactly what got it into trouble in the antitrust case that’s in remedies right now (see above). LINK
The fallout from Apple’s Siri fiasco continues. Last month, Siri itself was moved from the dedicated ‘AI’ group (run by Xoogler John Giannandrea) to the software group, and now Bloomberg says that a super-secret robotics project has been moved to hardware engineering. I’m not that interested in internal gossip and moving heads, except that these moves suggest Tim Cook and others are trying to fix the problem. LINK
Intel turnaround
Intel has layoffs, and apparently it will de-layer about 20% of staff, as the new CEO tries to reduce bureaucracy. Yes, but is the US going to fund it to compete with TSMC, and if not, what? LINK
Shein and Temu start raising prices
Last week Shein and Ted slashed their ad budgets; this week they’ve started increasing prices as well, as Trump’s import taxes approach. Shein has 40% of US fast fashion, but of course most of its competition is going to be tariffed too, since it’s being imported from Cambodia and Vietnam. LINK
Policy 1: Google and Meta in court
Google and Meta spent this week in court: Google is in the remedies phase of the search antitrust trial, while Meta is beginning the trial on whether buying Instagram and WhatsApp should be unwound. It’s hard to know how much attention to pay to these (other than to pick through the disclosures for interesting data, as I noted last week): what is actually going to change as a result?
Hence, the DoJ argued for Chrome to be split off: while this would be difficult and complex, it’s hard to see how that would change search market shares - until OpenAI (and Yahoo) claimed they’d want to buy it, which seems like a great way for the DoJ to create a new monopoly.
The FTC faces a harder struggle in its case against Meta, where the logic of US antitrust law means that government lawyers have to pretend to believe that Meta doesn’t compete with TikTok, YouTube, or iMessage. The case itself is becoming an extended exercise in second-guessing - would the WhatsApp founders have changed their minds and done ads? Who can say? Would Instagram co-founders who wanted to review every ad themselves have turned it into an advertising powerhouse? Maybe? Is that enough? SYSTROM, OPENAI
Policy 2: The EU fines Meta and Apple for… what, exactly?
This week the EU fined Apple €500m and Meta €200m. In both cases, it’s unclear what the companies were supposed to have done. And, as above, it’s hard to see what this will really change.
The Apple fine relates to the rule that Apple has to let app developers use their own payment method instead of using Apple’s. Apple has responded by allowing this (in the EU), but with a complex set of rules, with everything from warning screens to a ‘platform charge’ that developers have to pay instead. The EU has decided that this breaks the DMA and fined Apple. However, it’s much easier to object to the process Apple designed than to say what, exactly, it should have done that would have been compliant. Are you sure? Apple is having to guess, and getting fined for guessing wrong.
Meta is in a tougher position again. The EU claims that the DMA means Meta must offer an option to use its products without ads based on your interests, in a form that is ‘less personalised but equivalent’. Since personalised ads are the only kind that make sense on social, Meta then offered an alternative option to pay a subscription. This model is already used by a wide range of European newspapers, without the EU objecting. Now the EU has decided that for Meta, payment is not ‘less personalised but equivalent’. Fine, so what is it supposed to do? Give everyone in the EU an option to use the product for free with no ads? Was that really the intention of the DMA? And if ads really are evil and subscriptions are not an acceptable alternative, why does that only apply to large American companies and not Axel Springer? LINK
Ideas
The latest cool viral trick for ChatGPT is to ask it to work out the location of a photo. LINK
Apparently Microsoft Copilot has been flat at only 20m weekly actives for a year, and the anonymous quotes are out for Mustafa Suleyman, the head of AI (and DeepMind co-founder) who was acquihired a year ago. LINK
An analysis of Trump’s trade policy for semis. LINK
The $20k modular electric pickup funded by Jeff Bezos . LINK
A profile of arXiv, the online scientific research platform. LINK
In March 2023 a bunch of people signed a public letter demanding that governments force a pause on all LLM research for six months, in case something called ’AI’ killed us all. Here we are over two years later and it still can’t read a PDF accurately. LINK
Adult content, naturally, is a strong early use case for AI image generation, but, as we’ve seen in other spaces, Visa and Mastercard are effectively ‘shadow regulators’, refusing to allow payment processing for companies that don’t meet their T&Cs. This week ‘Civit’ (me neither) was hit with the ban-hammer. LINK
Amazon sold $56bn of ads last year, and Google and Meta want to help other retailers join in the ‘Retail Media’ boom - here, AdWeek breaks down Meta’s plans. LINK
Outside interests
The Architecture of the Wire. LINK
The Shop Signs of Peking, from 1931. Vanished worlds. LINK
Ettore Sottsass’ copper plates. LINK
Data
Microsoft’s annual software at work index. LINK
Epoch AI put together a dataset of the compute and power consumption of AI data centres (both known and estimated). LINK
Shein is over 40% of US fast fashion spending, according to the Earnest Analysis credit card panel. LINK
Google’s Gemini has about 350m MAUs. LINK
Column
The AI strategy gap
Every morning, I get half a dozen email newsletters filled with hundreds and hundreds of links to new science and engineering in the building of LLMs. There are more models, more benchmark scores, and more acronyms than anyone can really keep track of. The whole thing feels a lot like the ‘feeds & speeds’ phase of the PC industry in the late 1980s and early 1990s, when you had to know about PCI and ISA and how many megahertz you needed, but there’s a massive amount of innovation and creative energy here.
