A blog post from a developer in SF went massively viral across tech, arguing that the latest models are effectively a second generation of AI, that everything has changed since last autumn, and things will now change much faster. You should read it, not to agree, but to see the psychology, and then read this week’s column. LINK
Meanwhile, after software stocks crashed last week, markets have been swirling around all sorts of other sectors to sell out of fear they might be the next field where people panic about AI. The latest ‘scare trade’ is real estate services. None of these make much sense, at least, but people are worrying both that AI investment won’t pay off (the $650bn that the big four platform companies announced they’ll invest this year) and also worrying it will demolish other industries (both can be true). Again, see this week’s column. LINK
All the main chatbots now have embedded e-commerce checkouts live now, with several competing sets of standards (standards are one of the main ways that tech companies try to take control of the agenda). The adtech and measurement side of this is also being built out, or at least the first generation, although actual consumer use is tiny so far. Everyone says ‘agentic’ a lot too, but that really doesn’t exist yet. GOOGLE, SHOPIFY, OPENAI INTERVIEW
Anthropic raised another $30bn in capital (at a $380bn valuation) and said its revenue is now at a $14bn run rate (OpenAI said it had ~$20bn revenue in 2025). LINK
Meanwhile, though Anthropic is less loud about its infrastructure plans than OpenAI, the Information says it has a Xoogler infra team and is looking at commitments for at least 10 gigawatts of capacity in the next few years. (OpenAI ended 2025 with 1.9 gigawatts and its future plans are… lots). Given the rule-of-thumb right now is (say) $30-50bn per gigawatt, it looks like they will need to come back to investors. LINK
OpenAI hired Peter Steinberger, the developer behind OpenClaw. Great proof of concept but not a consumer product, and that difference isn’t the fun part. LINK
Both Google and OpenAI say they are fighting off massive distillation attempts from Chinese models. Obviously. LINK, OPENAI
T-Mobile USA has launched a beta of live phone-call translation, for calls originating on its network. No apps or fees. It’s a very long time since a telco could offer a new experience - all the innovation moved up the stack when the iPhone launched. This is already rolling out to the apps that are replacing PSTN calls, of course (especially for international calls): Apple added this to the iPhone last year (for both PSTN and FaceTime) and it’s also available on some Androids, but if it’s in the network you know it’s there. T-MOBILE, APPLE, PIXEL
Waymo confirmed that it has remote staff who can take some action when Waymos get stuck, but not actually drive them remotely, and that some of these staff are based outside in the US, for example in the Philippines. LINK
Goldman Sachs is using Anthropic models to build tools to automate back-office functions, with Anthropic engineers working there full-time. It would be surprising if they weren’t. LINK
Half of the 12 cofounders of xAI, Elon Musk’s lab (just bought by SpaceX), have quit recently, with two leaving this past week. Reporting varies as to why: Elon Musk himself claims he wanted them out, others say the SpaceX acquisition meant they could cash out and move on, and others again suggest that the work at xAI was too similar to the work they’d do elsewhere. Given that most of the leading researchers in the field are now ‘post-economic’, you need to give them another reason to stay. LINK
Illustrating tthat point, someone also quit Anthropic... to do a poetry degree. LINK
A small but indicative note of where things are going - Google released an open-source library for extracting structured data from unstructured text (medical notes, say) using LLM APIs. No hospital will (or at least should) use this to build their own record-keeping system. Rather, LLMs will commoditise a broad class of capability that today needs a lot of time, money, and domain expertise, and that will change the competitive landscape. In other words, there will be a lot more software. LINK
Remember when everyone in tech was talking about MrBeast? He’s still there, his videos still get 150m views each, and he just bought a bank. LINK
A useful AI taxonomy. What do you actually mean when you say AI - what are the kinds of problems that you might be trying to solve? LINK
Steven Sinofsky, former head of Office and then Windows, who grew up there last time we were doing this, on the ‘death of software’: “Nope.” LINK.
