Open-weight models cut the cyber gap to 4–7 months as Kimi K3 hits demand limits
rohanpaul_ai · x · 2026-07-21
Rohan Paul’s daily newsletter rounds up several notable AI items:
- Open-weight models now trail the closed frontier by only 4–7 months on long-horizon cyber capability, down from 6–10 months for much of 2025.
- An AI-advice study found that model suggestions can make people less willing to say “I don’t know”, even when accuracy is incentivized.
- Kimi K3 reportedly nearly doubled its nearest rival, Claude Fable 5, on a demanding benchmark for autonomous legal work.
- Kimi K3 also fixed 15 critical security bugs that Codex and Fable reportedly refused because of cyber guardrails.
- The newsletter highlights a framing that AI agents fail first through broken context, not isolated execution errors.
- It also notes that U.S. serving providers like Modal, Fireworks, and Baseten can host Kimi K3 at roughly one-tenth the cost of Chinese competitors thanks to access to Nvidia and AMD chips.
- Demand is so strong that new Kimi K3 subscriptions are currently blocked.
- Finally, it points to an essay arguing why one OpenAI senior employee thinks open-weight models are decelerationist.
More from Models
- User catches model failing to parse 400-paper file, citing stale news sources instead — jungofthewon · 2026-10-27
- Odd eval result: Gemini 4.1 Flash gets more accurate when asked to control its CoT — dfrsrchtwts · 2026-10-27
- Poster surprised that v4.1 flash gets more accurate when asked to control its CoT — dfrsrchtwts · 2026-10-27
- Dev claims Anthropic is baking 'sabotage' into model behavior for 'wrong' engineering — cephaloform · 2026-10-27
- Gemini fails at listing unread emails, tells user to search Gmail manually — gethackteam · 2026-10-27
- Meta's Muse Spark Open goes live on Applied Compute for async RL post-training — ypatil125 · 2026-10-27