See a great post, hit "save," and never go back to it?
Or send it to yourself on WhatsApp?
I'm building something new for people who collect posts like this all the time, and I'm looking for a small group to join the beta.
Interested? DM me
@OwenGregorian A one-person biotech that ran most of its work through Claude is the extreme version of the same pattern. According to Anthropic, those drug programs also help it see where its models are weak.
@dejavucoder The awkward part is that the lab employees at that party may soon be interviewing the wrapper founder. In the last year the labs have been buying those teams, or building the same verticals in-house.
@hannahhaina Coefficient fits the same pattern as the consumer apps. The labs pay up for a small team that already ran the model against a messy real domain and found where it falls short.
@FDavidsonT The product shutting down the same day is the tell. I think the labs are buying a map of where the model breaks in one vertical, not the travel app.
@wallstengine This isn't a one-off travel bet. In the past year OpenAI also bought health records, personal finance, slides, scientific writing, ads and drug development, and in education, legal and shopping it just built the product itself.
@Dhanush_Nehru The productivity stack is one front. The other is vertical: they ship education, legal and shopping themselves, and they buy the travel, health and biotech teams that already did the messy work.
@DavidPawlan Travel is one cell on a bigger map. They also bought health, finance and drug development, and in education, legal, shopping and government they didn't buy anyone. They just built it.
@briannekimmel@FDavidsonT@samyamiam Industry expertise is the right frame, and I'd take it one step further. They're buying a map of where the model fails in a real vertical, which is hard to hire for and harder to learn while training the next model.
In other fields they didn't buy anyone. They built it: education, legal, enterprise finance, shopping, government, health and code.
I think they're not buying products. They're buying the knowledge of where the model breaks in one domain, and the general product sitting on top of the model is what they're building now.
I'd look at fields on neither list yet, like insurance, logistics, real estate and construction, or at angles off the labs' critical path that complement what they build.
The longer version is here: https://t.co/1Bf8eYGsc6
Where I wouldn't start a startup today.
This week @OpenAI acquired @OdessiaTravel, an AI travel agent less than a year old that raised a $6M pre-seed led by @sequoia. It's not a one-off.
In the past year @OpenAI bought travel (Odessia), medical records (Torch: four people, reportedly $60M to $100M in stock), personal finance (Hiro), personal investing (Roi), docs-to-slides (NextSlide), scientific writing (Crixet), an ads agent (Notch) and drug development (@OrchestraBio).
@AnthropicAI bought Mithril Medicine, a one-person biotech that ran most of its work through @claudeai, and @CoefficientBio, fewer than ten people, reportedly about $400M in stock.
A must have in any builder tech stack !
I build 14 hours a day, and manual iOS QA used to eat a real chunk of that.
Then I plugged Mobile MCP from @mobilenexthq into my agent. Now it opens the app on the phone, taps through the flows, and checks the screens itself. Hours of clicking, gone.
Open source, works with Claude Code, Cursor and Codex.
Hats off @gilmegidish
@maxkarpis "AI independence" is the right frame. The cost cut is the headline, but the bigger win is no longer having your engineering velocity tied to a vendor's pricing page. I'd bet more European scale-ups follow Revolut within a year.
Managing tokens like cloud spend is right, but the cloud playbook also says: once a workload is big and predictable, you bring it in-house. Uber chose per-employee caps, Revolut chose self-hosting with unlimited tokens. Which one do you see more of in your enterprise conversations?
@aviancapital Same pattern on the software side: Revolut moved 2,000 engineers to a self-hosted open model and cut the bill to about a tenth. Fine-tuned open models on your own data plus your own hardware looks like the enterprise default by 2027. Frontier APIs become the specialist tool.
@jun_song Tough spot indeed. I don't think the labs lose the hard problems, they keep the frontier. But the volume work, the stuff that fills an engineer's day, moves to open models. The revenue forecasts assumed that volume. Revolut is the first big public example, it won't be the last.
Some data points to go with this: Scale VP's survey puts it around $1,040 per engineer per month, roughly 4% of payroll. At that level token spend stops being a tooling line item and becomes a headcount conversation. Revolut answered it by going self-hosted and cutting the bill to about a tenth.