Mars
136 posts
The dream state would be models within a few IQ points of the frontier labs (probably task specific first, depth before breadth), with a "training-as-a-service" arm to achieve breadth (inc. RL), all on TEE with true privacy, served at a fraction of the price of frontier models with significantly less hardware, thereby commoditizing and democratizing intelligence to the maximum extent possible. The entire world's knowledge trains these models, they should in turn be trained by and accessible to the entire world. The thing about privacy is, it's not really just about privacy. YOU are the product (builders) by ignoring privacy concerns. People are building the massive golden trillion+ dollar moats for the various labs for them by giving free RL data constantly. It's great, for a few, not so great for everyone else. Need to make sure AI remains accessible and available to everyone, always. If AI gets concentrated into a handful of players, it would mean those teams and maybe even a single person within ultimately gets to decide if you are allowed to multiply matrices in particular ways or not, not to mention constant surveillance etc. Parallax aims to be the destroyer of moats.
Lamida has acquired 3 subnets in Bittensor, and will be publicly announcing at @proofoftalk event. > 3 subnets will have projects incubated in @ycombinator like model. > 7% equity of project will be dedicated to Bittensor so that if any projects make to IPO, imagine the value back to $TAO community. > 5 months of incubation, graduation within 5 months of launch. > Subnet slots will be given to next great project if "revenue of project" is less than "emission provided to team" within 5 months of launch. We genuinely felt that we need to expand our dedicated efforts in Bittensor's mission of decentralized AI. And we need to bring public team, professional business structure and value aligned to Bittensor for longer term. Lamida will now directly participate in subnets, not just outsourcing or consulting or running infra for others. More details on Proof of Talk event.
Just for fun I kicked off a run of a 176b parameter model on ~140 steps to prove feasibility - 4 separate nodes across the internet using the "Parallax" method, works like a charm. Still need a more concrete plan on dataset curation, phases, context elongation, etc. etc. before a full run is ready of this scale, but at least we know 176b should be no problem at all.
While exploring the bittensor subnets, one idea kept getting clearer to me: @leoma_ai - Subnet 99 is not trying to be “just another bittensor subnet.” It is aiming to become the video generation company that proves bittensor can compete with the biggest closed AI labs. Thanks to @MarkCreaser, CEO of DSV and investor in @leoma_ai - Subnet 99, and @vex020900, CTO of Vex|Rendix, for taking me through the vision. Also, huge respect to the rendix team for building the leoma subnet. After hearing the direction directly, it became very clear to me why this matters. One thing i want to highlight is vex. The CTO of leoma is a giga brain in AI. From what i have seen, his technical knowledge is on another level, and i believe his skill, vision, and execution will be a major reason leoma can actually build what it is aiming for. Today, @leoma_ai - Subnet 99 miners are training and fine-tuning wan 2.2, one of the strongest open-source video generation models in the world, but that is only the starting point. During the call, the team shared that leoma is preparing an initial free-version launch on April 30. This is not the final vision. It is the first public product step: showing the community that leoma has built the infrastructure to sync with the top miner’s model and turn subnet competition into a real user-facing product. That's what matters the most. Because a subnet should not only produce emissions. It should produce a product. The real goal is much bigger: a 25B+ leoma-native video generation base model, which will be built by the rendix team, released as open source, and then improved through bittensor’s miner competition. This is where things become serious. Building a 25B+ video model will cost millions of dollars. GPUs, training infrastructure, datasets, storage, inference optimization, evaluation, and scaling are not small expenses. This is the level of infrastructure normally reserved for google, openai, xai, runway, and other giant AI labs. That is why I believe funding leoma’s infrastructure matters. ⇢ Without serious infrastructure, there is no serious model. ⇢ Without serious model training, there is no real challenge against closed-company SOTA. ⇢ Without subnets willing to take that level of challenge, bittensor will never prove its full potential. So yes, I will help support leoma’s infrastructure push, this is not just a support. But to help give leoma the resources it needs to build the 25B+ base model, support miner fine-tuning, scale evaluation, and move toward a real video generation product. Remember, no bittensor subnet has ever pushed past the largest open-source models in its field and gone directly after closed-company SOTA, leoma is aiming to do exactly that. ⇢ Not by writing another roadmap. ⇢ Not by pretending decentralization alone is enough. ⇢ But by combining capital, infrastructure, miners, validators, and open-source development around one objective: THE BEST MODEL WINS. If leoma can deliver video quality competitive with closed AI companies while offering generation at 3x–5x lower market cost, that is not just a win for leoma. That is a proof-of-concept moment for all of $TAO. Imagine a bittensor subnet producing video quality that competes with the biggest closed labs, while being cheaper, open, and continuously improved by miners. That is the kind of subnet bittensor needs. That is the kind of subnet that changes the narrative. ⇢ I am not promising @leoma_ai - Subnet 99, becomes the top subnet overnight. ⇢ I am not promising they beat every closed AI lab tomorrow. ⇢ But I believe this is one of the most important attempts happening on bittensor right now. TIME WILL TELL. But if leoma pulls this off, it will not just be a game changer for leoma. It will be a game changer for bittensor. To all $TAO holders: I strongly suggest you take a serious look at leoma. Read their papers, study the vision, and understand why this subnet could matter for the future of bittensor: leoma.ai/Leoma_Whitepap… leoma.ai/Leoma_Litepape…
