Mark's Reports

AI Radar — 2026-05-22

Scored for Mark Barnett: all-encompassing, personal AI newsletter — creative tools + model news + speech/STT + agent/coding news, in your voice.

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

  1. Google I/O 2026 dropped on May 19 and it was a lot — Gemini 3.5 Flash is now the default model across Google's stack (outperforms Gemini 3.1 Pro on coding + agentic benchmarks at Flash speeds), Gemini Omni creates anything from any input including video, and Managed Agents in the Gemini API let you spin up a fully-provisioned remote Linux agent with a single API call. This is the whole ecosystem shifting from "assistant" to "autonomous agent" in one keynote. Your Gemini embedding + TP3 stack just got a lot more powerful around it.

  2. Andrej Karpathy joined Anthropic this week — The co-founder of OpenAI (ex-Tesla FSD lead, ex-Eureka Labs) announced he's joining Anthropic's pre-training team to use Claude to accelerate pre-training research. This is an enormous talent signal. Anthropic has now pulled in some of the most respected technical minds in the field. The Claudes you're using in the next 12 months will be shaped by this hire.

  3. Claude Managed Agents shipped Dreaming, multiagent orchestration, and MCP tunnels — "Dreaming" lets an agent review past sessions and self-improve over time. Multiagent orchestration lets a lead agent break work into pieces and hand each chunk to a specialist (different model, prompt, tools) working in parallel on a shared filesystem. MCP tunnels let agents reach your private MCP servers without exposing them to the public internet. This is TP3-directly-relevant: your private MCP servers + Managed Agents is now a real architecture path.

  4. Veo 3.1 is the new all-arounder + ElevenLabs now hosts every major video model in one workspace — Veo 3.1 landed this week as the best overall AI video generator: 1080p 24fps, native audio sync, strong physics. ElevenLabs built a hub that puts Veo, Sora 2, Kling, Wan, and Seedance all in one place — you generate the video, add speech, lip-sync, music, and SFX without leaving the platform. This is the creator workflow you've been waiting for.