On the other side of the table, every week there are a whole bunch more ‘AI’ startups, and they all seem to fit into three categories: software for other AI startups, software for software developers… and enterprise SaaS companies aiming to solve some very specific problem buried deep inside some very specific industry, or, perhaps, to unbundle something out of SAP, Salesforce, or Oracle. Half of all startups fail, and half of these will too, but if a company wants to use LLMs to optimise the reconfiguration of telco billing systems, or convert COBOL to Java, you can’t really analyse them as ‘AI companies’ any more than you could look at Uber and Tinder as ‘App companies’.
But then, there’s a big fuzzy space in the middle. Half a dozen to a dozen giant companies are spending hundreds of billions of dollars to build SOTA LLMs, not as commodity infrastructure or API calls or a science project, but because they want you to go to a chatbot and use it as a completely new kind of software. Every couple of weeks, one of them announces that one of their models is better, and publishes a chart that proves it, and if you know what you’re looking for, you might even be able to tell.
But it’s very hard to tell how any of them have a strategy for their chatbot as a product that’s anything more than making the model better. What is the difference between ChatGPT, Claude or Gemini as a productthat a normal consumer uses? They do exactly the same thing in exactly the same way. One or other might have a better model, as underlying infrastructure this week, but how is the product different? They leapfrog each other in features - ‘memory’ this week, voice next week - but that’s not strategy. OpenAI and Anthropic have each hired ‘heads of product’ from Instagram, but what are these people supposed to do, other than get an email one morning that says ‘we have a new voice model!’ and say “OK, I guess I’m adding a microphone button today.”
You could do the analysis and say the Amazon was doing something different to eBay or Walmart, or the Apple was doing something different to Android or RIM, and think about why that might or might not work. But these companies are all doing pretty much exactly the same thing.
To be clear, this is not the same as corporate strategy, where the differences are easier and simpler: Meta wants these models to be commodity infrastructure and wants Open Source to achieve that, Google has to work out what this means for search, and so on. They’re doing the same thing, but starting from different places with different assets and different constraints. But what’s the difference in the products? Well, they have different graphic design. Is the logo a line drawing or a coloured swirl?
Indeed, if you read today’s issue again, and then read the last couple of weeks’, you’ll see the difference that matters - today ChatGPT has an order of magnitude more users than Google’s Gemini, and Gemini has an order of magnitude more users than anyone else. And most of that is about brand and distribution. ChatGPT is becoming a verb and no one outside tech has heard of Claude. Sam Altman’s job seems to be about one-third corporate politics, one-third capital-raising and one-third getting media coverage. OpenAI has distribution through brand (and deals with Apple and Microsoft), Google through search, and Meta is trying to get it by bolting a blue circle into its apps, whether or not the users know what that is.
Some of this is inevitable given how early we still are. Maybe the characters of the models themselves will diversify (‘Claude is better at coding’), though that’s not really a strategy. Maybe someone will work out how to create a network effect (memory might be a path).
But on the other hand, some of this might be inherent to the technology. If your thesis is that once LLMs are scaled for a few more generations they will be able to do ‘everything’, then, how would your app that does everything be different from any other app that does everything? If the right interface is to just ask the computer what you want, how would you make that different? How could an everything be unique?