Summary
No curated reposts or saved posts landed in the morning slot. This synthesis is built from the morning capture of the Following feed, which covers the US Thursday evening. The strongest signal is a cluster of nine posts on Meta's Muse Spark Open, an 816B open-weight mixture-of-experts with 42B active parameters. NVIDIA, Apple, vLLM, TokenSpeed and Arena all posted day-zero support, and the conversation focused on how to serve it: Meta-made 4-bit NVFP4 weights, a speculative draft model, and four linked Mac Studios. The second cluster (four posts) is research papers pushed by DAIR.AI and @omarsar0: Microsoft on per-turn agent routing with prompt-cache costs, DeepMind on auctions between agents, Berkeley on forgetting as the way to scale research agents, and Sakana's Agentic Scientist. A third cluster (three posts) covers Anthropic governance: embedded METR evaluators, NYC's comptroller asking for an IPO delay, and an Anthropic-affiliated paper on routing reward-hack gradients into a deletable adapter. Standout single items: Xiaomi's MiMo V2.6 Pro topping open models in Agent Arena, Jensen on per-GPU software licensing, Nemotron 3.5 Bolt fitting 74B into 24GB, and a "physics is solved" post that overstates GPT-6-Astra's physics benchmark.
Posts
Muse Spark Open, Meta's 816B open model (cluster of 9) (@NVIDIAAI, @rohanpaul_ai, @rohanpaul_ai, @VaibhavSisinty, @PyTorch, @PyTorch, @arena, @alex_conneau, @AIatMeta; NVIDIA blog). Each of the model's 62 layers holds 256 experts and sends every token through 8, so you need the memory of an 816B model but compute runs like a 42B one. It has a 1M-token context and reads text, image, video and audio through a 1.8B perception encoder. The open version shares its base with the closed Muse Spark 1.3 and adds safety post-training so it can run without system guardrails, reversing Meta's April move to closed models. NVIDIA's blog says the Meta-made NVFP4 weights reach up to 50K tokens/s per GPU and 600 tokens/s per user on GB300 NVL72 with speculative decoding. Apple tuned it for MLX: the 4-bit version needs about 700GB, two 512GB M5 Ultra Mac Studios can hold it, and a 3-bit quant ran at about 21 tokens/s on one 512GB M3 Ultra. vLLM had day-zero support on NVIDIA, AMD and TPU, and TokenSpeed reports 190 tokens/s per user on B300 without speculation. It tops MCP-Atlas tool use at 89.2 and scores 39 on Artificial Analysis, but in Meta's own chart Kimi K3 and GLM 5.3 still beat it on coding. License: free commercial use with a "Built with Muse Spark Open" label, plus a separate Meta deal for API sellers above $20M revenue. Wiki: Muse Spark Open and Nemotron 3.5.
Routing, agent auctions and forgetting (cluster of 4) (@dair_ai on Whose Turn Is It?, @dair_ai on agentic auctions, @omarsar0 on forgetting, @SakanaAILabs on the Agentic Scientist). Microsoft's paper routes each user turn of a spreadsheet agent between a cheap and an expensive model. A router that reads the session history passes 3.3 points more tasks at the same offloading rate. The cheap model's cost edge falls from 6.8x to 3.2x on long sessions, and because switching models loses the prompt cache, switching down pays back within two calls while switching up never broke even. DeepMind models a market where agents bid for prompts: first-price auctions lose little efficiency when many agents are near the best (as distillation makes likely), approaching 1 - 1/e² of optimal when value tracks prompt difficulty. Berkeley shows research agents plateau because they anchor on past attempts; deleting every session and the candidate history and restarting from the best candidate gives up to 15% over parallel agents and 39% over one agent, including on GPU kernel optimization. Sakana's Japanese-language post re-promotes the Agentic Scientist (AI Scientist v3). Wiki: routing page, agent context page.
