davidad bets model alterations would weaken qwen3.8-27b's multi-agent planning
davidad · x · 2026-10-29
Replying to @nlpnyc and @ArthurB, AI safety researcher davidad says he would bet this type of model alteration even makes qwen3.8-27b weaker than default at making plans that actually work — especially multi-agent plans. The exchange touches on side effects of model modifications on planning and reasoning, though no concrete experiments are detailed.
Related event: Neuron Steering Fails to Create Stable Model Personas(2 posts)→
More from Models
- Mistral: le Chonk beats GLM 5.3 and Kimi K3 in 64% of blind human code preference tests — b_roziere · 2026-10-30
- TRL v1.17 ships Mistral Large 4 support: LoRA fine-tune a 1T model on a single B300 node — SergioPaniego · 2026-10-30
- GPT-6.1 Sol costs $1.91 per task, 85% less than Claude Opus 5.5 — arena · 2026-10-30
- Agent Arena: GPT-6.1 Sol costs $1.91 per task, 85% cheaper than Claude Opus 5.5 — arena · 2026-10-30
- UK user gets £0 'Paid' ChatGPT Plus trial invoice but Plus never activates, OpenAI won't clarify billing — Afraid_Gazelle1184 · 2026-10-30
- LlamaParse Agentic scores 179/179 on trap-page parse benchmark, beating frontier VLMs — llama_index · 2026-10-30