Two interesting AI stories in the FT this week. First, in the last decade, there has been a boom in private equity industry firms buying out enterprise software companies, with up to a trillion dollars invested, $300bn last year alone. Some of those deals were done at inflated ZIRP prices, but now all of them look expensive even with the rerating of the software industry in the last month. While very few serious people think that AI will replace enterprise software, a lot of people think this could reset the margin structure and competitive landscape. LINK
Second, the FT also reported on a surge in investors looking at legal structures to allow them to buy big law firms… which are another candidate for AI-led rerating. LINK
On the other side, Martin Sorrell’s ad agency Monks says that a quarter of its revenue will have moved to flat-rate subscriptions by the end of the year. Ad agencies, like a lot of professional services industries, charge on a ‘cost plus’ basis, but with AI taking a lot of that cost out, absolute margins collapse, so you need a new pricing model (or a smaller company). LINK
Corning’s sharing price spiked in the Dotcom bubble selling fibre, and now it’s spiked again, even more, selling fibre interconnects for AI data centres. LINK
Om Malik does some useful back-of-the-envelope maths on just how many cars in how many cities Waymo would need to justify its latest valuation. LINK
Apparently, whenever a news story runs with leaked information from inside, OpenAI uses ChatGPT to try to work out who might have access to all of that information and might be the leaker. LINK
Bloomberg has a long piece on the strength of Andreessen Horowitz’s connections into the Trump administration and its role in influencing tech policy, especially on AI and crypto, which the two founders said was their concern when they endorsed Trump before the election. I worked at a16z from 2014 to 2019, and I mostly agree with Ben and Marc on AI and crypto policy, but I won’t comment further. LINK
Apple has changed the model for its pro video and music software, bundling them (oddly) with its word processor, spreadsheet, and presentation apps and including a scattering of generative AI features across the apps - interestingly, though, these all have a usage metre, since they are all running in the cloud and have marginal costs. I suspect this will go away pretty quickly as inference scales and becomes more efficient (and as Apple gets it onto the device), but at the moment, this is very 1998, or indeed 1978. LINK
The Fed ran the numbers on who’s paying for Trump’s import taxes. LINK
EssilorLuxottica says it sold ‘more than 7m’ of Meta’s AI glasses in the last 12 months - a year ago, it said it had sold 2m since 2023. LINK
An HBR study suggests that AI tools mean people work harder because they can do more. LINK
Ramp’s corporate spending data (which I suspect has a pretty skewed sample) says that OpenAI has been flat and Anthropic surged since last summer. LINK
Apollo points out that Chinese outward investment now exceeds inward investment. LINK
I started my career working for an investment bank in the spring of 1999. One of the things I remember most clearly as the bubble accelerated through the year was the sense of euphoria. People said not just that this was going to change everything, and was going to be huge, but that it already was, that it all worked now and everything was going to happen this afternoon. That came with an enormous sense that you should capitulate. You should let go. Throw away your reservations. Stop telling yourself that most people don't have broadband yet. Streaming video will work! Stop telling yourself that the stocks were expensive. This is going to change everything. Does this sound familiar?
Of course, it did change everything, but not that afternoon. My old boss Marc Andreessen used to say that all the failed ideas from the dot-com bubble would work now, but ‘now’ was 20 years later.
I remembered a lot of this reading Matt Shumer’s piece, and a lot of similar discourse in the last few weeks. Clearly, these models are a step change in what software engineers can do. They can spin up a working website. They can make a prototype in an afternoon in less time than it might have taken to plan what you’d do three years ago. And then you ask it to put the red box in the green box, and it can't. You can ask it what the last computer revolution did to the audit industry, and it will give you a fascinating description, full of numbers and industry dynamics, except that some of the industry dynamics didn't happen, and if you ask it for a source on some of the numbers, it will say that made them up.
Hallucination is fixed? Bullshit. We don’t even know if that’s possible. This passed the bar exam! Sure. That doesn’t make it a lawyer. It’s extremely useful to lawyers, once it’s wrapped in a lot of carefully considered product and information architecture, but it’s not a lawyer. It can replace journalists. Really? So, it can hang out in city hall and look for sources? It can go to Minnesota, knock on doors, and win the confidence of people terrified and brutalised by masked government thugs?
The euphoria is real, and the technology is amazing, and it can sort-ofdo a vast range of things, but often only sort-of. It doesn’t all work yet, and more importantly, even if it did, jobs can be a lot harder to automate than they look to an engineer in Silicon Valley.
The other side of the dotcom bubble, of course, was its mirror image. Something I remember very clearly from the crash was that a story would go round that didn't make any sense, even if you thought the whole internet was nonsense, and the stocks would go down 10% or 15%. And then, six weeks later, a version of the same story would be reported, and the stocks would go down again, and we would scratch our heads and say “But that was already in the price!” The truth was that something can be in the price more than once. We went from unquestioning optimism to unquestioning pessimism.
I think you see a lot of this in the so-called scare trade. While the internet did end up disrupting a bunch of industries, most obviously newspapers, its primary characteristic was not destructive of other industries and jobs, at least not in 1999 or 2001, whereas the obvious place you can see change from AI is in taking out cost, and in removing the need for a whole class of companies’ products.
As I wrote last week, I'm very sceptical that it's quite as easy as that: I think it's more likely that there's more competition than that we won’t have software anymore. But the enormous amount of surge of capex into the field last year and this year, and the acceleration of the models (or, perhaps, the clarity in what they can already do), concentrates minds and creates a very febrile atmosphere. We have elements of both a bubble and a crash at once.
Hugh Trevor-Roper supposedly said that the only thing history teaches us is that something will happen. People were calling the dotcom crash in '97, and they've also been calling for a crash in tech since 2016: you can’t call the timing. But I think the only thing one can say with certainty about AI is that if we're not in a bubble now, we will be, and perhaps, that this is the kind of thing that happens in bubbles.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。