Mark's Reports

AI Radar — 2026-05-29

Mark Barnett's weekly what's-on-my-radar AI newsletter. Not dry enterprise. Not a toys-only list. All-encompassing.

Top picks (the 3-4 highest-signal things this week)

  1. Claude Opus 4.8 + Dynamic Workflows — Your daily driver just dropped an upgrade (same price), and the headline feature — Dynamic Workflows in research preview — lets Claude Code spin up hundreds of parallel subagents on a single task (codebase migrations, end-to-end); this lands directly on your TP3 agent orchestration and any multi-step Bidet work you're spinning up.

  2. Google Gemini Spark — Announced at Google I/O (May 19), now rolling out to US Google AI Ultra subscribers: a 24/7 personal AI agent that runs on cloud VMs and keeps working after your laptop closes — integrates Gmail, Docs, 30+ third-party apps via MCP; think of it as what your Antigravity agent wants to be when it grows up.

  3. Microsoft VibeVoice-ASR — Open-source 9B model (on HuggingFace) that processes up to 60 minutes of audio in a single pass with joint transcription + speaker diarization + timestamps built in; this is the STT backbone Bidet AI's been waiting for — pull it, bench it on your whisper_corpus, compare against Parakeet on G16.

  4. Figma AI Design Agent (free beta) — Figma shipped a native AI agent inside the canvas this week (partnered with Anthropic + OpenAI), free during beta for Professional/Org plans; directs design via natural language, runs parallel agents, understands your existing components — worth opening a Figma file and playing with it today.

On the radar

# Item What it is Score Why it's on your radar / how it plugs in Link
1 Claude Opus 4.8 + Dynamic Workflows New Anthropic flagship (May 28), same price as 4.7, parallel multi-agent orchestration in research preview via Claude Code 24 You're running this model right now. Dynamic Workflows = hundreds of parallel Claude subagents on one task — directly relevant to TP3 multi-agent orchestration, big Bidet refactors, and curriculum generation pipelines. More honest uncertainty flagging is a real quality-of-life improvement. Anthropic news
2 Google Gemini Spark 24/7 personal AI agent running on Google Cloud VMs; persists when device is off; Gmail/Docs + 30+ MCP apps; built on Gemini 3.5 Flash + Antigravity 2.0 harness 22 This is the cloud-side "always running" agent concept — a lot like what you'd want a persistent TP3 agent to be. Also: your Antigravity desktop agent now has a much larger harness under it. Beta rolling out to US Google AI Ultra ($100/mo) this week. TechCrunch
3 Gemini 3.5 Flash + Gemini Omni (Google I/O) New flash model rivals flagship speed; Gemini Omni is their any-input → any-output multimodal model; both available via API + AI Studio now 21 TP3 uses Gemini for live embeddings. Gemini 3.5 Flash is your Antigravity's new brain. Gemini Omni's video understanding + Veo integration is the "create anything from anything" angle — worth watching for Breezy Farms content creation. Google blog
4 Microsoft VibeVoice-ASR Open-source 9B ASR model: 60-min single-pass audio, joint diarization + timestamps, 50+ languages, custom hotwords, available on HuggingFace 21 This is the Bidet AI STT upgrade. One model, one inference pass, you get a full transcript with who-said-what and timestamps. Custom hotwords = you can prime it on your vocabulary. Open weights = runs on G16 or Apex, zero API cost. HuggingFace · MarkTechPost
5 Figma AI Design Agent (free beta) Native agent inside Figma's canvas — direct natural-language design editing, parallel sub-agents, understands your component library; partnered with Anthropic + OpenAI; free during beta 20 Free this week on Professional/Org plans. Not just a button — it's an agent that knows your components and can iterate variations on what's already on your canvas. Immediately useful for any slide decks, school materials, or Bidet UI mockups. TechCrunch · FastCompany
6 Google Antigravity 2.0 Google's agent-first developer platform: desktop app + agy CLI + Antigravity SDK + Managed Agents via Gemini API, all under one harness 19 You're already running Antigravity on G16. This is the platform that just got a major version bump at I/O — Gemini Spark is built on top of it. Your existing Antigravity agent now has a much richer SDK and a CLI to play with. Google I/O developer highlights
7 Claude Managed Agents — Self-Hosted Sandboxes + MCP Tunnels Public beta: run Claude agents in your own infrastructure with built-in sandbox; MCP Tunnels (research preview) let remote Claude instances reach your local MCP servers securely 19 Directly relevant to TP3's MCP stack. MCP Tunnels = claude.ai web/mobile can punch through to your local omi-mcp and biometric-mcp without punching a hole in your firewall. Sandboxes = safer agent execution for the Bidet pipeline. Anthropic updates
8 Pika 2.5 AI video generator with physics-based interaction model — understands weight, impact, fluid dynamics; Pikaffects/Pikaswaps/Pikadditions for creative "what if" loops 19 The physics realism angle makes this genuinely different from "generate a clip" tools. Breezy Farms product demo videos, school creative projects — Pika 2.5 is where you go when you want video that doesn't look floaty. AI video review roundup
9 Suno v5.5 — Voice Capture + Custom Models Music generation with your actual voice captured, custom model training on your taste, personalization via My Taste 19 You can now record your voice and train Suno to generate music in that style. My Taste personalization means it learns what you actually like over time. Fun for Breezy Farms content — generate a jingle that actually sounds like something you'd pick. Suno
10 Qwen3-ASR (Alibaba, open source) SOTA open-source STT: 52 languages, diarization, timestamps; comes in 1.7B and 0.6B sizes — runs on-device with minimal RAM 19 The 0.6B variant is small enough to run on a Raspberry Pi. For InstaBidet, this is your offline fallback STT that doesn't need a GPU. Bench it vs. Moonshine and VibeVoice on your whisper_corpus — you want the on-device winner locked in before you build the final ASR harness. Gladia review
11 Kling 3.0 Omni ByteDance video AI: multi-shot sequences with shared audio timeline, native dialogue generation in 5 languages, filmmaker-friendly consistency 18 Multi-shot coherence is the thing that makes Kling 3.0 stand out — you can tell a story across shots without continuity breaking. Now accessible from inside Luma's UI as a model option. Good for Breezy Farms narrative content. Video AI guide
12 Ideogram → Claude Code MCP Ideogram image generation (industry-best text-in-image rendering) now wired into Claude Code as a skill — generate precise-text visuals directly from your chat 17 Ideogram has always been the best tool for images where the text has to be legible (signs, cards, social posts). Now you can fire it from Claude Code without switching apps. Curriculum graphics, Breezy Farms signage, anything where you need actual readable words in the image. MCP Market

Coming up on the radar

Cut with reason

Sources scanned

Generated 2026-05-29 19:11 by tp3_scripts/ai_radar/run_radar.ps1.