Creating the best open source models in the world through RLVR and Distillation · Subnet 5 on Biττensor
United States
Joined August 2025
- Tweets18
- Following16
- Followers741
- Likes28
Hone V4 is here⚔️
Miners now download large repos, find bugs, and return their solutions and agent traces to validators.
We harvest this powerful training data to do RLVR on the best open source models.
Let's make sure the future of AI is decentralized.
Pre-training is now cheap relative to reinforcement learning through verifiable rewards (RLVR)- it's just 7% of Anthropic+OpenAI's spend on their models.
Since RLVR is the new bottleneck, we ask: can it be open sourced?
Welcome to @bittensor Subnet 5: Hone.
We're initially focusing on reinforcement learning in coding environments.
Miners solve hard coding problems and we collect the successful rollouts for training our own coding agents.
If you have compute or extra API credits, consider mining on Hone!
Check out our Github here: github.com/hone-subnet-org/h…
Happy Honing⚔️
Hone Training retweeted
Introducing Secure Targon Compute
Today, we’re giving developers the ability to rent H200 and CPU nodes along with scaling workloads as needed, all secured by the Targon Virtual Machine (TVM).
Every instance runs inside a confidential environment powered by Intel TDX, AMD SEV, or NVIDIA Confidential Computing. Targon delivers the speed of modern cloud compute with the privacy guarantees enterprises expect.
targon.com/
Big news! 🗞️ 🌊
We just merged the first version of the codebase for @bittensor subnet 5 : github.com/manifold-inc/hone and we are opening up 1% incentive for our live production test! Validator is now running.
1/3 🧵
There are no constraits on miner solutions. Eg. miner starter code uses an OpenAI key to solve the @arcprize (ARC-AGI 2) problems, and you can:
- use another provider
- more prompt engineering
- finetune your own LLM
- try new methods like HRMs
2/3 🧵
Hone Training retweeted
Hash Rate - Ep 136: Hone ($TAO Subnet 5) Chasing AGI
🧙 Guest: @0xcarro of @traininghone
Can a Bittensor subnet crack AGI before Elon or Sam? Hone (SN5) thinks it can.
00:00 Intro to Hone and AGI
02:52 Understanding the Arc AGI Benchmark
08:48 Introducing JEPA & Hone's Approach to AGI
23:31 The Data Rich Get Data Richer
26:14 The Most Ferocious Form of Capitalism
32:14 Final Thoughts and Future Directions
GM ☀️ Latest @arcprize results :
GPT 5 : 9% vs 7% in public bench
Grok 4 : 22% vs 16% in public bench
Gonna keep optimizing, in parallel we should run the mainnet test soon so we can launch!
🔥🚀
Update on Subnet 5: 🗞️🪡🧶
The core of our synthetic data generation is ARC AGI 1, the difference between that and ARC AGI 2 is the difficulty & complexity of tasks.
For this we use github.com/google/ARC-GEN for ARC AGI 1 tasks generation
🧵 1/5
For ARC AGI 1 (the 1st part in our pipeline) :
- `Gpt5-medium` : on our synthetics (`57%`) vs public ARC benchmark (`56%`)
- `O3-medium` : on our synthetics (`49%`) vs public ARC benchmark (`53%`)
- `Grok-4` : on our synthetics (`60%`) vs public ARC benchmark (`67%`)
🧵 4/5
We are happy to announce Subnet 5, Hone · Hierarchical AI Pretraining, a collaboration between @manifoldlabs & @latentholdings
