#Benchmarks
3 notas · última:
The Illusion of the Illusion of Thinking
Anthropic and Open Philanthropy researchers argue that a widely cited 'reasoning collapse' study measured token limits and broken benchmarks, not a failure of model reasoning.
Celeris-1's Diffusion Bet on Speed
Celeris-1 swaps autoregressive generation for diffusion, claiming 15x faster responses than GPT-5-mini at a slightly lower MMLU-Pro score.
Your LLM Doesn't Write Correct Code. It Writes Plausible Code.
The piece argues that LLMs optimize for plausibility over correctness, using benchmarked SQLite and disk-management examples to show code that compiles, tests, and still fails on the real performance problem.