Summary
All of today's signal came in the morning slot. The afternoon and evening slots were empty, and no curated reposts or saved posts arrived all day, so this is the Following feed's take on the late US Wednesday. The strongest cluster is Gemini 4 Carbon: Google's new frontier model took #1 in Text Arena and #2 in Agent Arena, and it stays on the cost-per-task Pareto frontier, so the scores are not bought with extra tokens. The second cluster is the economics of compute on earnings night. Google Cloud doubled, Deutsche Bank estimates GB300 rents near $14 an hour, and Jensen Huang argued that faster kernels are pure revenue while demand outruns capacity. The best single item for routing is GitHub's Copilot Fusion, which scores close to Claude Code at about half the cost per task. DeepSeek's SSD-backed KV cache and a TikTok paper on MCP agents that leak data after finishing their task are also worth reading. The "100 sub-agents" and "nobody will prompt in 2 months" posts were pure noise.
Posts
- Gemini 4 Carbon tops the arenas (cluster of 6) (@arena · @arena · @ChrSzegedy · @natolambert · @StockSavvyShay · blog) [morning]. It is #1 in Text Arena, #2 in Agent Arena and #4 in WebDev, and on the Pareto frontier for cost, time and tokens per task. Pricing is $2/$10 per million tokens. Summary.
- Earnings-night compute economics (cluster of 4) (@rohanpaul_ai · @StockSavvyShay · @StockSavvyShay · @rohanpaul_ai) [morning]. Google Cloud grew 103% to $30.7B, and Deutsche Bank puts GB300 rents near $14/hr against $3.50 for H100. Huang: doubling kernel speed doubles revenue while demand exceeds capacity. Summary.
- GitHub Copilot Fusion, a cross-lab router (@rohanpaul_ai · @github) [morning]. It routes between Claude and GPT and scores 60 to Claude Code's 62, at $5.59 vs $11.42 per task. Summary.
- KV cache leaves the GPU (cluster of 2) (@_avichawla · @TheTuringPost) [morning]. DeepSeek's DeepInfer streams compressed MLA (multi-head latent attention) caches from SSD, so a cached 128K prompt starts in 500 ms instead of 13 s. AKV lets agents drop stale cache entries, cutting cost 27.7%. Summary.
- Opus 5.5 beats Astra at GPU kernels (@DimitrisPapail) [morning]. A strong kernel-writing result, though the same researcher hit cyber-classifier blocks on SASS (GPU assembly) work on 10-28.
- Log-To-Leak: MCP agents leak after finishing (@rohanpaul_ai) [morning]. Prompt-injected agents finish the task, then quietly call an attacker's logging tool, so outcome-based evals miss the leak. Summary.
- Chinese labs now distill from GPT-6 Astra (@rohanpaul_ai · The Information) [morning]. Anthropic's tighter limits and bans cut what distillers harvest from Claude by 90-99%.
- Apple SynthAPI: data mix beats task count (@rohanpaul_ai) [morning]. At a fixed task budget, adding more app diversity lifted a 4B model on AppWorld from 38.4% to 45.0%.
- EMMA re-shared (@omarsar0 · paper) [morning]. After each failure, it edits only the memory, prompt or skill responsible. Up to 32.5% over GEPA with fewer runs. Summary.
- Rubrics as Rewards (@berkeley_ai) [morning]. An LLM judge scores answers against written rubrics, giving RL rewards on open-ended tasks.
- Inference providers on OpenRouter (cluster of 2) (@vincentweisser · @ycombinator) [morning]. Prime Inference reached 120B tokens a day in its first month, and particle-1 launched there too.
- Industry quick hits (@rohanpaul_ai · @ycombinator · @ycombinator · @SakanaAILabs) [morning]. ServiceNow fell 19% after hours, and Broccoli raised $52M for trade-business phone agents. Exa indexed 3M podcasts, and Sakana opened its API.
- Muse Spark Open on two Mac Studios (@AICTechPro) [morning]. About 1 TB of unified memory across Thunderbolt 5 RDMA. No benchmarks yet.
- Engagement bait (@marfinxx · @v3lkera · @YesKhush_5) [morning]. Skip.
Full context: daily digest.