The Pulse — February 28, 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.
Unsloth Dynamic 2.0 GGUFs (new quantization method)
WHY IT ENTERED THE RADARQuantization is now the bottleneck for “real” local deployment. Unsloth claims Dynamic v2.0 improves KL divergence / “flip” behavior vs standard GGUF approaches and works beyond MoE models.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Quantization isn’t just ‘smaller’—it changes answers. Here’s what ‘flips’ mean and why Dynamic 2.0 is a big deal for local inference.”
Anthropic: Detecting and preventing distillation attacks (DeepSeek/Moonshot/MiniMax)
WHY IT ENTERED THE RADARThis is unusually specific attribution + scale (millions of exchanges) and ties distillation directly to national security and export controls. It’s a strong “upstream narrative” many creators will summarize later.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Distillation is normal… until it’s industrial espionage. What patterns make it detectable, and what defenses will become standard?”
Anthropic: Statement on the comments from Secretary of War Pete Hegseth
WHY IT ENTERED THE RADARConcrete policy boundary lines from a frontier lab (no mass domestic surveillance; no fully autonomous weapons) + a claim they’ll litigate. This will cascade into procurement, compliance, and competitor positioning.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The real AI product story is ‘terms of use meet government procurement.’ What does ‘supply chain risk’ even mean here?”
Cursor: agents can now control their own computers (cloud agents w/ isolated VMs)
WHY IT ENTERED THE RADARThis is the next step from “agent writes code” → “agent tests, records artifacts, and ships PRs.” Cursor claims 30% of merged PRs are made by agents operating in cloud sandboxes.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Self-driving codebases are real—here’s the missing piece: a sandbox where agents can run the software and prove it.”
Microsoft Copilot Tasks (preview): from chat to actions
WHY IT ENTERED THE RADARMicrosoft is packaging agentic automation for non-devs: “tasks that do themselves,” recurring runs, background work, consent gates. This is a mainstream distribution channel for agent workflows.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The ‘agent UX’ war: how Microsoft is turning agents into a consumer product (and what that implies for autonomy + safety).”
Google DeepMind: Nano Banana 2 (Gemini 3.1 Flash Image)
WHY IT ENTERED THE RADARFast image generation with “Pro-like” capabilities + stronger subject consistency + 4K outputs, plus provenance push (SynthID + C2PA). This is a big signal on the production direction: quality + speed + watermarking.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Image gen is splitting into two product features: (1) speed+iteration, (2) provenance. Nano Banana 2 shows both.”
ContextCache: persistent KV cache for tool schemas (29× TTFT speedup claim)
WHY IT ENTERED THE RADARTool calling keeps getting slower as tool lists grow. This proposes a practical, middleware solution: content-hash + group-cached KV for the whole tool block (and documents a negative result for per-tool caching).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Tool calling is secretly a latency tax. KV caching for tool schemas: how it works, why per-tool caching breaks, and what to try in your stack.”
Micro Diffusion (text diffusion in ~150 lines) — educational repo
WHY IT ENTERED THE RADARText diffusion is resurfacing as an alternative generation paradigm (parallel-ish generation, natural editing). This repo is a clean explainer you can riff on for a ‘teaching’ video.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Autoregressive vs diffusion for text—why ‘generate everything then refine’ is interesting, and where it still loses today.”