Researcher finds frontier models' confident arguments full of holes, Fable still unusable
anshulkundaje · x · 2026-09-26
Anshul Kundaje continues his critique of frontier models' confident-but-vague output: probe them with deeper critical questions and both you and the model quickly realize the original rationale is full of holes, despite "Ultra" thinking. He adds that he still finds Fable largely unusable, hasn't tried much of 5.5, and doesn't know if Mythos or the new GPT model does better — all in the context of AI-written grant proposals slipping past reviewers.
More from Models
- Anthropic's science blog shows Claude pulling off 'Nine Loops' particle physics calculations — rohanpaul_ai · 2026-09-26
- Xiaomi's MIT-Licensed MiMo-V2.6-Pro Hits OpenRouter at $0.87/M Output Tokens — VraserX · 2026-09-26
- Pangram AI detector only catches sloppy AI text, human-edited output passes — soumitrashukla9 · 2026-09-26
- PhD student one-shots 3Blue1Brown-style paper animations with Opus 5.5 in one hour — hsu_byron · 2026-09-26
- Calling fast models "System 1" misappropriates Kahneman, critic argues — emax · 2026-09-26
- Kimi K4 leak: Moonshot's next open-weight model reportedly targets GPT-6 in August — airesearch12 · 2026-09-26