THE AI PULSEEN

The Pulse — February 19, 2026

The signals that entered our radar, organized with sources and context to understand what changed.

ModelsAgentsAnthropic
LISTEN TO THIS EDITION

The audio script is ready; narration will appear after voice generation finishes.

  1. 01StepFun blog

    StepFun — Step 3.5 Flash (open-source MoE reasoning/agent model)

    WHY IT ENTERED THE RADAR

    StepFun is positioning “intelligence density” (196B total / 11B active) + speed (MTP-3) as the new wedge: fast enough for real agent loops, not just chat. They also push 256K context via hybrid SWA/full attention and claim strong SWE-bench Verified + Terminal-Bench results.

    SUGGESTED EDITORIAL ANGLE

    “The new open-source playbook isn’t bigger models — it’s faster reasoning per dollar. Here’s what ‘11B active out of 196B’ really buys you (and what it doesn’t).”

    Open original source ↗
  2. 02OpenAI (primary)

    OpenAI + Paradigm — EVMbench (smart contract vuln detect/patch/exploit benchmark)

    WHY IT ENTERED THE RADAR

    This is a rare benchmark tied to real money outcomes (EVM exploits) with a harness that tries to prevent “grader gaming.” It also reveals a key pattern: agents do better when the goal is explicit (“drain funds”) than in exhaustive auditing/patching.

    SUGGESTED EDITORIAL ANGLE

    “Benchmarks that matter: why ‘exploit-mode’ is the closest thing we have to measuring agentic cyber risk with receipts.”

    Open original source ↗
  3. 03OpenAI (primary)

    OpenAI — GPT‑5.3‑Codex‑Spark (real-time coding model; Cerebras serving)

    WHY IT ENTERED THE RADAR

    The claim isn’t just model quality — it’s interaction latency (1000+ tok/s) + infra changes (persistent WebSocket path, reduced time-to-first-token). This is an upstream “UX + serving stack” shift that will reshape how coding agents feel.

    SUGGESTED EDITORIAL ANGLE

    “The next war is time-to-first-token: why super-fast inference changes what ‘vibe coding’ even means (interrupt, steer, iterate).”

    Open original source ↗
  4. 04GitHub repo (primary) + Reddit discovery

    KittenML — KittenTTS v0.8 (tiny open TTS, practical deploy)

    WHY IT ENTERED THE RADAR

    TTS that’s small enough to ship (15M params; “works literally everywhere”) changes the feasibility of on-device voice agents, local call centers, and embedded UI narration.

    SUGGESTED EDITORIAL ANGLE

    “Forget ‘voice is expensive’—here’s a tiny Apache-style TTS you can actually bundle. Where does this break SaaS voice pricing?”

    Open original source ↗
  5. 05Zyphra blog (primary) + linked repo/HF

    Zyphra — ZUNA (EEG foundation model; ‘thought-to-text’ direction)

    WHY IT ENTERED THE RADAR

    Upstream signal: foundation-model thinking is expanding beyond text/vision/audio into biomedical time-series (EEG). ZUNA is positioned as a diffusion autoencoder for denoise/reconstruct/upsample EEG, released with Apache 2.0 + code.

    SUGGESTED EDITORIAL ANGLE

    “The next modality gold rush: why EEG ‘foundation models’ may be the real moat—not chat. What’s legit now vs hype?”

    Open original source ↗
  6. 06Claude Code docs (primary)

    Anthropic — Claude Code: new enforcement language around OAuth credentials

    WHY IT ENTERED THE RADAR

    The doc explicitly says OAuth tokens from consumer plans (Free/Pro/Max) can’t be used outside Claude Code/Claude.ai; third parties can’t route via Claude.ai credentials. This is upstream “platform boundary-setting” that will impact wrappers and unofficial tooling.

    SUGGESTED EDITORIAL ANGLE

    “The silent war behind agents: credentials + distribution. Why ‘login-based’ wrappers are getting squeezed (and what devs should build instead).”

    Open original source ↗
  7. 07Y Combinator YouTube (creator-watch)

    YC (creator-watch) — Inside Claude Code With Its Creator Boris Cherny

    WHY IT ENTERED THE RADAR

    Long-form creator interview content often contains upstream details (design constraints, why terminal UX matters, how they chose verbosity, how subagents evolve). These are the “next features” before they become product launches.

    SUGGESTED EDITORIAL ANGLE

    “Claude Code’s creator explains the terminal-first thesis. What people miss: verbosity, autonomy boundaries, and why ‘plan mode’ is a trap.”

    Open original source ↗
  8. 08YouTube RSS (creator-watch)

    Matt Wolfe (creator-watch) — The AI Stack That Saves Hours Every Day

    WHY IT ENTERED THE RADAR

    This type of video is often an early signal of what ‘default stacks’ are becoming for creators (Perplexity/Comet, WisprFlow, Cursor, etc.). The upstream value is: which tools are maturing enough that creators stop experimenting and start standardizing.

    SUGGESTED EDITORIAL ANGLE

    “The ‘creator AI stack’ is consolidating. Here’s the workflow that’s winning (and what’s just sponsor fluff).”

    Open original source ↗
TAKE THIS PULSE TO YOUR AI

Continue the analysis where you already work.

Copy this prompt into ChatGPT, Claude, Gemini, or whichever AI you use. It includes the signals, sources, and a guide for turning them into decisions.

No account is connected and no data is shared automatically.
PROMPT.md