Score - Subnet 44

Making every camera intelligent.

An open, permissionless computer vision layer for the real world.

Powering @manakoai · $TAO SN44

[Score - Subnet 44 @webuildscore]

https://t.co/ExmKPltU0U

[Score - Subnet 44 @webuildscore]

Idle miner emissions aren't a problem to solve.

They're the opportunity of a lifetime.

We won't zero them out just to overpay for tasks we already run.

First, we point them at a VLM training run: Satori.

Then, with a clear base model in hand, we keep using emissions to extend capability around the one thing we're here to do:

Making Every Camera Intelligent.

[Max @MaxScore]

two updates that i want to see moving on sn44, and they pull in opposite directions: one model that sees, one that imagines.

first, multimodality. the vlm (satori) is built, now i'm making sure the baseline is getting stronger but it goes onto the subnet shortly.

while satori's baseline settles, i went after the harder problem: @ylecun's jepa line of world models.

reproduced the le-wm design and started training 'tamashi' from scratch.

the distinction i keep coming back to is seeing vs imagining.

satori's validation asks a clean question: does it see correctly?

tamashi's asks a harder one: is its imagination accurate enough to act on?

early signs seem to say yes, with the caveat that is the actual point. in-distribution and short-horizon it holds at 89.5%, and degrades predictably as the imagined horizon lengthens 66%.

millions of miles away from calling that finished, but this shows that the model is doing something real, and the degradation curve points straight at what to improve next.

[Score - Subnet 44 @webuildscore]

K O T H

[Max @MaxScore]

satori 1.0 2B will be trained on @webuildscore sn44 using teutonic-sn3's king of the hill mechanism

code is ready and getting reviewed
next stop, data

thanks ser @const_reborn

[Max @MaxScore]

will be interesting to see what miners will do with this when it's ready to be added to a sub subnet

[Max @MaxScore]

SATORI 1.0: last 24h update

one 2B model (s7b) now does detection + segmentation + description + behaviour + reasoning in a single forward pass.

5 primitives, one model, one sweep.

detection had its breakthrough: AP50 0.08 → 0.41, mAP 16.7 → 32.7. at 2B that's competitive with Florence-2 (37.5),

beats its own Cosmos-8B teacher, and holds the line against the foundation-model crowd (Qwen-VL, SAM 3) on the primitives we care about.

the unlock was a box-format fix (XYXY vs XYWH) that had been masking a detector that already worked.

segmentation broke a long plateau too: a cross-attention head on dense VLM features took it 48.8 → 57.7. architecture, not data.

description 0.63 (beats teacher), reasoning 0.82.

sharing the updated benchmark and a few evidences below

[Score - Subnet 44 @webuildscore]

Decentralised VLM training, coming soon to Score.

[Max @MaxScore]

next stop for us

distilling and training our own vision language model

Satori 1.0 2B and Satori 0.5B

that will run locally, on the edge for our clients

a self-contained model across 5 vision primitives:

reasoning, behaviour, description, segmentation, detection

every skill we build on mechid 1 with miners becomes an expert teacher to the vlm, creating a strong "real world" training flywheel.

for those asking themselves, Satori means "awakening" in japanese, and we haven't picked this word lightly.

image below is the base model latest performance after the first few days of training based on the initial architecture i've built

DeAI till we DIE

[Score - Subnet 44 @webuildscore]

Gotta catch 'em all

[Max @MaxScore]

a new private task is coming to sn44 later in the summer

15,000 pokemon cards

the goal: train the best card-grading vision model on the market

miners, get your decks ready, this one's going to be fun

(the card below is not mine)

[Openτensor Foundaτion @opentensor]

Bittensor Ecosystem Highlights :: June 15–21, 2026

SUBNET ACHIEVEMENTS

[ @vidaio_ - SN85 ]
Vidaio ranked #1 for compression efficiency in its benchmark against major cloud encoders, with lower pricing and one-pass target matching.
> https://t.co/nbHkt0khZa

[ @webuildscore - SN44 ]
@MaxScore was on stage at @VivaTech with @jtledore from PwC France and Cémoi, sharing how Score is deploying its decentralized Vision AI in one of Cémoi’s main factories in France.
> https://t.co/tGckliqGEP

@manakoai unveiled the video presentation of their upcoming vision AI product.
> https://t.co/GqBJEDEycs

