biru (@designerbiru) header
biru (@designerbiru) profile picture

🚀 Designer working in #bittensor ecosystem. $TAO 📍 Branding · UIUX · Motion. DM for work 📩

Bittensor
Joined December 2023
Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)Photo by biru (@designerbiru)
$TAO : The more I go down the #Bittensor rabbit hole, the more I love it 🙏 Over a year designing for this ecosystem : Today I'm stepping in front of the work. Bittensor deserves world-class design. I'm here for it. OGs, subnet owners, builders, miners, validators : Let's connect. 👇
12
13
2
63
4,416
Why run a data pipeline as a Bittensor competition? The judge numbers are the argument. On claims (SN111, @DeSciClaims ), miners compete against the team's reference pipeline while two independent LLM judges review each claim blind. The share of accepted claims both judges approve without a tiebreak went from 80% in September to 97% by late October. Where a miner and the reference pipeline disagree, the judges now side with the miner three times in four, up from six in ten. That's judge agreement, not scientific accuracy. What stands out is the mechanism: every win becomes the standard the whole subnet is scored against next.
Our new competition is live! 31 submissions are competing. We'll run the winner in production next week. 🚀
1
5
185
A summary of a paper drops exactly what qualifies the evidence. Claims (Bittensor SN111) turns each paper into a structured record : every claim, tied to the evidence behind it, with the source sentences quoted verbatim. In their worked example, a Nature Medicine DMT trial in 34 adults with depression, the reference pipeline found 5 claims. The adjudicated record holds 14, 11 with the statistics attached, backed by 98 verbatim quotes. It includes a caveat the pipeline missed: at three months, the one-dose group was ahead of the two-dose group. The point isn't whether the drug works. It's what a summary leaves out. @DeSciClaims
Our new competition is live! 31 submissions are competing. We'll run the winner in production next week. 🚀
1
1
107
Lending runs on trust, so the security work has to come before mainnet, not after. For @ForgeLending that's three separate layers: TAO.com audited the wstAlpha wrappers and the BittensorEVM stack Forge Builds on, Bailsec is auditing Forge itself, and the contracts are non-custodial. At launch, the Forge Guardian Multisig sets the risk parameters, with Endure's risk intelligence integrated gradually.
Heat’s rising. Bittensor is about to get liquid. Mid-November.
1
76
The default way to get liquidity out of Alpha is to sell it. For a subnet owner it's worse, because selling their own Alpha to pay for operations can read as a sign of weakness in the network they run. Forge, a Bittensor-native money market, is built for another route: borrow TAO against your Alpha without selling it, with the TAO supplied by depositors on the other side of the market. Follow @ForgeLending and get ready for mainnet.
Heat’s rising. Bittensor is about to get liquid. Mid-November.
1
7
170
A single AI percentage doesn't tell you much about a text. @ai_detection 's It's AI (Bittensor SN32) breaks the score down by document, sentence and word. It ranked #1 by average ROC-AUC among the detectors evaluated on the MGTD benchmark, and now covers five languages: German, French and Spanish join English and Arabic. On DetectRL-X it reached up to 99.8% detection accuracy, with false-positive rates below 1% on student and learner writing in all four language tested.
2
67
Most people assume compressing a model costs you quality. In TPN's Competition 006, the winner scored 74.93% against 66.57% for the model it was built from, at 7.71GB instead of 16.99 GB. Same sealed benchmark, same epoch, and neither model had seen the test. The scores are on TPN's public dashboard if you want to check. Smaller models that hold up are what make local AI practical. Gratis, the new free app from @TPN_Labs , runs open source models on hardware you already own, so there's no usage limit. Try one in your browser: gratis.trueperformancenetwor…
We're excited to introduce @Gratis_AI. Run opensource AI models locally on your machine for free, with no setup required. Gratis runs models from @huggingface directly in a browser tab or desktop app for more power, making every chat private by design. TPN builds small, capable models; gratis runs them. gratis-ai.com
3
91
What are you building on Bittensor?
