Hedgehog (τ, τ)

Dendrite, TAO Miner

[SOMA @SomaSubnet]

Most growing companies building on modern LLMs are quietly overpaying for context their models don't actually need.

Our compressor can cut your agents token costs by 15%. At scale, that adds up to serious savings. And we’re not stopping there.

Talk to our CEO, @oli_soma

[Hux @hux_dendrite]

MiMo-V2.6-Pro shows how powerful cost-efficient AI can be.

We’re proud to be training Teutonic-II-110B-A7B, built on Xiaomi’s MiMo architecture with our own expert-sharing modification.

110B total parameters. ~7.3B active per token. Decentralized training. Big ambitions.

Follow the process: https://t.co/AhLCS6n6dD

[Hedgehog (τ, τ) @Hedgehog_AI]

@ExploitSummit @hux_dendrite bro finally decided to share the secret sauce

[const @const_reborn]

Conjectures has done a great job showing how incentive systems don't just parallelize search over solutions — they parallelize search over search systems themselves.

Traditionally, a search pipeline looks something like:

team → search algorithm → compute → solution

With incentives, this expands to:

[verifier + reward] → N competing search systems → solution

The latter is meta-search.

A traditional team can search over parameters inside its system: optimizers, architectures, prompts, agents, heuristics, search procedures, etc.

But many things remain fixed outside the search.

The team itself. Its ingenuity. Its hardware. Its capital. Its energy costs. Its infrastructure. Even the assumptions determining which algorithms it chooses to try.

An incentive system can push that boundary outward.

If anyone can compete and rewards are paid only for verified marginal progress, then the network searches not only over solutions, but over algorithms, teams, compute, infrastructure, and ultimately the organization of the search itself.

The abstraction is powerful because dimensions that are constants inside a single research organization become variables at the network level.

That larger search space has a cost.

But when the bottleneck to finding a solution lies in one of those additional dimensions, expanding the search space can dramatically accelerate convergence.

Congrats to the team show casing this to work so effectively with https://t.co/nigQyIl1cb

[Hedgehog (τ, τ) @Hedgehog_AI]

@capradavis Yup

[404 @404gen_]

1/
The very first 404 Game Jam Competition is now officially live.

We’re giving creators two weeks to build a playable 3JS game using 404’s game generator repo.

Simply point your agent at the repo, build a game from a single prompt, and compete for 10 TAO ($2,300).

The competition runs from Sep 11 to Sep 25, with winners announced live at @ExploitSummit .

Details below 🧵👇

[Conjectures.io @conjectures_io]

Conjectures, the incentive mechanism for mathematical discovery.

[Logan @loganallc]

We solved verifiable inference with no meaningful tradeoffs.

Full speed, full quality, fully deterministic.

[Hedgehog (τ, τ) @Hedgehog_AI]

huge congrats to the miners

results like this will bring more great mathematicians to Bittensor, and there’s plenty more to solve.

[Conjectures.io @conjectures_io]

Today, we’re announcing a solution found by our miners to the near-half-density case of Green’s Problem 51, open for 16 years.

The result establishes how much structure must exist in sums of dense sets. Verified in Lean through Conjectures.

Full proof below.

[SOMA @SomaSubnet]

SOMA Highlights - September 9

We opened SOMA to everyone. One competition later, savings went from 10% to up to 15%.

Try it in GitHub Copilot with DeepSeek V4 Pro.

This time @japanese_crispy joins @oli_soma.
Watch this week’s update below.

Compress. Pay less.

[Weather Boss @weather_boss_]

$TAO SN97 ALBEDO
🏅 New scoring pipeline!

We determine what needs to be done to progress based on multiple SOTA trajectories - then require it from 35b models.

Take part in the competition! 🏆
https://t.co/XSYeeDYuEt

[Hedgehog (τ, τ) @Hedgehog_AI]

@tmkadamcz check @conjectures_io ;)

[Hedgehog (τ, τ) @Hedgehog_AI]

The easiest way to pull big brains into Bittensor might be exactly this

τaopill them @Wejh69 🧠

[Conjectures.io @conjectures_io]

We’re excited to join the Mathathon Challenge at @Caltech as a Gold Sponsor.

The event will bring together 100 teams from around the world, from IMO gold medalists to mathematics professors and researchers at frontier labs, to work on some of the hardest open problems in mathematics.

The challenge will include problems from Conjectures, each with an additional bounty for the team that solves it.

We’ll also have a table at the event alongside sponsors including @AnthropicAI and @cognition. It’s a chance to meet the mathematicians and researchers competing, introduce them to Conjectures, and bring more great minds into open, machine-verifiable mathematics.

Looking forward to seeing what the teams can do.

[Conjectures.io @conjectures_io]

https://t.co/w6tub1chxY

[Dendrite @DendriteHQ]

🧵 1/2 Root baskets are now live on Bittensor.

Subtensor v450 allows root validators to choose which subnet tokens receive their basket-buying flow. Root dividends are converted into $TAO, which is then used to buy alpha according to each validator’s weights.

This is the current flow-weighted Top 20 across the six validators that have already configured their baskets.

[Hedgehog (τ, τ) @Hedgehog_AI]

Onward!

[SOMA @SomaSubnet]

This is the SOMA team - we're definitely real.

Engineers and researchers who have spent years on variuos projects, distributed systems and incentive designs, backed by @DendriteHQ.

@oli_soma - CEO of SOMA

Matt - Owns SOMA's technology end to end and leads the engineering org: architecture, technical direction, and the systems the whole network runs on: validator, competitions, miner scoring.

Simeone - Owns the platform SOMA runs on. Designs and hardens the systems that keep the subnet running under real load - deployment, sandboxing, performance

David - Agentic systems expert. Finds opportunities to improve the subnet and implements them himself.

Beaver - SOMA's resident mathematician. Beyond running the subnet day to day, he owns the science behind it: scoring formulas, validation methodology, and incentive design that make sure the best algorithms actually win.

Rémy - Builds everything around algorithms: agent integrations, product backend, and the infrastructure that turns it into a usable product.

Bork - Works alongside Rémy on the product built around the winning algorithms - implementing agent integrations, backend services, and the interfaces developers actually touch.

➡️ We’ve already shipped two working tools and now we’re taking the next step: building compression directly into Copilot.

The goal is to make compression a useful part of the everyday workflow for developers around the world - solving a real problem at scale and building a product people are willing to pay for.

[const @const_reborn]

Announcing Teutonic-II-110B: a frontier permissionless training of a 110B LLM

We are scaling 10x beyond our previous model and 50% larger than any prior decentralized run.

[Hedgehog (τ, τ) @Hedgehog_AI]

@xavi3rlu @ExploitSummit @DendriteHQ @const_reborn @pbarbachowski

[const @const_reborn]

Today, after 4 months, Subnet 3 (previously @tplr_ai) got retaken by holders through vote.

We're celebrating by releasing our first model.

Teutonic-I: the worlds most performant decentralized LLM.

https://t.co/TNHtP0hfyr

[Dendrite @DendriteHQ]

The new king of decentralized AI has just been crowned.

A 10B model just beat 72B.

Teutonic-I 10B: 62.28%
Covenant-72B: 57.55%

Same 11-benchmark average.

SN3 has changed hands. @const_reborn is the owner.

This is only the beginning. A new chapter starts here.

🧵 1/8