Weekly Links: Watermarking AI, Full Duplex Realtime, and Google's Westinghouse Bet

This week: Anthropic explores multiagent behaviours (heads up, there is a lot of literature on this!), PBS may be losing its data archive, and Meta is back in the open weights game.

On to the main stories:

• How Claude’s text watermark works. In what's been one of the main AI stories over the past week, in this article (which actually pointedly does not explain "how it works"), Anthropic announced that their newer models already introduce subtle indicators into text that enable AI detection. The feature is, at least in part, an update to comply with the EU AI Act's requirement that AI outputs be identifiable. You probably already know what I think about watermarking (see "Why AI Watermarking Makes No Sense". I can see the logic in Anthropic's rollout to comply, but it's hard to tell if they actually expect it to work. There are already services to strip out watermarks, but even more than that, it feels fundamentally wrong to apply this to all text generated by an LLM. A set of ideas a person generates and renders with AI has AI involvement, but it's not "produced by Claude" even if it was only a light copyedit. Fundamentally, I think the EU regulation is flawed, but there is likely no way around it for now. In the long run, I suspect we'll have very little content that has not been touched by AI. What will matter, though, is which ideas and statements humans stand behind. That's something you can only track by capturing human provenance.
• AI prices are plunging. That feels bad, but something else is happening. OpenAI reduced pricing for some of its smaller models and has seen a surge in usage on platforms like OpenRouter. In the article, the author (reluctantly) talks about the much-quoted Jevons Paradox (cheaper things get put to more use). That's probably part of it, but I think part of the answer is more basic: OpenAI is taking part of the open-source pie and also some of Google's pie. The price points for Luna are now such that for applications that previously used an open-source model, it likely makes more sense to use a distilled frontier model instead.
• ByteDance Seed Introduces SeedRealtime: a Native Audio-Visual Full-Duplex LLM That Watches, Listens and Speaks in One Model. This new model is closer to what one might imagine a true multi-modal conversation partner would be like, with an always-open communication channel between the user and the model. ChatGPT's new voice mode has elements of this, but it's not multi-modal. ByteDance's new model is only available in an app at the moment, but the architecture is public for people to try to replicate. It does seem that consumer applications for AI will become much more fluid soon.
• Google’s Westinghouse Bet. Great read from Tim O'Reilly on Google's DeepMind changes and what it means for their strategy. I think he has it right; Google is focusing on having "good" models they can serve for many non-frontier use cases at great margins. I agree with O'Reilly's take; there's a good chance that pays off massively more than being at the frontier. Further, if the cash machine keeps printing, then in the long run, there will be money for frontier research again. I guess the final factor is that it's not even really about money. It's about compute resources. Compute will be the bottleneck in the next 3-4 years, and all the major players are trying to ensure they lock it up. Spending huge amounts of compute on ever bigger models doesn't fit that need.
• AI assistant hacks gym website in first known Australian autonomous cyber attack. This is a great story which perfectly captures that we need to be careful what we ask for and ... how many weak systems are still out there. I especially love the part of the story where the agent bumped other people off the waitlist. If you want your spinning class slot, better make sure you have a powerful agent on your side!

Wishing you a great weekend.