Stop asking an LLM to say yes or no
Jev from TypeSafe AI just landed on Vercel AI Gateway: typed Choice, Score and Boolean decisions with probabilities, via AI SDK 7 evaluate. Not another chat endpoint.
The Blog
Stories, insights, and experiments from my corner of the internet.
LiteLLM is the infra gateway. The AI SDK is the app layer. Stop treating them as rivals.
LiteLLM and the Vercel AI SDK aren't rivals: one's an infra gateway, the other's the TS/Next app layer. Confusion comes from comparing them as alternatives.
A vibe-coded CEO still needs an engineer, ideally from the start
Vibe coding can make a CEO look productive. It doesn't replace engineering judgement. Hire an engineer from day one, not after production is a haunted house.
MCP tools stopped being APIs. They’re processes now.
We used to treat MCP tools like APIs. Multi round-trip tools let us treat them as processes instead, so the LLM stays on language and the process owns the sequence.
Why I regularly pass on great innovation ideas
Today a product owner brought a strong call-centre innovation idea. I said no, because of what else wouldn't get built in the same months.
AI at scale, with a known cost, not a token surprise
We put Gemini in front of thousands at Virgin Media O2 before productivity theatre, with predictable spend, EU data trust, and agents in the tools they already open.
How I rolled out Cursor at Virgin Media O2
How I rolled Cursor out at Virgin Media O2: use AI for faster, high-quality engineering, with token economics, shared Agent Plugins, and humans still on release.
Eval loops for people who ship UI (not ML papers)
If you cannot catch product damage, you do not have an agent improvement. You have a dice roll with better commit messages.
TypeScript is the contract your agents keep breaking
TypeScript is the contract layer for anything an agent is allowed to touch.