The Pulse — March 31, 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.
axios compromised on npm — malicious versions dropped a RAT via postinstall
WHY IT ENTERED THE RADARThis is a high-sophistication supply-chain attack against one of the most-downloaded npm packages (100M+/week). The clever part: axios itself contains zero malicious lines—damage happens via a staged dependency with a self-deleting installer.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The axios hack explains why ‘scan the diff’ is dead—how attackers hide in dependencies + what your CI should alert on (OIDC publish, outbound network, install hooks).”
Ollama preview: now powered by Apple’s MLX on Apple Silicon (big TTFT + tokens/s gains)
WHY IT ENTERED THE RADARLocal inference on Macs gets meaningfully faster by moving the stack to MLX + unified memory assumptions. Also interesting: explicit mention of NVFP4 + cache reuse/snapshots aimed at agentic/coding workloads.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Local agents on Macs just got a free turbo: what MLX changes practically (TTFT vs decode), and why caching is the secret sauce for coding agents.”
Claude gets “computer use” in Cowork + Claude Code (research preview) + Dispatch handoff
WHY IT ENTERED THE RADARThis is the mainstreaming of ‘agent does the clicking’ as a first-class feature (with explicit permission prompts + injection mitigations). The killer UX is Dispatch: task from phone → executed on desktop.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Computer-use agents are finally productized: what workflows become viable, where they still fail, and the 1 safety trap (prompt injection via webpages).”
Claude Code “Auto mode”: long-running tasks without --dangerously-skip-permissions
WHY IT ENTERED THE RADARA concrete ‘middle path’ for agent autonomy: per-tool-call classifier gating destructive actions. This is a pattern we’ll see copied everywhere (policy engine sitting between model + tools).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Auto-mode is the blueprint for safe autonomy: agent runs for an hour, but a classifier is the real boss—how to copy this architecture in your own agent.”
Gemini 3.1 Flash Live: Google’s real-time audio model (benchmarks + SynthID watermarking)
WHY IT ENTERED THE RADARAudio agents are the next UX battleground. Two things jump out: (1) function-calling-through-audio benchmarks, and (2) watermarking audio output with SynthID (anti-misinformation baked in).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Voice agents aren’t just ‘talking LLMs’ now—they’re function-callers with watermarks. Here’s what that implies for call centers + real-time assistants.”
Mistral releases Voxtral TTS (4B), multilingual + voice adaptation (3s reference) + low latency
WHY IT ENTERED THE RADARThis is a serious shot at the TTS “voice agent stack” with emphasis on emotional expressiveness, streaming latency, and customization. Also: they publish architecture details and positioning vs ElevenLabs.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“TTS wars are here: Mistral’s Voxtral vs ElevenLabs—what’s real (latency, multi-language, voice cloning) vs what’s marketing.”
Smallest.ai Lightning V3: TTS optimized for conversational streaming, not ‘read-aloud’
WHY IT ENTERED THE RADARGood framing: ‘speech fidelity ≠ conversational naturalness.’ They argue streaming context + turn-taking changes what matters, and that classic TTS evals (MOS) don’t measure the thing users feel.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Why most TTS demos lie: streaming voice agents need thinking + listening prosody. The benchmarks you should actually care about.”
TimesFM (Google Research): time-series foundation model repo (TimesFM 2.5, 200M params, 16k context)
WHY IT ENTERED THE RADARTime-series is a huge under-served market for ‘foundation models’ (forecasting in retail, infra, finance). TimesFM 2.5’s longer context + smaller size hints at a “small-but-specialized” trend.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“LLMs aren’t the only foundation models: why forecasting models (TimesFM) might be the most monetizable ‘boring AI’ in 2026.”