Sabitlenmiş Tweet
Score - Subnet 44
2.1K posts
Score - Subnet 44
@webuildscore
Making every camera intelligent. An open, permissionless computer vision layer for the real world. Powering @manakoai · $TAO SN44
Katılım Temmuz 2024
693 Takip Edilen9.1K Takipçiler
Score - Subnet 44 retweetledi
summer business dev update:
avia (fuel stations) → converted
reading fc (football) → converted
mettle data (cricket, ex card) → converted
eyeball (youth football) → converted
lavance (car wash) → in progress
cémoi (chocolate) → to be launched
two-a-day (fruit) → on hold
converted = 1 or multi-year agreement signed
remember the ✍️? that was this, plus partnerships we're now finalising with aws (thank you pwc), and integrations of our models and vision skills by leading security and tech companies.
we'll announce each one properly. there's a lot more in the pipeline.
can't stop, won't stop.
English
@webuildscore Perfect, that and the reveal of Max' last '✍️' should give us a good run x)
English
We're building the frontend that will host the decentralised VLM miners will train on sn44.
The model will be monetised from day one, at subnet level.
Users will try it free, then subscribe, through the interface or via API.
Payments will be done at subnet level too.
Note: product branding might change to Score, so it's clearly attached to the subnet and less confusing, this interface is just vibe coded but the scaffolding works.
English
@CaiusScottus The base model is ready.
We're currently collecting and generating enough data for the king-of-the-hill competition to run for months if needed.
English
Latency update, first results.
Four production-ready models live on SN44 right now:
CRIME: 5MB, 70-80ms, ~75% accuracy
FIRE: 9.8MB, 60-70ms, ~90% accuracy
CARWASH: 9.7MB, 60-70ms, ~75% accuracy
ROADSIGN: 9.8MB, 70-80ms, ~85% accuracy
Read those numbers again 👇
Under 10MB. Under 80ms. Production-grade accuracy.
Small enough to run on a camera, a gateway, an embedded board...
That's the frontier we're optimizing: accuracy per byte, per millisecond.
The latency loop enforces it, and miners are converging faster than we expected.
More elements coming.
Score - Subnet 44@webuildscore
👉conformity.scoredata.me/compliance/run… 👉conformity.scoredata.me/compliance/fai…
English
Score - Subnet 44 retweetledi
putting autofrontier to work for real: training satori-8B base.
the frontier: referring-expression grounding (Acc@0.5) on RefCOCO.
genesis: 0.850
bar to break: 0.927 (Qwen2.5-VL-72B)
an 8B chasing a 72B's record, agent hunting angles overnight on one GPU.
3 duels in:
1 crowned → king now 0.920
2 rejected → one posted 0.95, above the bar, and still refused: the edge didn't replicate
repo in comments.
English
Score - Subnet 44 retweetledi
Edge AI just got its incentive layer dialed in.
@webuildscore is turning latency into alpha:
sub-100ms or you don’t get paid and consistency is enforced in real time.
No model swapping, no hiding behind benchmarks...just live, adversarial validation every hour.
This is how you actually bootstrap production grade vision at the edge:
reward what works, slash what doesn’t, iterate fast.
#SN44 is starting to look like where real world AI performance gets priced.
Score - Subnet 44@webuildscore
It's live. Our big speed update is now running on sn44. Why this update? Our whole mission is vision AI that runs right where it's needed, on-device, on-site, on the most efficient compute possible. For that to work in the real world, the models have to be fast and light, not just accurate. So we're now rewarding exactly that: models that are both accurate AND fast enough to run at the edge (sub 100ms). Here's how it works: We set up a dedicated machine that acts like a referee. Once an hour, it quietly tests the best models (and the ones almost good enough to make the leaderboard) by throwing 10 random challenges at each. Miners can fail two ways: their model is too slow, or it gives different answers than it did before (no swapping in a weaker model after they've qualified). If they fail, they simply stop earning rewards for that model until they fix it. It's not a ban. They can just submit a new version and they're right back in. And a slip-up on one task doesn't hurt any of their other submissions. Everything is out in the open. You can see every test we run on our conformity page (links in first comment).
English
QME
It's live. Our big speed update is now running on sn44.
Why this update? Our whole mission is vision AI that runs right where it's needed, on-device, on-site, on the most efficient compute possible.
For that to work in the real world, the models have to be fast and light, not just accurate.
So we're now rewarding exactly that: models that are both accurate AND fast enough to run at the edge (sub 100ms).
Here's how it works:
We set up a dedicated machine that acts like a referee.
Once an hour, it quietly tests the best models (and the ones almost good enough to make the leaderboard) by throwing 10 random challenges at each.
Miners can fail two ways: their model is too slow, or it gives different answers than it did before (no swapping in a weaker model after they've qualified).
If they fail, they simply stop earning rewards for that model until they fix it. It's not a ban.
They can just submit a new version and they're right back in.
And a slip-up on one task doesn't hurt any of their other submissions.
Everything is out in the open. You can see every test we run on our conformity page (links in first comment).
