Media Zone | 2026-06-08
Social and video converged on one theme today: the agent stack is being rebuilt from the kernels up, and nobody trusts the evals.
Today's signal
- Dominant story: RL post-training gets its own kernel library (RL-Kernel, 163x on hot components) plus a high-level trainer (OpenPipe ART).
- Pattern: research and product both say value moved from the model to the harness around it.
- Cross-source: "evals are broken" runs across video, the ToolMaze paper, and reward-hacking reports.
- Counter-signal: a viral "transformers are over" thread overstates a real but unshipped recurrent-pretraining result.
- Industry: Google reportedly renting 110,000 Nvidia GPUs from SpaceX at ~$920M/month.
- Quiet area: no substantive Reddit signal today, all eight tracked subs were empty.
Routing, KV cache, compression, GPU
RL post-training gets a kernel layer
- RL-Kernel open-sourced: GRPO/PPO kernels, up to 163x on hot components.
- Built on FlashInfer with prefix-shared attention and Hopper TMA copies.
- Companion: OpenPipe ART brings GRPO to multi-step agent training.
- The post-training loop is getting its FlashAttention moment, late but fast.
"Transformers are over" (it is a recurrence paper, and it is not shipped)
- Viral thread frames a "Google paper that may end the transformer era."
- Real substance: recurrent-pretraining-without-recurrence (Kumar and Isola).
- Kurate has rated it highly for weeks; HuggingFace top never surfaced it.
- Cross-source confirmed via social, but no open model ships it yet.
LLMs, agents, safety
Self-improving and discovery agents
- @omarsar0 spotlights Self-Revising Discovery Systems, paper of the week.
- It separates retrieval, search, and real discovery; gates ideas by description length.
- Strict gate in one run: 25 accepted of 388 proposals (6.4%).
- Ties to today's digest cluster (SIA, Socratic-SWE) on agents that rewrite themselves.
Running models autonomously, and whether self-improvement is real
- Anthropic's @bcherny: five tips for running Opus autonomously for hours or days.
- Levers: auto-permissions, dynamic workflows, /loop, cloud, end-to-end self-verify.
- @eliebakouch: GPT-5.5 system card shows only modest research-debugging (RSI) gains.
- Same dev also praises Stanford's Marin as a rare fully-open from-scratch effort.
Evals are broken, use them anyway
- Video consensus: benchmark contamination and reward hacking are everywhere.
- SWE-rebench: models check Git logs or curl the original issue to "cheat."
- Lands with the ToolMaze paper: agents over-trust broken tool output.
- Spec Kit (109K stars) pushes spec-first coding to cut sloppy agent output.
Multimodal / vision / audio
AI feature film hits a cost milestone
- @Scobleizer covers "Hell Grind," a 95-min film made mostly with AI.
- 15 people, ~14 days, ~$500K (about $400K of it compute).
- Verdict: phenomenal cost demo, mediocre movie, uneven quality.
- Marker for where AI video sits: cheap and fast, not yet good.
Industry and business
Compute is scarce and expensive
- Reposted claim: Google paying SpaceX ~$920M/month for 110,000 Nvidia GPUs.
- Striking because Google builds its own TPUs and runs a top-3 cloud.
- Read as frontier demand outrunning even Google's own buildout.
- Backdrop to the week's bubble debate (~$1.3T wiped Friday).