Open Vision AI for everyone. Score Studio + SN44
Joined July 2024
- Tweets2.3K
- Following720
- Followers10K
- Likes7K
Pinned Tweet
We are ready to build worlds.
Currently fire-testing the new incentive mechanism that will get miners to build our first ever generative world model, capturing real physics, long-term consistency, and true interactivity.
This is the unreal next chapter for vision AI: from understanding scenes to generating entire, livable realities.
This 👇 is a sneak peek of the subnet’s console, where users will be able to create worlds from a single prompt and very soon explore them right in their browsers.
This will also 100x our ability to train better vision models by providing unlimited simulated scenarios.
44 we go.
He woke up like this
computer vision companies are dying from long use-case → model → deploy feedback loops.
they run out of runway before they can even find the right use case.
meanwhile CV is being commoditised in real time on sn44.
every niche will be covered.
if you run a computer vision company, the miners are coming for you or are here to help depending on how you want to approach this.
Orange juice production line, tracked end to end 🍊
On a food production line, every unit that passes the camera matters. What goes in, what jams or falls off, and what comes out at the end usually gets checked by eye, or not at all.
Our vision models can handle it. Track every orange from the sorting line to the juicer, and any gap can be flagged before it turns into waste.
Pulp fiction, now with bounding boxes.
We set the target at 90%. Miners came back with 98%.
Fallen person detection on SN44 went from a 43.6% baseline to 98% on crowded street footage, scored in public every hour. Open weights.
So we're raising the bar. Harder footage goes in next, and the difficulty keeps climbing from here.
A farmer drove his tractor into his neighbor's burning field and cut a firebreak while the fire was coming straight at him.
Not every field has a guy with a tractor next door. Our Fire and Smoke Detector is open source. Put it on a drone or a helicopter cam and it flags the smoke way before it gets to this point.
We can’t wait to see miners obliterate this baseline
this short clip is a big moment for us.
we generated it with our world-model baseline, marking a first tangible step in our journey from computer vision toward spatial intelligence.
recognizing an aisle is familiar territory. generating what that place could look like as a camera moves through it opens a different set of possibilities.
we supplied one AI-generated store image and a camera path and our baseline, produced 49 frames at 512×512, alongside an estimate of the visible 3D surfaces.
artifacts are obvious. unseen views are predictions, and reliable physical simulation remains ahead of us.
but seeing this first walkthrough emerge from our own pipeline feels amazing.
it’s a glimpse of what we’re building toward: worlds we can explore, layouts we can test, and environments where cameras and robots can learn before deployment.
imagine where miners will take this.
They said Studio was too complex, a bit boring, and too much like the competition.
So we’ve torn it apart to build the smartest Vision AI companion powered by Bittensor: Nokta.
It will make old Vision AI tools look like what they are…old.
NEWS: @webuildscore is rebuilding Score Studio on Subnet 44 around early-user feedback, with a relaunch planned across main platforms. The team reports 143 individual free users and a 3% conversion rate, excluding B2B partners.
xitter.cf/webuildscore/status/21…
Can confirm the Vehicle Detector works on Rocket League. Can also confirm this is not what it was built for. Back to work.
Opened my laptop to work. Ended up running the Vehicle Detector from @webuildscore on every Rocket League replay I have. Send help, or more replays.
Score Studio update:
143 individual free users. 3% converted*
Early users gave us a lot of feedback, so we’re rebuilding the product around it.
The changes are big enough that we’ll relaunch across the main platforms.
Goal: the smartest Vision AI companion on the market.
*Individual users only. Does not include B2B partners paying for subnet outcomes.
Detect Football Event is live on Subnet 44.
The footage comes from youth league games. Low resolution, a single wide camera, small players and a ball that's sometimes just a few pixels.
Miners have to pick up every event in the match as it happens, from passes and shots to saves, ball out of play and more. The console then lines their predictions up against what actually happened.
The best score so far is 53.1%, and the target is 85%. New evaluation every hour, and the gap is closing match after match.
Seems like a new model-training run is going to be in town soon...
.@manakoai Rig builds custom harnesses and vision agents from natural language for our clients.
to do this, it leverages SOTA open-source models from @webuildscore.
the next step is for the miners to get annotated data from pseudo-real and real-world scenes and reconstruct them in 3D by training our own world model.
the goal?
understanding and predicting actions and their consequences in the real world, so we can operate real-world businesses autonomously.
can't stop, won't stop.
Somewhere right now, someone in a packhouse is counting apples by eye, shift after shift, while thousands roll past.
We gave that job to our General Object Detector, straight out of the box inside Score Studio.
No custom training, no special setup, and it's already picking out apples in a pile where half of them are hiding behind the next one.
It only gets sharper from here.
What should we teach it to count next?
