The Pulse — May 21, 2026
The signals that entered our radar, organized with sources and context to understand what changed.
The audio script is ready; narration will appear after voice generation finishes.
An OpenAI model disproved Erdős’s “unit distance” conjecture (discrete geometry)
WHY IT ENTERED THE RADARThis is a rare, clean “AI did real frontier math” story: a decades-old conjecture is disproved with an infinite family giving a polynomial improvement (mentions δ≈0.014 in follow-up refinement).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“What does it mean when an AI finds a proof humans didn’t? (And why number theory showed up in geometry.)”
Codex “from anywhere”: mobile + remote connections + Remote SSH + hooks (enterprise)
WHY IT ENTERED THE RADARThis is the productization of long-running coding agents: live state, approvals, diffs/tests/screenshots streamed to your phone; plus Remote SSH + Hooks (guardrails, validators, logging) + scoped enterprise tokens.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The new workflow: ship code while you’re away from the desk—what’s real vs hype, and where the security traps are.”
Interaction Models: real-time, multi-stream “micro-turn” human–AI collaboration
WHY IT ENTERED THE RADARStrong claim: “interactivity should scale alongside intelligence,” so they train models that natively handle interruption, overlap, time-awareness, and concurrent tool use—less harness, more model.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Turn-based chat is the bottleneck: here’s the architecture shift (interaction model + async background model) and what it unlocks.”
Agent view in Claude Code: manage parallel agents from one UI
WHY IT ENTERED THE RADARIt’s a pragmatic UX upgrade that makes “parallel agents” actually usable: single place to view sessions, see which need input, peek/reply, and background runs (/bg, claude --bg).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Agent orchestration is becoming a first-class developer skill—here’s what changes once you can supervise 5–20 agents sanely.”
Command A+ open-source (MoE, Apache 2.0) aimed at agentic workloads
WHY IT ENTERED THE RADARA 218B total / 25B active MoE, Apache 2.0, big context (128K in / 64K gen) with practical deployment emphasis (quantizations; runs on 1×B200 or 2×H100 at W4A4).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Open-weights for enterprise agents: what ‘25B active’ actually means for speed/cost, and who this is for.”
Google Search: ads in AI Mode + “independent AI explainer” inside sponsored results
WHY IT ENTERED THE RADARThis is the monetization blueprint for AI search: conversational discovery ads, highlighted answers, AI shopping explainers, and even an “agent for leads” embedded in ads.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“AI search is becoming an ad format: how ‘AI explainers in ads’ could reshape trust, SEO, and affiliate ecosystems.”
Hermes agent hype wave (creator-watch → upstream)
WHY IT ENTERED THE RADARSignals what the “mass market” of agents is converging on: persistent memory, self-improving skills, Telegram front-ends, and $5 VPS deployments.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“What’s real about ‘agents that write their own skills’—and what breaks once you scale beyond demos.”
Qwen rumor mill: “another 27B soon” (watchlist)
WHY IT ENTERED THE RADAREven as a rumor, it’s a market signal: 27B-ish is a sweet spot for local + agentic workflows, and the community treats release cadence like “mini-frontier.”
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Why 27B is the new default ‘serious local agent’ size (and what to test immediately if a new Qwen drops).”
Harness vs model: same model, different coding-agent harnesses give very different results
WHY IT ENTERED THE RADARA concrete reminder that tooling scaffolds (edit tool schemas, web access defaults, iteration loops) can dominate perceived model quality.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Stop comparing ‘models’ without comparing harnesses—here’s a simple benchmark setup you can replicate.”