Hello from Austin this week, with the Austin City Limits music festival in full swing. This week, we saw OpenAI sell shares at $500B, quantum computing prove a result we long assumed was true, and a cool look inside Mana's drone delivery service in Ireland (recommend 2x audio speed). On to the five most interesting stories:
- Rebuilding Devin for Claude Sonnet 4.5: Lessons and Challenges. This post from the Cognition team on the work required (and benefits) of upgrading the Devin development agent to use the new Anthropic Claude 4.5 release is pretty interesting. The new Claude model takes a lot more internal notes and manages its own context window to some extent. It also highlights that new model releases aren't just "better / faster" but also "different," which means that rebuilds might be needed for each new major release. There's a great discussion on this in one of the latest Latent Space podcast interviews with one of the founders of Langchain.
- Supabase raises $100 million at $5 billion valuation as vibe coding soars. Supabase is really the hidden engine behind the vibe-coding revolution (certainly behind lovable). The service is a cloud-hosted database on steroids with a wide range of useful functions, such as asynchronous worker jobs. Lovable recently introduced a cloud layer (also see Nico Grenie's X post on the launch here), which could be a first step towards becoming independent of Supbase, but I doubt that will happen any time soon. A lot of Lovable and Bolt.new's wows are really Supbase wows. Vercel (another company with a vibe coding application) also raised a new funding round this week at a $9.3B valuation.
- It’s time to prepare for AI personhood. One of the interesting topics I track is mainstream mentions of "rights" for AI systems. This article probably looks like it is before its time, but I'm sure we'll see more. I'm of the belief that if we do produce systems with many of the properties of sentient beings, we're better off proactively treating them as first-class citizens than waiting until we are forced to do so.
- Microsoft Sets the Tone for 'Vibe Working' With New Agent Mode in Word, Excel. Agents are coming to your office work environment. Microsoft's new additions this week allow users to prompt AI within Excel and other applications and then walk through solutions that are being proposed. The idea is to complete really quite complex tasks, such as building new Excel models. A question, though, is how trustworthy the results will really be. If users use the tools to build models that they themselves don't understand well, how will validation be done? Another interesting fact is that the agent uses Anthropic's Claude model under the cover, now OpenAI's models, despite the latter having such a close relationship with Microsoft.
- Major airliner to cut thousands of jobs by 2030 using AI. Lufthansa announced this week that it expects to cut around 4000 people from its global workforce (of around 100,000 today) due to AI efficiencies. The airline also expects to hire new staff in other roles. What is interesting, though, is to see an organization that isn't obviously defined by knowledge work promising such a reduction. No doubt there are plenty of IT systems that could be modernized, but it is still surprising.
In non-AI news, Strava sued Garmin. Seems like a wildly bad idea to me. Plus, possible life on Enceladus.
Have a great weekend!
