AI research wiki
media-zone · 2026-06-05

Media Zone | 2026-06-05

Media Zone | 2026-06-05

Anthropic's "AI builds itself" report and NVIDIA's Computex blitz split the feed; coding tools and open models fill the gaps.

Today's signal

  • Dominant story: Anthropic's recursive-self-improvement report ("Claude writes 90% of our code") ricochets across the feed.
  • Pattern: NVIDIA shipped agent substrate (Nemotron 3 Ultra, Cosmos 3, RL environments) as papers map self-evolving agents.
  • Counter-signal: @eliebakouch warns Anthropic's 8x-code plot may be confounded by model-version access.
  • Open-weights push: Nemotron 3 Ultra and Step 3.7 Flash both dropped free in Kilo.
  • Quiet area: no curated retweets, all eight Reddit subs empty, no practitioner ground-truth check.

Routing, KV cache, compression, GPU

Open frontier models go free, marketed for agents and long context

  • Nemotron 3 Ultra: 550B total, 55B active, 1M context, "top US open-weights."
  • Step 3.7 Flash: 198B/11B-active, 256k context, 400 tok/sec, near-Sonnet coding.
  • Both free in Kilo "for a limited time"; pitch is open beats vendor lock-in.
  • NVIDIA frames Nemotron 3 Ultra for long-running agents: 5x faster, ~30% cheaper.

@kilocode · @nvidia (Nemotron 3 Ultra) · @eliebakouch (Nemotron Coalition)

NVIDIA AI-cloud buildout on Vera Rubin

  • Worldwide AI-factory expansion: CoreWeave, Nebius, Lambda, Nscale, SK Telecom and more.
  • Pitch: lowest token cost, best throughput per watt for frontier and open models.
  • Lands against SemiAnalysis's space-datacenter economics (4x terrestrial cost today, parity ~2040).

@nvidia (AI cloud) · @ns123abc (SpaceXAI orbital compute)

LLMs, agents, safety

Anthropic's "AI builds itself" lands, and so does the pushback

  • Claim: Claude writes 90%+ of Anthropic's code; engineers ship 8x more per quarter.
  • Preview model beat human researchers' next-step calls 64% of time (22% in 2024).
  • Sholto Douglas: "models will only come to us for hard calls... eventually not at all."
  • @eliebakouch: 4-week-average plot may hide model-version confound; "a bit misleading."

Hermes self-evolving agent desktop app

Anthropic report · @_sholtodouglas · @ns123abc · @eliebakouch

Coding tools converge: memory, semantic indexing, design canvases

  • OpenAI's "Dreaming": background memory that updates user context without save requests.
  • Kilo brings back semantic codebase indexing: "find code you cannot name yet."
  • Cursor ships Canvas Design Mode plus an interactive context-usage report.
  • @eliebakouch: Codex first-UI in ~7 min vs Claude Code ~20 min.

Semantic code retrieval on Claude Code GPT-5.6 leaks and Microsoft MAI-Thinking

@eliebakouch (CC vs Codex) · @cursor_ai · @kilocode (indexing)

Biosecurity moves up the agenda

  • Altman, Amodei, Hassabis sign letter urging Congress to raise biosecurity.
  • Anthropic's Logan Graham: red-teamed bio-risks since 2022; "clear reason for caution" now.
  • Frames the fix as deploying biodefenses, not just restricting models.

@logangraham

Multimodal / vision / audio

Cosmos 3 and ambient health sensing

  • NVIDIA Cosmos 3: open omni-model unifying text, image, video, sound, robot action.
  • Pitch: physical AI "thinks before it acts" via mixture-of-transformers backbone.
  • Google: passive heart-rate from front camera, accurate across all skin tones.

@nvidia (Cosmos 3) · @GoogleResearch

Industry and business

  • Pinterest signs $4B AWS deal through 2031, using Trainium + Graviton for 600M users (@mattsgarman).
  • Prime Intellect joins NVIDIA's Nemotron Coalition with 2,500+ open RL environments (@eliebakouch).
  • Quantinuum IPOs on Nasdaq (QNT), $1.68B raised, 20x oversubscribed, largest quantum listing ever (Quantum Zeitgeist).
  • Kilo argues the free-compute era is over after GitHub Copilot's usage-based billing and the SpaceX-Cursor deal (@kilocode).