Epoch index shows open-weight models closing the gap with closed frontier to under 10 points
SpencrGreenberg · x · 2026-10-31
Following his post on forecasts of open-weight cyber capabilities, Spencer Greenberg shares an Epoch Capabilities Index chart tracking the open-vs-closed gap since 2023. Closed frontier models progressed from GPT-4 (March 2023) to GPT-5.5 Pro (159), while open weights climbed from Llama 2-70B through DeepSeek-R1/V3 and Qwen3-235B to Kimi K2.6 (151) — a gap now under 10 points, lending weight to the 96% forecast of open-weight superhuman hacking by mid-2027.
Related event: Forecasters See Open-Source AI Reaching Superhuman Hacking by 2027(2 posts)→
More from AGI Musings
- Fields medalist Cedric Villani: by 2026 mathematicians will find AI a replacement, not an assistant — IgorCarron · 2026-10-31
- Consciousness Computability Debate Flares on X Over Penrose's View — fkasummer · 2026-10-31
- Revolut automated 90% of its 5,000-person support; CEO says teams should shrink to 2-3 people — davidsenra · 2026-10-31
- UPI deep-dive podcast with PhonePe and Razorpay founders debates AI shopping agents — sachdh · 2026-10-31
- Nebius CTO: Why the Promise of Agents Doing Everything Is BS — HarperSCarroll · 2026-10-31
- Gary Marcus reports from the Vatican: the Pope's five points on AI — GaryMarcus · 2026-10-31