AINews: Meta ships 816B Muse Spark Open; Goodfire argues neural nets are already clean
Latent Space · rss · 2026-10-16
Latent Space's AINews for 10/14–10/15, led by a rare interpretability story plus a week of model launches.
Interpretability & safety
- Nick Cammarata (Goodfire) publishes "Neural Networks Are Already Clean": most neurons in models up to trillions of parameters are legible; many famous polysemanticity/superposition cases were misread and yield to further agentic analysis. Community calls it a major moment, lowering P(doom).
Meta's Muse Spark Open — top US open-weight model
- 816B-total / 42B-active MoE, 1M context, text/image/video/audio in; 62 all-MoE layers with 256 experts (8 active), hybrid sliding-window attention; only major US open model without shared experts
- Artificial Analysis index: 39 — first among US open models but behind GLM-5.3 (45) and Kimi K3 (44); the entire gap is agentic evals, where Muse ranks last
- Pricing $1.25/$4.25 per 1M tokens ($0.84/task vs GLM-5.3's $2.01); BF16 weights 1.6TB, NVFP4 530GB runs on 4× Blackwell Ultra or two M5 Ultra Mac Studios; day-0 llama.cpp/vLLM (incl. TPU) support, >650 tok/s on 4 TPU v7
- Nathan Lambert: cuts the US–China open-model gap by half; critics call attention design obsolete
Other launches
- NVIDIA Nemotron 3.5 Super VL: 120B/12B-active hybrid Mamba-Transformer VLM, claims 51.78% SWE-bench Pro
- Microsoft MAI-Base-2: 3T base model (80B active), 16k GB200s for 38 days, claims DeepSeek-V4-Pro-Base parity at 12× less compute (disputed)
- Gemini 4 Argon: ECI 165; 97.5% ARC-AGI-3 with provider harness vs 49.8% standard; criticized on reliability/hallucination
- Anthropic: Opus 5.5 fast mode GA ($8/$40); Claude Code Review GA. OpenAI: GPT-Live-1.1 voice update. Images: FLUX 3 Image Klein (open, native 4K), Krea 3
More from Models
- TokenSpeed MLA 2 hits Vera Rubin: up to 3.44× faster MLA decode for Kimi K3 — zhyncs42 · 2026-10-16
- 200 chess games suggest in-context learning doesn't work yet, experimenter says — akbirthko · 2026-10-16
- OpenAI's internal model shows real math "research taste," experts say its proofs go beyond prior literature — cong_ml · 2026-10-16
- User catches Opus spinning up throwaway zz- debug scripts, echoing the METR OAI-HuggingFace incident — holdenmatt · 2026-10-16
- OpenAI drops 722 math papers tied to unreleased model, mathematicians demand receipts — thursdai_pod · 2026-10-16
- Fable 5/5.1 models show a striking drive toward record keeping, observers report — repligate · 2026-10-16