there’s one team on earth turning vision foundation models into deployable specialists at industrial scale.
it isn’t OpenAI.
it isn’t Meta.
it’s us.
and it’s happening on sn44 on bittensor $TAO
Imagine a widget on your online store that says "You might also like..." — except instead of being based on basic rules, it's powered by AI models like ChatGPT, Claude & Gemini constantly competing to make better picks.
That's @Bitrecs. More relevant suggestions = more add-to-carts = more revenue. Simple.
https://t.co/zIsRdMiKSS
You know how Amazon shows "Customers also bought" and it's somehow always spot on? That's a recommendation engine — and it drives a huge chunk of their revenue.
@Bitrecs brings that same tech to your Shopify store. A widget that shows shoppers the right products at the right moment, powered by the world's best AI models competing to get it right.
https://t.co/zIsRdMiKSS
Vidaio x Pip Studios — Joint Venture
We are beyond excited to announce this joint venture! This is bigger than a partnership. It’s a gateway to the global media ecosystem.
Through this joint venture, VidaioOS is now positioned directly within Pip Studios’ network spanning major studios, platforms, and content owners, including Netflix, Amazon, Sony, Universal, Paramount, and more.
This means:
→ Direct access to enterprise workflows
→ Real-world video workloads entering the network
→ Faster adoption at scale
Pip Studios is part of the TPN (Trusted Partner Network), the industry standard for secure content handling. That places VidaioOS inside trusted pipelines used by the world’s biggest players.
This isn’t about selling individual tools, its about embedding an AI-native video infrastructure layer across an existing global client base. Think of this as a gateway to 50+ major partnerships.
We’re not just building the future of video infrastructure; we’re now connected to where it already operates.
SN97 Arena v3 is live.
Instead of pretending miners won’t optimize the eval, we designed the eval so optimization becomes productive.
That is the Goodhart problem in reverse. If the benchmark is broad, procedural, adversarial, and tied to real model behavior, then “overfitting” starts to look a lot like building a better model.
KL still matters, but it is no longer the center of gravity.
SN97 is moving toward evals that reward durable capability, not leaderboard tricks.
Everyone told us real estate was too relationship-driven to automate.
Today, RESI closed its first week with:
→ 120 users on the portal
→ 1,000+ AI appraisals
→ A newly signed partnership with a nationwide lender
Changing the real estate industry one appraisal at a time.
Unfortunately, things didn't work out for @HermesSubnet. Which is a shame because their Pilot interface could have become:
→ the default Web3 client
→ the main on-ramp for new Web3 users
→ the reasoning layer unifying fragmented Web3
→ the world's primary interface for crypto data, for both users and other agents
→ a Google for agents, and the benchmark for A2A
4/
Unforunately, the terminal I rented crashed several times and I was forced to spend days retraining the model from scratch each time (newbie user error probably?). After 7-8 terminal crashes wiping everything, and having to start over each time, the final crash was one too many. Proud to say I did finally get a few scores, if only very low ones.
Happy to share insights and battle scars if anyone else is working on Score mining. I learned a lot along the way but renting a H100 can get expensive, especially since I didn't earn a cent! Sometimes you have to know when to stop.
1/
I've spent the past month or so trying to build a competitive miner on @webuildscore's Detect Football Event track. I failed. Here's a few things I tried.🧵
I've been trying to mine on SN44 @webuildscore using a @TargonCompute GPU rental for a couple of weeks now and running into a frustrating pattern — my container keeps resetting and wiping everything, losing all my progress each time. Has happened at least 5 times now 😅
My setup is a jump server tunnelled to a Targon H100 (for security). Each reset means reinstalling the entire Python stack, re-downloading training data, and rebuilding from backups.
Still learning the ropes and vibe mining at this point, but making some progress between crashes!
Has anyone else dealt with Targon container instability? Any tips for making miners more resilient to resets? Would love to hear how I might prevent further crashes.
$TAO
3/
Things I tried that didn't work:
• Used Gemini AI to create a live timestamped football commentary of the validator's challenge video clip
• Pose estimation to detect players going down (players are 30 pixels tall)
• Ball tracking (the ball is 3-5 pixels, basically invisible)
• Optical flow as a filter (killed more true positives than false positives)
• VLM pseudo-labels built from behavioural cues — 11,000 events, almost entirely noise
I also spent a couple of weeks manually annotating match events, which showed some promise. But...