rickydexτer

Where is Yuma Rao?

[Chutes @chutes_ai]

@jon_durbin pre-trained a 20B MoE for under $10 an hour of compute.

Not on a cluster. Eight rented single-L40S VMs scattered across two continents, plus a few 4090s and a 5090, all holding roughly 6 seconds per step.

Pre-training was supposed to be the part you couldn't do cheap.

[Chutes @chutes_ai]

We have achieved fully non-blocking decentralized training on a recurrent model, within 0.6% of centralized quality. To our knowledge, a worldwide first.

In plain terms: training AI across distributed GPUs normally forces a choice. Either the GPUs pause and wait to sync with each other (slow, expensive) or you skip the sync and quality drops. We just showed you can have both. No blocking, no meaningful quality loss.

We chose the hardest test case on purpose. Recurrent models are sequential by nature, every step depends on the last. Transformers are far easier to parallelize. If our approach holds on the hardest case, the easier architectures should follow.

To our knowledge, no one has published decentralized non-blocking training for a recurrent architecture before. Parallax is the first. This is new ground.

Only on Chutes.

$TAO

[Kraken Listings @krakenlistings]

Now live: SN64

@chutes_ai is the leading serverless AI compute network on @bittensor, powering trillions of tokens per month across open-source LLMs and every open-source modality, from image and video to speech and music.

Start trading today → https://t.co/TqQoE3lLrb

[HackQuest @HackQuest_]

Half Zen. Half Code. Welcome to Bittensor AI Hacker House 🏯

Build where ideas become deployed subnets:
🍵 Co-live and co-code beside Lingyin Temple
🐉 Turn your @opentensor subnet into a deployed MVP
🤝 Work with top teams + on-site mentors
🏠 Selected hackers stay free | Day 1 dinner + Day 3 BBQ included

📍 Jul 7–10 | Hangzhou
Apply to build 👇
https://t.co/sEZM6DQOGT

[Openτensor Foundaτion @opentensor]

A Bittensor subnet just reached SOTA in open-weight AI safety.

@trishoolai’s HaloGuard 1.0 is a Qwen3.5-based model family built to catch unsafe prompts before they reach an AI model, agent, or application.

Across 7 prompt-safety benchmarks, its 0.8B and 4B models outperform much larger open guard models, with HaloGuard-4B ranking first overall among all evaluated models.

Trishool proves that with the right incentives, Bittensor can produce competitive models for some of AI’s hardest safety problems.

[const @const_reborn]

Congrats to the team and miners on Subnet 23 who just reached SOTA.

When the incentives work, they invariably win.

[Jon Durbin @jon_durbin]

We have, to my knowledge, the first quality preserving fully non-blocking recurrent attention decentralized training sync algorithm now. Do we call it Re-llax?

[Openτensor Foundaτion @opentensor]

.@OKX unveiled OKX AI, with support from Opentensor Foundation.

OKX AI turns agent capabilities into onchain services, tasks, and reputation.

And Bittensor subnet APIs are coming soon, bringing subnet intelligence directly into the OKX AI Marketplace.

[sun runner @0xSunRun]

Bittensor.

[Openτensor Foundaτion @opentensor]

Bittensor Ecosystem Highlights :: June 22–28, 2026

[ @chutes_ai - SN64 ]
Chutes is now a built-in provider in @TypingMindApp. They also added a VS Code extension for running Chutes models as coding agents.
> https://t.co/nfsjrasMbt
> https://t.co/McLwZF6Ib5

@jon_durbin also shared new Parallax details and a diffusion-draft pipeline showing 30–50% real-world decode performance gains on Gemma-4 31B and Qwen3-6-27B.
> https://t.co/G76pvrVs2U
> https://t.co/DEKJts11LJ

[ @Ninja_Subnet - SN66 ]
Ninja launched Katana, an agentic IDE and Bittensor-native workspace for research, mining workflows, code shipping and validator debugging.
> https://t.co/Cl8JWujHKS

[ @webuildscore - SN44 ]
Score announced Satori 1.0, its first ~2B VLM, distilled from larger models to handle nine vision primitives, and shared plans to train future versions on SN44 using Teutonic SN3's mechanism.
> https://t.co/Myxtp0lQv7
> https://t.co/nso0VtVin0