On the radar

# Item What it is Score Why it's on your radar / how it plugs in Link
1 Google I/O 2026: Gemini 3.5 Flash + Omni + Managed Agents Full-stack Google AI launch: Gemini 3.5 Flash (new default), Gemini Omni (multimodal generation), Managed Agents API (single-call remote Linux agent) 24 Your primary embedding model's whole ecosystem just leveled up. Managed Agents is the most direct TP3 upgrade path — private MCP + agent orchestration without exposed ports. blog.google
2 Karpathy joins Anthropic pre-training team OpenAI co-founder, ex-Tesla FSD, joins Anthropic May 19 to use Claude to accelerate pre-training research 20 Direct signal for Claude's future capability trajectory. If you're betting your stack on Claude (and you are), this matters. TechCrunch
3 Claude Managed Agents: Dreaming + multiagent orchestration + MCP tunnels Dreaming = self-improving agents; multiagent = parallel specialist delegation; MCP tunnels = private network reach 20 Dreaming is relevant to TP3's self-improving memory thesis. MCP tunnels let agents hit your private MCP servers from Managed Agents cloud without a public URL. Orchestration = Claude Code can delegate to specialist agents in parallel. Anthropic
4 Veo 3.1 + ElevenLabs all-in-one video workspace Veo 3.1 = best overall AI video (1080p, native audio, physics). ElevenLabs integrates Veo/Sora/Kling/Wan/Seedance + speech/music/SFX in one platform 20 The "generate video + add your voice + sync music" workflow is now one tab. No API juggling. Relevant to anything you want to produce for Legacy Soil pitch decks, Bidet demos, or teaching content. ElevenLabs video hub, Veo 3.1 review
5 Antigravity 2.0 (May 19) + Cursor 3 Composer 2.5 Antigravity 2.0: dynamic subagents, scheduled background tasks, Go CLI, public SDK, Gemini 3.5 Flash. Cursor 3: Build in Parallel mode, Composer 2.5 frontier coding model at 200+ tok/s 19 Both tools you're already running got major upgrades the same week. Antigravity 2.0 is installed on your G16 — the Go CLI and background task scheduling make it a much stronger async agent. Cursor 3's parallel agents is the answer to "run multiple code tasks at once." Lushbinary comparison
6 Kling 3.0: 4K 60fps, Music Sync, multi-shot storytelling Kuaishou's video model now does physics-accurate motion, 6 connected shots, 4K 60fps via Omni One architecture, and Music Sync that matches motion to beat 18 The most production-ready cinematic AI video generator right now. Music Sync is unique — no other model auto-locks motion to music tempo. If you're doing anything with Suno-generated tracks + video, this is the pairing. Kling comparison
7 Claude Mythos Preview: Anthropic briefing global financial regulators Anthropic's most advanced (restricted) model found zero-day vulns in every major OS/browser. Now briefing finance ministries + central banks on what it found. Project Glasswing has ~40 partner orgs 17 You're not getting access to Mythos — but the fact that Anthropic is using it to hunt real-world vulns and brief governments is a massive signal about where capability is. The "AI that autonomously finds 17-year-old FreeBSD RCE bugs" version of Claude exists and is being used by defenders right now. The Decoder
8 SeedVideoAI + Seedance 2.0: one ecosystem for song → art → video SeedVideoAI platform integrates Seedance 2.0 video + GPT Image 2 + Suno music — generate a song, make cover art, produce the music video, all in one workflow 17 End-to-end creative pipeline from audio idea to finished video without switching apps. Relevant for teaching content, Bidet product demos, or just creative fun. Worth a test drive this week. Zapier best AI video generators
9 Qwen3-ASR: 52-language on-device ASR (0.6B + 1.7B) Alibaba's Qwen team released Qwen3-ASR in January 2026: language ID + recognition + timestamp prediction, two sizes (0.6B = fast, 1.7B = accurate), fully local 17 This is a direct Bidet/TP3 speech candidate. 0.6B would run on a phone; 1.7B would run on Apex. 52 languages + diarization-friendly timestamps = possible upgrade path from current Moonshine if you want multi-language classroom capture. Free, open weights. Best STT models 2026
10 NVIDIA Parakeet TDT 0.6B v3: fastest on-device English ASR NeMo Parakeet at 0.6B parameters, optimized for low-latency English dictation, recommended for local/offline ASR alongside Whisper Large v3 Turbo 16 Your Bidet voice-typing pipeline runs faster-whisper large-v3 right now. Parakeet v3 is worth a head-to-head — it's built for fast dictation, not just transcription accuracy. If latency on the G16 Bidet is ever a pain point, this is the swap. Resemble AI open-source STT
11 Google Pics (Nano Banana image model) Google's new image creation + editing tool built on their Nano Banana model — party flyers to infographics, rolling out to limited testers now, globally this summer for AI Pro/Ultra subs 14 You're already paying for Google AI (Gemini embeddings). When Pics rolls to Pro this summer it's included. Nano Banana is also what powers Gemini Omni's image leg — so this is the same model you'll be using anyway. Limited beta now, waitlist open. Google I/O announcements
12 WebMCP: open standard for browser AI agents Chrome origin trial in v149 — a proposed web standard that lets websites expose structured tools so browser-based AI agents can execute complex tasks natively 13 Not immediately actionable for your stack, but this is the infrastructure layer for "AI agents that browse the web for you" becoming a real web primitive. The MCP ecosystem you're already running gets a browser-native extension. Worth watching. Build Fast with AI

Coming up on the radar

Cut with reason

Sources scanned

Run date: 2026-05-22 · Model: Claude Sonnet 4.6 · Est. cost: ~$0.12

Generated 2026-05-22 19:26 by tp3_scripts/ai_radar/run_radar.ps1.