This week: Anthropic tugs at public perception with a large-scale survey, AI makes friends with Fermat's last theorem, and Unitree inexplicably persists in pushing further into robot combat demos.
On to the main stories:
- Claude Cowork is coming to mobile and web. At last week's AI Engineer conference, the OpenAI keynote featured a funny segment about the common scene in AI-pilled offices of half-cracked-open laptops. This is because long-running agents need the thread that prompted them to still be active on the machine that started the job. OpenAI teased the ability to dispense with this. I guess Anthropic was watching and decided to ship first. No doubt OpenAI will follow. It's actually a pretty important capability. Most users don't want to carry a device around just to keep a long-running agent active.
- Muse Spark — Meta's Most Powerful AI Model. Meta gets back in the model game with a new (and powerful) multi-modal LLM. The model is particularly targeted at image creation, but its aims are broader. The company also made the model available across its properties, including Instagram, where it had to pull back the ability to create variants of existing Instagram posts after a backlash. The ability to have Instagram images remixed was "switched on by default" by Meta, so it is unsurprising that people reacted negatively. On the whole, the new model seems strong, and Meta will have a powerful business advantage if it controls its own leading-edge model. They are clearly optimizing for different things (e.g., ad-creative). Shipping the "default on" feature in Instagram seems like a mistake, but was it a calculated publicity grab to lay the groundwork for sensitizing people to more and more AI?
- OpenAI Just Declared ChatGPT as You Know It Dead (also see the Spyglass take). In the other frontier lab camp, OpenAI's app update this week confused many users. The company's new push into ChatGPT Work clearly follows Claude's success with Claude Code and Claude CoWork, but, in the process, it created a very strange app experience. The new ChatGPT desktop app is effectively the previous Codex app with updates. While this makes sense, as Codex is the agentic product that underpins long-running work, there is no readily accessible first-class chat experience in the new app. You can still get this experience with an app called "ChatGPT Classic," but there is no obvious way to download this if you don't already have the ChatGPT app and want to migrate/upgrade. The company will likely need to bring the chat experience back or risk losing its user base on desktop devices. I'm currently in the UK, and this update doesn't seem to have rolled out yet, so for now I'm still enjoying the previous App experience.
- Humanoid robots controlled by surgeons did world-first operation on live pigs. I'm not sure how I'd feel if I were a pig in this scenario, but it's impressive to see that humanoid robots are now dextrous enough to perform some types of surgery. The robots were remote-controlled by human surgeons, but no doubt this is a precursor to learning routine surgery types in more and more autonomous modes. The key value here is that it reduces the need for specialized surgical robots. It's also a bit dissonant to see the same type of robot performing surgery one minute, and kung-fu kicks another.
- Has China obtained the world’s most important machine? The Economist covers claims that an ASML advanced EUV lithography machine has made its way to China. The machine is central to manufacturing the world's most advanced computer chips, and export of the machines to China has been prohibited. ASML denies the claim, and there seems to be no proof that this has happened. My take would be, though, that China will surely obtain the machines it needs eventually and, indeed, create its own. Within 5, 10, or 15 years, it seems inevitable that lithography technology will be available to whoever can pay. The economic and strategic imperatives are just too strong for this not to be the case.
Wishing you a great weekend.
