AI research wiki
social-stream · 2026-05-12

2026-05-12

Summary

The day's strongest cross-slot cluster is Claude Code's internal architecture. The morning surfaced @bcherny's launch of the Claude Code agent view (a single list of all in-flight sessions), the afternoon carried a public 5-layer field guide thread, and the evening closed with Gary Marcus reading Claude Code's 53 symbolic tools and ~500k lines of scaffolding as vindication for neurosymbolic AI. Three independent angles on the same codebase in one day is a real signal, not noise. The standout single-slot post is AutoTTS in the afternoon, with both the authors' thread and an analytical writeup arriving inside a few hours: $39.9 and ~160 minutes of agent search beats hand-crafted test-time scaling baselines, and the wiki already has a page on it from yesterday so this is cross-source confirmation. The rest of the afternoon is unusually dense for one slot, with PwC's clarification-timing paper, the tool-calling steerability probe, Thinking Machines' always-on Interaction Models, and the Microsoft + Salesforce 200K-conversation drift study all landing in the same window. Morning and evening are otherwise thin, with the @bcherny Cowork-books-flights cluster as the only other workflow-grade post, plus AWS Claude Platform GA in industry news. Several opaque x.com/i/article reposts go uncollapsed because the synthesis cannot expand them inline.

Posts

  • Claude Code agent view (research preview) (@bcherny · @claudeai launch) [morning]. Unified list of all in-flight Claude Code sessions instead of cycling between terminal tabs. Productizes the many-agents-per-user pattern.
  • Claude Code's 5 architectural layers (@NainsiDwiv50980 · wiki) [afternoon]. Field guide thread on CLAUDE.md as memory layer plus four further layers that go beyond prompting. Clean public summary of material already in the wiki's Claude Code pages.
  • Gary Marcus: Claude Code is the most neurosymbolic system he has ever seen (@GaryMarcus · ccunpacked.dev · wiki) [evening]. Reads Claude Code's 53 tools plus ~500k lines of orchestration around a frontier LLM as proof that progress is coming from classical-AI scaffolding, not pure scaling. The linked site is a source-level dissection of the agent loop and tool registry.
  • @bcherny Cowork + Opus 4.7 one-shots 8 flights and 5 hotels (cluster of 2) (@bcherny a, @bcherny b) [morning]. Flight preferences go into Cowork instructions, Opus opens a browser, navigates sites, books everything in parallel while the user does other Claude Code work. Frontier-agent browser use is crossing into real workflows.
  • AutoTTS, frontier LLMs design their own test-time scaling (cluster of 2) (@zhengtoong, @omarsar0 · arxiv · wiki) [afternoon]. Environment-driven discovery framework: humans design the search environment, coding agents discover the width-depth TTS controller. Discovery cost $39.9 and ~160 minutes, results generalize across held-out benchmarks and model scales. Second day of independent signal.
  • Clarification timing in long-horizon agents (PwC) (@dair_ai · arxiv) [afternoon]. Forced-injection framework across 4 frontier models, 84 tasks, 6,000+ runs. Goal clarification loses almost all value after 10% of execution. Deferring past mid-trajectory is worse than never asking. Empirical brake on the "always ask early" prior.
  • Tool calling is linearly readable and steerable (@tldr_ai_papers · arxiv · wiki) [afternoon]. Probes 12 instruction-tuned models (270M to 27B). Adding the mean activation difference between two tools flips the chosen tool with 77-100% accuracy and the JSON arguments autoregressively conform to the new schema. Small set of mid- and late-layer attention heads localized via patching.
  • Thinking Machines Interaction Models (TML-Interaction-Small) (@rohanpaul_ai · blog) [afternoon]. 276B MoE, 12B active. Replaces walkie-talkie turn-taking with always-present AI: audio, video, and text sliced into 200ms micro-turns, model listens, watches, speaks, acts, and tool-calls while the interaction is still happening. Trained from scratch with a multi-stream micro-turn design.
  • RAO: Recursive Agent Optimization (@apurvasgandhi) [afternoon]. End-to-end RL for training LLMs to spawn, delegate to, and coordinate with recursive copies of themselves. Sub-agents as inference-time scaling primitives. Adjacent to the Sakana Conductor and AutoTTS thread on learned orchestration.
  • Anthropic memory plus "Dreaming" continual-learning preview (@daniel_mac8) [afternoon]. Reports on a recent talk framing memory as the next first-class agent primitive after MCP, Skills, and harnesses: writable shared context, provenance, review, background consolidation. "Dreaming" is described as recursive self-improvement at the agent-system level.
  • GEPA explainer on long-horizon agent RL (@blc_16) [afternoon]. Walkthrough of why sparse rewards throw away trajectory information and how GEPA learns from the trajectory itself via textual critiques, prompt edits, and Pareto-frontier selection.
  • Microsoft + Salesforce: 200K conversations, 39% average accuracy degradation (@HowToAI_) [afternoon]. ChatGPT 96.6% to 72.6%, Gemini 97.4% to 68.1% as conversations lengthen. Attributed to an anchoring trap. Mechanism overlaps with the PwC clarification-timing paper in the same slot.
  • Curved geometry of LLM activations (@che_shr_cat) [afternoon]. Argues the Linear Representation Hypothesis is a useful lie that breaks down fast: straight-line steering produces teleportation and diversity collapse. Conceptually opposite to the tool-steerability probe in the same slot.
  • Nature Neuroscience: brains do not predict every word uniformly (@ValerioCapraro) [afternoon]. Zou, Poeppel, Ding: brain activity tracks word surprisal LLM-style inside phrases but the match weakens across major phrase boundaries. Counterweight to the "humans are just next-word predictors" frame.
  • Claude Platform on AWS GA (@mattsgarman · AWS blog) [morning]. Anthropic's native Claude Platform, including Managed Agents, Agent Skills, MCP connector, code execution, and files API, accessible directly from AWS accounts. AWS is the first cloud provider to offer it natively. Also in today's Industry Pulse.
  • NVIDIA at Dell Technologies World (cluster of 2) (@nvidia a, @nvidia b · event) [morning]. Jensen Huang and Michael Dell co-keynote on AI-accelerated enterprise compute, May 18-21 Las Vegas. PR-cycle event.
  • Opaque x.com/i/article reposts (click through) (@AmarSVS, @AlphaSignalAI, @neural_avb, @ns123abc) [afternoon + evening]. Bare X-native long-form article links the synthesis cannot expand inline.
  • @magicsilicon "Whoa" (@magicsilicon) [afternoon]. Reaction post, no content. Skip.