TOPIC
3 Stories Published
AI evaluation
A.I.
9m
Inference Compute and Benchmark Scores: Why a Fixed Token Budget Understates Newer Models
Inference compute changes AI benchmark scores. UK AISI and EvalEval released verified results showing how token budgets and retries shift results.
Research
8m
The Erdős Unit Distance Conjecture Falls: What OpenAI’s Model Proved and What Mathematicians Say
An unreleased OpenAI model disproved a 1946 Erdős conjecture about unit distances in the plane. The proof, the outside checks, and the…
Research
8m
Mechanistic Interpretability Under Test: Do Circuit Explanations Tell the Whole Story?
A master's-thesis preprint scores circuit-based explanations in GPT-2 Small. They read well and are sufficient, but cover only about 22% of the…