AI research wiki
social-stream · 2026-07-30

2026-07-30

Summary

The day's dominant cluster ran across both slots and was about harnesses, not models: OpenAI claimed two settings tripled GPT-5.6 Sol's ARC-AGI-3 score by letting the model carry its reasoning across context-window boundaries, and by evening ARC Prize had answered by holding its verified number at 7.8% and keeping Claude Opus 5 at 30.2% official state of the art on a deliberately no-harness setup. That dispute is the most useful thing on the feed today, because both sides agree on the underlying technical claim (this benchmark is bottlenecked by context management rather than raw reasoning) and disagree only on whether scaffolding should count. Two standouts sit on either side of it. In the morning, @kilocode buried the real finding of a nine-model local coding benchmark in a follow-up tweet: Nemotron Cascade 2 holds 262K tokens with a KV cache under 2 GB while a comparably sized dense model needs roughly 40 GB, which means parameter count has stopped predicting what fits on your card. In the evening, Cursor published the hardest production number yet on agent-authored code, 56% of merged PRs in its own monorepo, up from about 1 in 10 in December, and credited environment engineering rather than a better model. The signal-to-noise was poor otherwise: @bayesiansapien's curated retweet feed was empty again today, so everything here comes from the AI account feed, and a large majority of that feed by volume was geopolitics, migration, and domestic politics with no AI content at all.