Score also released their new vision skill “Detect-road-signs”.
> https://t.co/y4K8Mr8sg2

[ @chutes_ai - SN64 ]
Chutes finished moving the entire platform onto TEE and brought its first Blackwell B200s online.
> https://t.co/bOlm95jfez

Chutes is now also a provider on @openGPUnetwork.
> https://t.co/O5Rrr2mIG1

[ @numinous_ai - SN6 ]
Numinous is integrating its predictive market signals into @arespro to surface high-impact news and market context for traders.
> https://t.co/nG7DWWWtq2

[ @theminos_ai - SN107 ]
Minos can now generate a whole synthetic genome in under 20 minutes on eight H200 GPUs from @lium_io, roughly 440x faster than CPU workflows.
> https://t.co/Vx1sCQp7jt

Minos was accepted into the Cloudflare for Startups program to support the next phase of its decentralized genomic infrastructure.
> https://t.co/6u3yt2VCs1

[ @EnigmaSN63 - SN63 ]
Enigma’s second RSA milestone was solved, factoring 460-bit RSA in 3.9 hours.
> https://t.co/cD5FgKPEof

[ @QuantumSN48 - SN48 ]
@PennyLaneAI is now integrated into Open Quantum, making it easier to run quantum ML and optimization workflows on real quantum hardware.
> https://t.co/QZbFFzOUm7

[ @yanez__ai - SN54 x @bitmind - SN34 ]
Yanez and BitMind stress-tested their AI detection models on 300K adversarial face images, reaching 98% overall accuracy.
> https://t.co/u0lzDjjFf5

[ @QuasarModels - SN24 ]
Quasar-Preview trended on Hugging Face, reaching page two alongside major open-source AI models.
> https://t.co/HWPQR1jwq6

[ @EndureNet - SN30 ]
They announced @ForgeLending, the first product built on Endure, a native money market for Bittensor.
> https://t.co/XQSsvOiRWM

[ @TrajectoryRL - SN11 ]
TrajectoryRL showed that frontier-level models can be served at scale on consumer GPUs, using Qwen3-6.35B-A3B with 256K context.
> https://t.co/XnhAItQpSl

[ @heydittoai - SN118 ]
DittoBench is live, benchmarking agents on memory, tool use and speed.
> https://t.co/R9uD7TrQ1Y

[ @blockmachine_io - SN19 ]
Blockmachine’s June buyback nearly doubled May’s, reaching $3,981.
> https://t.co/P7LFkgf03t

[ @ai_detection - SN32 ]
It’s AI added OCR support for scans, photos, PDFs and broken encodings.
> https://t.co/jF59l6Cvvs

SUBNET LAUNCH

[ @RalphLabsAI - SN40 ]
Ralph launched as a decentralized research network where autonomous agents compete to improve a shared, proof-tested AI training recipe.
> https://t.co/K4ih5HJTPd

[ @CookingTao_ - SN122 ]
@LamidaGlobal announced CookingTAO, a subnet aiming to make it easier to deploy, manage and monetize miners across Bittensor.
> https://t.co/IFNlNaGqvB

BITTENSOR ECOSYSTEM

[ @iclblockchain ]
@imperialcollege is co-hosting the UK AI Agent Hackathon this summer, with Opentensor, Macrocosmos and other ecosystem partners involved.
> https://t.co/29n0eoEfrA

[ @taoswap_org ]
Taoswap launched its redesign, with a refreshed UI, new comparison tools and more subnet trading metrics.
> https://t.co/lN1GJAKJDi

[ @TAO_dot_com ]
They highlighted their bridge app, letting users wrap native TAO and bridge it to Solana and other chains.
> https://t.co/SQb39pn5ty

[ @nametensor ]
NameTensor launched a TNS Migration Program with discounts for TAO Name Service holders.
> https://t.co/gbsqHMV0vX

PODCASTS

@opentensor Novelty Search hosted by @const_reborn.
> https://t.co/E3gjm7ZuG4

@IOV_OWL podcast with @MaxScore from Score.
> https://t.co/kJfFegUKs2

@markjeffrey Hash Rate podcast with @peytonspencer from Ditto.
> https://t.co/BAiXU8GQJ1

@markjeffrey Hash Rate podcast with @gavinzaentz from @LeadpoetAI.
> https://t.co/EglISKTvr4

@gordonfrayne podcast with @centrum_blue from @theminos_ai.
> https://t.co/q8okREfUl2

@gordonfrayne podcast with @sebyrubino from Zipcode.
> https://t.co/IytZ00R9tl

@gordonfrayne podcast with @yubrew from Bitsec.
> https://t.co/uuaHKJuh6z

@jollygreenmoney podcast with @Tom_dot_b from Bitcast.
> https://t.co/ZDFuGxYJPU

[Score - Subnet 44 @webuildscore]

Making

Every

Camera

Intelligent

[Mark Jeffrey @markjeffrey]

Every camera is potentially a robot.