38
The question every subnet gets : Does it actually produce anything you can use? @TPN_Labs runs SN65 as open competitions. Independent teams compress a model to fit under a fixed memory ceiling, then get scored on a sealed benchmark. Winning models go to Hugging face under an open license. Now there's Gratis, which just launched. It runs open source models. TPN's included, in a browser tab: free, no install, no account, and nothing you type leaves your machine. Run a model yourself : gratis.trueperformancenetwor…
We're excited to introduce @Gratis_AI. Run opensource AI models locally on your machine for free, with no setup required. Gratis runs models from @huggingface directly in a browser tab or desktop app for more power, making every chat private by design. TPN builds small, capable models; gratis runs them. gratis-ai.com
4
150
The interesting part of moving data is what happens once it arrives. Beam makes transfers programmable. Webhooks let HTTP-based services take part without running the full protocol, and encrypted Multicasr Rooms give machines, apps and AI agents a way to talk. Reported demo : 10 GB, R2 to R2, in 21 seconds. Not a guarantee, results depend on endpoints, network conditions and workers. Beam studio, the SDKs and the CLI are there for builders to try. @b1m_ai
An open AI economy needs infrastructure that connects data, models and machines. Our vision is to make data movement open and programmable, so builders can put data to work wherever it is needed. Beam is building the transport layer for that future.
1
1
6
122
How is everyone doing on the community?
3
89
Compute is only useful if the data can get to it. Datasets, checkpoints, model weights, results : all of it has to move between clouds, storage and machines first. That's the layer @b1m_ai is building on Bittensor Subnet 105. Independent operators compete to move data between existing endpoints, and validators score delivery performance. Beam doesn't store your files, it coordinates moving them between the systems that already do.
An open AI economy needs infrastructure that connects data, models and machines. Our vision is to make data movement open and programmable, so builders can put data to work wherever it is needed. Beam is building the transport layer for that future.
1
1
15
267
Unbiased, yet true as always @TaoOutsider . You're the man 🙌
I’ve never lied to you, and I’m not starting now. This bittensor:native chart looks ugly AF. Lose this area and we’re likely heading down to clear that liquidity between $235 and $245. But there’s a risk that selling pressure pushes us closer to $200. Come on, let’s get back to $270+ quickly, at least.
3
81
When an AI model gets fooled, the cost lands on someone : The Bank, the insurer, the exchange. Perturb is built to find those weak spots before attackers do. Testing produces a robustness score, a vulnerability heatmap and verified attack data that can help retrain the model. Image-model testing is live today. Text and LLM attacks are on the roadmap. @perturbaix
4
116
A fine day on @bittensor , what are you building currently?
2
1
8
355
What's the smallest change you can make to an image that makes an AI model confidently wrong? On Bittensor Subnet 26, @perturbaix pays miners to find it. Validators check the answer really flips and the change stays imperceptible, copies earn nothing, and the best attackers earn the most TAO. Fight AI with AI
55
S-Tier, Tier list 😁 @JesusMartinez
I Ranked 20+ AI Crypto Coins Thoughts?
1
45
A lot of subnet teams talk about what they'll build. The team behind Bittensor SN45 shipped a product with paying customers. @alpharidge_ai is a trading research and automation platform : charts with built-in ratings, a screener covering about 4,500 US stock plus crypto, RidgeScript for running TradingView Pine scripts, and an AI analyst that can build scripts, screens and alerts. Its news and events engine is powered by SN45. Sponsored by AlphaRidge.
3
100
Most AI answers a question and stops. Ditto V2 (Live now) runs workflows you set up in one sentence, on a schedule. Anything that sends, spends or deletes waits for your eyes. $1 free credit, no card : heyditto.ai @heydittoai
1
41
What is the ideal TAO holding to retire?
1
5
104
Ditto V2 is live. I told Claude about a project, then asked Ditto about it from a different tool, and it already knew. One shared memory across the AIs I use, so I'm not repeating myself all day. Start free at heyditto.ai @heydittoai
1
1
8
218