The Pulse — March 29, 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.
ARC-AGI-3 launches (interactive benchmark for agentic intelligence)
WHY IT ENTERED THE RADARThis is a benchmark shift: from “answer questions” to “explore unknown environments, infer goals, build world-models, plan.” Humans at 100%, frontier AI reported <1% (0.26% on the ARC Prize page).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“LLMs aren’t failing at ‘reasoning’—they’re failing at goal discovery + exploration. ARC-AGI-3 makes that painfully measurable.”
Anthropic: large-scale Claude distillation campaigns (DeepSeek / Moonshot / MiniMax)
WHY IT ENTERED THE RADARConcrete numbers + tactics (24k fraudulent accounts, 16M exchanges) and a clear claim: distillation at scale can bypass safety/export-control intent by copying capabilities faster/cheaper.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Distillation is normal in ML—but now it’s an industrial espionage primitive. Here’s how they try to extract chain-of-thought + tool-use behaviors.”
Claude “Dispatch + Computer Use” (phone → desktop handoff; Claude controls mouse/keyboard)
WHY IT ENTERED THE RADARThis is the next UX wedge for agents: the model doesn’t need a perfect API integration—if it can use your apps like a human, the product surface expands massively (with new security issues).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Computer-use agents are the new ‘browser automation’, but with a model in the loop—what’s actually useful vs demo bait?”
Claude Code: “Auto mode” (permission classifier instead of --dangerously-skip-permissions)
WHY IT ENTERED THE RADARThis is quietly huge for real-world agent workflows: fewer human interrupts while keeping a safety layer that blocks destructive actions.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The real moat in coding agents isn’t model IQ—it’s runtime governance: permissions, sandboxes, and action classifiers.”
OpenAI releases GPT‑5.4 mini + nano (small models tuned for high-volume + subagents)
WHY IT ENTERED THE RADARThis reinforces the ‘system design’ trend: big model plans, smaller models execute in parallel. Also notable: OpenAI explicitly positions mini as strong for tool use + “computer-using systems.”
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Stop comparing single models. Start comparing stacks: planner model + swarm of mini/nano subagents.”
OpenAI launches a Safety Bug Bounty (abuse + agentic prompt injection / MCP risks)
WHY IT ENTERED THE RADARThis is upstream signal that “agentic prompt injection + data exfiltration” is now a first-class category (they explicitly mention agentic products and MCP-related risks).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“We’re entering the era where prompt injection is treated like a vulnerability class—what’s a ‘reproducible’ agent hijack?”
LiteLLM supply-chain incident (malicious PyPI releases 1.82.7 / 1.82.8)
WHY IT ENTERED THE RADARThis is an “AI tooling layer” compromise: gateway/orchestration libs sit in dev machines + CI + production. The incident shows how fragile the agent ecosystem is to dependency attacks.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Your AI agent stack is now part of your security perimeter. Here’s the failure mode: transitive dependency → CI secrets → cloud compromise.”
TurboQuant (vector quantization w/ random rotations) + community adaptation for weight quantization
WHY IT ENTERED THE RADARQuantization is becoming algorithmic again (not just “try GPTQ/ AWQ”). The random rotation intuition is simple enough to explain, and the OSS repo shows near-lossless “4+4 residual” style compression.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The 60-second intuition for TurboQuant (and why ‘polar coordinates’ is the wrong explanation). Then: what changes if you quantize weights vs KV-cache?”
BDH (Dragon Hatchling) fast-weights write-back: first open-source working implementation
WHY IT ENTERED THE RADARThis is a rare example of “model rewrites its own weights during inference” being implemented and debugged in the open, including consolidation experiments.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“RAG vs fine-tuning vs fast weights: three memory paths—and why fast weights are attractive (and scary) for agentic systems.”
LawBreaker: symbolic-math graded adversarial physics benchmark (anti-sycophancy / anti-anchoring)
WHY IT ENTERED THE RADARA practical evaluation pattern: generate infinite adversarial variants + grade with symbolic math (no LLM judge). This is a blueprint for ‘hardening’ evals in other domains.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Benchmarks are getting gamed; symbolic grading is a way out. Build a ‘LawBreaker’ for your niche.”