18 May 2025
News
OpenAI makes a coding move
Earlier this month it was reported that OpenAI would buy WireSurf, an AI-for-coding company, but this week it announced Codex, its own coding product. As I’ve written previously, AI for software development is a very hot space, with lots of companies talking about 30-50% efficiency gains, and it seems this might be ‘the new AWS’ - see this week’s column. LINK
Data centres in the Gulf
Trump went on a tour of the Persian Gulf and announced new AI datacenter deals with the UAE and Saudi Arabia, while cancelling Biden’s ‘three tier’ restriction system that limited those country’s access to GPUs. Nvidia will sell ‘several hundred thousand’ of its most advanced computers to KSA, AMD has a $10bn deal, and Amazon, Cisco and Super Micro also announced things. And Bloomberg reports that OpenAI is considering an enormous (10 square mile, 5 Gigawatt) datacentre in Abu Dhabi.
This is an obvious trade: the AI infra buildout needs (or at any rate can spend) hundreds of billions of dollars, and these countries are eager to turn oil money into something that takes them past Peak Oil and past the transition to EVs. But how much of those GPU fleets will end up running Chinese models for Chinese companies that can’t buy the GPUs directly? Check back in 5 years… OPENAI, DEALS, POLICY, ANALYSIS
DeepMind does more science
DeepMind is no longer the centre of the AI research universe, but it’s still there doing important primary research, and this week it announced a new model that can be applied to an open rage of problems as opposed to being focused on one thing (cf Alphafold). COVERAGE, LINK
Remember when Klarna replaced customer service with AI?
Last year Klarna got a lot of attention by saying that it was replacing entire departments with LLMs, including customer support. Maybe not - it’s hiring CS agents again. Automation rarely means you have no people at all. LINK
Grappling with copyright
It remains pretty unclear quite how copyright law should manage both the input and outputs of generative AI. Is it OK to train on copyright work if you’re not competing with or reproducing that work? If you use these systems as tools to make something new, where is the copyright in that? The answer is not to say that it’s simple, nor to pontificate about what US copyright law says today. Hence, the US Copyright Office recently released a 100+ page draft policy, but Trump then fired both the head of the Office and the head of the Library of Congress, which oversees the office. POLICY, TRUMP
Meanwhile, in the UK the government is running a consultation procession to change the law and make all data available for training unless the owner has explicitly opted out (note for Americans- it is actually possible to change laws). Elton John is very unhappy about this. LINK
Shein and Temu rebalance
Sensor Tower told Reuters that Shein and Temu have increased their ad spend in Europe by a quarter to a half, as tariffs lock them out of the US (at least at the prices they built their businesses on). LINK
Everything is hacked, Coinbase edition
Cryptocurrency people have the best security there is (they have to), but there are always ways around, and it turned out someone was bribing Coinbase customer support staff to steal user data. LINK
Meanwhile, the WSJ rounds up a spate of recent ‘wrench attacks’, with kidnappings and two severed fingers (as XKCD points out, it doesn’t matter how strong your password is if someone hits you with a wrench until you give it up). This is a combination of people perceived to be rich and tokens that are a bear currency (though not, necessarily, anonymous or untraceable). A16Z’s head of crypto security, formerly of the Secret Service, wrote up a primer on personal security. LINK, A16Z
Everything is hacked, Chinese edition
Reuters reports that people have been finding undocumented cellular radios in power inverters and other electrical infrastructure built in China. LINK
Ideas
Bloomberg has a long profile of Liang Wenfang, the highly technical founder of DeepSeek. Apparently his nickname in the industry is ‘tech madman’. All very interesting, but I still think the base case is that DeepSeek represents commoditisation of LLMs, not a breakthrough. LINK
OpenAI published a very long guide to ‘prompt engineering’ for GPT-4.1. I find this kind of stuff unintentionally hilarious: if your thesis is that these models are replacing software, why do I need to memorise incantations and learn what JSON means to get the best results? All of this should be abstracted away. LINK
A decade ago Geoff Hinton, one of the pioneers of machine learning, famously said that we should stop training radiologists. It turned out out that radiology is about more than image recognition: the Mayo Clinic has since increased its radiology staff by 55%, but also hired a 40-person AI team to build tools for them. It’s worth remembering, in this and broader contacts, that knowing a lot about the science of neural networks doesn’t necessarily mean you have a good understanding of how they will be used in a broader society. LINK
Last week I linked to a story about college students using AI to ‘cheat’ - this week, when their teachers use this, it can’t be cheating, obviously. LINK
The WSJ says there’s an ‘epidemic’ of criminals using ads to drive scams on Instagram and Facebook. Meta has a huge number of small business advertisers, but it’s always wanted to get the friction lower to help more small traders onto the platform. That comes with trade-offs. LINK
