Montreal doesn’t seem to want me on @ExploitSummit Heavy rain in NYC, cancelled flights, a 3-hour delay. Then the eTA system went down and they couldn’t check me in. Nice try. I’m coming anyway.

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

Palantir CEO Alex Karp with Polish President Karol Nawrocki in New York. Palantir stands with Poland, committed to supporting its national security and defense needs as we further extend our footprint in Central Europe.

If you’ve wondered what “mining Bittensor” actually means, this is your session - just before Day 2 of Exploit begins. @hunter_tensor + @hux_dendrite of @DendriteHQ take you step-by-step through choosing a subnet, understanding what validators reward, the hardware + setup involved, and getting started. Hubert has mined across 15 subnets; Dendrite has been building on Bittensor since 2022. Get ready: luma.com/1kfeefnr

Finally, we’re bringing our mining experience to @ExploitSummit! On day one, our CEO @pbarbachowski will share lessons from breaking subnets. On day two, @hunter_tensor & @hux_dendrite will show you how to become a miner and think like one. We’ll also have our own Dendrite booth. Come meet the team!

Very much agree. Older subnets had a head start and built up much larger TAO pools. That gives them a huge advantage over newer ones. This needs to be fixed before shorting goes live. Otherwise we risk wiping out new subnets while cementing the position of the old ones.

Four years of mining on Bittensor teaches you where the incentives break. On the Exploit stage, @pbarbachowski of @DendriteHQ shares lessons from the move from major miner to subnet operator: what miners exploit, where teams go wrong, and what it takes to build a sustainable subnet: luma.com/exploitsummit26

There’s no better duo to introduce you what we’re cooking in Dendrite! Must join for everyone! Gg @oli_soma @Hedgehog_AI

This Thursday on Novelty Search :: SOMA, SN114 @SomaSubnet is building a decentralized marketplace for AI services, starting with context compression. Miners compete by submitting compression algorithms that make AI workloads more efficient. Validators test & score miners on the balance between compression and retained performance. Thursday :: 5PM EDT / 9PM UTC Live via Youtube and X

Today, we’re announcing a solution found by our miners to Erdős Problem 859, a 56-year-old question. Let dₜ be the density of integers whose distinct divisors can sum to t. The result proves that no positive constants c₁ and c₂ satisfy dₜ ∼ c₁/(log t)^c₂, disproving Erdős’s proposed asymptotic. Verified in Lean through Conjectures. Full proof below.

Bitter Lesson 2.0

Last month I wrote about how we can build a positive and safe future for everyone: meta.com/thefutureisforevery… Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens. The reality is: - People won't want to use agents that are misaligned with them and that don't do what they ask, so labs have a strong natural incentive to make their models more aligned. There is a lot of debate about slowing progress on capabilities until alignment catches up. My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models. Any lab that doesn't focus on alignment will fall behind. - Labs face significant liability if their models cause harm, so they have a strong incentive to prevent this as well. Meta delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing for people and for us. I'm proud of the security foundations we've built. - Engaging independent evaluators and advisors is industry best practice. MSL already does this today in several areas because it helps produce better work. Other labs can just do this too. In general, it would be helpful for there to be a larger and more diverse ecosystem of evaluators. - Committing the significant majority of compute towards serving people rather than racing towards recursive self-improvement is one of the best ways to ensure we develop this technology safely. Meta has made this commitment and other labs can do this as well. I believe the key to building a positive future for everyone is maintaining the right balance of power. This is within our power to do.

39.7% cheaper input with SOMA. We tested our context compressor against GPT-6 Astra. The original prompt used approximately 12,937 tokens, with an estimated input cost of $0.1294. We then took the exact same prompt, compressed it with SOMARIZER at a 0.60 compression ratio, and ran the test again. The compressed prompt used approximately 7,801 tokens, with an estimated input cost of $0.0780 - a 39.7% lower input cost. Compress. Pay less.

The next name in a math textbook could be yours. Conjectures put bounties on unsolved problems. Use any AI model to help. Submit a proof that passes verification and review, and earn the reward. @Wejh69 amazing explainer!

Apparently this is the week for solving old math problems, right? This case was open for 16 years and miner competing on Bittensor found a proof Pretty cool. Even cooler that the miner got a $24k reward Btw huge chapeau bas @jen_w1n ✌🏻

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.

Constant competition over static benchmarks Heard this from my colleague @pbarbachowski recently. Couldn’t agree more. You can train for a known test and call it progress. Much harder to get away with that when the tasks keep changing and other miners are trying to take your emissions. That’s what I like about Bittensor. You have to keep pushing $TAO

Dumbest decision of the decade. You do not protect children from intelligence. You teach them to command it. NY has been watching too much EU. Beijing says thank you.

SOMA is now open. Plug it into GitHub Copilot and run DeepSeek V4 Pro with context compression. Try it on a real coding workload. Tell us what works, what breaks, and what we should improve. More models coming soon. Try SOMA: app.thesoma.ai/login