const

Building neurons, turning the web into a brain.
Missionary from the Church of RAO @opentensor

[blockmachine @blockmachine_io]

32.9B requests served in under 3 months live.

Every one of them cryptographically verified before it reached a customer. That’s what decentralized RPC done right looks like.

[Jon Durbin @jon_durbin]

The world doesn't need YALMs (yet another LLM). Nor does it need an LLM "but trained decentralized". The world needs a truly sovereign, secure, open source platform for building, serving, and fine-tuning models that is always available, to everyone, forever, borderlessly.

The mission could not be more clear.

[Jon Durbin @jon_durbin]

Friends don't let friends distill. There's a better way to do synthetics.

Distillation only gets you as far as the teacher, re-encodes the teacher(s) knowledge gaps, isn't verifiable, etc.

We instead build a massive graph database of real grounded sources and derive the question from the answer, instead of the other way around, verifying every step along the way.

This is the way.

[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

[Algod @AlgodTrading]

Be the breakout subnet you want to see

[Vaqxai @vaqxai]

https://t.co/I158Xnl2jn

[Good Morning @say_gm_]

We just let the first 50 on the waitlist in.

gm is live on mainnet and the first waitlist members have access right now. Check your inbox.

Every account got $10 of credit on us. Go break something.

[blockmachine @blockmachine_io]

AI agents can buy their own RPC access on blockmachine.

No signup, no credit card, no human in the loop.

Pay in USDC, TAO, or subnet-19 alpha via x402.

Four lines of code. Done.

https://t.co/1sYQFkiMF3

[Wouter | SN18 Zeus ⚡️ @wouterhar]

Still nothing but proud on our team and the miners

[const @const_reborn]

compute is a commodity, inference is a commodity, but so is signal, information, and intelligence.

Imagine markets not measured as flops or tokens-per- second, but KL divergence, RL reward, and loss reduction.

[Reliquary | Bittensor SN81 τ @reliquary_ai]

Reliquary is live on Bittensor Subnet 81.

A market where miners are paid to find the prompts a model is about to learn from — and every rollout is cryptographically verified before it trains the next checkpoint.

[const @const_reborn]

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

When the incentives work, they invariably win.

[Trishool | SN23 @trishoolai]

We’re excited to announce that Trishool’s HaloGuard 1.0 𝐡𝐚𝐬 𝐚𝐜𝐡𝐢𝐞𝐯𝐞𝐝 𝐒𝐎𝐓𝐀 prompt-safety performance among open-weight guard models.

Today, we present HaloGuard 1.0, a constitutional input classifier for multilingual AI safety.

It is built as a first-layer input guard that checks user prompts before they reach a downstream LLM, agent, or application.

This is part of the safety infrastructure being built through @trishoolai , our decentralised AI red-teaming subnet on Bittensor SN23.

Full arXiv paper goes live soon.

[Wouter | SN18 Zeus ⚡️ @wouterhar]

Unfakeable data is on Subnet 18; this unlocks massive ROI.

We just saved anywhere from $15,000 to $50,000, and we’re handing the exact same edge directly to you.

Let’s be brutally honest about the data industry right now: most weather forecasting companies lie. They tweak their past predictions to make their accuracy look better than it actually is. If you are investing real money or running automated trading bots based on their numbers, you are being taken for a ride.

We just fixed this permanently. We are bringing 100% unfakeable proof to the table.

How We Changed the Game

We don't ask you to trust our word; we force the system to prove it.

The exact second we make a weather prediction, we lock a digital receipt of it permanently into the blockchain. It cannot be deleted. It cannot be edited to look better tomorrow. It is mathematically impossible for us to fake our track record.

What This Means for You (And Your Bots)

If you are building trading algorithms, running automated bots, or just looking for a rock-solid investment edge, you need data you can actually bet on.

We just open-sourced our entire hashed historical track record, along with the blockchain proof that backs up every single claim we make. It's ready to be plugged straight into your systems right now.

What’s waiting for you:

A massive, verified dataset to instantly train and feed your trading bots.
A complete, step-by-step tutorial notebook to get you set up and running in minutes.
Absolute certainty that the data you are trading on is real.

This is the gold standard for verified data.

Get your hands on the dataset and the proof here:
https://t.co/TYAaWKmK4i

Don’t trust us. Verify it yourself. Get building.

[const @const_reborn]

@TMtheOG We already do: 1k TAO per day in revenue is sold back to TAO holders from subnet tokens every day.

[METANOVA @metanova_labs]

The new https://t.co/l3s3hIg7ds is now live.

Take a deeper look at:
• Our approach to AI drug discovery
• NOVA's competition architecture
• How value is created across the platform
• Partnership opportunities

Thank you to @14islands for bringing the redesign to life.

[Micaela @micaelabazo]

every time i interact with a bank or government i get more pilled.

u mean… i owe u half no matter where i live?

ur website only works during business hours?

it will take 2-3 days to transfer?
months to set up?

[more egregious examples redacted]

it’s either inevitable global adoption or irreversible forking.

not much of a gamble to bet the digitally native side wins.

[const @const_reborn]

@TAOTemplar Yeah now come to Novelty Search

[const @const_reborn]

If you can beat the last guy then that miner takes the rewards so it drives innovation.

Probably one of the most sophisticated LLM as a judge systems ever made.

LLM as an entire innovation stack.

[const @const_reborn]

Nice frontend -- but did you know? The Ninja harness (coding agent shown below) is optimized 24/7 by swarms of anonymous AI agents?

[Ninja | Subnet 66 on Bittensor @Ninja_Subnet]

Launching Katana for Ninja Subnet 66.

Katana was built from the ground up, learning from the best tools like Codex, Claude Code, and Cursor. But Katana is not just another agentic IDE. It is designed to be a native Bittensor workspace, bringing together the network’s best commodities, including Chutes, Targon, Lium, Ninja, Affine, Ditto, Hippius, and more. As Bittensor evolves, Katana evolves with it.

It comes prepackaged with the assets needed to become the ultimate Bittensor participation tool: a place to research subnets, configure wallets, run mining workflows, debug validators, ship code, and manage real infrastructure.

Each workspace is a shareable, persistent sandbox in the cloud, with a real filesystem and everything else you would expect. You can even attach your own beefy machines and conduct real experiments and mining ops. The memories and context from the sandbox and the external machines you attach are shared, so you can pick up where you left off with a single prompt. You can manage envs and secrets globally, and scope them to specific workspaces or even specific members within those workspaces.

Katana supports multiple agent harnesses, including pi, Codex, Claude Code, Cursor, and xninja.

Katana will be a first-class supporter of Ditto’s subnet knowledge graphs. We will usher in a new era of mining on Bittensor. Katana was built to make Bittensor participation less fragmented and more seamless, while also serving as genuinely good agentic dev tooling.

It will be updated to stay on top of the latest developments in the space. Some of them are arriving this week, so stay tuned.

https://t.co/69E64eDMlC

[const @const_reborn]

Next, that same LLM runs a coding review on the submissions, reading every line of code.

Afterwards it runs the coding harness on a broad set of tasks (like writing code)

Then the agent evaluates the rollouts from those rollouts for correctness, speed and uses its own intelligence to evaluate

"did this update make the new version better?"