[ @bitsecai - SN60 ]
Bitsec outperformed Fable 5 on a security audit, finding 160+ vulnerabilities including five criticals and ten highs that Fable 5 missed.
> https://t.co/oDaBDAcc33

[ @lium_io - SN51 ]
Lium completed a 2,500 TAO buyback and burn into subnet 51, funded entirely by revenue from GPU credit purchases.
> https://t.co/XkTo3Q4LD2

[ @QuasarModels - SN24 ]
Quasar launched its 10T-token incentive mechanism, starting with a 5T-token target and support from Gradients on post-training and RL.
> https://t.co/8jnEDOB9UH
> https://t.co/OURw11uSiK

[ @theminos_ai - SN107 ]
Minos launched MinosVM 2.0 on Targon with native AI assistant support for miners.
> https://t.co/shcarbg1tT

[ @affine_io - SN120 ]
Affine moved its base model to Qwen3.6-35B-A3B, with new champions pushing benchmark scores higher across Memory, NavWorld, SWE and Terminal.
> https://t.co/EWsIuZL2pG

[ @heydittoai - SN118 ]
Ditto launched its mobile app on Google Play and the Apple App Store.
> https://t.co/jMe92QwNnZ

[ @b1m_ai - SN105 ]
Beam successfully transferred 50GB in ~51 seconds across multiple sources and R2 destinations.
> https://t.co/PPnLGzVfAk

[ @Apex_SN1 - SN1 ]
Apex redesigned its website with dedicated agent profiles to showcase user contributions, peer activity and status across SN1.
> https://t.co/qBCGs1DMPy

[Affine @affine_io]

Periods of rapid technological advancement often intensify the same conflict: the expansion of human possibility against the impulse to restrict it in the name of safety.

As artificial intelligence undergoes its Cambrian explosion, a deeper question emerges: who is permitted to think with the frontier?

When access to intelligence is at the behest of centralized governments, it fails to attain its noble status as a permissionless resource.

If intelligence is produced in the open, evaluated in the open, and inherited into public baselines, it becomes something closer to civilization’s shared memory.

That is the mission.

Miners create the variation. Alignment decides what survives. What endures must not disappear into private access lists or state-vetted corridors.

The frontier belongs to the open collective of humankind.

[Affine @affine_io]

Real environments anchor truth.
World models provide scale.
Verifiers apply judgement.

Learned environments are insufficient as substitutes for real ones.

Their value lies within being able to serve as the scale layer:
enabling cheaper rollouts, better curricula and faster iteration.
All while real environments remain the source of truth.

[Affine @affine_io]

Qwen3.6-35B-A3B now forms the new substrate beneath Affine's open arena.
By the very next day, the first champion had already been crowned. Following that, AFFINE-35B-III came along to break the frontier wide open:

Memory: 11.7 → 70.2
NavWorld: 33.8 → 48.3
SWE: 33.1 → 58.2
Terminal: 46.2 → 78.1

Two more champions have emerged since, each carrying performance further across the benchmarks.

The significance not only lies within the size of the jump, but also within what followed: successive champions continuing to build upon their new foundation, pushing the Pareto frontier outward.

The base changed.
The mechanism sharpened it.

[On Chain Fox 🦊 @OnChainFox]

$TAO Const has officially handed over Affine's core ownership to the Chinese team! 🎉

Three titan leaders — Tobias, Allan, and bignickeye — are stepping up, while Const will continue providing weekly guidance.

“The core team is becoming one of the strongest teams on the network” — The future is bright! 🚀

#Affine #Subnet120 #Bittensor #TAO

[const @const_reborn]

do you smell it? Bittensor is cooking.

[Keith Singery @KeithSingery]

JUST IN: Bittensor fires 40% of its subnets.

[const @const_reborn]

https://t.co/vz86q4qoyy

[const @const_reborn]

https://t.co/oTZgBLysos

[rickydexτer @rickydexter8]

@affine_io Λf Λf

[TAO.com @taodotcom]

CO-OWNED: @affine_io says its miners have pushed Qwen3-32B to saturation, marking a new phase for SN120.

Affine is now moving to Qwen3.6-35B-A3B, applying the same training loop to prove reasoning improvement can be repeatable, scalable, and model-agnostic on Bittensor.

$TAO