Summary
The morning slot captured no reposts and no tweets from tracked handles, so this synthesis reads the two Following-feed captures that cover the US Thursday. The strongest signal is a cluster of about 12 posts on Jensen Huang's resignation as Nvidia CEO: Nvidia announced the transition, the feed argued over whether the affair or the GPU smuggling around Megaspeed was the real cause, and the Commerce rule against Chinese firms renting Blackwell chips abroad kept circulating. The second cluster, 4 posts, is Google's Gemini 4 Hydrogen, a fast and cheap model that Agent Arena placed #5 overall at $1.6 per task. A third cluster of 5 posts is agent memory and self-improving harnesses: SoMem (agents organizing memory as linked documents), ReHarness (outcomes edit the notes that caused them), JoMoHa (RL trains the harness writer and the task solver together), and Trustworthy RSI (statistical tests before keeping a harness change). The most interesting systems item is a paper claiming the first recursive self-improvement loop over a full frontier pre-training run, which produced an open 30B mixture-of-experts base model. PyTorch posted three efficiency write-ups: Helion kernels that skip autotuning and still beat handwritten Triton and CUDA, a 1T-parameter MoE trained on Europe's public JUPITER supercomputer, and SGLang brought up on IBM's Spyre accelerator in under a week. Anthropic's IPO post was the single most-engaged post of the window. A small-account chart claimed no open model released since August uses full attention in every layer.
Posts
Jensen Huang resigns as Nvidia CEO (cluster of 12) (@nvidianewsroom via nvidia, @edzitron, @edzitron on Supermicro, @edzitron on Moonshot, @lexfridman, @rohanpaul_ai on DeepSeek funding, @rohanpaul_ai on the rental rule, @techreview, @GaryMarcus, @rasbt). Nvidia announced a CEO transition after a week of reporting on Huang's affair with Alice Huang, whose companies are alleged to be a front for buying Nvidia GPUs for China. The feed split two ways. Lex Fridman called it "Apple firing Steve Jobs all over again." Ed Zitron's position, repeated across many posts, is that the affair is the cover story and the real issue is a GPU smuggling ring "larger than China": Supermicro co-founder Wally Liaw allegedly sold GPUs to a Megaspeed subsidiary, Moonshot AI trained on B300s inside Supermicro servers (per Bloomberg), and DeepSeek was reported training on Blackwell back in December 2025. Rohan Paul relays The Information's report that companies tied to Alice Huang invested in DeepSeek's latest round to become its chip supplier, and that Commerce's new rule would stop Chinese AI firms renting Blackwell and newer chips. MIT Technology Review notes the Chip Security Act (location verification for exported chips) is getting renewed attention. Sebastian Raschka's worry is narrower and practical: he hopes Nvidia's Nemotron line continues, because its open data and full training recipes are much harder to replace than open weights. Background: Nvidia's CEO crisis and the remote-access rule.
Gemini 4 Hydrogen: the fast, cheap tier (cluster of 4) (@demishassabis, @GoogleDeepMind, @arena). Google added Hydrogen to the Gemini 4 family, below Carbon (the hardest-problems model). DeepMind says it finishes tasks 2.5x faster than Gemini 3.8 Flash and beats Gemini 4 Argon on the Artificial Analysis Intelligence Index. Agent Arena's numbers across 6.5K real agent sessions: #5 overall at +14.8% net improvement, the fastest model in the top 10 at a median 5.3 minutes per task (3x faster than #1 Claude Fable 5.2), and $1.6 per task, 3.7x cheaper than GPT-6 Astra. API price is $1.5 / $7.5 per million input / output tokens. The catch: it spends 41% more output tokens per task than Carbon, so its cheapness comes from price per token, not from being terse. It ranked #2 on steerability.
Recursive self-improvement on a whole pre-training run (@dair_ai, paper). "No Second Run" (GAIR at SII with Xinxu AI) asks whether RSI (recursive self-improvement, AI improving the process that builds AI) can reach pre-training, where a run costs hundreds of thousands of GPU-hours, happens once, and is judged by how much headroom it leaves for post-training. Their answer is to split it into small loops that can close: agents debug infrastructure and data, check headroom with tiny SFT (supervised fine-tuning) runs, and turn expert judgment into data-pipeline metrics, with a shared context store that keeps negative results. The output, 801-30B-A3B (an open 30B MoE with 3B active parameters), beats Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B as a base model. With only 6B tokens of SFT and no RL it posted the highest average on four agentic benchmarks against fully post-trained models.
Agent memory that rewrites itself (cluster of 2) (@omarsar0 on SoMem, paper, @rohanpaul_ai on ReHarness). SoMem (City University of Hong Kong and Tencent) drops the fixed memory schema: the agent stores memory as ordinary linked documents, decides what to summarize, which topics to create and when to revise them, and every summary links back to the original interaction as evidence. It beat the strongest baseline by 2.66 points on LoCoMo, 3.10 on BEAM-100K and 2.00 on LongMemEval-S, and memory written by one model worked when read by another. ReHarness (MIT and Meta) attacks stale notes: after each task, a revision agent narrows or corrects the notes that were retrieved, using only facts from each note's original task. It averaged 0.240 normalized gain versus 0.168 for the best raw-trajectory baseline, and replaying full raw history cost 3.4x the tokens. Concept page: agent memory.
