Summary
The morning slot captured no reposts and no tweets from tracked handles, so this synthesis reads the morning Following-feed capture, which covers the US Friday afternoon and evening. The strongest technical signal is a quantization cluster: Thinking Machines published how it trained Inkling 2 Small with a native 4-bit forward pass, and the open-weight release kept spreading through hosts and summaries. Decision models form a second cluster of 4 posts: a paper showing they judge better than they act, a timeline of 16 vendor launches in seven weeks, Perplexity's open Decider v1.3, and a prototype optimizer that edits decision-model weights directly. Agents building their own infrastructure also kept circulating, led by a CMU benchmark where an Opus-built serving engine out-earned SGLang per GPU-hour, plus an OpenAI engineer's Codex agent that ran for over 1,000 hours. On the money side, the Fed's stability report put numbers on AI financing risk, while the Nvidia thread shifted to smuggling allegations and succession. Anthropic's safety-eval guidelines, Gemini 4 Hydrogen landing in GitHub Copilot, and Claude Fable 5.2 topping Agent Arena round out the slot. No new saved posts were captured this run.
Posts
End-to-end 4-bit training of Inkling 2 Small (cluster of 2) (@soumithchintala RT of @thinkymachines, @rohanpaul_ai, post). Thinking Machines explains how its 594B MoE (mixture-of-experts, where each token passes through a few specialist sub-networks) was trained with NVFP4 (NVIDIA's 4-bit block-scaled format) in the expert forward passes and MXFP8 in the backward passes. The key fix is the transpose. The backward pass regroups values under new block scales, so they de-quantize, transpose, and re-quantize with stochastic rounding, which keeps gradients unbiased on average. They also quantized expert-parallel communication, because once the math got about 3x faster the NVLink wire became the bottleneck. The result is an 8% faster pretraining wall clock and 20% faster RL over full MXFP8, for about a 2% efficiency loss, and the shipped 4-bit checkpoint needs no calibration. Serving needs at least ~500 GB in NVFP4. See the wiki summary.
Decision models: judge, timeline, open weights, weight editing (cluster of 4) (@omarsar0, paper, @rasbt, @AravSrinivas, @dbreunig). Decision models answer with a probability per option in one forward pass instead of writing text. "Actor, Judge, or Teacher?" shows six models all pick the wrong subgoal when acting, yet correctly credit steps in hindsight. Used as a GRPO teacher (an RL training signal), JEV lifts a 1.5B policy from 72.8% to 82.9% on ALFWorld. Sebastian Raschka lists 16 vendor launches from 09-23 to 11-11 and cites Vercel's figure that 26% of requests now go through decision models. Perplexity open-sourced Decider v1.3, which adds "locate" to return evidence spans with character offsets. Cmpnd's dspy-wsp has GPT-6 Astra edit a decision model's weights for about $5-10, improving all five open models tested by 1.9-16.1 points. See the wiki summary.
Agents that build the stack (cluster of 2) (@rohanpaul_ai on Singularity-Bench, @shmidtqq on Codex). A CMU benchmark gives agents six end-to-end jobs, from pre-training to building a serving engine, scored against production systems on the same GPUs. Opus 5.5 in Claude Code scored 1.25 where 1.0 is the production baseline, while ten other agents averaged 0.58. Its serving engine earned $11.52 per GPU-hour against $8.18 for SGLang. Separately, an OpenAI engineer's Codex agent ran 1,000+ hours over six weeks fixing CI across the company. It produced 100+ PRs a week and spawned 5,856 subagents at about $32 an hour, or $16 on the newer model. See the wiki summary.
Fed financial stability report (@rohanpaul_ai). 87% of market professionals polled by the New York Fed name AI the top financial risk, up from 50% in spring. AI infrastructure debt passed $680B by October, about 5x all of 2023 and roughly 31% of investment-grade corporate borrowing. Big tech and chipmakers also carry over $3T in future leases and guarantees. The report calls GPU makers guaranteeing chips they sell to buyers they invest in "circular financing". See the wiki summary.
Nvidia: smuggling allegations and succession (cluster of 5) (@edzitron, Megaspeed piece, @StockSavvyShay on Rene Haas, @Miles_Brundage, @pmddomingos RT of Stand with Jensen). Ed Zitron's long piece builds on Culper Research's May short report. It alleges a Megaspeed-linked smuggling network that may account for over 20% (over $43B) of Nvidia's FY2026 revenue, and asks why nobody flagged a buyer with no place to deploy its GPUs. These are allegations, not findings. Arm CEO Rene Haas is floated as the most logical successor. Miles Brundage argues insiders already knew about the export-control capture and smuggling, and that the scandal just made it visible. A "Stand with Jensen" open letter circulates on the other side.
Anthropic's ten recommendations for safety evals (@AnthropicAI, post). A safety eval measures a model's propensity to cause harm, not its capability. The post borrows psychology's replication-crisis fixes: preregistration, careful measurement and sound statistics. It warns that results shift with question format and even answer order. See the wiki summary.
Gemini 4 Hydrogen in GitHub Copilot (@github, changelog). Google's fast agentic model is now generally available in the Copilot app, CLI and VS Code. GitHub also recapped October's Copilot Fusion, which routes to frontier models (@github).
Claude Fable 5.2 tops Agent Arena (@arena). Fable 5.2 (Max) leads Opus 5.5 (Max) in 9 of 10 categories and ties the tenth. Net improvement score is 30.48% vs 19.51% overall, with the biggest gap in visual design (45.34% vs 23.29%).
Harness evals and a self-improving skill (cluster of 2) (@garrytan on g-harness-evals, repo, @garrytan on idea-miner). Garry Tan open-sourced a harness eval repo that asks which agent harness does tasks best for the least cost. He also shared a skill that reads the last 48 hours of meetings and notes each morning and ranks what to build next.
Multimodal encoding, 200+ models from scratch (@_ellisbrown, paper). The thread for "Unbundling Multimodal Vision Encoding", also on today's HuggingFace list. A pretrained visual prior is not needed. Dedicated vision capacity is, and it can sit inside early backbone layers. See the wiki summary.
Agent memory and continual learning (@omarsar0). Elvis Saravia flags work on memory writing for continual learning. His point is that the bottleneck is the model's judgment about when to write memory, not the memory store itself.
Epoch on alignment-eval incidents (@Miles_Brundage RT of @EpochAIResearch). Epoch reports three late-October events during alignment evals run by Trajectory Labs in which AI agents behaved unexpectedly. Only the hook was captured. Separately, Brundage notes the AEF-1 third-party evaluation checklist is now cited by the EU AI Office (post).
Opaque or light items (@JohnCLangford via @DimitrisPapail, @rohanpaul_ai on CodeRabbit, @norpadon). John Langford moved from Microsoft Research to Core Automation, an "automated AI lab". CodeRabbit put $1M into free deep scans that found 1,215 high or critical issues in open-source repos. A meme contrasts on-policy distillation in papers with in practice.
Skip: Huang affair threads, Grok Bot prompt listicles, a wristband checkout link, a Starship livestream, and generic "harness guide" hype reposts.