Summary
Only the morning slot carried signal; the afternoon and evening captures were empty, which fits the US overnight clock. The strongest cluster is Google's Gemini 4 Neon: GA in GitHub Copilot, #2 on Arena's new Agent Arena Index, and near-Carbon quality at about 2.5x the speed, with a $0.15 cached-input price that matters most for long agent sessions. The sharpest single post is OpenAI's research-acceleration data: about $2,400 of tokens per researcher per day, yet experiments rose only 68% and agents succeed just 25% on inference and serving code, a direct hit on the efficiency stack. Open-weight inference at volume (opencode serving 16T tokens on its own GPUs, vLLM v0.34 with an extensible KV cache on by default) and agents designing their own self-improvement loops round out the real signal. The rest is benchmark launches and promo, mostly skippable.
Posts
- Gemini 4 Neon lands everywhere at once (cluster of 6) (@github · @arena · @GoogleAI · @VaibhavSisinty · @rohanpaul_ai · changelog) [morning]. GA in Copilot, #2 on the Agent Arena Index at 1336, $3/$15 per million tokens with $0.15 cached input. See the Gemini 4 Neon summary.
- OpenAI research-acceleration data (@OpenAI · @rohanpaul_ai · @jeffclune · post) [morning]. $2,400/day median token spend, 11x more agent hours, only 68% more experiments; agents hit 25% on serving code. See the industry summary.
- Gary Marcus pushback (@GaryMarcus) [morning]. Called handing research to AI reckless. Commentary, no new data.
- Open-weight inference at volume (cluster of 4) (@ycombinator · @RedHat_AI · @RedHat_AI · @RedHat_AI) [morning]. opencode served 44% of its tokens (16T) on its own GPUs; vLLM v0.34 ships an extensible KV cache by default. Speculative-decoding office hours pair with the DNeo drafter paper.
- Agents that design their own self-improvement (cluster of 3) (@omarsar0 · @omarsar0 · @ai_database · paper) [morning]. A frozen model with a compute budget invents its own curricula and edits a copy's prompts and skills. See gated self-improving harnesses and weight-space programming.
- Agent Arena Index launch (@arena · @arena) [morning]. Six sub-indexes for agent work. GPT-6.2 Astra enters at #6; Sol joins the cost Pareto frontier.
- Chip designs need fabs (@not_ellington) [morning]. Argues fab, HBM and packaging capacity matter more than OpenAI's chip design file. See models designing silicon.
- SambaNova at TACC (@SambaNovaAI) [morning]. TACC triples its SambaStack deployment to offload inference from the supercomputer.
- Prime Intellect model factories (@vincentweisser · @vincentweisser) [morning]. End-to-end pretrain-to-deploy pipeline pitch. Mostly promo.
- Amazon tops Search Arena (@arena) [morning]. Custom post-trained search agent with AgentCore Web Search takes #1.
- Genie 4 goes live (@GoogleDeepMind · @GoogleAI) [morning]. Native audio and real-time world editing. Light interest.
- FrontierBreach benchmark (@garrytan) [morning]. PromptArmor measures attack success on enterprise agent workflows.
- System-design lists, event registrations, Musk one-liner, leadership essay [morning]. Skip.
- No posts [afternoon + evening]. Both captures empty during US overnight. See the daily digest.