The Pulse — May 17, 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.
Interaction Models (real-time, micro-turn, multi-stream human–AI collaboration)
WHY IT ENTERED THE RADARThis is a concrete architectural proposal for “real-time collaboration” models (simultaneous perception + output) instead of bolting interactivity on with harnesses (VAD, turn-taking, etc.). It also explicitly frames interactivity as a scaling axis alongside intelligence.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The next LLM jump isn’t bigger context — it’s time (micro-turns, interruption, overlap). Here’s what changes when models are trained for interaction natively.”
Codex “from anywhere” (mobile steering + remote connections + hooks)
WHY IT ENTERED THE RADARThe product direction is “agents that run long” + human check-ins as the core UX. The important detail is the workflow primitives: remote connections, approvals, and enterprise hooks (validators, secret scanning, logging, memories).
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The real story isn’t ‘Codex on your phone’ — it’s the emerging agent operating model: approvals, hooks, and remote sessions as the new CI/CD.”
$δ$-mem: efficient online memory for LLMs (delta-rule state + low-rank attention corrections)
WHY IT ENTERED THE RADARA clean middle path between ‘just increase context’ and ‘full fine-tuning’: keep the backbone frozen, add a tiny online associative memory state (e.g., 8×8) that writes continuously and modulates attention. This is exactly the kind of mechanism agent builders can watch for in next-gen “memory-native” models.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“You can give an LLM memory with an 8×8 matrix — here’s the trick, and what it implies for assistants that learn over time.”
llama.cpp: MTP (multi-token prediction) support testing on Qwen3.6 + RTX 5090
WHY IT ENTERED THE RADARLocal inference speed is still the biggest gating factor for “always-on” assistants. MTP can change perceived latency dramatically (especially for short outputs), and community benchmarks show where it helps and what flags matter.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The fastest path to a better ‘assistant feel’ might be MTP + good decoding, not new models. Here’s what to toggle in llama.cpp and what to expect.”
NVIDIA SANA-WM (minute-scale world modeling; 720p video claim)
WHY IT ENTERED THE RADARWorld models + longer horizon generation is creeping from research into “almost productizable” claims (minute-scale, 720p). Even without full details in the page fetch, it’s a strong upstream breadcrumb worth tracking for releases (paper/code) and the next wave of video generation tooling.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“World models are quietly eating video generation: what ‘minute-scale world modeling’ could enable (consistent scenes, controllable camera, long actions).”
New personal finance experience in ChatGPT (account linking + financial memories)
WHY IT ENTERED THE RADARThis is a template for how “agentic” features will ship to normal users: connect to real systems (Plaid), add domain-specific memories, add stronger guardrails + deletion controls. It’s also a distribution wedge for deeply context-grounded assistants.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Finance is the first mainstream ‘connected agent’ category. What it teaches us about data permissions, memory, and trust UX.”
Anthropic Labs: Claude Design (visual work: slides, prototypes, one-pagers)
WHY IT ENTERED THE RADARLabs products are where interaction patterns show up first. “Design” is a good proxy for where multimodal + tool use is heading: structured outputs, iteration loops, and publishing-ready artifacts.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“AI is moving from ‘generate’ to ‘ship’. Design workflows are the canary — because the output quality is instantly obvious.”
Zerostack: Unix-inspired coding agent written in Rust
WHY IT ENTERED THE RADARThe agent ecosystem is fragmenting: more local-first, compiled, permissioned agents instead of one monolithic cloud assistant. Rust-based agents are interesting for sandboxing and distribution.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Why coding agents are turning into Unix tools: composable, local, strict permissions — and what that means for dev workflows.”