howdy, it’s Barsee again.

happy thursday, AI family, and welcome back to AI Valley.

here are the biggest things worth knowing today:

• Dyson made a $499 AI toothbrush

• Perplexity is bringing local models to Computer on Mac

• Meta released Muse Spark 1.3

• Gemini can now decide which bits of a video it actually needs to watch

• Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

BRIEF AI

Customers can’t use prototypes

PMs can build strong prototypes around a clear customer need, but a prototype isn't a system. It doesn't carry the constraints, tradeoffs, or paths engineering tried and rejected.

Brief closes that gap. It keeps customer evidence, product decisions, and engineering constraints attached to work. So prototypes that reach engineering already reflects what the codebase can do.

Get started for free

_{*This is sponsored}

THROUGH THE VALLEY

1/ Dyson made a $499 AI toothbrush - the CameraJet has a tiny camera in the brush head that takes 28 images a second, uses on-device AI trained on 470K dental images to spot the gaps between your teeth, then shoots mouthrinse into them within 100 milliseconds. So you basically brush normally and it does the flossing-ish bit at the same time. (watch the launch video here)

Dyson spent 6 years, 661 engineers and 38 patents building this thing. They even made synthetic plaque with the National University of Singapore and special low-foam toothpaste so the camera can actually see inside your mouth.

Dyson says it removes up to 69% more plaque in hard-to-reach areas than premium electric toothbrushes (in lab tests). A 4-week study also had 74% of CameraJet users reach “healthy gums” vs 6% using a manual toothbrush.

$499 for a toothbrush is pretty wild, but putting a camera + AI + tiny mouthrinse cannon inside one is also pretty wild.

2/ Perplexity runs a model locally on your Mac - a new Hybrid Mode will let Computer use models running directly on your Mac for some tasks, while still using cloud models for the harder stuff. The local steps don’t use Computer credits and the data involved stays on your machine.

There are three local models being prepared, including a ~19GB Perplexity model, Qwen 32B and a smaller Gemma model that works on Macs with 16GB of memory.

There’s also a Privacy Gate that uses another local model to check what Computer is about to send to the cloud. If it spots personal/sensitive stuff, it can ask you what to do with it first.

Basically, Computer is becoming less of a cloud agent and more of an orchestrator between frontier models and whatever you can run locally. No release date yet though.

3/ Meta says its models are finally catching up - Muse Spark 1.3 is Meta’s new best model, with a big jump in coding and agentic stuff. Alexandr Wang says it’s competitive with Claude Fable 5.1, better than GPT-5.6 Sol (especially at coding) and better than any of the current open-source models. Obviously, those are Meta’s claims.

It also uses 25% fewer tokens, can keep multiple workflows running at once and is better at following long instructions/remembering stuff across tasks. It’ll ask before doing anything irreversible too, which seems useful for an agent.

Developers can use it through Meta’s API now, and it’s coming to Meta AI, Instagram, and Facebook soon. Meta hasn’t decided whether to open-source the weights yet, but Muse Spark 1.2’s weights are still coming.

And Watermelon, Meta’s bigger model, is apparently still on track. No date yet.

4/ Gemini can now decide which bits of a video it actually needs to watch - Google added “agentic” video understanding to Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Instead of processing a whole video at a fixed frame rate, Gemini can jump around between different moments, frames, audio and transcripts, then go back and look closer when it needs to.

Turns out you don’t need to watch everything to understand a video. Google says this uses up to 88% fewer tokens, costs up to 66% less and actually improves accuracy by up to 7%.

It gets more interesting with long videos. Gemini can search through hours of footage for one specific moment, rewatch fast movements at a higher frame rate, count objects/actions and look closer when it spots something weird.

Basically, Gemini now watches videos a bit more like we do: skip the boring bits and pay attention when something matters.

It’s available through the Gemini API now, and coming to the Gemini app and Ask YouTube later.

TRENDING TOOLS

• Gemini Agentic Video - searches long videos by deciding which moments, speeds, frames, audio, or transcript it needs to inspect for your question

• Cohere Parse 5 - turns unstructured documents into structured, AI-ready data using vision language parsing across nine languages

• Muse Voice Transcribe - Meta’s live speech-to-text AI, tracks 20+ speakers

• Manus - resumes independent operations away from Meta

• Ato - A screen-free voice companion for older adults that offers friendly conversations, proactive check-ins, helpful reminders, and family messages

• Conveo AI - AI interviewer holds in-depth video conversations with thousands of real consumers, so brands learn in days what used to take months

• Higgsfield Genjutsu - upload your video, add your character, and transfer the exact motion in a few clicks: acting, lip sync, camera movement, and VFX, with every detail held consistent

• Harvey Horizon Scanning - it monitors 12,000+ sources across 100+ jurisdictions and surfaces updates with impact and risk assessment

• ThunderPhone -builds reliable AI phone agents starting at 2¢ per minute with models included across 47 languages

WHAT I'M CONSUMING

• Making the internet boring

THE VALLEY GEMS

What’s trending on social today:

THAT’S ALL FOR TODAY

Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee

👍️ New reader? Subscribe here

Thanks for being here.

REACH 100K+ READERS

Acquire new customers and drive revenue by partnering with us

Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.

If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.