Generative interfaces 'Generative interfaces', like 'generative search', is a phrase that immediately suggests all sorts of profound change, but also massive uncertainty. What would it mean for an LLM to generate a user interface dynamically, in response to a question or context? How does that compare to all the accumulated decisions and institutional knowledge that go into working out what options you should see - what happens if an LLM tries to work it out? This week we saw two interesting launches that point to how some of this might work, from different directions. On one hand, Google launched 'Disco', a browser that can generate simple apps based on the tabs you have open, letting your filter, combine and remix information across different websites. Will this work? Is this a scalable model? Is the a new take on generative search? Maybe, but it points to how much will change in the web as database front-end. On the other hand, Cursor, the red-hot AI coding assistant, launched a visual editor, letting you drag-and drop UI elements to build your own app or website GUI. One reaction is that this kills Figma, but Figma is a design ecosystem and platform, not just a GUI design tool. A jaded cynic might say that this is just Visual Basic all over again, or indeed Notion - a useful tool for simple things that bigger use-cases will outgrow (Webflow should be worried, though). But both of these point to pretty profound changes in the ways we separate interface, design, data and website: these aren't the answers, but they point to the questions. GOOGLE, CURSOR OpenAI distribution OpenAI did two interesting distribution deals this week: it licensed Disney characters and IP for its Sora social video app (with Disney taking a $1bn stake), and added Instacart to its app platform, so you can (try to) shop for groceries inside your chat - for example, take a photo of a recipe, a photo of your fridge, and tell ChatGPT to buy all the ingredients you don't already have (in previous issues I've pointed out a few reasons why I think flows like that are easier said than done, though). The common thread to both of these deals, of course, is the trade-off between control and distribution. Instacart is giving up control of the UX and the customer in exchange for OpenAI putting it in front of ChatGPT's tens of millions of US daily active users. And Disney is thinking about how much media companies regret selling shows to Netflix in the past, building up a new competitive threat. So it's putting Disney characters (and not actors or live-action footage, sidestepping those sensitivities) in a new place, for new money, but also getting equity. DISNEY, INSTACART OpenAI Enterprise ChatGPT still has far more consumer adoption than any of the other chatbots, but Gemini and Meta are coming up fast powered by their parents' distribution, and meanwhile (apparently) only 5% of ChatGPT's 800m WAUs are actually paying. Anthropic, on the other hand, is practically invisible in surveys of consumer use but has a healthy enterprise API business, as does Gemini, and so now OpenAI has decided to make a more aggressive push into enterprise, hiring the CEO of Slack as CRO. ENTERPRISE, SLACK Capex wobbles Oracle is a leveraged play on AI deployment capex, borrowing against its legacy cashflows to build data centres for Sam Altman, and so it's also a bet that OpenAI will be to raise the capital to fund the $1.4tr and counting of capacity that it's signed up for. The markets are getting more and more nervous about that, and as a consequence, Oracle's bonds are now trading like junk bonds. LINK Meanwhile, Broadcom reveals that the mystery buyer of $10bn of chips that it disclosed last quarter is Anthropic, buying Google-designed TPUs, and it placed another order for a further $11bn this quarter. LINK Agentic standardisation Anthropic, OpenAI, and Google (plus others) launched a standards body for interoperability of AI agents, with Anthropic contributing its MCP protocol. Meanwhile, a consortium of adtech companies has launched Agentic Advertising, which aims to be a non-profit setting standards for ads in LLMs, run by a former head of the IAB. Standards bodies are a classic part of a platform cycle, and a classic strategic tool, and there are three ways that they typically evolve. 1: One company's proprietary tech becomes the de facto standard and all the losers create their own to try to break in (generally this doesn't work); 2: the field is nascent and everyone values broader adoption over proprietary lock-in and so they make a standard, or 3: the same as 2 but it's too early and that doesn't work either (think of the long and painful history of smart home). AGENTIC AI, ADVERTISING Meta's AI reset Following Meta's multi-billion dollar AI hiring spree, there's a lot of arguing about strategy going on. Apparently the new arrivals want to switch from open to closed source, and to prioritise building a SOTA model over building more incremental and directly revenue-driving capabilities, and meanwhile (as you would expect) there are plenty of 'healthy exchanges of views' with the existing exec team. Open source is an interesting question: Meta's open source LLM strategy was a way to try to turn LLMs into cheap commodity infrastructure that it could build on top of; since then we have a bunch of very good open Chinese models, models do seem to be commodities, and costs fall (say) 50x a year for a given result. Meanwhile, it has to compete with companies with (for now) deeper pockets: it's having to find budget for AI from the metaverse project. So should it go closed and go for more control and revenue? BLOOMBERG, NYTIMES The week in AI OpenAI released 5.2 of ChatGPT, bumping performance on a few metrics and taking it back ahead of Google, for now. LINK President Trump finally signed an executive order attempting to prevent US states from passing their own AI regulations, to avoid fragmentation and accelerate US deployment. That seems sensible on its own terms, but it presumes that the US could actually pass national regulations, which seems unlikely. LINK Waymo is expanding to London. LINK Meanwhile, Rivian is expanding its autonomy offering, LINK The US military signed a Google Cloud deal for Gemini. It's not that long ago that Google employees staged protests at the idea of a similar deal: now Silicon Valley is getting back to its roots. LINK | 
No comments:
Post a Comment