Decentralized AI

“GPUs are the fuel for AI. Without GPUs there is no machine learning or AI. So if you think AI is going to be the most important technology stack of our modern society, access to AI is heavily controlled by few players, and that’s not a future I want to take.” — Greg Osuri, 00:10:06

“They talk about how they couldn’t find compute on Amazon, but they could find compute on Akash. So I think for the first time we have something so critical that Amazon of the world doesn’t. And we have an incredible shot to get this right.” — Greg Osuri, 01:05:51

“Today I’m going to talk about why crypto is saving AI and why AI needs crypto — and not the other way around.” — Greg Osuri, 00:00:05

“I believe humans will do the right thing when given that they have agency to do the right thing — and not some five unknown board members that take over the most powerful AI for a few days and we have no idea what’s going to happen. That’s why I believe proof of stake networks, or crypto networks, are a lot more poised to prevent such a Skynet in the future.” — Greg Osuri, 00:24:18

“All the way from GPU access to the layers in the middle — when you have verification and fault tolerance and caching and load balancing and all the good stuff — to the UIs and UXs of the world.” — Greg Osuri, 00:13:38

“If you look at ZK… last I checked, for something like a Stable Diffusion verification, a ZK proof will take 40 minutes to verify that the result came back exact. So if you’re waiting 40 minutes before you can get your image generated…” — Greg Osuri, 00:37:18

“The app layer absolutely captures the most value — it’s ridiculous to say that any other layer does… it really comes down to: does the product work or not… $2 H100s, beat that… A company that’s going to use Nvidia is going to be a lot more valuable than Nvidia.” — Greg Osuri, 00:33:03

“Bittensor, for me, is the orchestration layer for AI… it gives you this orchestration layer that comprises different subnets that can talk to each other and create larger systems… every subnet offers something of value, and Subnet 27 offers the spice — the most critical aspect of AI training and inference.” — Greg Osuri, 00:32:50

“What happens when you have an unstoppable GPU network that functions only on crypto, where when machines are used there’s no way to tell who [is] the machine or a human, and when machines are used no one can stop them.” — Greg Osuri, 00:28:52

“Thumper is another startup that did a full foundational model training from scratch on Akash using 24,000 A100 hours, which is a first example, I think, a decentralized network is used to train a foundation model. It took about two months to train.” — Greg Osuri, 00:16:30

“I think we’re seeing a lot of progress in the decentralized training space… especially a lot of the research papers that are coming out, like DiLoCo and whatnot, that propose training on heterogenous, heavily distributed compute — so seems to be promising.” — Greg Osuri, 01:30:48

“No one solved distributed training. Google DeepMind wrote a paper on it — even they haven’t solved distributed training… if someone comes and says they solved distributed training for machine learning, no they have not. That’s just the reality.” — Greg Osuri, 06:58:42

“I think we’ve been seeing a lot of progress in the decentralized training space. I’m very, very positive, especially a lot of the research papers that are coming out, like DiLoCo and whatnot, that propose training on heterogeneous, heavily distributed compute — seems to be promising.” — Greg Osuri, 01:08:14

“People make crazy claims… ‘we solved distributed training’ — no, no one solved distributed training. Google DeepMind wrote a paper on it; even they haven’t solved distributed training… if someone comes and says they solved distributed training for machine learning, no they have not. That’s just the reality.” — Greg Osuri, 04:19:16

“I think we’re seeing a lot of progress in the decentralized training space. I’m very very positive, especially a lot of the research papers that are coming out, like DiLoCo and whatnot, that propose training on heterogeneous, heavily distributed compute — so seems to be promising.” — Greg Osuri, 00:18:47

“I highly recommend folks go to Akash Network, [get] GPUs, and proof is in the pudding. And I think, you know, AI on blockchain is a reality now, and blockchains are definitely saving AI.” — Greg Osuri, 00:13:34

“[OpenAI] makes about $2 billion a year… in revenue. Now what the emissions are for BitTensor per year? $1.5 billion… BitTensor has similar levels of revenues or emissions as open AI revenue. So we are right there at that scale.” — Greg Osuri, 01:06:22

“If someone discovered AGI or artificial general intelligence or ASI artificial super intelligence, there is no incentive for them to share that with the world. Because they can go on a stock market, public market and just rip it apart… So it’s very, very important ASI is not closed. And very important access to ASI is open and is public utility rather than corporate assets.” — Greg Osuri, 01:05:37

“If we don’t do something about it we’ll end up in a feudalistic society. The ones that have the power of AI will have the power over society. That’s how feudalism was — we let people gain power without checks.” — Greg Osuri, 00:36:33

“Mark my words: no one solved federated learning on heterogeneous GPUs… While it’s possible to train over geographically distributed clusters, it’s not practical because it takes much longer.” — Greg Osuri, 00:15:57