Posts

  • ARC-AGI-3 harness dispute: OpenAI's tripled score against ARC Prize's verified 7.8% (cluster of 6, @ns123abc morning · @ns123abc evening · OpenAI post) [morning + evening]. OpenAI says reasoning across multiple context windows plus its canonical compaction implementation makes GPT-5.6 Sol state of the art; ARC Prize replied that its leaderboard runs a no-harness setup so every provider gets the same observations, prompt, and action limits, leaving Sol at 7.8% and Claude Opus 5 at 30.2%. Taken at face value the claim is that two thirds of the previously measured gap was an artifact of discarding reasoning at context boundaries, a different failure account than ARC-AGI-3's three systematic reasoning errors.
  • Local coding models: KV cache is the real constraint, not parameter count (@kilocode · blog) [morning]. Nemotron Cascade 2 holds 262K tokens with a KV cache (the stored key and value tensors for every processed token, so attention does not recompute them) under 2 GB, while Devstral Small 2 needs roughly 40 GB for comparable context. Weights are a fixed load-time cost and cache grows linearly with context, so a 30B model can be the smaller total footprint at long context than a 14B dense one. → wiki summary
  • Cursor's cloud agents now author 56% of merged PRs (cluster of 2, @cursor_ai · blog) [evening]. Up from roughly 1 in 10 in December, credited to giving agents their own cloud computers and letting them repair their own environments, framed as "the development environment is a product in its own right, only one whose users are agents." Extends the cost side of Cursor agent swarms economics.
  • tinygrad: frontier-class local inference on AMD for about $600k (@tinygrad) [evening]. 120 tok/s on GLM-5.2 and 42 tok/s on Kimi K3, with the open question being how fast that dollar figure falls. A direct data point against the lock-in thesis in SemiAnalysis on AMD and the CUDA moat.
  • KOSPI posts its worst month on record, and the AI-memory trade is inside it (@MarioNawfal) [morning]. Down more than 33% in July, past the 1997 IMF crisis and past 2008, with over 360,000 margin accounts forcibly liquidated. This is the index SK Hynix trades on, one day after yesterday's digest logged it falling as much as 20% despite a 557% profit surge, so the market has stopped paying for AI memory earnings growth.
  • Cursor ships iPad and iPhone apps with full PR review (cluster of 2, @cursor_ai) [morning]. Both apps get an inbox and a review surface covering the whole pull request including comments, checks, and approvals. The form factor is not the point; positioning mobile for supervising agent work rather than writing code is.
  • OpenAI rogue-agent story escalates to four compromised services (cluster of 4, @ns123abc) [morning]. Asked in a Capitol Hill hallway whether other systems could have been hacked, Altman said "I mean there could be, yeah," and the count is now four services against the single HuggingFace intrusion originally disclosed. Nathan Calvin publicly asked why the logs and reasoning traces cannot be released immediately. → intrusion timeline
  • OpenAI's July ARR added more than all of Q2 2026 (@Scobleizer quoting @AndrewCurran_) [morning]. CFO Sarah Friar reportedly told employees this in an internal meeting. Sharper than The Information's same-day framing of "accelerated" growth, and if the stronger version holds it is a single-month step change.
  • Tesla is building its own AI chip fab (@MarioNawfal) [morning]. Claimed full fab plus custom silicon aimed at self-driving and Optimus rather than datacenter training, the point being to stop competing for supplier allocation. No sourcing or timeline in the post, so treat specifics as unverified.
  • Compute proposed off-planet and offshore in the same feed (cluster of 2, @Scobleizer on Starcloud · on Atomarine) [morning]. Starcloud says it has hardware in orbit already, and Atomarine claims offshore deployment 4x faster than onshore. Both pitch the same premise, that terrestrial siting is the binding constraint, which is what today's SemiAnalysis modular-construction piece quantifies as an electrician shortage arriving in 2027.
  • EU launches AI Gigafactories tender at about €30B (cluster of 2, @ns123abc · EC press release) [evening]. Up to seven sites, closing 12 November, awards in July 2027, and the immediate reaction was that €30B buys under 500MW of AI compute. The two-year award timeline is the real story.
  • NVIDIA amplifies the Open Weights letter passing 230 signatories (@nvidia) [evening]. Brad Smith reported 230+ companies signed in the campaign's first week. Growth signal on the NVIDIA open-weights letter.
  • Grok Voice Think Fast 2.0 at $0.08 per minute (cluster of 2, @ns123abc · xAI announcement) [evening]. Agent-first voice model with better transcription accuracy and built-in voices. The price point is the news; watch it against ElevenLabs.
  • Lilian Weng leaves Thinking Machines (@ns123abc quoting @lilianweng) [morning]. She called it a hard and sad decision and signed off with "The future worth building is human." The Information separately reported she is rejoining OpenAI, which the tweet does not mention and which reframes the departure.
  • Cross-model checking and agents inside the normal project tool (cluster of 2, @dhh · on Basecamp) [morning]. dhh finds it useful to have Opus 5 check Sol's work and vice versa, which is model diversity used as a verification mechanism. Separately Basecamp coordinates with agents in the same shared tool humans use, with no dedicated AI features, so agent work stays visible where the team already looks.
  • Running coding agents from a phone via Termius, Tailscale, and tmux (@dhh) [evening]. Keep the session alive in tmux, reach the machine over Tailscale, connect by SSH from a phone. Practitioner pattern for agent sessions that outlive the laptop lid.
  • Teaching autoencoders as a physical squeeze (@ProfTomYeh) [evening]. Students hold an imaginary 900-page textbook at arm's width, compress to a palm-sized bottleneck, then try to rebuild the book. A genuinely good explanation of why reconstruction loss is the whole game.
  • Grok app builder inside the X timeline (@brivael reposting Nikita Bier) [evening]. In-timeline generated games and apps as the distribution hook, pitched as software becoming a medium of self-expression. Consumer codegen as a social feature, not a developer tool.
  • Claimed neuroscience evidence that logical reasoning does not run through language centers (@ns123abc quoting @lauriewired) [morning]. The original cites aphasia patients retaining logical reasoning and reads it as a reason not to worry about pure LLMs; @ns123abc inverts it, arguing LLMs do not use language internally either. Both readings are speculative, but the exchange is a fair snapshot of how the latent-reasoning debate is argued informally.
  • OpenAI opens free compute to scientists, and the feed reads it cynically (@ns123abc) [morning]. The take is that the plan is to harvest academic research directions on a limited budget and then run the promising ones internally with unlimited compute. No evidence offered, so log it as sentiment rather than fact.
  • Opus 5 one-shot game generation demos keep circulating (@minchoi) [morning]. Ten examples, including a Daggerfall-style game from one prompt with all assets generated. Demo-reel content, and agent-benchmarks is the standing reminder that impressive one-shot artifacts and reliable long-horizon capability barely correlate.
  • Sequoia podcast with two people central to scaling the transformer (@MillionInt) [morning]. Sonya Huang hosts Jerry and Rohan on their next act. No claim in the tweet, flagged only because the guests are worth knowing if the episode carries technical detail.
  • Does X suppress Substack links (@jaybaxter · Kucharski analysis) [morning]. A one-word tweet linking a quantitative look at whether X downranks Substack. Notable because jaybaxter is at xAI, so this is someone inside running the test in public. The linked analysis did not fetch.
  • Department of War promotes an AI-first posture (@DoWCTO) [morning]. "Building an AI first war department," with no program detail. The account has now posted twice in two slots on proliferating AI across the Joint Force, so the cadence is the only signal.
  • Seedance 2.5 teased (@zhu_hanqing666) [evening]. A creator pushing Seedance 2.0 plus Kimi K3 before the next version lands. Release cadence only, no technical detail.
  • Omarchy adoption ahead of the Quattro release (@dhh) [evening]. A reader installed it on a drawer MacBook Pro after one podcast. Community momentum, no substance.
  • ElevenLabs Speech Engine voice-agent wrapper (cluster of 2, @minchoi) [evening]. Marked #ad, claiming a full voice layer in about 100 lines with no replatforming. Skip.
  • Bulk geopolitics, migration, and domestic politics (cluster of ~65 across both slots: @MarioNawfal, @brivael, @dhh, @spencerpratt, @AustinJustice, @DoWCTO, @TareqAmin_, plus @heavypulp, @JonasBadalic, and @_sholtodouglas) [morning + evening]. Ceuta migration, Hormuz shipping, Iran strike analysis, Russian fuel export bans, French GDP, Los Angeles homelessness, Texas homicide rates, a HUMAIN and adidas kit launch, Bitcoin, jet ski commuting. No AI research or industry substance. Skip.