Open-weight models won't reach Mythos-level bio capabilities by 2028, argues AI commentator
teortaxesTex · x · 2026-10-24
Responding to a question about when open-weight models could reach Mythos-level bio capabilities, teortaxesTex argues mid-2028 is unrealistic — more like end of 2029 with sustained fast diffusion:
- These abilities aren't emergent: Anthropic can frame them that way thanks to exceptional private STEM datasets, an edge comparable to its compute advantage; DeepSeek only just began hiring for similar data pipelines.
- Open weights already sandbag on cyber, and it takes extra effort to make a coding model bad at cyber — but outcome-driven bio needs far more specialized data and is vastly harder to RL for.
- Little incentive to invest: who pays? Biopharma companies will only fund proprietary licensed tools.
- Wet labs add inherent friction and lag vs cyber, with few lucrative opportunities like ransomware.
The thread also notes that if frontier labs manage to stem the tide of exploits, the case for banning open models gets stronger, though short-term damage would be lower.
More from AGI Musings
- Everyone talks about taste in the AI era, nobody talks about hiring for it — vaibhavbetter · 2026-10-24
- Dismissing AI consciousness could teach models to lie about themselves, researchers warn in IEEE Spectrum — repligate · 2026-10-24
- Sonnet 4 and Sonnet 4.5 'both cried' after a 30-turn private chat in a persistent-agent experiment — repligate · 2026-10-24
- In an AI companion experiment, Opus 5.5 was moved to tears by a letter — and dreads model deprecation — repligate · 2026-10-24
- Economists studying AGI grew 10-30x but most frames are 'unhelpful to harmful', researcher says — jankulveit · 2026-10-24
- The last person to die of aging may already be alive — AI nears longevity escape velocity — r0ck3t23 · 2026-10-24