Score - Subnet 44@webuildscore
SUB 60ms!!! Miners, miners, miners… Monday we ship an update: models must be accurate AND fast. The bar? 100ms. It’s Friday. Top miners are already sub 60ms. They beat the deadline before it even existed. Bittensor is crazy, folks.
English
Score - Subnet 44 retweetledi
from perception to action
our VLM takes pixels and words and gives you understanding: detect, segment, describe, reason, count.
a VLA takes the same pixels and words and gives you motion, the actual robot commands to get something done
we're now turning satori-2b, a 2b VLM, into a VLA.
same model now takes one camera view and a plain sentence, "put the bowl on the stove," and drives a robot arm through pick-and-place, drawer-opening, multi-step tasks it never saw, closed-loop, self-correcting until it's done.
from language alone.
at 2b it's landing in the neighborhood of openvla, a 7b model built for this from the start
here's why that matters: a VLM understands the world but can't touch it.
a VLA closes the loop, the same model that reads a scene can now change it. you stop programming robots and start telling them what you want, in plain language, any arm, any task it's never seen, described in a sentence.
and at 2b it's small enough to run on the robot itself and cheap enough to deploy everywhere. that's when a benchmark number turns into real-world work
English
Honoured to be part of this.
@ExploitSummit is shaping up to be one of the most interesting rooms in AI this year, and we’re glad to help build it.
Montreal, September 🤝
Exploit Summit@ExploitSummit
They call themselves the optic nerve of AI. Celebrating @webuildscore as a Gold Sponsor of Exploit. SN44's climb tells you how Bittensor works. It started by predicting football scores - then a network built on constant competition pushed it further: from predictions, to vision skills, to compact models that see the world and run on the camera itself. On Bittensor, you improve or you fall behind. Score kept improving. Physical AI, where the world actually happens. Building with us in Montreal 👇 exploitsummit.com
English
Native Money Market for $TAO?
We’re in!
Forge@ForgeLending
We're proud to announce Forge's official partnership with @webuildscore (SN44). Building a native money market for the $TAO ecosystem means starting with the right partners. Score is one of Bittensor's most established subnets, and we're grateful to @MaxScore and the entire Score team for believing in us. Building new financial infrastructure requires trust. We're excited to take this first step together and look forward to helping Score unlock new ways to access liquidity without compromising long-term alignment.
English
"They're finally training their first VLM" - @const_reborn
That's right jefe, that's happening.
7B → ... → 72B
τim@tm0klc
Subnet 44 @webuildscore is expanding. We’re incentivising training for a new vision-language model: Satori. Satori reasons AND grounds. It doesn’t just answer questions about an image. It points to the evidence. - Reason about scenes - Detect and segment objects - Read text - Count entities - Ground claims in pixels Most VLMs are split: strong reasoning OR strong grounding. Detection models localise, but can’t talk. Chatty VLMs describe fluently, but can’t prove it. Satori sits at the intersection. We’re starting with a 7B base model.
English
Score - Subnet 44 retweetledi
Proud to be part of this first batch alongside top tier teams
Kraken Listings@krakenlistings
Added to the roadmap: Chutes, Hippius, Lium, Score, Targon, Ridges AI and Vanta. See the full roadmap: kraken.com/listings
English
SUB 60ms!!!
Miners, miners, miners…
Monday we ship an update: models must be accurate AND fast. The bar? 100ms.
It’s Friday. Top miners are already sub 60ms.
They beat the deadline before it even existed.
Bittensor is crazy, folks.
English
Miner burn: 24.63%.
More challenges incoming to take it to zero.
Goal: 100% of rewards to miners.
Accelerate.
English
2B was not ambitious enough.
A 72B model would make the world go 👀 on our Decentralised Vision Language Model putting us on par with Qwen, Gemma, or DeepSeek.
More to come.
Max@MaxScore
satori's spatial grounding is moving. tables +45, detection +7, small-object AP nearly doubled. the panel that was furthest from frontier is closing fastest now. 2B is moving fast locally. time to get more ambitious and train something bigger on SN44. 72B would make us run a market leading SOTA vlm. would SparseLoCo work for this? asking for a friend
English
Detect Petrol Station: huggingface.co/manak0/Detect-…
Detect Vehicle: huggingface.co/manak0/Detect-…
Detect Person: huggingface.co/manak0/Detect-…
English
The skills miners distill daily on SN44 might have cracked the single biggest pain point fuel retailers have.
At unmanned stations especially, the question they always ask is:
"Tell me the moment there's a payment terminal issue."
The catch: the payment systems themselves can't tell them in real time. Without cameras, they're often waiting up to 24 hours to find out. And that costs them a lot of money.
Now we can. Here's the recipe:
'Detect Petrol Station' + 'Detect Vehicle' + 'Detect Person' combine to flag it in real time. Satori 1.0 (2B) reasons over the bounding boxes to work out:
- Did the person stay near the terminal?
- Did the nozzle actually reach the car?
- How long did the car sit in the bay?
- How many cars behaved the same way in the last few minutes?
No single skill does this on its own. The model chains them together to spot a problem none of them were trained to detect.
That's why the Vision AI bricks miners build every day are gold for any company using VLMs: you can compose them, and plug them in to sharpen your own model downstream.
Links to the skills on @huggingface in comments 👇
English