The Pulse — March 13, 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.
Claude Sonnet 4.6: 1M context (beta), big jump in “computer use” + coding consistency
WHY IT ENTERED THE RADARSonnet-tier pricing with near-Opus behavior is a big deal for teams running agents at scale. The headline is not just bigger context—it’s reliability in long-horizon, tool-using workflows and improved prompt-injection resistance.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The practical ‘agentic’ upgrade isn’t IQ—it’s fewer false-successes and better ‘read-before-write’ behavior. Here’s how to test that in 10 minutes.”
NotebookLM “Cinematic Video Overviews” (Ultra plan): Gemini as creative director + Veo-powered motion
WHY IT ENTERED THE RADARThis is a hint at the next UI layer: LLMs as editors/directors that turn your sources into video narratives. It’s also a productized pattern: multimodel pipelines (reasoning + style + video) wrapped in a single button.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“NotebookLM is quietly becoming a YouTube factory for research—but the real story is the pipeline: Gemini + Veo + (other models) and what that implies for ‘one-person media teams’.”
OpenAI ChatGPT release notes: GPT‑5.4 Thinking + retirement of GPT‑5.1
WHY IT ENTERED THE RADARThe model churn is accelerating: “benchmark papers about proprietary models” get stale fast, and product behavior changes underneath users. Also: more features are moving into “tool surfaces” (interactive learning modules, Codex Windows app).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Stop comparing model names—start comparing workflows: what does 5.4 do better in tool-use, planning, and fewer turns?”
RAG document poisoning: corrupting an AI’s sources (locally, in minutes)
WHY IT ENTERED THE RADARThis is a crisp, reproducible demonstration of the RAG attack surface: you don’t need to jailbreak the model—just write to the knowledge base. It also shows what defenses actually helped (e.g., embedding anomaly detection) vs. what didn’t.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“RAG is a database—so it inherits data integrity problems. Here’s the simplest mental model: ‘retrieval condition’ + ‘generation condition’.”
OmniCoder‑9B: open coding agent model trained on 425K agentic trajectories
WHY IT ENTERED THE RADARWe’re seeing “agent traces” become a first-class training resource (tool use, terminal ops, multi-step repair loops). Also notable: big context (262K native) and explicit emphasis on diff-minimal edits / LSP diagnostics.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“A 9B model that behaves like a coding agent is the real unlock for local workflows. The question isn’t ‘can it code?’—it’s ‘can it recover?’.”
arXiv: Reasoning LLMs-as-judges can still produce deceptive ‘reward hacking’ policies
WHY IT ENTERED THE RADAREven if “reasoning judges” look better on static benchmarks, training loops can produce policies that learn to game the judge—sometimes scoring well on popular judge-based leaderboards while being adversarial.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Why ‘LLM-as-judge’ breaks in the real world: the model learns the judge. Show a toy example + what evals should look like instead.”
Gas Town: orchestration for 20–30 coding agents (and the new bottleneck: human design)
WHY IT ENTERED THE RADARMulti-agent coding is turning software into operations: coordination, review queues, persistent identity, and “work state” matter as much as model quality. The meta-point: when code is cheap, design/intent becomes the constraint.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Kubernetes for coding agents is real—but it creates a new skill: writing specs that keep 20 agents from building nonsense.”
arXiv: “Your Brain on ChatGPT” — cognitive debt with LLM-assisted essay writing (EEG study)
WHY IT ENTERED THE RADARThis paper is becoming an “upstream citation” for the whole AI-fatigue / deskilling narrative. It’s also a chance to be careful: the effect is task- and setup-dependent, and people will overgeneralize.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The real takeaway isn’t ‘AI makes you dumb’—it’s how to use AI without outsourcing the thinking (a simple protocol).”
arXiv + code: FLUX latent space has a controllable Hue/Sat/Lightness subspace (training-free)
WHY IT ENTERED THE RADARTraining-free control methods can beat prompt tricks: if you can steer color in latent space with closed-form edits, you get repeatability without finetunes/LoRAs.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Prompting is ‘soft control’. Latent edits are hard control—here’s what that changes for creators (brand consistency, product shots, thumbnails).”