Today I Learned (TIL)
Things I've learned or things I find interesting will be logged here. For long-form content, you might want to check out my newsletter.
TIL: Claude Shannon's six ways to get unstuck
TIL: Six ways to cut LLM costs
TIL: The AI Wall
Gen AI helps people in adjacent roles perform almost like experts, but it can't close the gap for people far from the domain. Here's what the AI wall means.
TIL: What’s actually different about GGUF and ONNX?
What GGUF and ONNX actually contain, what they expect from the runtime, and why choosing between them is not simply a question of quantisation or hardware.
Speculative Decoding
Trading cheap guesses for expensive forward passes
Query, Key, Values
How to think about Q, K, and V vectors in the Attention layer of a Large Language Model
On Durable Objects, Orleans, and prior art for the agentic web
TIL: Ads in AI chatbots are not just a UX problem
How "Thinking" Models Actually Work
Lilian Weng's Why We Think is a survey of test-time compute and chain-of-thought reasoning. Here's what I pulled out of it.
We’re Being Too Loose With the Term “World Model”
I think we are still too loose with the phrase “world model”.
TIL: Quantisation
Quantisation is really a precision-allocation problem.
How the Claude Code team designs agent tools
Make Claude Code Review Its Own Plans
Stripe's coding agents: the walls matter more than the model
Deep Blue
TIL: The real bottleneck in AI coding isn't speed
TIL: Markov Language
TIL: Learning New Tech With AI Assistance Might Backfire
A study of 52 developers found that using AI to learn a new Python library led to worse comprehension scores, with no speed improvement. Here's what actually works.