ideasthesia

Sign in Subscribe

Ryan Collison

I write about cool science and share it here.

Model self-verification and refinement loop

Self-Refinement and Verification: Models That Check Their Work

A model that checks its own outputs catches mistakes a single forward pass would miss. Self-refinement loops — generate, verify, critique, revise — turn inference into an iterative process. The result: measurably better answers on tasks where first-shot generation reliably fails.

Monte Carlo tree search applied to language model reasoning

Tree Search in Language Models: Monte Carlo Meets GPT

Language models are fluent but essentially reactive. Monte Carlo Tree Search wins at Go by imagining futures. Combining them gives LLMs deliberate reasoning: the ability to branch, evaluate, and backtrack before committing to an answer.

Trade-off visualization between training and inference compute

The Compute Trade-Off: When to Train vs When to Think

Every flop spent training is a flop not spent thinking — and vice versa. The balance between pre-training and inference compute isn't just technical; it's an economic and architectural bet with enormous consequences for AI capability, cost, and deployment.

Extended chain of thought reasoning process

Chain of Thought on Steroids: The Mechanics of Extended Reasoning

How do reasoning models actually work? From simple chain-of-thought prompting to search, self-refinement, and tree exploration — the full mechanics behind extended AI reasoning.

Evolution from o1 to o3 reasoning models

From o1 to o3: How OpenAI Discovered Inference Scaling

From o1 to o3: How OpenAI Discovered Inference Scaling Series: Test-Time Compute Scaling | Part: 2 of 9 The story of how OpenAI discovered test-time compute scaling isn't a story of planned research proceeding according to hypothesis. It's a story of surprising results, paradigm violations, and researchers

Inference-time reasoning scaling with extended computation

The New Scaling Law: Why Thinking Harder Beats Training Bigger

For years the recipe was simple: more parameters, more data, better model. Then researchers found that giving a trained model more time to think outperformed simply training a larger one. That result cracked open a new scaling law, and the field hasn't been the same since.

Synthesis of Graph RAG insights about structured coherence

Synthesis: What Graph RAG Teaches About Structured Coherence

Standard RAG finds relevant passages. Graph RAG finds relevant relationships — traversing a knowledge graph to retrieve not just facts but connections between them. That structural difference illuminates something deeper about what coherent knowledge actually is.