AI research wiki
social-stream · 2026-05-20

2026-05-20-afternoon

Summary

The afternoon slot is unusually research-heavy on the curated side: five of the seven bayesiansapien reposts link to substantive arxiv work or surveys, three of which sit squarely in cere-bro's deep-interest zones. The standout is SpectralQuant, a KV cache compression claim of 5.95x on Mistral 7B at +7.5% perplexity overhead, pitched directly against Google's TurboQuant (+22% at the same ratio). Behind it sit the AIRA Meta paper on agent-designed neural architectures (cross-source confirmed: this was already in HuggingFace and now has social amplification) and a 100+ page Code-as-Agent-Harness survey arguing executable, inspectable, stateful, governed code is the substrate for agents. Two opaque x.com/i/article/ reposts can't be fetched and need a click-through. The AI-account feed is mostly noise: Scobleizer monetization drama and French commentary from @brivael on grandes écoles and tech-as-game-theory, with one decent retweet about a Mistral interpretability researcher moving to Anthropic SF.

Posts

  • SpectralQuant claims 5.95x KV cache compression with 3x less degradation than TurboQuant (@ashwingop). On Mistral 7B, 5.95x compression at +7.5% perplexity overhead vs TurboQuant's +22% at the same ratio. 15-second per-model calibration, then drop-in for any HuggingFace LLM, ViT, ESM, AlphaFold Evoformer, or VideoMAE. Tier 1 if the numbers replicate. Direct shot at the TurboQuant baseline the wiki already documented. See also KV cache concept page.
  • AIRA: Meta's dual-agent system that autonomously designs sub-Llama-3.2 architectures in 24h (@omarsar0 · arxiv 2605.15871). AIRA-Compose searches macro architecture (11 agents), AIRA-Design implements low-level attention mechanisms (20 agents). Outputs AIRAformers and AIRAhybrids that beat Llama 3.2 by 2.4-3.8% accuracy at 1B and scale 54-71% faster. Already in the wiki as AIRA-Compose / Design, now cross-source confirmed via social.
  • Code as Agent Harness: 100-page survey arguing code is the agent substrate (@omarsar0 · arxiv 2605.18747). Frames agent infrastructure around three layers: harness interface, harness mechanisms (planning/memory/tool use), and scaling from single to multi-agent. Argues future systems need four properties: executable, inspectable, stateful, governed. Worth a Tier 2 summary page given the agentic-systems track.
  • On-Policy Mix: a data mixing algorithm for continual learning across pretraining, midtraining, and instruction tuning (@michahu8). Pitched as solving the open problem of finding the right data mix when distributions drift. Thread teases the full method (only 1/6 captured here). Click through for the algorithm. Tier 2, data-mixture territory.
  • Mechanistic interpretability claim: LLMs have multiple non-overlapping circuits per task, not one canonical circuit (@che_shr_cat). If this holds up, it undermines the canonical-circuit framing that a lot of mech-interp work assumes. Only the thread opener is captured. Relevant to the responsible-ai track and to recent interpretability work like monitoring internal monologue via probe trajectories.
  • Two opaque X-native articles reposted by bayesiansapien (@DimitrisPapail, @AmarSVS). Both are x.com/i/article/ long-form posts that the farmer cannot fetch without cookies. Click through to read.
  • Scobleizer on Aligned News finances and paid-post disclosures (@Scobleizer). $15K spent in month one running alignednews.com/ai, $4K of that on X API fees. Meta-commentary on the cost of running an AI news aggregator, with a side jab at competitor Digg. Skip unless tracking AI-media business models.
  • Scobleizer on Soma's passing, paid-post disclosure, robot dexterity reposts, sensor-glove demo (@Scobleizer, more, more, more). Mix of tribute, business disclosure, and robotics-demo reposts. Tier 4. Skip.
  • @brivael: a French tech-philosophy thread on the four games of tech (survival, status, control, creation) (@brivael, companion). Girardian framing of why entrepreneurs default to status and control games rather than the creation game once they exit survival mode. Opinion essay, not AI research. Skip unless you read French and want a perspective piece.
  • @brivael: grandes écoles polemic, retweet of a Mistral-to-Anthropic interpretability researcher (Soizig), Demis Hassabis admiration RT, Hassabis-as-Nobel piece, France socialism dig (@brivael, Anthropic move, Hassabis RT, other, other, other) (cluster of 6). Cultural and career commentary in French. The Mistral-to-Anthropic move is the only data point worth a tag: another Paris-to-SF interpretability hire. Otherwise skip.
  • DAIR.AI Academy promo (Vibe Coding with Claude Code course) (landing page). Bundled into two of the @omarsar0 reposts above as a footer link. Pure promo. Skip.