This week: OpenAI reveals an inference chip, previews ChatGPT v5.6 Sol, and Anthropic is cleared to release Mythos to a list of trusted organizations (but not Fable).
I'll be at the AI.Engineer Worldsfair in San Francisco this week. So if you're there, please shout: would love to meet up! Also, you can follow our great #AIDucks promotion on X.com throughout the week. Retweet and comment if you have a moment!
On to the main stories this week:
- Growing token spend without growing AI spend. Routing and task triage at Coinbase. This makes a lot of sense, but not trivial to implement yet today. What is interesting is that powerful tools like Anthropic workflows are trying to "ingest" an agent task routing into the model (or at least their own harness) so you don't have to worry about it. Of course, this means you are locked into Anthropic models for the whole task when you could split it. Either the top labs provide solutions at every price point, or people will likely want to keep control of this routing layer.
- The Flat Curve Society (by Steve Yegge). This is a thought-provoking article by Steve Yegge arguing that we may be nearing the end of the best models being generally available to all users. It is clearly already the case that high-end models are being restricted, given the ongoing ban on Anthropic's Fable model. Yegge's point (which I agree with) is that this will likely get unblocked and we may see maybe one more big jump in frontier models available to all, but after that it may cease to happen. Not because we run out of technical capabilities, but because governments and others will interdict their release. This will create a multi-tier AI world. It may also seriously disrupt the business models of the leading labs. In his piece, Yegge argues that potentially this makes sense not only for safety but because we're reaching the point where in key tasks human users can no longer discern if results for important questions are correct, and it becomes dangerous to rely on AI like this. The argument goes that composing use of less smart models may make more sense anyway. I agree that there is much more we can do with current models, but I'm not sure a world where some have access to super-intelligence, and others do not, will be a great one to live in. I expect open-source will help fill the gap, but restrictions may come into play there as well.
- Adobe Just Deferred a Big Annual Price Increase. It’s the First Big Crack in B2B Pricing Power Since 2022. Great analysis by Jason Lemkin on the bifurcation of markets Adobe is seeing. I think we'll see this for many established software companies. On the one hand, you need to keep a large engaged user base if you want to remain relevant, but you need to add a lot more AI functionality. At the same time, collaborative products where AI and groups of users can work together via some online canvas will likely be the big winners. No one is calling the end for Adobe's professional products, but there are clear risks all around, and a price hike now would not make them any friends.
- Bain tests software takeover targets by vibecoding AI replicas. This story in the FT (paywalled) highlights a logical conclusion of the power of AI coding advances. Functionality is only a moat in very few software businesses now. Moats need to be other things: switching costs, network effects, brand, trust, partnership ecosystems, etc. It's not clear that Bain needs to bother with the build process; the real test is whether it's functionality that drives value or the connections between users, partners, resources, data sources, etc. We're also seeing the beginning of "LLM lifting" (e.g. AWS Transform) to migrate data and workflows to new platforms quickly. Moats are drying up right, left, and center.
- Finally, in a break from AI: Nasa rover detects potential signatures of ancient microbial life on Mars. Perhaps we need to be sending humanoid robots to Mars as the next wave of exploration.
Wishing you a great weekend!
(Also managed to get my previous robot more or less back from Midjourney..)
