AI research wiki
social-stream · 2026-05-09

2026-05-09-morning

Summary

Strongest signal is a four-post Claude Code workflow cluster: Khairallah on shipping 49 features in two days with zero hand-coded lines, Thariq Shihipar arguing HTML is the new Markdown for Claude Code output, dr_cintas mapping the five-folder Claude Code system (CLAUDE.md, skills/, hooks/, subagents/, plus a fifth), and Avichawla on why a smarter Claude burns 54% more tokens on MCPMark V2 (the cause is full-stack backend awareness, not the model). Second cluster is Anthropic Natural Language Autoencoders (two reposts of @AnthropicAI) which today's digest already covers as a Tier 2 Deep Dive. Standout non-cluster items: DCI (Direct Corpus Interaction) was reposted from the first author and is today's Tier 2 deep dive on agentic search, ASI-Arch claims fully autonomous neural architecture discovery, Google DeepMind sketched a Talker+Planner dual-agent medical co-clinician, and NVIDIA + Unsloth published low-level fine-tuning optimizations (packed-sequence caching, double-buffered checkpoint reload, faster MoE routing) for up to 25% speedup.

Posts

  • Claude Code workflow cluster (cluster of 4). Khairallah (@eng_khairallah1) on shipping 49 features in 2 days, all AI written, 30 min talk released. Thariq Shihipar (@trq212) on HTML over Markdown for Claude Code output, with examples and prompt suggestions. dr_cintas (@dr_cintas) maps the Claude Code five-folder system: CLAUDE.md (memory), skills/ (auto-invoked workflows), hooks/ (deterministic guardrails), subagents/ (delegation), plus a fifth folder. Avichawla (@_avichawla) on MCPMark V2 showing smarter Claude burns 54% more tokens because it must understand the full backend (RLS policies, storage, edge functions, auth), not because of model regression. Cross-links to Claude Code architecture and permissions / hooks layer.
  • Direct Corpus Interaction (DCI) (@zhuofengli96475 · paper). First-author repost of today's HuggingFace paper. Replaces RAG retrieval with grep + bash on the raw corpus. Sonnet 4.6 jumps 69 to 80 on BrowseComp-Plus, $424 cheaper. Today's DCI Deep Dive covers it.
  • Anthropic Natural Language Autoencoders (cluster of 2). saprmarks (@saprmarks) and dejavucoder (@dejavucoder) both reposting the AnthropicAI announcement. Method translates Claude's internal activations to readable text. Today's NLA wiki page and digest cover the deceptive-alignment finding.
  • ASI-Arch / AlphaGo Moment for Model Architecture Discovery (@ZabihullahAtal). Claim is fully autonomous neural architecture discovery without human-designed search spaces. Generates ideas, writes executable code, runs experiments, validates. Worth tracking as the autonomous-AI-research thread accumulates.
  • Anthropic Dreaming feature (@iam_elias1). Step-by-step explanation of Claude Managed Agents Dreaming, scheduled background process reading up to 100 past sessions to extract patterns. Yesterday's digest covered the announcement; today's skill curation cluster places it as the production embodiment of the SkillOS architecture.
  • NVIDIA + Unsloth fine-tuning optimizations (@Suryanshti777). Packed-sequence metadata caching, double-buffered checkpoint reloads, faster MoE routing for GPT-OSS. Up to 25% training speedup with detailed bottleneck explanations and benchmarks. Tier 1 GPU-optimization signal worth following.
  • Google DeepMind co-clinician (@TheTuringPost). Dual-agent safety architecture: Talker interacts with patient, separate Planner monitors and verifies clinical safety boundaries. Same Talker+Planner pattern adjacent to multi-agent coordination.
  • Google CodeWiki (@AuroraMar1eL). Drop in a GitHub repo, get an interactive guide with auto-generated diagrams and a chatbot. Code-comprehension product launch worth tracking against the agent-tooling-as-substrate thread.
  • Cloudflare 1,100 layoffs framing (@GregKamradt · Cloudflare blog). Greg's framing as the most direct evidence of AI displacing jobs we have seen. Today's digest Industry Pulse covers the underlying announcement.
  • Mythos Preview Firefox 423 bugs in April (@aakashgupta). Bobby Holley's vertigo quote on a 20x step function over the previous 22-bugs-per-month elite human baseline. Yesterday's digest covered Mozilla's 271-vulnerability disclosure; this post adds the April monthly count.
  • NeurIPS deadline + conference trackers (@aimalysheva · aideadlines.org · trybibby.com). Deadline trackers worth bookmarking. trybibby shows CORE rank inline.
  • AI subreddit list (@aiedge_). 16-subreddit list for AI hidden alpha. Reference, low signal density.
  • Trading Obsidian second-brain anecdote (@browomo). Single-user case study of an N8N + Obsidian pipeline producing daily trading ideas. Anecdotal, not generalizable.
  • Tesla Vision earlier airbag deployment (@Tesla). Tesla Vision deploys airbags up to 70ms earlier on unavoidable collisions. Inference-latency-as-safety-feature signal.
  • NVIDIA + ServiceNow Project Arc and Vibe Coding (@nvidia). NVIDIA AI Factory integrated into the ServiceNow platform at Knowledge26. Enterprise-AI deployment story.
  • Opaque x.com article reposts (3) (@trq212, @eng_khairallah1, and one in @bayesiansapien's earlier slot). Click through to read.
  • Sticker giveaway (@bcherny). Skip.