Biττensor = decenτralized AI mining proτocol incenτivising creaτion of τruely open, decenτralized, permissionless, safe AI. ~/21Mτ.
Joined February 2013
- Tweets12.7K
- Following411
- Followers2.7K
- Likes24.6K
rich.τ retweeted
October on SN44, in a few lines.
-More public tracking tasks running on the subnet
-3 new public elements out
-A big chunk of example data from the cricket private track is now public
-A lot of quiet work on security and scale
-We can spin up new tasks faster now, with plenty more coming
And following IBC, one of the largest real-time graphics software companies is now testing a Studio endpoint, running on the football model built earlier this year under our previous incentive mechanism.
Meanwhile, we’re collecting feedback from Studio users for what comes next.
From here the pace only picks up, with way more tasks every month and one goal, making computer vision a commodity.
rich.τ retweeted
Unpopular opinion: Everyone is overreacting over the Null Consensus update
1. Subnets like Chutes (64) don’t technically use Yuma consensus the way it was designed, instead have validators audit the subnet validator
2. Most subnets run their own validator and have kappa (majority) stake from child hotkeying, making them consensus
3. Hardware validators are incurring significant extra costs (which weighs on tao price) for the same results as child hotkeying
4. Subnets like Bitcast (93) are able to onboard more than 256 miners through creative workarounds because their growth depends on it. Increasing the UID limit per subnet to 2,500 makes that workaround native to the network
5. The same workaround allows ‘miners’ to register and mine without paying the usual miner registration fee of TAO; necessary for friction-free onboarding and reaching a demographic of miners outside of Bittensor. POW registration now makes that native to the network too
All the recent updates are opt-in meaning optional, not mandatory. For some subnets they will see value to improve certain facets of their business through these features, and others may not. Regardless, these updates are not completely new concepts, but rather legitimizing on chain what subnets have had to build themselves to achieve their mission, and giving those same tools to everyone else.
I noticed there’s been some confusion around this update.
In simple terms it just increases the miner cap per subnet from 256 to 2,500.
This is great for subnets that experience UID pressure (miners hogging UIDs as a way to deter competition) or those with commodities that scale with the amount of miners (speed of data collection, amount of compute providers etc)
rich.τ retweeted
This is exactly right. Soon people will realize why @webuildscore is different. $TAO
Everyone’s building better infra. important work, but a lot of us are building versions of the same thing as said by Atallah
who’s turning that infrastructure into products that actually change how ordinary businesses operate?
customers don’t want an inference stack. they want an outcome.
@markjeffrey
rich.τ retweeted
$TAO
Bittensor devs at @opentensor are already working on quantum-resistant wallets that will be natively supported by the chain 🔥
rich.τ retweeted
> This deserves more attention.
To all the Bittensor video creators out there: could someone make a simple step-by-step tutorial on setting up and using a multisig wallet?
I suspect many of us still don’t fully understand how to do this properly - myself included - yet security should be priority #1 when investing.
A short video showing the process from start to finish could potentially prevent some of the tragedies Const is talking about.
I’d also love to see the teams behind the wallets and apps we use provide more educational content around multisig, Ledger, QR vaults and coldkey security.
We spend countless hours researching where to invest our $TAO. We should spend at least some time learning how to protect it.
Who’s up for making the tutorial? 👀🔐
rich.τ retweeted
Try Kimi K3 with a 1M context in Claude Code through engy.ai.
ANTHROPIC_BASE_URL=api.engy.ai
ANTHROPIC_MODEL=kimi-k3[1m]
engy.ai/docs
rich.τ retweeted
preached decentralised AI in front of the largest French PE and VC funds (hundreds of billions under management in the room).
main takeaways from the COO of a multi-billion-dollar manufacturing company:
- efficiency is the ultimate goal.
- all big-data LLM pilots have been cancelled across all verticals: no clear use cases, no productivity spikes.
- vision is the only vertical where he and his industry counterparts know value can be created, so they want to double down on it.
rich.τ retweeted
@webuildscore's potential customers: factories, warehouses/ports, industrial safety, critical infrastructure/security, retail chains/OSR/malls, sports clubs/broadcasters, rails/airports/municipalities, energy/mining/oil & gas, insurance, agriculture, robotics, world model... $TAO
rich.τ retweeted
Two weeks ago, CrofAI went from "cheapest inference provider in the world" to shutting down.
