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.
Swap 'spider' for 'ant' in Claude's mid-layers and 8 legs becomes 6. Anthropic's global workspace paper, the J-lens readout, eval-awareness detection, and its limits.
RL gains sit at 40-60% depth: on Qwen3-8B, training only layer 16 beats full-parameter RL (67.1 vs 66.5). Notes on arXiv 2607.01232 and what it doesn't claim about efficiency.
Fine-tuned Qwen3-4B on 799 of my own edit pairs, quantized to a 2.3GB GGUF at 38 tok/s on an M4 mini. Eval loss looked fine, but it barely removed slop — and the real fix was feeding paragraphs, not single sentences.
Traced Gemini's "Do NOT issue search queries to the google search tool" message to a leaked system prompt, plus fixes: new chat, forced search, Saved Info.
1,560 edit pairs mined from Claude Code logs, 622 posts scanned with Sonnet, ModernBERT-ja-130m trained on a 16GB M4 Mac mini: zero false positives, 60% recall.
Three LLMs converted the same 10 Japanese scene briefs into Anima (Qwen-DiT) prompts, generated as 60 fixed-seed images on an M1 Max with a merged 3-character LoRA. The Qwen-to-Qwen affinity hypothesis did not survive; a strict formatter brief with character-count locks is what actually moved the results, and two failure modes survive any prompt.
Fujitsu's PHOTON claims up to 475x over Transformers, but that's tokens/s/GiB (multi-query memory throughput), not faster single responses. What the 1.2B paper tables, the quality drop, and 9-query integration really show.
DeepSeek-V4-Pro-DSpark isn't a new base model. It's the same 1.6T V4-Pro checkpoint plus a DSpark speculative-decoding head (~893GB). What config.json and the DeepSpec repo reveal, and why there's no speed benchmark yet.
Sakana Fugu trains no base model: a learned conductor routes GPT-5.5/Claude/Gemini. How it compares to PLaMo (scratch, closed) and LLM-jp (fully open), how it differs from OpenRouter, and its biggest risk.
Tested Qwen3.7 Plus on ModelScope: native function calling and parallel tool calls work. I built a tool loop, skills, and error recovery with just the openai SDK, then had it ship a working Flask BBS.