A study on using a credit card instead of Apple’s payment inside iPhones apps, as permitted by last week’s US court order. Unsurprisingly, this has much lower conversion, so while the developer isn’t giving up 30% to Apple, they might still be worse off. LINK
Meanwhile, Apple is adding new ‘scare screens’ to app store listings for apps that use external payment in the EU. This seems like a very easy way to get another EU fine. LINK
The FTC/Meta trial rolls on, and every day more exhibits are uploaded: there are now over 300 files. There’s a lot of email on the daily struggles of running social networks (people worrying that US growth is slowing (2022), or that WhatsApp is overtaking Messenger (2013), and I didn’t know that Meta tracks consumer sentiment towards Meta the company. There also a few tantalising ideas - a project for a paid, premium version of Facebook from 2020, for example. And there are charts of how the ad load has doubled. But really, what you see across all of these documents is a relentless focus from the top on making the current thing work better while worrying about the next thing. LINK
Bloomberg has another detailed piece with lots of anonynous quotes on Apple’s struggles to build LLMs. The only really interesting part is that the head of AI (hired from Google), John Giannandrea, doesn’t really believe in AI assistants anyway, apparently. LINK
Outside interests
The Kaleidoscope House. LINK
Data
The IEA’s 2025 EV forecasts. LINK
The Mail Online says yes, Google’s AI overviews have hit its traffic. LINK
LG’s smart TV ad team has some interesting data on how their devices are used and what people watch. LINK
A generation of globalisation means that GM is now the largest importer of cars into the USA (which will be reflected in the Trump import taxes it has to pay). LINK
Column
LLMs as the new AWS
When Facebook bought Instagram for $1bn in 2012, it had 30m users, and yet it achieved that with just 13 staff, and had raised less than $10m in capital (it closed a $50m round days before the deal). That would have been impossible in the dotcom boom, and a lot of the difference was AWS. Startups could get infrastructure at marginal cost instead of having to buy, build, and manage their own server farms.
Because of AWS (and open source, and a bunch of other accumulated infrastructure), the cost of getting an idea to market fell by an order of magnitude or more. Combine that with a vastly bigger TAM (we forget that there were only 20-30m consumer PCs on Earth when Netscape launched) and we got far more startups (including successful startups!) with much lower initial funding needs (which also reshaped the venture capital industry).
In the last year or two, everyone in software development has started talking about 25-50% efficiency gains from AI coding assistants, and that’s when they’re still really only just getting started. Y Combinator got a lot of attention a few weeks ago by saying that for some of its startups 90% or more of the code is written by AI. Every VC I talk to says that all of their investments are using this, and the companies building these tools have rocketing valuations and, more importantly, rocketing ARR.
Of course, this is how every new abstraction layer works. As my former colleague at a16z Martin Casado pointed out, sure, the AI writes 90% of your code, but if you write an iPhone app, 90% of the code your app uses was written by Apple, and if you write a Windows app, 90% of the code your app uses is written by Microsoft. We stand on the shoulders of giants, who stand on the shoulders of giants, all the way down. Now we get another acceleration. It might be that this means fewer jobs for software developers, but it seems more likely that if it becomes cheaper and easier to write software, we might get much more software, especially given that LLMs open up huge new opportunities for new kinds of software. Hence the title - ‘LLMs are the new AWA” - a step change in what it costs to create software that leads to a ‘Cambrian explosion’ in how much software we have.
However, there’s a contradiction. Every launch for every big new model starts out by showing how this new model can do complex generalised tasks without needing dedicated software… and then says ‘look, it’s great at writing code!’ If you think that the LLM is going to be able to do ‘the whole thing’, then isn’t building tools to write code the equivalent of shorting the future?
Well, maybe, but the idea that the LLM will do ‘the whole thing’ is the big unknown question, and as I’ve written a few times, I’m on the sceptical side of this - I think these systems need to be wrapped in GUI and go to market to reach broad utility. More pragmatically (or cynically), we should probably expect that over time, OpenAI’s developer tools will be best at writing code to integrate with OpenAI’s systems and APIs, and the same for Microsoft, Google, and Anthropic. Developer tools have always been a strategic weapon. And in the meantime, it’s money and OpenAI needs cash, especially cash of its own.
I think this also reflects a broader contradiction, or at least dualism: there are hundreds of startups, big companies, and consultants building things with LLM APIs right now, where the thing they’re building is a single-purpose point solution. Yet those APIs plug into underlying systems whose premise is that they might grow to become completely general purpose, and subsume all of those point solutions, or at the very least that they’ll become capable of doing much more and going much further up the stack. I don't take that very seriously. But, the models do keep scaling.