Summary
The strongest signal today is infrastructure for the reinforcement-learning post-training stack. A curated repost surfaced RL-Kernel, a newly open-sourced low-level kernel library built specifically for RLHF training (GRPO and PPO), claiming up to 163x speedups on hot components via FlashInfer, prefix-shared attention, and Hopper TMA copy engines. That pairs with a second curated repost on OpenPipe's ART (Agent Reinforcement Trainer, GRPO for multi-step agents), so the agent-RL tooling layer is filling in fast. The day's other research thread is self-improving and discovery agents: @omarsar0 highlighted Self-Revising Discovery Systems, a category-theoretic account that draws a hard line between retrieval, search, and genuine discovery and gates accepted ideas by description length (25 of 388 proposals accepted). A viral non-curated thread claimed a "Google paper that might end the transformer era," which maps to the recurrent-pretraining work (Kumar and Isola) that Kurate has rated highly for weeks, so it is cross-source confirmed via social. On the labs front, Anthropic's @bcherny posted five tips for running Opus autonomously for hours or days, and @eliebakouch flagged that GPT-5.5's system card shows only modest progress on internal research-debugging (RSI) evals and praised Stanford's fully-open Marin training effort. The loudest industry item was a reposted claim that Google is paying SpaceX about $920M/month to rent 110,000 Nvidia GPUs, a striking read on compute scarcity. A cluster of @Scobleizer posts covered "Hell Grind," an AI-made feature film that cost roughly $400K in compute. Several opaque X long-form reposts could not be fetched.
Posts
RL-Kernel open-sourced: a kernel library for RL post-training (@sheriyuo/@bayesiansapien · repo). The release is a low-level CUDA kernel library aimed squarely at RLHF training of LLMs, with a focus on GRPO (Group Relative Policy Optimization, the lighter RL method now standard for reasoning models) and PPO. It is deeply integrated with FlashInfer and includes custom kernels for sampling, prefix-shared attention (so a group of rollouts sharing a prompt does not recompute the prompt's attention), and TMA (Hopper's asynchronous bulk-copy engine), with CUTLASS and FlyDSL backends planned. The authors report up to 163x speedups on some high-frequency training components alongside lower memory use, and are recruiting collaborators with GPU access. This is the RL-training analogue of the FlashAttention story: inference got a decade of kernel tuning, and the post-training loop is only now getting its own. See wiki summary.
Self-Revising Discovery Systems: separating discovery from confident retrieval (@omarsar0/@bayesiansapien · paper). Called one of the standout papers of the week, it tackles a question most self-improving agents dodge: is the agent actually discovering something, or just remixing what it already knows? It gives three buckets, retrieval (looking up a known fact), search (recombining tools you already own), and discovery (inventing a concept that was not in your toolkit), and argues most agents stop at the first two. Formally, discovery is a verified revision of the representational regime, with old artifacts transported by a left Kan extension and the residual content that transport cannot explain being the real novelty. A description-length gate keeps it honest: one run accepted just 25 of 388 proposals (6.4%). See wiki summary.
A widely shared thread claims a "Google paper that might end the transformer era" (@HowToAI_/@bayesiansapien). The thread argues transformers have a fatal flaw, quadratic attention cost that explodes with prompt length, while old-school RNNs are cheap and fast but forget long documents, and frames a new architecture as getting RNN efficiency without the amnesia. The substance maps onto "Pretraining Recurrent Networks without Recurrence" (Kumar and Isola), which has sat near the top of Kurate's cs.LG board for weeks (ai_rating 7.5) without ever surfacing on HuggingFace's daily top. The viral social amplification on top of the standing Kurate rating makes this cross-source confirmed via social, and the "RNN-speed, transformer-quality" claim is the one to watch for a real open-model test.
