After our midweek interlude post, we're back to normal programming!
This week, Google Cloud accelerates (with Anthropic chipping in), Richard Dawkins weighs in on AI consciousness (already achieved), and Microsoft is working on Project Lobster.
On to this week’s bigger stories:
- I, robe-ot: the android monk working to reboot the faith of South Korea’s Buddhists. Douglas Adams predicted this in his 1987 novel “Dirk Gently’s Holistic Detective Agency”, and robot priests are a venerable tradition in sci-fi that probably makes a lot of sense. Amazing that we have gotten here so fast.
- A recent experience with ChatGPT 5.5 Pro(aka GPT 5.5 pushes the Math frontier). This detailed post is a great walkthrough of how GPT5.5 solved a frontier-mathematics problem with little human input. There’s clearly novel reasoning going on. Impressive to see. I also like this comment from the author: “So maybe there should be a different repository where AI-produced results can live.” Yes, and it will probably quickly become more interesting than ArchivX. If I were ArchivX, I would definitely allow AI publication as well, but I would ensure such content was clearly flagged. You’d clearly want moderation (or scoring), but realistically, if the experience of Agentic coding is anything to go by, it’ll need to be AI-moderated. An interesting thing to consider is the token cost of work done this way. If I spend the effort to generate new work with AI, I’m expending tokens. If I then publish it for public review and others run agents to verify, I get the benefit of that verification for free (and so does everyone else). The author also says this: "The lower bound for contributing to mathematics will now be to prove something that LLMs can’t prove, rather than simply to prove something that nobody has proved up to now and that at least somebody finds interesting.” Making true research contributions across many domains has just gotten harder.
- ServiceNow, SAP, and Workday Make AI Agents Pay to Play. Several of the larger established SaaS infrastructure players are joining Salesforce (Slack) in restricting third-party agent use of their APIs. The public story is to ensure compliant usage and for infrastructure costs. However, it seems obvious that this is an attempt to ringfence external AI access, so customers are more likely to use internally provided AI Agents. There is a logic to this move for incumbents (concretely, why would you need SAP AI if you can have Claude go over the top and do everything for you). However, I very much doubt the bet will pay off in the long run for all but the most embedded companies. They may lock in existing customers, but new customers will be extremely wary of joining the party.
- Andrej Karpathy: From Vibe Coding to Agentic Engineering. Great podcast interview with Andrej Karapathy on the current state of AI coding. He's someone who is really good at looking ahead and calling things as he sees them. Worth a listen. I especially like his point that the LLMs we use today are more like spiky ghosts of animals rather than actual animals. Agents even more so.
- Higher usage limits for Claude and a compute deal with SpaceX. Anthropic has been slightly behind in its acquisition of compute capacity, and this has been starting to create some user frustration. Now they can breathe a bit easier since they have a large chunk of new capacity sleeping on Grok's couch in Texas. It's unclear what this means for Grok, but it's smart of SpaceX to sell capacity for top dollar when the opportunity presents itself.
Have a great weekend!
