Weekly Links: More Hacks, Amazon Blocks Muse, and Risky Personal Agents + See You In London!

This week: no camera smart glasses from Meta (also VR/AR glasses...), openAI models posting poorly anonymized user photos on the public Internet, and we get DNA Computers.

• Now AI is trying to gobble up dinner reservations. This story is a little microcosm of the multi-front clash happening between agentic ways of doing things and interfaces designed for humans. Appointment spamming isn't new; it is now just open to a whole new class of users. What is scarier is what comes next: there is already a forest of MCP servers that access restaurant booking (see this one on Apify, for example) which provide web automations for bookings; ultimately, Resy will likely need to launch its own MCP server. So far, so normal... but we're still left with the competition to get seats at high-end restaurants; what comes next might just be restaurant-to-agent negotiations on just how much someone is willing to pay.
• I Think I Found an AI Agent Worth the Risk. In another personal agent story, a Wired researcher used the Instinct personal agent and came away concluding that the utility might outweigh the risk. I'm really not sure that's true yet unless the task set is very constrained. Underlying web infrastructure just isn't set up for Agent use. I'm not sure people are pricing in the time cost of worrying about what the agent might or might not have done.
• Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war. The new model releases from both Anthropic and OpenAI this week came with price cuts attached. Anthropic reduced effective token costs on Opus models, and within 90 minutes, OpenAI had cut Astra and other prices effectively in half. Market share is still clearly uppermost in the minds of both companies. With new approaches to some tasks (like Jev), the blended token costs across frontier and easy tasks will likely fall even further.
• How would AI actually ‘kill all humans’? Here are the top five most likely scenarios. Humanity tries to figure out exactly how AI superintelligence would wipe out civilization. One of the five answers is still paperclips; another two are nuclear and bioweapon strikes. I'm not sure whether this is a sign that we aren't very creative or a sign that we don't want to write solutions on the Internet in case agents find them. I'm not dismissing the problem, just pointing out that we should be a lot more specific about the failure cases.
• Amazon blocks Meta’s Muse AI assistant in new standoff over agentic shopping. This news was almost inevitable as soon as Meta's personal agent service Muse launched. Amazon is the aggregator of choice for eCommerce and benefits heavily from users coming to its online store (users often buy multiple items and also see Amazon's on-platform advertising). Muse agents could buy directly on Amazon with the user never seeing any of the flow. This is a fight for the last mile to the customer. However, the shape of how this may play out can be glimpsed in Amazon's deal with ChatGPT, which makes Amazon a larger advertiser on ChatGPT. These ads, in part, still drive users to Amazon. Perhaps a similar deal is possible with Meta. However, it may be harder since Muse is specifically designed as a "hands-off" experience, and there really isn't much need for users to see an advert. In the long run, the disintermediation risk is clear. Meta could well partner with other players (Walmart being the obvious player) instead of Amazon and simply drive business there. This doesn't matter when few users are on Muse, but will matter a great deal if it becomes widely adopted.

Finally, I'm excited to share that I'll be speaking at APIDays/GenerationAI in London this week (thank you to Mark Boyd and Mehdi for the kind invitation. Here's my lovely speaker card:

If you don't have a ticket yet, what are you doing? Check out the whole FOST program here and reach out to me; I still have some passes available to hand out!