LLMs can't originate a KRAS-style insight, but excel at cross-domain hypothesis synthesis
Ghost_Pilot_MD · x · 2026-08-28
Using the story of the "undruggable" KRAS target and Daraxonrasib as a thought experiment, the author asked Grok and ChatGPT whether the idea could have been invented by AI at the time Gregory Verdine created it (with scant literature) — both said no or unlikely.
The takeaway: teams working on undruggable targets today mostly apply brute-force AI to molecular structures and benchmarks, with little hypothesis-driven creativity, and this hasn't been very successful. LLMs' real value lies in combining knowledge across fields plus human-driven hypothesis pressure-testing and generation — truly novel ideas aren't in training data and must be validated in the real world.
More from Research
- AI Book Club to host live chat with author of 'Build a Reasoning Model' — sophiamyang · 2026-08-28
- Tech Comparison: Sparse Attention Implementation in Qwen vs. Minimax — stochasticchasm · 2026-08-28
- Analysis compares Qwen and MiniMax sparse attention implementations — stochasticchasm · 2026-08-28
- Feeding Real Photos to the Distillation Critic: Krea2 LoRA Stays Sharp at Just 2 Steps — TimeTruth2490 · 2026-08-28
- Sharing the cleanest GDN diagram seen so far — stochasticchasm · 2026-08-28
- Can AI Optimize AI? New Benchmark Tests Recursive Self-Improvement — shehio · 2026-08-28