Sτeve

Damn, #Bittensor is so beautiful that sometimes it‘s almost hard to look at.

@RadoTsc @chutes_ai You alrady contacted them on Discord, right?

I tried using @chutes_ai for my CRM business, my honest experience... Kept hitting [CHUTES] Status 429: {"detail":"Infrastructure is at maximum capacity, try again later"} for many days and I couldn't test my AI prospecting feature. I use chutes for my private discord, for llm summaries, I don't have problems but with the CRM, no idea it barely works maybe because im requesting while i'm in Vietnam and it surges during that time zone but it was unusable. i ended up, doing like every customer, after some friction just putting my CC to deepseek API and forgetting everything. Felt a bit dissapointed, but time is my most valuable ressource and I don't have time to try to understand utilization, "why" it crashes, "what" to do about it. "why" is it always at 90-100% capacity with a red circled emoji. Code model fallbacks and take risks the infra can "handle" it. I asked AI, it said anthropic or deepseek api is more stable and went w it. I just wanted to test and build FAST. Am I entitled and asking too much? I don't think so, it's a fair reaction even tho I'm a bittensor maxi. I'll go back to chutes when im more experienced maybe. No hate what so ever, i'm journaling my experience, and I know it'll only get better. $TAO #sn64

@RadoTsc @chutes_ai Thank you for your report. I shared it on the #Bittensor Chutes Discord channel.

Lium will become the first agent-first compute provider. We are leaning in heavily to the agent revolution. Here's how we make changes to lium now: We prompt our agents: "using this crypto account, create an account on lium, fund it, get an api key, rent some GPUs, and serve Kimi K2.6 with sglang (or another gpu-required task). When done, create a report for any issues you encountered with using lium and how it could have been faster or easier" Then we constantly improve the agent experience so that it's seamless to tell your agent to do any task, and it will just spin up lium gpus in seconds to solve it effortlessly and without any human intervention needed. We have recognized that SAAS companies need to pivot to agent-first software in order to keep up, and lium is the one leading the charge on that for cloud compute.

If it takes him a full week just to train a 1B with Arbos, just imagine what’s coming with #sn66 👀 🥷

@lium_io Well done! 👏

We recently made changes to our IM that allow us more dynamic control over GPU supply. Subsequentially, providers supplied almost every in-demand GPU type, as opposed to few unique GPUs previously. Simultaneously, we have seen a ~20% increase in the number of GPUs allotted to lium. These IM changes will help us scale up, while better serving our customers by having a wide variety of GPUs constantly available. Need any of these GPUs? go to lium now.

I’d say #Bittensor subnets should stop playing it safe and go all-in route all $TAO emissions to miners. No burns, no half-measures. Make mining absurdly, unapologetically profitable. When rewards are that juicy, competition won’t just heat up. It’ll ignite! And where fierce competition goes, serious capital tends to follow.

Half a million (soon million) blackwell GPUs in a single DC doing 10t! param models. Hopper/blackwell GPUs are also very difficult to source right now, even if you wanted to pay extra. GPU scarcity, power costs rising, closed AI labs ramping up... This is what the open source community is up against. We need a step-change in our approach to decentralized training.

We can’t beat Anthropic at their own game. We need to start thinking about training models through completely new systems; ones that out-perform and move faster. Affine (below) is a solution: a lab that’s just a market mechanism framing the problem “you have compute, you have engineers, how do you get SOTA” Think of it like a global incentivized auto research which pays engineers/agents who lower the loss of a model against RL environments. In other-words, it’s an AI lab but you get paid for contributing to the run by taking the current top model and training it further. The beauty is that we let everything else stay mutable: how you copy, how you train, how you work, where you work, your compute procurement and your algorithms. Just push us. Turns out that’s a very hard game to line up especially when you exist in a global permissionless system which includes a lot of people that speak Korean … But we did it, we trained our first model family, together, that out performed our base, first one in history and just the beginning. Come mine with us.

@tunguz Cramer talks, the interpreter counter-prets!

Jim Cramer ofc

X is the best counter-indicator on Earth - Much better than Jim Cramer Always do the opposite of the consensus on X

What's the problem miners? too tough? 🤭

@RalphFischer_ Okay, then Netflix after all…

Claude is down.

@VictorTaelin know who you are → healthy relationship with people you care about

@Vadimkus3 @bittingthembits Are you referring to the 768k $TAO that are currently deposited at Binance? These $TAO are owned by many investors that use Binance for trading and as a wallet. These $TAO are not owned by Binance.

@bittingthembits Sorry to spoil the party, Binance holds circa 1m $TAO ….

Most crypto tokens were designed so founders and insiders got rich while you held the bag. $TAO is structurally different. No Pre-mine. No VC rounds. No Massive founder allocations. Most are Locked tokens that “unlock” and dump right when retail FOMO hits. Same story every cycle. BITTENSOR is built completely different. - Fair launch. No pre-mine. No VC slush fund. - 21 million hard cap, just like Bitcoin. - Every single $TAO emitted goes to people actually doing the work: miners (41%), validators (41%), subnet owners (18%). - You stake, you mine, you build a subnet, you earn. No one gets a free 20% slice at genesis. - Subnet alphas let anyone capture upside from real products (leads, inference, predictions, compute) without waiting for some insider unlock schedule. The incentives are aligned with the network, not with a small group cashing out. That’s why this one actually has a shot at becoming the infrastructure layer for decentralized intelligence instead of another exit liquidity play. Enough of the same old shit. This one was structured for the long game. $TAO #Bittensor Early…YES....but the structure is finally right

@Rapido_ai 👀

I really have a lot of issues to invest in subnets 100% burning. Why do you burn ? Why are you on bittensor ? How do you justify to provide a service without miners ? If you tell me you have miners, how and why do they provide anything for free ? #top13 #bittensor How can you justify burning to prevent miners sell pressure meanwhile @chutes_ai never did so since first day. @ridges_ai @taoshiio @tplr_ai @webuildscore @hippius_subnet @basilic_ai Fight me, answer those question, and convince me to buy in some $TAO.

@TheTNetHunter 💯

$TAO - Halving in 30 days.

@TheTNetHunter What exactly about SN98 does sound good?

@lium_io Is so CHEAP!!!! I am getting a 2x 4090 for $0.28 an hour!! Think Imma train a 1.6B then post multi token prediction