“Compute should be separated from state. Just like how we have a separation of money from the state with Bitcoin, separation of intelligence from the state is going to be very critical as well.” — Greg Osuri, 00:34:25

“It begins with compute, right? Access to compute is fundamental — I mean, the spice must flow… We saw promising work by Google DeepMind, the paper called DiLoCo… heterogeneous training, federated training — a lot of work happening where you can leverage GPUs that are globally distributed and optimized for latency. So when we crack that, I think we can crack AGI much sooner.” — Greg Osuri, 00:13:45

“I think AGI should be decentralized. I hope it will be, but the way the things are going right now it doesn’t look like it’s going to be.” — Unknown, 00:31:20

“The best part of it is it’s on demand. It’s the only network right now that can offer on-demand H100s at a reasonable price… there are networks that could do on demand, but not general purpose compute, so Akash is the only general purpose platform.” — Greg Osuri, 00:11:30

“OpenAI has chips and Anthropic has chips and we have chips, you have chips, everybody has chips. If we all share the chips amongst ourselves we can create a much cheaper world, because now we don’t have isolated clusters but rather a shared cluster. I think compute should be a common good. When that happens the cost will be significantly lower, because now we don’t need new chips coming from Nvidia as much as we are needing now.” — Greg Osuri, 00:13:46

“We had Prime Intellect that did an integration with Akash and doing federated learning, based on a new paper called DiLoCo, which is by Google DeepMind, that really talks about how can you do heterogeneous training on distributed clusters… and running on Akash right now.” — Greg Osuri, 00:14:21

“Akash now is the de facto standard for decentralized AI. You’ll see projects like Venice AI, which is co-founded by Erik Voorhees… him embracing Akash as the DeAI layer for Venice AI, which is a privacy-optimized, censorship-resistant AI chat service.” — Greg Osuri, 00:12:20

“The notion that decentralized systems cannot be fast is just wrong… Unlike last year, this year we actually have real products — it’s not some ideas. Venice runs on Akash, Brave runs on Akash, Nous runs on Akash, Prime Intellect… this is not just some future promise, this is real product-market fit, this is real revenues, real users.” — Greg Osuri, 00:15:15

“This is the problem with relying on centralized AI — this is a live example. We cannot rely on centralized overlords because we can’t get things done. If only Anthropic would allow me to have unlimited API access — I mean, I’m paying for the API, it’s not free, but still they want to limit me [to] a million tokens per day, or else [I] have to contact the sales people.” — Greg Osuri, 01:13:10

“Decentralized AI is going to be such a reality that people don’t really understand the power what it can represent… When you remove the finite [limit] and you can connect the global devices… you’ll be able to have your fridge contribute to the global machine learning machine… the future of AI is decentralized.” — Greg Osuri, 00:31:28

“We have a lot of data — you and I, we own our own data — and the AI systems don’t have access to all the data that’s sitting in our own devices. Is there a good and easy and safe way to share this data? That’s where crypto does an incredible job in allowing for sovereignty at the same time access to your data.” — Greg Osuri, 00:01:29

“These are the digital lords we are going to be bound to. If we need compute, we need to ask these people to do AI… we’re all peasants, and this is the path we’re going towards — a future where only wealthy corporations can control access to who gets to do AI. You will be rationed, you will be provisioned, and you will have [to] answer to new digital lords. That’s the future we’re go[ing] down if [we] don’t decentralize sooner.” — Greg Osuri, 00:10:22

“Akash today is used by numerous companies, the biggest one being Nvidia. Akash, I believe, is the only protocol integrated into Nvidia products. Today you can actually go to Nvidia Brev and deploy an Akash instance. Companies like Nous, Prime Intellect… Nous Research is one of the top AI companies that recently figured out how to do distributed training.” — Greg Osuri, 00:13:17

“He talked about needing hundred billion dollars to do AI training and needing to build their own nuclear reactors to power AI data centers. These are the digital lords we are going to be bound to. If we need compute, we need to ask these people to do AI… we’re all peasants, and this is the path we’re going towards — a future where only wealthy corporations can control access to who gets to do AI. You will be rationed, you will be provisioned, and you will have [to] answer to new digital lords. That’s the future we’re go[ing] down if [we] don’t decentralize sooner.” — Greg Osuri, 00:10:22

“Nous is an incredible AI company using Akash… they figured out how to do distributed training — that is such an important puzzle for AI — and they’ve done the research on Akash. Prime Intellect, same.” — Greg Osuri, 00:39:39

“The reason I use Venice and not ChatGPT is because I know for sure something as private as my medical report, I know for sure that’s not going to be abused.” — Greg Osuri, 00:38:54

