This week, Microsoft shut down its internal library, someone beamed energy to Earth from space, and McKinsey is now beginning to evaluate candidates on AI prompting. On to the main stories for the week:
- Our approach to advertising and expanding access to ChatGPT. Probably the biggest news of the week (despite not getting that much coverage) is ChatGPT announcing they will be testing showing Ads in the free and Go versions of the product in the US (the Go version gives access to memory and better models, but was previously only available outside the US). The reason this is such a big step (well, the launch will be) is that it is a clear step across the Rubicon of becoming a mass consumer company. ChatGPT already has a large user base, but to be able to keep serving those users a better model experience, it needs to monetize. The only viable model to monetize at a truly global scale is through advertising the way Google and Facebook do. ChatGPT aims for its ads to be independent of the queries, and maybe they will be. Still, the experience likely won't be great to begin with. It was only a matter of time before the company activated the most powerful consumer business model there is. For what it's worth, I do see this as a positive – AI is already becoming a key life and productivity tool. The fact that there is an increasingly good version that is free is truly important. Maintaining AI accessibility at a non-degraded experience seems likely to be an important social goal in the future.
- Elon Musk on the Moonshots Podcast: AI Abundance, AGI Timeline, and other topics. [Podcast 2h37m]. Elon Musk is definitely a controversial figure; he has done some brilliant things, but he also says and does other things that many (including me) have very negative views on. This podcast is also very hard to listen to: the interviewers and interviewee interrupt each other, don't stay on topic, and often don't challenge important points. So why is this link here? Because, as is often the case, Elon says interesting things about the future, and on quite a few, he's likely directionally right. On other topics, he very likely sees things that other people do not (e.g., that chips won't be the bottleneck very soon, and then again sometime in the future). I'll definitely say, though, on the topic of UHI (Universal High Income - which is a line floated by multiple people in AI and AI investing), the point being made (AI and Robitics will make everything very cheap or free) ignores a lot of fundamental points of economics and reality. Firstly, because these actions are deflationary - that's good for consumers of those goods and services - but potentially apocalyptic if it were to happen in every sector of the economy at the same time. It's also unlikely to go to zero so fast: IP claims, goods still have physical costs (which may even rise). There is a real challenge that society faces if humans are moved out of the labor force en masse without price drops, and even questions about what happens if the price drops occur. This deserves a longer blog post, but the simplistic economics here are frustrating from such smart people.
- Langfuse joins ClickHouse. Clickhouse, the Databricks and Snowflake competitor, this week reportedly raised $400M at a $15B valuation and also acquired Langfuse. Langufuse is a European-based engineering platform for LLM-based applications, providing tracing, evaluations, and prompt management. The acquisition, no doubt, is designed to help Clickhouse boost its customer offerings for LLM applications to similar levels as the other players in the space. LLM-based application on top of enterprise data is a compelling pattern in 2026.
- AI Has Basically Killed Stack Overflow. I've mentioned this before in various posts, but the problem for many pre-AI websites isn't that users are satisfied with AI answers. It's those same questions (and the answers to them) that created the virtuous circle that made the site and the community work. No questions means no answers, which means less relevance, and right round again to fewer questions. This is the silent death spiral of any user-driven content site for which AI now provides credible answers. The questions and answers going to AI are no longer public and simply disappear down the proprietary black holes of each model provider. It's really not clear how this dynamic can be undone unless some players begin to voluntarily start to publish Q&A data again (sanitized for privacy), but the disincentive is strong since by doing that, they would become a training source for all their rivals.
- Anthropic’s Claude Cowork Is an AI Agent That Actually Works. Hot on the heels of its latest Claude Code release, Anthropic introduced a variant of the agentic tool for office, document, and file-system work. This allows users to ask Claude to run long-running, complex tasks over documents and data on their own computer. A significant amount of trust is required to allow it to do that, but there's definitely a lot of excitement amongst users. Digital workspaces quickly become cluttered; having a power tool to reorganize and use the data is a dream for a lot of people. It's likely quite scary for Microsoft and Google in particular to see this. Up until now, they probably assumed they'd have a key lever in AI by sitting on top of office-type file data for many users, but Claude will be aiming to turn them into a dumb filesystem.
Wishing you a great week.
