Summary
The scheduled morning slot captured no curated reposts, no tracked-handle tweets and no new saved posts. This synthesis reads the morning Following-feed capture, which covers the late US Thursday conversation. The strongest signal is a cluster of 5 posts on self-improving agents that only keep edits they can verify: Google's GUARD, Amazon's stale-skills paper, Safe Personal RSI, and Google's gated-harness paper. All of them say the judge matters more than the proposer. The top paper post of the night was QDA from MIT and NVIDIA, where an agent designs quantization recipes and beats expert ones. It is also on HuggingFace and is a DAIR pick. A cluster of 3 posts covered infrastructure that cuts idle GPU time: NVIDIA Dynamo for RL rollouts, OpenAI's Performance Factory, and Positron's DRAM-based servers at Parasail. Anthropic's IPO letter from Dario Amodei led the feed by reach, and token-spend numbers from OpenAI ($23,000 a day for its 90th-percentile researcher) kept the cost conversation going.
Posts
QDA: agents design quantization recipes (@omarsar0, paper). Post-training quantization recipes differ for every model and bit width, so experts rebuild them by hand. QDA gives that loop to an agent. Each candidate is graded in minutes by KL divergence from the original model on reasoning traces the agent never sees, and only promising ones get long generations to catch repetition loops. It beats the strongest public recipes by 0.9, 1.4 and 2.2 points at W4A16, W4A8 and W3A16 using 64 GPU-hours, and recovers 98.3% of FP8 accuracy at NVFP4 W4A4 on the 320B GLM-5.3-Flash. See the QDA summary.
Self-improvement that knows what it broke (cluster of 5) (@omarsar0 on GUARD, @rohanpaul_ai on stale skills, @dair_ai on Safe Personal RSI, @rohanpaul_ai on gated harness edits, @rohanpaul_ai paper link). GUARD (Google, 2611.23476) accepts a harness edit only if its gains outweigh the behaviors it newly breaks, and beats Meta-Harness by 6.3 points at 54.5% fewer tokens. Amazon's stale-skills paper (2611.22525) shows a saved skill worth +4.7 points for Opus fell 11.4 points below no skill after an app changed, and selective repair lifted Haiku's tax-form accuracy from 30.5% to 97.5%. Safe Personal RSI (2611.23651) tests each self-proposed instruction update for a health assistant and rejected 20 of 21 risky ones. Google's gated-harness paper (2611.22299) found an unchecked full rewrite matched accuracy but cost nearly 10x the tokens of one-module edits. Read with the Resolution Hypothesis summary, which shows selection in agent evolution is often no better than random.
Infrastructure that cuts idle GPU time (cluster of 3) (@NVIDIAAI, @rohanpaul_ai on OpenAI, @rohanpaul_ai on Positron). NVIDIA Dynamo routes RL rollouts to workers that already cache the prompt and ships weight updates as deltas from GPU memory; CoreWeave reports up to 15x faster weight updates. OpenAI's Performance Factory turns ChatGPT latency alerts into Codex investigations: about 17 a day at 70% accuracy, but only about 15 of 600+ agent fixes shipped unchanged. Parasail will deploy Positron Titan servers at 5MW, which hold models up to 32T parameters in commodity DRAM instead of HBM. See training and serving plumbing.
Step-level model routing (@omarsar0). A short post arguing routing deserves more attention in agent development, pointing at Drex Router, a small model that picks which model to use at each step. It lines up with AnyS1 on HuggingFace the same day, which cut agent episode cost 75-91% with a 4B decision model. See the AnyS1 summary.
Anthropic's IPO letter (@AnthropicAI). Dario Amodei framed the listing around the company's founding purpose and his background in biology, the hope that AI compresses medical progress. The feed's highest-reach post of the night; no technical content.
Token spend keeps doubling (@AndrewCurran_ via @garrytan). The 90th-percentile OpenAI researcher now uses about $23,000 of tokens a day, up 105% since September. Context: yesterday OpenAI said its median researcher used about $2,400 a day.
Hopper prices go up, not down (@StockSavvyShay). A thread to Michael Burry asks why Hopper GPUs are recontracting at up to 2x their original prices while token prices fall. Read it as a sign that capacity, not chip generation, is the binding constraint.
Sakana AI Security Research Group (@hardmaru, blog). An independent, public-interest R&D team in Tokyo, run apart from Sakana's business, that treats security as an upstream quality-engineering problem and promises public benchmarks and guidelines.
Small-model and open-model notes (@evilmathkid via @fchollet, @OVHcloud via @huggingface, @googlegemma via @huggingface). New ARC-AGI public-eval records and 70% on ARC-1 with a 19M-parameter model; OVH released its first LLM, Olala mini 7A1B; Airbnb benchmarked Gemma 4 E2B for production use.
Policy (@Rep_Whitesides via @Miles_Brundage, @XavierBecerra via @Miles_Brundage). A bipartisan AI Safety and Innovation Caucus launched in Congress, and California's governor-elect says he will strengthen the state's AI safety laws.
Skip. Higgsfield GPU-rental promo, an IT-help-desk agent stunt claim, Karpathy "graphs" engagement bait, video-editing repo lists and course recommendations.