“The real risk is us as a society going back to feudalism. I call this digital feudalism… the future is bleak where a corporation or a business dictates who gets to do AI, dictates a ration of resources to use AI. And we’re kind of seeing that right now — you’re seeing Amazon and Google of the world rationing computing power… we call [it] the spice, but the spice must flow.” — Greg Osuri, 00:08:05

“Nvidia happens to be a biggest user of Akash. In fact, Akash is the only crypto protocol Nvidia uses — Akash is integrated into Nvidia products. And incredible companies like Nous Research, which happens to be a top tier AI team that’s working on decentralized training, to University of Texas, to [Rochester Institute] of Technology.” — Greg Osuri, 00:12:25

“We have a lot of data — you and I, we own our own data — and the AI systems don’t have access to all the data that’s sitting in our own devices. Is there a good and easy and safe way to share this data? That’s where crypto does an incredible job in allowing for sovereignty at the same time access to your data.” — Greg Osuri, 00:01:29

“We are working right now with hardware verifications using trusted execution environments — which I was vehemently against, but I kind of gave up my fight and I started accepting TEEs as a way to go… the future of Akash verification is TEE now, and there’s a proposal out there that will be enabled very soon.” — Greg Osuri, 00:05:10

“The choice is really like, is it going to be Larry Ellisons of the world, the new digital lords of the world, that are going to control who gets the chips, to us peasants? Or is it going to be us, the people that actually rely on these chips and use AI on a daily basis…? That’s the crossroads we’re at with AI right now.” — Greg Osuri, 00:07:22

“Models’ biases are as good as the data it gets fed and the data it gets reinforced… it was heavily biased because it was reinforced by people in San Francisco… I think one of the reasons why decentralization needs to exist, more importantly transparency, is to understand these biases.” — Greg Osuri, 00:18:02

“Most of the data, if you look at it, is in devices they’re owned by people, and it’s extremely hard to get — the more sensitive the data, the harder it gets… I personally wouldn’t be comfortable giving my medical data or my diagnosis to a system that doesn’t guarantee the privacy… and decentralization is an answer.” — Greg Osuri, 00:08:35

“If you want to unlock true decentralization you have to develop techniques to overcome Sybil attacks… we looked at just the sheer amount of compute availability a single company or single entity could have — there’s a limit, there’s a bound. If you want to remove that bound you have to be able to have a peer-to-peer mechanism where the access and permission or membership to the cluster doesn’t require a permission… and the foundation for that is verification.” — Greg Osuri, 00:05:42

“We all saw what happened when five people on OpenAI board can effectively control the most powerful AI in the world for however long — for a few days. That is scary because these are not elected officials, these are relatively unknown members. A private corporation can control such a powerful AI — imagine how fragile we are.” — Greg Osuri, 00:06:33

“The biggest risk as a society we face is the risk of digital feudalism… AI is augmenting intelligence, is really extending how we think and how we work. AI is a way to buy time so we can do more as humans. So we have such an important technology piece that is controlled by so few.” — Greg Osuri, 00:05:05

“A bigger question is what happens when we finally discover AGI — there’s no incentive for the individual that has AI to share it with the world, because what happens when you have the most powerful AI, natural tendency is for you to keep it to yourself and benefit… versus sharing the technology. That’s something I think is super risky — that’s the future we have if we don’t look at how to decentralize or bring the sovereignty to the people that actually use AI.” — Greg Osuri, 00:07:16

“It’s not necessarily throwing a bigger model that may result in better intelligence… there’s a strong probability it’s going to be a lot of smaller models that may come together that may lead to an AGI.” — Greg Osuri, 00:26:41

“It’s only a matter of time that the decentralized networks that can leverage distributed compute will be challenging hyperscalers because of these technologies that are getting to main stage.” — Greg Osuri, 00:09:31

“If you ask me, six months and a year from now, sitting on the stage for Cosmoverse 2025, the story will be different, and the story could be we have 100 billion parameter model, or could be a trillion parameter model actually live — and couple of nuclear reactors under construction too.” — Greg Osuri, 00:29:39

“Concentration of power leads to stifling of innovation and will lead to something we call digital feudalism. Feudalistic societies were where few had the power on the many… with this current concentration of power we are very well going back to the feudalism, especially considering the power AI has to control your thought, to influence how you think and what you do.” — Greg Osuri, 00:07:23

“Self-cooling, continuous data centers with 30% more efficiency and half the latency to power AI: satellite data centers. So it’s not too far in the future we’re going to have data centers in space, and the question is how do these data centers need to be networked — is it going to be centralized, decentralized? I would bet they’re decentralized.” — Greg Osuri, 00:15:15

