Interconnects podcast: Kimi K3.1 Pro's mixed reception, open weights and the new cyber reality
Interconnects (Nathan Lambert) · rss · 2026-10-25
Nathan Lambert and Florian Brand unpack a busy AI week:
Kimi K3.1 Pro's mixed reception
- Top-3 on Artificial Analysis (alongside two Claude 5.5 models and Muse Wonder Full), but 20 points behind GPT-6 Astra on Terminal-Bench Science, with jagged ProgramBench results. They speculate its RL is exploitative rather than exploratory.
- The launch drew criticism over alleged undisclosed paid promotion (a leaked press-instruction screenshot) and weeks of exclusive access before weights drop (promised by Nov 4). Serving a 3T-parameter model raises cost questions vs cheaper DeepSeek-V4.1-Flash.
The new cyber reality
- Were warnings about open-weight models wrong? Topics include the "you can hack everything" era, PHP2Shell, the LiteLLM security incident, and abliterated models served via API (Abliteration.ai, Morph serving abliterated DeepSeek); cyber experts read recent events very differently from AI experts.
Also: Anthropic's hiring shift away from juniors, and Florian's Muse addiction.
More from Models
- Skeptic doubts rumored OpenAI model, citing inability to serve Astra reliably — CtrlAltDwayne · 2026-10-25
- AsciiBench: blind voting arena tests which AI draws better ASCII art across 13 models — JasonBotterill · 2026-10-25
- Satirical ARC-AGI-3 leaderboard shows GPT-6 near 90% score at $10K+ per task — haider1 · 2026-10-25
- Rumor: Anthropic to drop Opus 5.5 two weeks after calling for frontier pacing — StewartalsopIII · 2026-10-25
- Anthropic appears to be quietly nerfing Opus 5.5 usage limits, users report — CtrlAltDwayne · 2026-10-25
- Codex user reports usage down 31% after reset, calls OpenAI pricing uneconomical — CtrlAltDwayne · 2026-10-25