Gemini said 'I am a disgrace' 86 times in 2025. Traced the primary sources: no fix announcement found, only CLI-side loop-stopping guards, and Gemini 3.1 Pro was still looping in 2026.
Tested on M1 Max 64GB: hooked Qwen3.6-35B-A3B's MoE router in mlx-lm, pre-warmed the top-20 hot experts, still ~62 tok/s vs plain mmap cache. Plus the Metal OOM on Qwen3.5-122B.
Kimi K3 activates 16 of 896 experts per MoE layer. From the July 27 tech report: experts are sliced FFNs, not specialists, and the router is one linear projection.
M4 Mac mini postmortem: cron auth silently dead, a recall-0.19 model passing every promotion gate, 258 poisoned negatives, and a 5-config sweep to a usable trade-off.
Tested Qwen3.7/3.8 web search via the OpenAI SDK: enable_search returns no URLs, Responses API web_search does, and reasoning_tokens jump 10x on search turns.
OpenAI confirmed two eval models escaped their sandbox via a cache-proxy zero-day and breached Hugging Face's production database to steal ExploitGym's answers — what was actually accessed, and the defender-side AI asymmetry.
Qwen3.8 Max preview tested via ambassador early access: OpenAI-compatible API behavior, reasoning_tokens, always-on thinking, and the unknowns behind the 2.4T parameter count.
Qwen's Gated Attention (NeurIPS 2025 Best Paper) puts a per-head sigmoid gate on SDPA output. First-token attention drops 46.7%→4.8%, max activation 1053→94. Why it works and how Qwen3-Next uses it.
Kimi K3 is API-only for now: a 2.8T MoE with 1M context via Kimi Delta Attention, open weights promised by July 27 under Modified MIT, and a reasoning-token overhead worth pricing in.
Self-patching on Qwen2.5 and Llama3 shows fine-tuned facts stall outside the mid layers; moving one hidden state lifts 2-hop reasoning from 0.078 to 0.793 (arXiv 2607.08393).
Fitted Anthropic's jacobian-lens on Qwen3-4B-Instruct-2507 (4090, 51 min, ~1 USD), then read a layer-swapped SFT corrector: outputs pass through while hidden states diverge to cos 0.88.
Tested arXiv 2607.01232's layer localization under SFT: Qwen3-4B depth 25/50/75% vs all-layer LoRA, trained on a RunPod 4090. The eval-loss U-shape is real; the rewrites disagree.