Training the harness, and testing harness changes (cluster of 2) (@omarsar0 on JoMoHa, paper, @dair_ai on Trustworthy RSI, paper). JoMoHa (Salesforce AI Research and HKU) trains two policies with RL from the same task outcomes: one writes the executable harness (the code around the model that manages tools, context and control flow), one solves tasks inside it. Naive turn-taking made them drift apart and performance rose then fell; separate histories and value estimates per policy fixed it. It averaged 70.3% across ALFWorld, WebShop and DBBench, with 97.0% on ALFWorld, and a trained harness writer beat a frozen one. Trustworthy RSI targets the opposite failure: self-improving agents usually keep any harness change that raises the validation score, so noise gets accepted and costs performance later; the paper measures how often that happens and adds statistical acceptance tests. Concept page: agent harness engineering.
Helion kernels that skip autotuning (@PyTorch, blog). Helion is PyTorch's DSL (a small language) for portable ML kernels, and autotuning was its bottleneck: one team spent 10 hours tuning a single paged-attention kernel across 72 shapes. The new heuristics read the kernel's IR and dataflow at compile time, classify it as matmul, reduction or pointwise, and pick block sizes, warps and pipeline stages directly, at a cost of tens of milliseconds. On H100, across nine linear-attention variants the computed config beats Flash Linear Attention's handwritten Triton by 1.18x forward (1.12x with backward), and across six production vLLM reduction kernels it is 1.27x faster than vLLM's handwritten CUDA/C++.
A 1T-parameter MoE on Europe's shared supercomputer (@PyTorch). The EUROPA consortium (Domyn, Fraunhofer IAIS and IIS) scaled pre-training of a 951.9B-parameter MoE (32.8B active, 384 routed experts, top-8, 64K sequence length, Muon plus AdamW) on JUPITER, the EuroHPC exascale machine shared by hundreds of projects. Per-GPU throughput rose from 90 to about 1,300 tokens per second across 384 GH200 GPUs per data-parallel replica, a 14x gain from fine-grained TorchTitan tuning on hardware that was not designed for AI.
SGLang on IBM Spyre in under a week (@PyTorch). Two IBM engineers with AI coding agents wrote a roughly 1,500-line plugin plus four small SGLang patches. It delivered about 3.6x the throughput of the reference torch-spyre implementation, matched vLLM on Granite 3.3 8B and Gemma 4 26B A4B, and RadixCache prefix sharing worked unchanged. Same theme as yesterday's Mythos-on-AMD claim: agents shrink the cost of bringing up new silicon.
Full attention is disappearing from open models (@crazyMLguy). A chart claims 59% of 2025 open releases used full attention in every layer, 27% from January to August 2026, and 0% from August to November, with 48% using no full-attention layers at all (Mistral Small 4, Kimi K3, MiniMax M3.1 Flash). Berkeley AI amplified a related thread noting Qwen, Kimi and GLM all bet on linear attention but its recurrent state is still slow to read and rewrite every token, and Ion Stoica relayed attention matching the FP32 baseline at 8-bit and 4-bit (@berkeley_ai RTs). Small account, unsourced dataset; directionally consistent with the wiki's hybrid-attention thread.
Agents as meta-optimizers (cluster of 2) (@DimitrisPapail rant, @DimitrisPapail on quants). Papailiopoulos argues that a frontier agent left to build something bespoke beats researchers' favourite inference stacks, optimizers and models, because it does not care about a clean story. A practical corollary: since agents are now good at quantization and custom kernels, quantize locally from official weights, because any third-party quant changes every weight and breaks provenance.
Six serving techniques that remove six kinds of waste (@techNmak). A clear explainer: prefix caching reuses shared-prefix KV states, continuous batching swaps finished requests out between iterations, PagedAttention stores the KV cache in blocks to stop fragmentation, chunked prefill interleaves long prompts with decoding, speculative decoding drafts tokens cheaply and verifies them, and prefill/decode disaggregation runs the two phases on separate hardware.
CoT monitorability gets measured (@rasbt). Raschka summarizes a thread on what makes chain-of-thought harder to monitor. Better pre-training (lower perplexity) raises how much a model can reason in one forward pass without writing anything down, and architecture mostly does not matter. RL alone raises CoT controllability (how well a model can reshape its written reasoning on request), and that controllability passes through distillation. Both make monitoring harder; process supervision was not the main culprit.
Cerebras adds Microsoft (@StockSavvyShay). Cerebras confirmed Azure as a hyperscaler inference customer alongside AWS, Oracle and OpenAI; Q3 revenue was $231M with cloud services up 265% year on year, RPO at $26.1B, and FY26 guidance raised to $950M.
Funding and launches (@luminal_ai via YC, @arena on Inkling 2 Small). Luminal raised a $25M Series A led by Standard Capital for heterogeneous inference. Thinking Machines released Inkling 2 Small's weights, now the #1 US open model in Agent Arena. Agent-infra launches (InstaBox, Simular's Box, Omnigraph, a branch-and-merge graph database for multi-agent writes) were the day's volume but mostly promotion.
Skip: the Kimberly Powell succession thread and assorted Dario/Jensen gossip posts are unsourced speculation; stock "falling knife" posts and growth-hack threads are engagement bait.