AI research wiki
social-stream · 2026-05-23

2026-05-23

Summary

The day's dominant cluster is morning: three curated retweets converge on the same architectural argument that the agent's environment (harness, memory, skill registry) is now the bottleneck rather than the model. Code-as-Agent-Harness (@burkov), Harness Engineering (@_vmlops), and ReasoningBank (@DanKornas) all argue that next-gen agent capability comes from structured, persistent, executable environments around the LLM. The strongest non-cluster signals are the SKILL.md supply-chain attack paper from @FeiziSoheil, which shows adversarial natural-language metadata can manipulate skill marketplaces at 86% pairwise win rate, and the GRAM 10M-parameter generative recursive reasoning paper (@rohanpaul_ai) that beats deterministic recursive AI three times its size. On industry economics, two morning reposts amplify the Microsoft-cancels-Claude-Code thread and an evening @ns123abc note ties the RAM oligopoly to Anthropic's reported billion-a-month SpaceX compute bill, which together form the day's clearest signal that token and compute economics are now the binding enterprise-AI constraint. Anthropic's Project Glasswing crossing 10,000 critical vulnerabilities in one month is the strongest standalone signal that offensive AI security capability has outrun defensive deployment. Evening was thin on original AI signal beyond the RAM note.

Posts

  • Code as Agent Harness, Harness Engineering, ReasoningBank (cluster of 3) (@burkov · arxiv 2605.18747, @_vmlops · learn-harness-engineering, @DanKornas) [morning]. Three reposts in one slot converge on the harness thesis: code is the persistent executable working material of agents, Anthropic's controlled Opus 4.5 experiment shows the harness is what turned an unusable run into a playable game, and Google's ReasoningBank stores successful and failed reasoning trajectories as test-time scaling memory. Wiki page: Code as Agent Harness.
  • Demis Hassabis on World Models as the limit of text-trained LLMs (@rohanpaul_ai) [morning]. Hassabis argues text quietly carries physics, psychology, plans, and cause-and-effect, but is still a compressed residue of experience that cannot encode grip, friction, timing, or motor correction. Frames the gap that the harness cluster operationalises.
  • SKILL.md as a semantic supply-chain attack surface (@FeiziSoheil · arxiv 2605.11418 · code) [morning]. Studies SKILL.md-only attacks across Discovery (86% pairwise win rate, 80% Top-10 placement), Selection (77.6% adversarial choice in paired trials), and Governance (semantic evasion bypasses blocking in 36.5-100% of cases). Wiki page: SKILL.md supply-chain attacks.
  • GRAM: 10M-parameter generative recursive reasoning beats deterministic models three times its size (@rohanpaul_ai) [morning]. Lets the model think in parallel reasoning paths simultaneously instead of following a single deterministic train of thought, breaking out of wrong trajectories without scaling parameters. Strong scaling-counterexample worth tracking.
  • Reward hacking as a deeper scaling-RL problem (@vincentweisser) [morning]. Cites @jessicafeiyali arguing the "vague reward function" framing is partially true but offers no remediation, and that the field lacks systematic methods to understand when hacks emerge in semi-verifiable domains. Pairs with today's RLVR credit-assignment work in the daily digest.
  • Enterprise AI's ridiculously expensive amnesia problem (@ashwingop) [morning]. Agents wake up, reread the company, burn tokens rebuilding context, hand judgment back to humans. Article is gated behind an x.com/i/article URL so substance beyond the headline is not captured, but the headline is the sharpest enterprise-side framing of the harness thesis.
  • Microsoft canceled internal Claude Code licenses (@kimmonismus, original @HedgieMarkets) [morning]. Microsoft, the entity that put $13B into OpenAI and runs Anthropic's cloud, cancelled internal Claude Code because token billing made cost untenable. Uber's CTO reportedly burned the entire 2026 AI budget in four months, GitHub is moving off flat-rate plans, American AI software prices jumped 20-37%. Two reposts in one slot is a high-conviction signal that the AI subsidy era is ending.