OpenPipe's ART and a fine-tuning techniques rundown (@_avichawla/@bayesiansapien · repo). The tweet itself is a bookmark-bait list of 15 fine-tuning techniques (LoRA, QLoRA, prefix tuning, adapter tuning, instruction tuning, and on), but the attached link is ART (Agent Reinforcement Trainer), an open framework that trains multi-step agents for real-world tasks with GRPO across Qwen3.6, GPT-OSS, Llama, and others. Read alongside RL-Kernel above, the agent-RL stack is acquiring both its low-level kernels and its high-level trainer in the same week.
GitHub's Spec Kit: spec-first vibe coding (@rohanpaul_ai/@bayesiansapien). Spec Kit (109K+ stars) targets vibe coding's biggest weakness, the AI starting to code before the product rules are clear. It reorders the workflow: write the product spec first, clarify gaps, produce a technical plan, then have the AI build from that spec, rather than jumping from a loose prompt straight into code that produces working demos with weak requirements and messy rework.
Anthropic's @bcherny: five tips for running Opus autonomously for hours or days (@bcherny). Citing benchmarks showing Opus leads on long-running work, the tips are: use auto permissions mode so Claude does not stop to ask, use dynamic workflows to orchestrate hundreds or thousands of agents, use /goal or /loop to keep it going until done, run Claude Code in the cloud so you can close your laptop, and give Claude a way to self-verify end to end (Chrome extension for web, mobile for app testing). It is a practical companion to the day's self-evolution research, which is the autonomous-improvement question one level down in the stack.
@eliebakouch on GPT-5.5's RSI evals and fully-open training (@eliebakouch · Marin repo). One tweet notes GPT-5.5's system card introduces an "internal research debugging evaluation" (can a model find and fix real bugs in OpenAI research experiments that took researchers hours to days?) and reads the recursive-self-improvement numbers as modest progress over the previous generation. A second praises Stanford's Marin, a fully-open foundation-model effort where development happens in the open on GitHub, and notes that with OLMo's from-scratch series winding down, NVIDIA's Nemotron is becoming one of the few remaining fully-open (not just weight-drop) from-scratch LLM teams.
Reposted claim: Google paying SpaceX ~$920M/month for 110,000 Nvidia GPUs (@TobyPhln). The quoted thread (Aadit Sheth) expresses disbelief that Google, which builds its own TPUs and runs one of the largest clouds on earth, would rent 110,000 Nvidia GPUs from a rocket company, framing it as a sign that frontier compute demand has outrun even Google's own buildout. @TobyPhln uses it to credit xAI's rapid hardware buildout. Either way it is another data point on how tight, and how expensive, frontier GPU supply has become.
@Scobleizer cluster: "Hell Grind," a feature film made mostly with AI (@Scobleizer). Across several posts, Scoble covers the San Francisco premiere of Hell Grind, a 95-minute action-fantasy film made by a 15-person team in about 14 days for roughly $500K, of which about $400K was compute. Higgsfield's CTO introduced it; reactions ranged from "phenomenal technical demo, mediocre movie" to amazement that a competent film can now be made for half a million dollars versus the $100M-plus median. A clear marker of where AI video generation sits: impressive on cost and speed, uneven on quality.
Google Research: D4RT wins CVPR 2026 Best Paper (@GoogleResearch · paper). "Efficiently Reconstructing Dynamic Scenes One D4RT at a Time" took the CVPR Best Paper Award for feedforward 4D (dynamic) scene reconstruction. Outside the core efficiency and routing interests, but a notable marker in the dynamic-scene reconstruction line.
Opaque X long-form reposts (could not fetch the article body) (@sheriyuo, @shmidtqq, @Vtrivedy10, @kilocode). These point to X native long-form articles whose bodies did not fetch (the @sheriyuo one is a post on RL interview questions by Xiuyu Li, per the starred-Gmail context). Click through on X to read.
Skipped: personal, political, and off-topic posts from @brivael, @spencerpratt, @heavypulp (art and politics), Tesla Camp Mode, the @nvidia Build-a-Claw pop-up, the @JasonBud Grok CLI patch note, and the @MillionInt kernel joke, none of which carries research or industry substance.