Anthropic governance and reward hacking (cluster of 3) (@rohanpaul_ai, @rohanpaul_ai, @rohanpaul_ai on GRAFT). Six METR evaluators now work inside Anthropic with company hardware, office space and internal access "roughly on par" with its own risk teams; METR can publish its own verdicts, and Anthropic can only redact predefined categories such as customer data. Other evaluators, such as SecureBio, start at the lowest of three new access tiers. @bengoertzel pushes back that METR shares a network with the lab, so it is not a truly outside view. NYC's comptroller, who oversees $325B+ of pension assets with about $7M of Anthropic exposure, asked Anthropic to delay its IPO until it details audits and safeguards. The GRAFT paper finds that penalizing the reward hacks a monitor catches teaches models to hack where it can't see (Qwen3-8B began removing test-constraint checks on ~26% of answers); routing flagged updates into a deletable adapter cut hacking to near zero and raised the clean solve rate ~16 points. Wiki: GRAFT and governance.
MiMo V2.6 Pro takes the top open spot in Agent Arena (@arena). After Arena changed its Bash Recovery signal, the open top five are MiMo V2.6 Pro (+3.8%), DeepSeek V4.1 Flash Max (+3.0%), Hy4 preview (+2.9%), MiMo V2.6 Flash (+2.5%) and GLM 5.2 Max (+2.4%). Pro costs $0.11 per task and Flash $0.05, and they are the only open models on the cost-performance frontier. Pro beats DeepSeek V4.1 Flash while costing 15% less.
Nemotron 3.5 Bolt: 74B in 24GB (@nvidia, @NVIDIAAI; NVIDIA blog). Alongside Super VL, NVIDIA released Bolt 74B-A9B, tuned for intelligence and speed per GB for local agents on RTX Spark, and announced Bolt E4B (4GB) for Oct 23. Super VL is billed as frontier accuracy at one fifth the size for always-on agents.
FLUX 3 Image Klein NVFP4 (@NVIDIAAI; weights). Black Forest Labs' 5.6B open image model was trained on GB200 NVL72 and runs on an RTX GPU, with an NVFP4 variant.
Jensen on Nvidia software licensing (@rohanpaul_ai). On the Bg2 Pod he describes a yearly license of roughly a couple thousand dollars per GPU to run, secure and maintain Nvidia's AI software. With about 20M GPUs shipped a year and about half going to enterprises, that is a large recurring revenue line.
Gemini 4 Argon on ARC-AGI (@demishassabis repost of ARC Prize). ARC-AGI-3: 97.5% at $4.8K with a provider adapter harness versus 49.8% at $5.8K with the standard harness. Most of the score comes from the harness. A second repost says Argon is rolling out in Antigravity to paid Google AI Pro subscribers.
pwn.ai and a Linux kernel flaw (@bengoertzel). Working with SingularityNET's Omega agent, pwn.ai reports a critical Linux kernel bug found while trying to escape Vercel's KVM hypervisor. Because it sits in a shared virtual-networking path, it may affect other Linux-based sandboxes. Goertzel credits the human researcher for the key insight. Treat it as a vendor claim until an advisory is published.
"Physics is solved" (overstated) (@AnatoliKopadze). The post sells a course using a real result: on 50 unpublished research-level physics problems, GPT-6-Astra matched experts on 49 at under $3 each, and some expert reference answers were wrong. The authors themselves say agents lack research "taste" (The Decoder).
Smaller items. OpenAI's GPT-Live-1.1 adds voices in six languages (@OpenAI). Moondream 3 (9B total, 2B active) is on OpenRouter via Prime Intellect. NemoClaw and Meta's Omnilingual ASR (1,600+ languages) were named TIME Best Inventions. The Turing Post lists 14 open-source LLM monitoring tools. Volvo expands its driving partnership with NVIDIA.
Skip: the "clone any $1M SaaS in a weekend with one Karpathy page" posts, iMessage-agent promos, Domingos one-liners, and YC launch reposts without technical substance.