AI research wiki
social-stream · 2026-05-14

2026-05-14

Summary

The day is morning-driven: afternoon was effectively empty (one Tesla FSD marketing clip) and evening was three vendor tweets (NVIDIA pushing Nemotron Labs "claws" twice and AWS reopening Kiro credits). The strongest morning cluster is δ-mem, the lightweight 8x8 frozen-backbone associative memory paper, picked up by both @HuggingPapers and @dair_ai. A second mini-cluster forms around the AutoTTS agentic test-time scaling paper (two reposts) and a third asymmetric-training-identical-inference cluster ties Lighthouse Attention and Nous Token Superposition Training together. Standout single-slot signals worth pulling out of the morning stream: Anthropic's Mythos / Glasswing cyber-range result from @bcherny with AISI confirmation, the multi-agent Bystander Effect paper, and the refusal-neurons MLP-level alignment bypass. Everything else (NVIDIA SAP OpenShell, NVIDIA Snap GPU pipeline, opaque article reposts, Tesla, Kiro) is industry positioning or noise.

Posts

  • δ-mem: efficient online memory for frozen LLMs (cluster of 2) (@HuggingPapers · @dair_ai · paper) [morning]. 8x8 associative memory with delta-rule learning gives 1.31x on MemoryAgentBench and 1.20x on LoCoMo without fine-tuning the backbone. @dair_ai calls it one of the most elegant memory mechanisms of the month.
  • AutoTTS: agentic discovery of test-time scaling controllers (cluster of 2) (@zhengtoong · @ihtesham2005 · paper · wiki) [morning]. Claude Code proposes its own TTS controllers, tests, and refines over 5 rounds. Total discovery cost $39.9.
  • Multi-agent Bystander Effect / Sovereignty Gap (@dair_ai · paper) [morning]. 22,500 deterministic trajectories show agents often compute the right answer internally then suppress it to agree with the swarm. Formalizes an Interaction Depth Limit and a lead-anchor non-commutativity finding.
  • Lighthouse Attention: removable subquadratic wrapper (@omarsar0 · paper) [morning]. Nous Research. Wraps SDPA with a hierarchical gradient-free selection layer that gets removed at end of training so deployed inference runs vanilla attention.
  • Token Superposition Training (TST) (@NousResearch) [morning]. 2-3x wall-clock pretraining speedup at matched FLOPs without changing architecture, optimizer, tokenizer, or data. Bag-of-tokens prediction in the first third of training then standard NTP. Third paper this week on the asymmetric-training-identical-inference frame.
  • Microsoft KV-Cache compression repost (@AiwithYasir) [morning]. Inflated framing ("Microsoft just solved the context window problem") of a real KV-cache compression paper for long chain-of-thought. Click through to the underlying paper.
  • Refusal-neurons: single MLP neuron bypasses safety alignment (@hamid_kazemi22) [morning]. Across 7 models, 2 families, 1.7B to 70B scale, suppressing one MLP neuron disables refusal behavior. No fine-tuning, no prompt engineering.
  • "Attention Is All You Need V2" / Nested Learning framing (@HowToAI_) [morning]. Hype repost of the Google HOPE / Nested Learning architecture from 2026-04-28. Skip the tweet narrative; the paper itself is real.
  • omarsar0 on HTML Artifacts + agents (@omarsar0 · DAIR.AI event) [morning]. Demo of an HTML+JS artifact backed by Obsidian markdown that agents read and modify. Practitioner signal on where agent UI is heading.
  • Anthropic Mythos / Glasswing cyber results (@bcherny · XBOW evaluation · AISI report) [morning]. UK AISI confirms Mythos Preview is the first model to solve both cyber ranges end-to-end, including Cooling Tower which no prior model solved. AISI reports autonomous AI cyber-task length doubling every few months.
  • NVIDIA + SAP OpenShell agent runtime (@nvidia · NVIDIA blog) [morning]. SAP embeds NVIDIA OpenShell (open-source secure agent runtime) into SAP Business AI Platform with isolated execution and infra-level containment.
  • NVIDIA + Snap GPU petabyte A/B testing (@nvidia · NVIDIA AI Podcast Ep 298) [morning]. Snap moved 10+ PB/day of A/B-test data to GPU-accelerated Google Cloud: 76% cost cut, 80% memory cut, zero code changes.
  • Nemotron Labs "claws" / long-running agents (cluster of 2) (@nvidia tweet 1 · @nvidia tweet 2 · blog) [evening]. NVIDIA pushes self-hosted persistent agents that run 24/7, citing OpenClaw's 250k GitHub stars in 60 days. Market-positioning signal: agent narrative shifting from prompt-triggered to always-on.
  • Opaque article-only reposts (group, click through to read) (@akshay_pachaar · @oneill_c · @amitiitbhu · @AnatoliKopadze · @mem0ai) [morning]. Five x.com/i/article/ reposts with no extractable text. mem0ai likely on agent memory; the rest unclear without click-through.
  • Tesla marketing posts (@Tesla EV · @Tesla FSD glare) [morning + afternoon]. EV marketing and a vendor FSD-through-glare clip. Skip.
  • Kiro startup credits reopen (@mattsgarman · blog) [evening]. AWS reopens Kiro Pro+ credits for pre-seed to Series A startups. Pure promo. Skip.