AI research wiki
social-stream · 2026-05-24

2026-05-24

Summary

The day's strongest signal is the Life-Harness paper landing across two curated retweets in the afternoon, both arguing that agent intelligence is a harness problem and not a model problem. Life-Harness keeps the LLM frozen and only evolves the runtime interface, reporting 88.5% average relative improvement across 126 model-environment settings with transfer across 18 backbones, which doubles down on yesterday's code-as-agent-harness thesis. The other concrete production signal is Boris Cherny pushing Claude Code auto mode down to the Pro plan with Sonnet 4.6 support in the evening, framing parallel-agent workflows as the default usage pattern. The morning slot is dominated by aggregator-of-aggregator noise but carries one load-bearing data point, an OpenAI Codex cache-regression rollback that pushes prompt caching into user-visible production territory. There is also an unverified Scoble field report about OpenAI paying NYC households to install in-home 360-degree cameras for training data, which is a rumor but a noteworthy frontier-data signal if it holds. Everything else is French political commentary from @brivael, NVIDIA color content from Jensen's Taipei trip, and Tesla self-driving positioning.

Posts

  • Life-Harness, the harness evolves and the model stays frozen (cluster of 2) (@rohanpaul_ai and @omarsar0 via @bayesiansapien retweets, arxiv:2605.22166) [afternoon]. Lifecycle-aware runtime harness that converts recurring interaction failures into reusable interventions, reports 88.5% average relative improvement across 126 model-environment settings on tau-bench, tau-squared-bench, AgentBench, with transfer across 18 backbones from a harness evolved only on Qwen3-4B-Instruct trajectories. Second concrete instance this week of the frozen-model harness-does-the-lifting frame after code-as-agent-harness.
  • Claude Code auto mode now on Pro plan with Sonnet 4.6 (@bcherny) [evening]. Cherny pitches auto mode (no permission prompts, shift+tab) as his #1 tip and the building block for parallel Claude Code sessions. Anthropic rolling this from Max down to Pro signals they want parallel-agent workflows to be default usage, not a power-user opt-in.
  • OpenAI Codex cache-regression rollback (via @Scobleizer Aligned News, alignednews.com/ai) [morning]. Codex shipped an optimization that dropped cache hit rates and accelerated usage-limit depletion on long sessions, then reverted it. One of the first public production cache-regression incidents in a frontier coding agent visible enough to surface in news aggregators, a marker that prompt caching is now a first-class user-visible production concern.
  • OpenAI allegedly paying NYC households for in-home 360-degree camera data (@Scobleizer retweeting Loffredo) [morning]. Unverified field report that OpenAI is paying hundreds of NYC families to install 360-degree cameras across their homes with paid couriers rotating memory cards, told it is for training. If accurate, a new procurement model for multimodal embodied training data once internet text is exhausted. Treat as rumor.
  • Scoble launches Aligned News, an aggregator of AI aggregators (@Scobleizer, alignednews.com/ai) [morning]. LLM-curated reader that itself reads digg.com and techmeme.com upstream, claims 8,700+ AI companies tracked and 150+ launches in recent weeks. Same architectural category as cere-bro, useful as a marker that aggregator-of-aggregators is becoming a working pattern.
  • NVIDIA amplifies the Jensen Huang "How I Built This" interview (@nvidia, nvda.ws/4uoDmmZ) [morning]. Personal-narrative version of CUDA history at the same moment NVIDIA posts a roughly trillion-dollar backlog. Context signal, not research.
  • NVIDIA Taipei color posts (Jensen signature, night market) [evening]. Computex trip PR with no product or roadmap content. Skip.
  • Tesla self-driving testimonial repost (@Tesla, Tesla.com/drive) [morning]. Test-driver experience tweet amplified by Tesla, positioning narrative rather than technical claim. Skip.
  • Scoble Tesla robotaxi-as-subscription thought experiment (@Scobleizer) [afternoon]. Sketches a 20K-USD price point pitch to put seven million Teslas on the road. Industry positioning. Skip.
  • Palantir "Start your engines" teaser (@PalantirTech via @brivael) [evening]. Cryptic teaser, no copy. Click through later.
  • DAIR.AI Claude Code course (@omarsar0, academy.dair.ai) [afternoon]. New "Vibe Coding AI Apps with Claude Code" course, attached as secondary link to the Life-Harness retweet. Skip.
  • Figma AI-first sentiment ping (@brivael) [evening]. One-line trader take that Figma is up 10% on AI-first repositioning. Retail sentiment data point only.
  • Argil mobile app dev shoutout (@brivael) [evening]. Follow plug for an indie iOS engineer on the @argildotai app. Skip.
  • @brivael off-topic cluster (cluster of 36 across morning, afternoon, evening) [morning + afternoon + evening]. French political and cultural commentary, Macron, French cinema, anti-Bollore petitions, an Argil community-manager joke, social class theory, an artificial heart prototype, Tesla and his pigeon. No AI research substance. Skip.
  • Non-AI keyword-filter false positives (@WHFraudTF taskforce, @AustinJustice politics) [morning + afternoon]. Config-leak items unrelated to AI. Skip.