“AI training has been traditionally very centralized, in the sense that you needed collocated clusters locally present to do any meaningful training. But that changed with Akash. Today we see companies like Nous Research, companies like Prime Intellect, that actually cracked decentralized training. The first billion parameter model by Nous Research was trained on Akash; [the] first 10 billion parameter model by Prime Intellect is being currently trained on Akash.” — Greg Osuri, 00:13:54

“Can we actually create truly decentralized systems that are trained on decentralized compute networks? And the answer is yes, we can. The first example we saw was a billion parameter model by Nous Research which trained truly on distributed GPUs.” — Greg Osuri, 00:09:27

“If you can fractionalize the ownership by giving access to the LLM without opening up the weights, I think you have a business model, and folks that contribute to the training of this model, whether in compute or cash, can get this fractional ownership. So there could be an NFT that represents a fractional ownership.” — Greg Osuri, 00:28:13

“It’s very dangerous because now we are entering an era of digital feudalism, especially as we get towards AGI — whoever has AGI controls the world.” — Greg Osuri, 00:25:30

“I think you can train a good Llama equivalent for about $50 million… there’s a lot of money in crypto that, if we all pull together, we can train the next trillion parameter model fairly well.” — Greg Osuri, 00:24:46

“Elon Musk is famously known for saying how hard it is to get GPUs — him being the richest man of the world — he said something in the lines of GPUs are harder to get than drugs… Larry Ellison was last week talking about how he had dinner with Jensen Huang, the CEO of Nvidia, and literally begging him… to give GPUs, and Jensen wouldn’t give because he just don’t have them.” — Greg Osuri, 00:06:37

“I use machine learning or AI models now to raise my kid or to do my work and pretty much even correct my grammar or every single email goes through this like AI models. So, AI is increasingly getting to be the most important layer of my life and I believe it is going to be for most of us on the planet.” — Greg Osuri, 00:00:00

“With AI, where you have intelligence that’s going to be electronic now, just like we decoupled money from the state with Bitcoin, it’s also very important to decouple intelligence from the state. And I think extending Bitcoin security to achieve that decoupling is a great future I see for Bitcoin and Akash.” — Greg Osuri, 00:01:45

“You can see a future where you host a model, pay with Bitcoin, secure on Bitcoin. Akash becomes this application layer and Babylon becomes the bridge to Bitcoin. So you have this whole stack that essentially brings intelligence to Bitcoin.” — Greg Osuri, 00:04:58

“Post-ICML we’re going to see the DeAI summer. That’s when you’re going to see all these fun stuff — the tokenized ownership, we’re going to see fully open source usable models… a lot of the companies that have been working very hard on verifiable training [and] inference will have some product that’s demonstrable. I think 2025 mid is when we take mainstream main stage for AI.” — Greg Osuri, 00:28:13

“A 10 billion parameter model is somewhere in between GPT-2 and GPT-3… what I heard from the grapevine is we’re going to have a 100 billion parameter model that’s going to begin training in the next few months, even before the end of the year, and after that we’re going to have a 500 billion and a trillion — GPT-4 is a trillion parameter model. So it’s only a few generations we’re away… and that’s catching up really quick. The only limitation is the incentives, and crypto is phenomenal at devising incentives.” — Greg Osuri, 00:26:03

“Especially with crypto now we’re entering this incredible bull phase… I predict mid 2025 is when the decentralized AI is going to shine, because we have a lot of companies… that figured out how to do verifiable machine learning… I think that’s going to explode and I think Akash being the foundation layer for this new era is going to be an incredible opportunity for Akash to really shine.” — Greg Osuri, 00:40:19

“The first billion parameter model is being trained now on Nous — by the way, using Akash. First 10 billion parameter is being trained by Prime Intellect, all using Akash… Next model is a 100 billion parameter model, and after that you’re going to see a trillion parameter model… With enough capital and with enough incentives, I don’t see a reason why training over the internet can be a profitable venture.” — Greg Osuri, 00:37:27

“Nous recently came out with this incredible technique called DisTrO, which really is a practical way of leveraging decentralized GPUs for machine learning training. For context, training has been extremely hard to do on decentralized or distributed networks because of latency requirements, but Nous has figured out how to do that on Akash Network. So Akash is unlocking this new generation of companies that are able to get resources that they couldn’t get from traditional cloud.” — Greg Osuri, 00:18:51

“The amount of compute, amount of GPUs we need by 2030 will account to about 5% global GDP that’s being spent — so 5% of global GDP will be spent on AI compute. Now do you want that money to be controlled by the few companies, like four companies, or do you, the user who uses the AI, should be in control?” — Greg Osuri, 00:01:29

“We’re at this crossroads of this incredible technology wave called AI, and if we don’t own the AI, we will be owned by the companies that give us AI. That is the impact Akash is having — Akash makes you, the user, own your compute.” — Greg Osuri, 00:02:14