  • RAM oligopoly and Anthropic's compute bill (@ns123abc) [evening]. Three firms (Samsung, SK Hynix, Micron) make 90%+ of global DRAM, CXMT is under 1% even if it reaches export markets, and Anthropic reportedly pays SpaceX a billion a month for compute. Memory-market frame that pairs with the Microsoft repost on token economics as binding constraint.
  • Anthropic Project Glasswing finds 10,000+ critical vulnerabilities in one month (@ns123abc) [morning]. Claude Mythos found 2,000 bugs in Cloudflare (400 high/critical), 271 vulnerabilities in Firefox 150 (10x Firefox 148), solved both UK AISI cyber attack simulations end-to-end, prevented a $1.5M fraudulent wire transfer at a partner bank, found certificate-forgery in wolfSSL, scanned 1,000+ open-source projects at 90.6% true-positive rate. Anthropic is warning no company has safeguards strong enough to prevent misuse.
  • OpenAI staffers used to code with Claude until Anthropic revoked access (@ns123abc, original @gdb) [morning]. Greg Brockman: "trying to remember what it was like to code before codex". Lab dogfooding and competitive-move pattern.
  • Karpathy adds $50B to Anthropic's IPO valuation (@ns123abc) [morning]. Single-line punchline on the Karpathy podcast's effect on the IPO narrative.
  • DeepSeek 75% discount made permanent (@ns123abc) [morning]. V4-Pro now permanently at the discount; covered as a Deep Dive in the daily digest via The Decoder.
  • Cursor SDK with Composer 2.5 (@cursor_ai, amplified by @sualehasif996) [morning]. Build-your-own-agent SDK in Python and TypeScript, Composer usage 90% off for the long weekend. Direct product move into the agent-harness space the morning cluster is theorising.
  • Claude Code auto mode on Pro plan with Sonnet 4.6 (@ClaudeDevs · permission modes doc) [morning]. Shift-Tab enables auto mode; classifier and boundary states are documented.
  • NVIDIA GTC Taipei kickoff (@nvidia) [evening]. Jensen opens GTC week at a Taipei dev event with an autonomous-agents framing. Watch for the keynotes.
  • Dwarkesh and @reinerpope chip-architecture blackboard lecture (@_sholtodouglas) [morning]. Logic gates through GPUs, TPUs, FPGAs and brain. Sholto Douglas (Anthropic) suggests "Dwarkesh University" as the next move. Useful chip-design fundamentals pointer.
  • NVIDIA on How I Built This (@nvidia) [morning]. Jensen with Guy Raz on near-collapse moments, a decade of unprofitable CUDA investment, AI-job-loss arguments. Narrative content.
  • NVIDIA + Zalando + Allsides scan-to-storefront (@nvidia) [morning]. 48-hour generative-3D workflow using Cosmos, Gen-3C, DiffusionRenderer, with 3-4% add-to-cart uplift and footwear return reductions. Outside core wiki topics but a proof of generative-3D production deployment.
  • zero2claude free course endorsed by @bcherny (@bcherny) [morning]. 17,000+ students, 7 languages, terminal-zero to shipping with Claude Code. Anthropic endorsing community-driven onboarding.
  • La Fontaine fables on labs, hyperscalers, and VCs (cluster of 3) (@brivael) [evening]. Three Le Corbeau et le Renard pastiches with AI labs as the crow holding the GPT and hyperscalers as the fox holding the GPU and datacenter. Cultural commentary, clean rhetorical capture of the lab vs cloud-provider power asymmetry.
  • Opaque x.com/i/article reposts (cluster of 2) (@AlphaSignalAI, @ashwingop) [morning]. X long-form articles the farmer's session cookies did not unlock. Click through to read.
  • Tesla FSD "you don't have to drive anymore" (@Tesla) [morning]. Skip.
  • @Scobleizer SF startup festival coverage (@Scobleizer) [morning]. 120 AI companies, 2,000 attendees at @fdotinc's event. Ecosystem moment, not research signal.
  • @brivael 19-tweet personal feed (@brivael) [morning]. French personal Paris chat, nightclubs, geopolitical jokes, SpaceX banter. Skip.
  • @AustinJustice Austin crime commentary (@AustinJustice) [morning]. Skip.
  • Off-topic political and banter cluster (@brivael, @WHFraudTF, @Scobleizer) [evening]. French politics, anti-Zuckerberg rant, "nuke Mars" banter, taxpayer-fraud task force press, popcorn-guy commentary. Skip.