Mark's Reports

AI Radar — 2026-06-05

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

  1. Claude Opus 4.8 + Managed Agents self-hosted sandbox (public beta) — Anthropic shipped a stronger Opus (coding, agentic reasoning, practical knowledge work) AND opened a public beta letting you run agent tool-execution in your own infrastructure — Cloudflare, Modal, Vercel, Daytona, or your box. For TP3 this is huge: your MCP servers are exactly the kind of private tool environment this is built for. This is the first real "agent runs here, I trust it" option from Anthropic.

  2. Wispr Flow — dictation that works everywhere, 4× faster than typing — Speak instead of type, across every app and field. This is the Bidet dream as a standalone product. Worth benchmarking directly against your Bidet pipeline: Wispr is polished, cross-platform, and reportedly shipping fast. If it works on Windows for you it may leapfrog the current Bidet input layer. Free tier available.

  3. TwelveLabs Rodeo — natural-language video editing copilot — Launched June 1. Describe what you want, Rodeo finds the clips, assembles the cut, and hands off to Premiere / Resolve / Final Cut in NLE-compatible format. Powered by Marengo 3.0 + Pegasus 1.5 (understands visuals, audio, speech, text simultaneously, up to 1-hour videos). If you ever want to do anything with video — lesson content, Breezy Farms footage, DC trip recap — this is the tool to have on your list.

  4. Microsoft MAI-Transcribe-1.5 + MAI-Voice-2 (Build 2026) — Microsoft quietly shipped a full in-house AI model family at Build 2026. MAI-Transcribe-1.5 is their new STT model; MAI-Voice-2 is their TTS with a fast Flash variant. This is a directly competitive shot at Whisper/ElevenLabs from Microsoft's own AI team. For your on-device STT interest and Bidet's transcription core — these are worth watching for benchmarks against Whisper, especially since you're on Windows.

On the radar

# Item What it is Score Why it's on your radar / how it plugs in Link
1 Claude Opus 4.8 + Managed Agents sandbox Stronger Opus model + public beta: run agent tool-execution on your own infra (Cloudflare, Modal, Vercel) 24 TP3's MCP servers are exactly the private-tool sandbox this is built for. Agents that call your endpoints without leaving your control. Anthropic release notes
2 Wispr Flow Dictation everywhere — speak instead of type, 4× faster, works in any app on any field 23 Direct overlap with Bidet's core job. Test it on your Windows stack. Free tier. If it clears your bar, it's the Bidet input layer, polished. Product Hunt
3 TwelveLabs Rodeo AI video copilot — natural language → clip search → edit → export to Premiere/Resolve/FCPx 20 Creative workflow unlock. Understand visuals + audio + speech simultaneously, up to 1-hour clips. Launched June 1, no enterprise setup required. TwelveLabs Rodeo
4 Microsoft MAI model family (Build 2026) 7 in-house models: MAI-Transcribe-1.5 (STT), MAI-Voice-2 (TTS), MAI-Thinking-1 (reasoning), MAI-Code-1, MAI-Image-2.5 20 MAI-Transcribe-1.5 is a direct Whisper competitor from Microsoft. MAI-Voice-2 challenges ElevenLabs. Both are relevant to your STT/Bidet pipeline and on-device interest. Microsoft Build 2026
5 OpenAI Codex Computer Use on Windows Codex can now see, click, and type in Windows desktop apps — GUI automation via natural language 20 You're on Windows, your whole Apex stack runs there. This is agentic automation for native Windows workflows. Potentially relevant to the teacher-workflow automation ideas. OpenAI Codex changelog
6 Gemini Spark rolling out to AI Ultra Google's 24/7 agentic assistant (I/O 2026 debut) — Gmail, Calendar, Docs, custom sub-agents, even authorized payments — hitting AI Ultra subscribers now 20 If you're on Google AI Ultra at $100/mo, this is live now in the US. The Gmail+Calendar integration is directly relevant to your teacher workflow. Worth a first-look before buying in. 9to5Google
7 Stability AI Stable Audio 3.0 4 new fully-licensed audio models; large model generates 6-min+ full compositions, maintains melodic structure throughout 20 Fully licensed (that matters for classroom/teaching use). 6-minute songs beats Suno's current default hard limits. Free to experiment. Great for Breezy Farms ambient or lesson-video soundtracks. TechCrunch
8 GPT-5.5 Instant OpenAI's daily-driver model update: 52.5% fewer hallucinations, more concise (30% fewer words), better personalization via past-chat + Gmail context 19 If you use ChatGPT at all, this is meaningfully less wrong than 5.3 Instant. The hallucination reduction on high-stakes prompts (medicine, law, finance) is the headline stat worth remembering. OpenAI
9 ElevenLabs Music Text-prompt music generation, 7 songs/day free. Create + remix, licensing partnerships, opt-in marketplace for voice/music creators 19 Free tier is genuinely playable. Pairs with ElevenLabs Voice (already in your world). Good for quick Breezy Farms or Bidet demo soundscapes — no cost to try. TechCrunch
10 Perplexity Personal Computer for Windows Local AI agent that works across your files and installed apps — runs on your machine, not the cloud 18 Right in your wheelhouse: local-first, Windows, works with your actual file system. Could complement TP3's local search for anything not yet in Postgres. Worth a look before buying hardware. Product Hunt
11 Microsoft Scout Personal work "autopilot" agent in Microsoft 365 — proactively handles meeting prep, scheduling conflicts, routine tasks in Teams + Outlook 17 Early Frontier access only right now. Less relevant if you're not deep in M365, but it signals where Windows productivity is heading. Teacher-workflow angle: if your school runs M365, this is coming to you. Microsoft Build
12 Seedance 2.0 (ByteDance) Currently the top-ranked AI video gen model on benchmarks — native multi-shot generation, synchronized audio in a single pass, best narrative coherence 15 The state of the art has moved. If you experiment with AI video, Seedance 2.0 and Kling 3.0 (native 4K/60fps/15sec/multilingual lip-sync) are the names to know right now over Sora (which OpenAI discontinued as a consumer app). Pixflow AI Video Guide
13 NVIDIA RTX Spark Arm-based superchip announced at Computex: AI agents + content creation + gaming in a single portable chip, on-device inference native 14 This is hardware that matters for your on-device direction. Not buying anything yet, but it's the clearest signal that on-device AI inference is a first-class design goal for consumer hardware, not an afterthought. Computex / AI news roundup

Coming up on the radar

Cut with reason

Sources scanned

Generated 2026-06-05 19:10 by tp3_scripts/ai_radar/run_radar.ps1.