Summary
The scheduled morning slot captured no curated reposts, no tracked-handle tweets and no new saved posts, so this synthesis reads the late-morning Following-feed capture, which covers the US Wednesday afternoon and evening conversation. The strongest signal is a cluster of 6 posts on Google's Gemini 4 Neon: GitHub made it generally available in Copilot, Arena ranked it #2 on its new Agent Arena Index, and early testers said it feels close to Gemini 4 Carbon at about 2.5x the speed. The most consequential single post is OpenAI's research-acceleration data, amplified by Kevin Roose and Rohan Paul: the median OpenAI researcher uses about $2,400 of tokens a day, yet experiments rose only 68%, and agents succeed just 25% of the time on inference and serving code. A second cluster of 4 posts covered open-weight inference at scale: opencode said 44% of its tokens (16T in a day) ran on its own GPUs, vLLM v0.34 shipped with an extensible KV cache on by default, and Red Hat promoted a session on what moves speculative-decoding acceptance rates. A third cluster of 3 continued the self-improving-agent thread from the MIT/Tübingen paper on agents that design their own learning loops, plus Google's gated harness edits. Standout items: the Japanese-language recap of models rewriting their own weights, the Agent Arena Index launch, and a pointed reply arguing chip designs mean little without fab, HBM and packaging capacity.
Posts
Gemini 4 Neon lands everywhere at once (cluster of 6) (@github, @arena, @GoogleAI, @VaibhavSisinty, @rohanpaul_ai, GitHub changelog). GitHub made Neon GA in Copilot, describing quality close to Gemini 4 Carbon at lower token cost per task, faster responses and fewer unrequested code changes. Arena put Neon (High) at #2 on its new Agent Arena Index with 1336 points from 4,200+ real agent sessions, ahead of Carbon (1303), Hydrogen (1271) and Argon (1229), which makes it Google's best agent model. Rohan Paul's summary: 61 on the Artificial Analysis Intelligence Index, third behind larger and slower models, at $3/$15 per million tokens. An early tester said it matched Carbon at about 2.5x the speed. The detail the threads skipped is the $0.15 cached-input price, a 95% discount, which is what long agent sessions mostly pay. See the Gemini 4 Neon summary.
OpenAI's research-acceleration data (@OpenAI, @rohanpaul_ai, @kevinroose via @jeffclune, post). OpenAI says researchers' AI use keeps rising exponentially and agents complete longer tasks, but humans stay central. The numbers people quoted: the typical researcher burns about $2,400 a day in tokens at API prices, and the heaviest 10% pass $23,000. Agents run as many hours per researcher as 5.7 people on 8-hour shifts, up 11x, but experiments rose only 68% over 2025 and capability gains sit just above trend. For GPT-6.2 Astra, pretraining researchers said Codex helped find key insights, which OpenAI calls AI's first reported foothold in research taste. Agents succeed 79% of the time monitoring training runs but only 25% writing inference and serving code, which could cap the overall speedup. Gary Marcus called handing research to AI reckless. See the industry summary.
Open-weight inference at volume (cluster of 4) (@opencode via @ycombinator, @vllm_project via @RedHat_AI, @RedHat_AI, @alexa_griffith_ via @RedHat_AI). opencode said 44% of its tokens that day, 16T, were served from its own GPUs, with 10x more hardware coming; YC partners reposted it. vLLM v0.34.0 shipped with 1,061 commits from 412 contributors, 10 new models and an extensible KV cache on by default. Red Hat's Speculators maintainers will break down what moves speculative-decoding acceptance rates (algorithm, data, epochs, loss) at vLLM office hours on November 19, which lines up with today's DNeo paper on drafter conditioning. A hands-on course shows how to wire llm-d, KServe, Envoy and vLLM into one inference stack.
Agents that design their own self-improvement (cluster of 3) (@omarsar0, @omarsar0 repost, @ai_database, paper). The MIT CSAIL and Tübingen paper gives a frozen model a compute budget and simple tools: it can write practice games, watch another copy of itself play, and edit that copy's prompts and skill files, with no real data and no imposed loop. Capable models do self-improve and invent strategies resembling self-regulated learning and self-generated curricula, and the text artifacts transfer to other games and 3D navigation. Elvis Saravia reposted his Google gated-harness thread alongside it. A Japanese account summarized the separate "models rewrite their own weights" paper: an open Qwen3.5 model inspected its wrong answers and internal states, then rewrote its own weights and got better. See gated self-improving harnesses and the weight-space programming summary.
Arena launches the Agent Arena Index (@arena, @arena on GPT-6.2). A new headline measure of agent performance with six focused sub-indexes: agentic work, coding, chat, task outcomes, tool use and steering. GPT-6.2 Astra (Max) enters at #6 with 1291 points as OpenAI's best, and GPT-6.2 Sol (Max) at #11 joins the cost Pareto frontier. In Code Arena WebDev, Astra is #3 and Sol #7.
Chip designs need fabs (@not_ellington). A reply to OpenAI's chip-design post argued that even a better tokens-per-watt inference chip than Rubin is worth little without Nvidia-scale fab capacity, HBM allocation and advanced packaging, which is why OpenAI's Broadcom deal matters more than the design file. See models designing silicon.
SambaNova at TACC (@SambaNovaAI). Rodrigo Liang on why TACC is tripling its SambaStack deployment: train on the supercomputer, run inference on SambaNova, and keep simulation capacity free.
Model factories (@jannik_stra via @vincentweisser, @0xsmac via @vincentweisser). Prime Intellect pitched a pipeline covering pretraining, SFT, RL, eval, deployment to production inference and trace capture, for teams that want their own model with their own harness and evals.
Amazon tops Search Arena (@arena). Amazon's custom post-trained web-search agent, paired with AgentCore Web Search, took #1.
Genie 4 goes live (@GoogleDeepMind, @GoogleAI). Project Genie now runs Genie 4 with native audio, real-time text editing of the world and an agent that helps build worlds.
FrontierBreach (@PromptArmor via @garrytan). PromptArmor released a benchmark for attack success rates on enterprise agent workflows.
Skip. AI system-design concept lists, PyTorch Day Japan and Open Nexus China registration posts, an Elon Musk one-liner, and a leadership essay with no AI substance.