20 April 2025
News
Netflix plays with AI recommendations
Netflix is exploring an OpenAI partnership using LLMs for recommendations. Today, search and discovery systems for anything from Amazon to YouTube have to reply 1: on whatever metadata people have typed in and 2: on how a critical mass of users behave - did a lot of people watch this and then watch that? This makes your search options and your suggestions pretty crude and pretty dependent on volume of use. But LLMs will let users ask different kinds of questions, and have a much deeper knowledge of what each item in the catalogue might be. LINK
Antitrust week
There are three big competition cases in tech this week. The FTC’s case against Meta came to court, the DoJ asked for remedies against Google in the TAC case (paying to be search default), and Google mostly lost its ad-tech case. See this week’s column. GOOGLE ADTECH RULING, GOOGLE TAC REMEDIES, FTC v META SLIDES, META v FTC SLIDES
Meanwhile, court cases always have lots of interesting disclosure. You can see all the Meta case exhibits online (updates every day). Amongst other things: Meta tried to buy Snap for $6bn in 2013, in 2014 it blocked competitor ads (where it didn’t later for TikTok), Brian Acton told Zuck he wanted to run WhatsApp as a lifestyle business comparable to Craigslist (DX-1102). Files PX-03602 and PX-02389 (deep linking doesn’t work) are fascinating insights into how Mark Zuckerberg thinks about social. EXHIBITS, REDACTIONS
OpenAI searches for product
It’s really hard to see a product strategy from the foundation model companies. There are plenty of features, though - this week OpenAI is expanding its ‘memory’ feature (which I think is stickiness, not a network effect, and other models do it too), explored buying the Windsurf coding tool for $3bn... and looked at building a social network. Well yes, but remember when it was doing an app store? SOCIAL, MEMORY, WINDSURF
The week in AI
Anthropic launched its own ‘Deep Research’ product and Google Workspace integration. Where are the moats for foundation models? LINK
Microsoft is finally relaunching its ‘Recall’ feature, which takes continuous screenshots of your Windows PC so that an LLM can answer questions about what you’ve been doing (‘where did I see that chart about XYZ?’) - this was announced last year but had to be pulled for security and privacy reasons. LINK
Wikipedia got fed up with AI companies scraping it for training data, and put a clean, structured training data set online instead. LINK
OpenAI launched some new models that got better benchmark scores. Feeds & speeds. LINK
Shein and Temu slash their ad spend
Unsurprisingly given they’re hit both by the Trump tariffs and the end of the de minimus loophole, Shein and Temu have slashed their ad spending for user acquisition. China-based advertisers in total are about 10% of Meta's revenue. Interestingly, Temu’s rank in the app store immediately collapsed, while other Chinese apps briefly shot to the top. Meanwhile Shein’s IPO (in London) was approved this week, but will be pretty hard to get out (though much of its business is outside the USA, of course). TEMU, IPO
Streaming piracy
Remember when streaming was supposed to kill piracy? Nope. In the UK, the FT reports there’s now a huge trade in Amazon Fire HDMI sticks pre-loaded with pirate streaming services. In Spain, the ‘LaLiga’ football league persuaded a judge to give it a very broad blocking order that resulted in ISPs having to block CDNs and hosting platforms including Cloudflare and Vercel. VERCEL, STICKS
EVs as phones
There’s a pretty widespread view that EVs will play out like Android, with a flood of Chinese OEMs producing a few Darwinian winners who crush everyone else. Part of that is actual Android players entering (see Xiaomi), but a big part of the question is whether cars could move to a contract manufacturing model as consumer electronics did. This is obviously a question as much for car experts as tech experts, but Foxconn is now planning to make its own EVs, as well as looking at buying Nissan. LINK
Ideas
The SemiAnalyst team says that Huawei’s answer to Nvidia is not as good, but good enough to be useful, and mainly behind on power consumption, where China is less constrained than the USA. LINK
Bain on SEO for LLMs - what happens to brands when people ask LLMs what to buy? LINK
Google’s cloud team listed over 600 ways that enterprises are using LLMs for point solutions. LINK
Generative AI in US military intelligence. LINK
Illegal charging of Indian e-rickshaws. LINK
Outside interests
Arbitraging the Texas lottery - when the combined price of all the tickets is less than the jackpot. LINK
Why flight search engines are so hard. LINK
Data
Sam Altman suggests that ChatGPT is used by 10% of the world’s population weekly - 7-800m people (a quote from a TED interview that was otherwise too boring to link). LINK
Alix Partners forecasts for media and entertainment. LINK
LG Ads data on smart TVs. LINK
A consumer survey saying that 37% of Americans have bought from TikTok Shop. LINK
The Information says Bytedance’s 2024 international revenue (i.e TikTok) grew 63% to $39bn. LINK
Column
Antitrust day
The one foundational problem with all regulation of tech today is that the market changes faster than the reaction cycle of regulators. You can wait until everything is clear and the market structure is settled, but then it’s too late, and the network effects are so strong that it’s hard to know what policy would change anything. Or, you can move early, when you really can change things, but then you’re speculating wildly on what might happen (as the FTC tried for Within). As a bonus, the places where it’s easiest to square this circle are generally the places that don’t matter much. We can see all of these problems this week, in an opening case against Meta, a judgement against Google, and an attempt at remedies (and penalties) in another case against Google.