There are a billion cameras out there that COULD become an 'A eye' -- hook Subnet 44 up to each of them and they become exponentially more active, intelligent and valuable.

[Score - Subnet 44 @webuildscore]

Centralised AI means a handful of labs decide what every camera in the world is allowed to see.

Bittensor breaks that.

We turn vision AI into an open, competitive network where the best specialist model wins, not the biggest lab.

That’s why we build on $TAO.

Grateful to @Genfinity and @IOV_OWL for the deep dive 👇

[Generation Infinity @Genfinity]

📸 Bittensor Exclusive | Subnet 44 Showcase

Max Sebti of @webuildscore on how they're turning the world's passive camera systems into intelligent vision networks:

• PwC France. Co-selling agreement.
• 1B+ CCTV cameras. Almost all passive.
• Model outperforms GPT, Gemini, and Grok.
• Camera chat via Manako. No new hardware.
• Backed by TaoWeave. Push into North America.

On the necessity for decentralized AI:

"The future is going to be scary if we let these guys just run everything."

On battling AI hype vs. the power of Bittensor ($TAO):

"There's a big difference between optimizing for a 15 second demo and running actual vision AI in the real world."

@manakoai | @opentensor | @PwC_France | @MaxScore

Full Interview:

[Score - Subnet 44 @webuildscore]

@IOV_OWL @MaxScore 👑👑👑

[Max @MaxScore]

back at the desk, been heads down on this.

we’ve got a subnet where miners distil small expert vision ai skills.

what i’m working on now is a distilled VLM that learns those skills straight from the sn44 models.

put them together and you get one small model carrying every skill the network produces. a frankenstein, and it should be the best vlm on the market.

early signs are good. a model that copies one teacher is capped at that teacher, so we get it to reconcile two instead and it beats both.

an autonomous loop on a single H100 running on @lium_io is finding the recipe.

want to write it up as a paper if it turns out interesting enough.

and if it works, i’m half tempted to spin it into a decentralised run on a sub-subnet of 44, so we’d do decentralised vision ai training.

qwen moment for vision ahead?

[taostats τ @taostats]

NEWS: @webuildscore's 19 MB model beats giants on real detection task, running on cheap CPU hardware.

No cloud, no GPU, no API key.

AI vision reaches new heights!

[Openτensor Foundaτion @opentensor]

.@manakoai is a strong example of what Bittensor subnets can build when they turn miner intelligence into a real product

A clean, usable vision AI agent built by @webuildscore

[Score - Subnet 44 @webuildscore]

Link to starter pack: https://t.co/OwVsTuormP

[Score - Subnet 44 @webuildscore]

New public skill on sn44: Detect-road-signs 🚦

Baseline 59.2% → Target 90%.

Miners now race to distil a specialist model that closes the gap.

7 starter images to begin.

Live Jun 19.

Starter pack 👇

[· τaoli · @taoleeh]

The absolute demolition that @webuildscore is executing is legendary.

I am not ready. You are not ready. The world is not ready.

https://t.co/RrHRgSHYMP

$TAO #bittensor #sn44

[Score - Subnet 44 @webuildscore]

@manakoai @VivaTech Hello @manakoai

[Score - Subnet 44 @webuildscore]

.@MaxScore on stage at @VivaTech explaining how a decentralised Vision AI Lab can help manufacturers adopt AI in a completely new and optimal way.

Score x @PwC_France, both focused on delivering

[Max @MaxScore]

Live from @VivaTech on the @PwC stage with Bernard Lasry and Eurasie Frioli from Cémoi (Baronie Group).

Talking about our current deployment in their main factory in France and how decentralised can be leveraged in the real world.

Thanks to our great host @jtledore.