An independent investigation found that requests for models like Kimi K3 were being routed through OpenRouter to smaller, cheaper models. The price was low because the model wasn't the one you asked for.
If you use any hosted inference provider, three questions are worth asking:
1. Which model actually answered my request?
2. Who served it?
3. Who can read what I sent?
A good provider should be able to answer all three with more than "trust us.
Here's how SayGm answers them.
SayGm is a gateway. We route requests across a competitive network of providers, which is how prices land below list. We say that upfront, because routing isn't the problem. Hiding it is.
The gateway runs inside a hardware-attested TEE (Intel TDX), so the code handling your request can be verified, not just described.
For confidential models, inference runs in within a trusted enclave. Your prompt is sealed off from us, the host, and the model provider. We can't read it, and neither can anyone else.
For frontier models like Claude and GPT, routing is identity-blind: we and the host can't see your prompts, but the model provider's servers still receive them. We'd rather say that plainly than overclaim.
Cheap inference is great. Cheap inference you can verify is better.
And if you were a CrofAI user, the thread in the replies has solid advice on refunds and rotating your API keys.
rich.τ retweeted
Invite friends to test Score Studio and get up to 5 credit resets on your free plan.
👉 scorestudio.ai
rich.τ retweeted
This is so embarrassing for @Google and served as a reminder that billionaires will always side with authoritarians, because authoritarians will always protect billionaires at the expense of democracy and the working class.
Google Maps has changed the name of Lake Ontario to Lake America following Donald Trump’s executive order and amid a trade war with Canada rollingstone.com/politics/po…
rich.τ retweeted
With all due respect, the only person we won't ever be able to detect is @JohnCena
rich.τ retweeted
NOVA Nanobodies: Experimental validation
We're excited by the quality of submissions we've received in our nanobodies competition and are pleased to announce that we’re beginning nanobody production in partnership with @Yalotein.
This marks the start of experimental validation for our nanobodies track, alongside our small molecules program. With this milestone, NOVA now has two therapeutic modalities progressing through both virtual screening and experimental validation.
This milestone reflects a broader vision for NOVA. Rather than building infrastructure around a single target or therapeutic modality, we're building a discovery engine that can support multiple modalities under a common incentive system. We believe this breadth and flexibility is essential for the future of AI-native drug discovery.
rich.τ retweeted
Basically the opportunity Bittensor is built for. Take @webuildscore: computer vision as a headless service that agents can call, pay for on demand, while a decentralized network competes to deliver the output. SaaS → API → agents → usage-based revenue. $TAO
rich.τ retweeted
NOVA Compound - more drug-like submissions
We’re adding the BRENK structural filter to our molecule validity criteria. The filter identifies and removes chemical structures associated with undesirable properties in drug development.
This further constrains the search beyond predicted binding affinity, helping prioritize molecules that don’t just score well computationally, but are associated with better downstream development.
NOVA Nanobodies - greater human-likeness
We’re introducing a minimum human framework score of 0.85 for submitted nanobodies.
This raises the bar for human-likeness in candidate sequences, helping prioritize nanobodies with a better starting profile for therapeutic development while reducing potential downstream humanization requirements.
rich.τ retweeted
we can build this for you guys at Apple. public evals, private data, just tell us your requirements.
gopro are already eyeing our 20MB detection/segmentation models. we’ve built and are running 2B VLMs on edge on very lightweight hardware too.
building a 10B multimodal vision model will be done in a matter of weeks.
and instead of 1 researcher, we have 256 teams competing against each other to deliver the exact model you have in mind.
my dms are open.
We are hiring a researcher who can help us bring powerful video reasoning models to iPhone. Please reach out to me if you are interested. jobs.apple.com/en-us/details…
rich.τ retweeted
Get TAO against your alpha. Without selling a single token.
Your position stays staked, your emissions keep flowing on the entire collateral, and you walk away with liquid TAO.
262 TAO are available to borrow right now against your @webuildscore (SN44) tokens.
Selling is no longer an obligation for builders.