Meta bought Instagram and WhatsApp over a decade ago. At the time, both deals seemed crazy to most people (Google ‘John Stewart Instagram’), and competition authorities reviewed and allowed both deals, but now the FTC wants a judge to let it second-guess its own decision. To do that, it has to persuade him that these deals allowed Meta to create or preserve an illegal monopoly.
But a monopoly of what? Market definition is the starting point for any competition case, and it’s probably significant that in its first filing the FTC did not include one, and was thrown out of court and told to come back and try again. Now it’s given us a definition, but the only way that it can reach the market share thresholds required is to argue that Meta does not compete with YouTube, TikTok or iMessage. Hence, the FTC’s argument begins with a quote from Mark Zuckerberg that described Facebook as about public sharing with your friends, as opposed to person-to-person messaging (iMessage) or consuming content from strangers (YouTube, TikTok). The trouble is, this quote is from 2006, almost 20 years ago and before the iPhone existed. That was indeed Facebook in 2006 - but today a big chunk of Meta’s use comes from a literal TikTok clone, another big chunk comes from Messenger, and three quarters of use is content from strangers. Meta’s defining characteristic is that has tried to absorb or co-opt every social behaviour online.
This creates a basic contradiction in the FTC’s position - “we didn’t realise Instagram would become this massive super-app dominating the market - and also let’s pretend Reels doesn’t exist.” But that’s where the law has taken it. Claiming that Meta doesn’t compete with TikTok doesn’t pass the sniff test. But if you accept that it does, then under US jurisprudence Meta doesn’t have market dominance and the Instagram acquisition wasn’t anti-competitive. And that… also seems problematic. But turning around again, the Instagram founders very probably would not, themselves, have made Instagram what it is now. That creates another contradiction - Instagram became a huge competitive force in large part because Meta bought it. The disclosure this week talks about plenty of other threats too - remember Path? If Meta hadn’t bought Instagram, maybe it would be the same. Regulation is hard, especially about the future.
Google’s TAC case shows the next step in the process: if the regulator (here the DoJ) wins the case, then what? I wrote a long essay last summer arguing that it was much easier to say that Google should no longer pay Apple $20bn a year to be the default search engine in Safari than to see an outcome that would change Google’s market share in search. Apple won’t make a search engine. Mozilla will go bust. Then what? Now the DoJ has come back with some of the options I discussed, most obviously forcing a spin-out of Chrome, but without really answering the questions I posed - hence, if Chrome can’t sell its default search slot to Google, what’s its revenue model? Should it take money from Microsoft to make Bing the default when everyone agrees that’s an inferior product? Is that good for consumers? And as if admitting its lack of any better ideas, the DoJ then proposes that a ‘technical committee’ should have oversight of all Google product integration for a decade or so. That’s not a great position to take after you’ve just alleged a lack of innovation.
Meanwhile, the third result this week is a finding that parts of Google’s ad network stack (about 10% of revenue) broke the law. This is both simple and complex: the simple description is that Google runs both sides of the market plus the marketplace and puts its thumb on the scales - the complex description makes most people’s eyes glaze over, including mine. And like getting Al Capone on taxes, it’s always struck me that these companies were much more likely to lose cases clearly and simply on the ad side than the product side. But that works both ways - a remedy here might be simpler, but at 10% of revenue it’s also much less consequential both for the company and for the broader tech ecosystem.
Stepping back, I have a lot of sympathy for regulators in these positions, and a few weeks ago I spoke at a meeting of the DRCF, a UK body that aims to put all digital regulators in the same room. It’s a problem to claim that Meta doesn’t compete with TikTok, but it’s also a problem to claim that Meta buying Instagram and TikTok didn’t give Meta dominance of the market for a decade. It’s easier to say that Google shouldn’t be allowed to spend $30bn a year locking out competition than to say how you could overturn the network effects that drive Google’s dominance, or indeed whether you should - but if you don’t, then what’s the point?
But meanwhile, TikTok did happen, and does pose a major competitive threat to Meta - unless it gets shut down. There is always something else. And for Google, generative AI is that something else. It feels slightly surreal to talk about how network effects lock in Google market dominance of web search even as OpenAI redefines what search might mean - at the minimum, we really don’t know what the web and web search will be in five years. This takes me back to the beginning - the speed of the market is faster than the regulators’ reaction cycle. But what are they to do? Throw up their hands and say that in the long term all monopolies will pass like Ozymandias? As Keynes tells us, in the long term we’re all dead.
All of this makes me think that the whole ‘neo-Brandeisian’ argument spearheaded by Lina Khan, that US competition law had become too narrowly focused on low prices as opposed to choice, needs some complements. How do you think about speed and speculation about the future? About network effects? About trade-offs with other regulators, most obviously privacy regulators? And is using lawsuits as the atomic unit of US competition law still the right approach, or is a shift to statutory regulators on the SEC model better? What other questions are there? But really, is it a good idea for the FTC be suing Meta over a deal from a decade ago on a market definition from two decades ago?