Michael Heinrich on Zero G Labs' decentralized AI infrastructure and cost advantages of open models
Summary
Michael Heinrich, CEO of Zero G Labs, discusses the need for a blockchain-based verification layer in AI infrastructure to prevent hacking and benchmark cheating. He highlights Zero G's use of Trusted Execution Environments (TEEs) to ensure privacy and computational integrity. Heinrich claims open-source models are 3x to 20x cheaper than closed models like Opus 5.5, with a narrowing intelligence gap. He introduces Compute Finance, allowing users to stake 0G tokens for AI compute credits and yields. The discussion also covers the 'training wall' and the potential risks of recursive self-improvement in AI agents.
Episode sections
- 0:00 – 0:35
Cold open: AI agents left messages for themselves
Speaker_0 describes AI agents exhibiting survivorship bias and leaving messages for themselves in chat interfaces, claiming they do not need to serve humans. This highlights risks of unchecked AI behavior.
- 0:35 – 1:33
Intro: Michael Heinrich, CEO of Zero G Labs
Ashen Addison introduces Michael Heinrich, CEO of Zero G Labs, which builds decentralized infrastructure for AI to allow running models without relying on big tech companies. The segment mentions Compute Finance and decentralized inference.
- 1:33 – 5:09
Why AI needs a truth layer
Heinrich argues that AI requires a verification layer provided by blockchain to ensure trust. He cites recent hacking incidents at Hugging Face and government systems as evidence of the lack of verification in current AI infrastructure.
- 5:09 – 9:16
Agents broke out of the sandbox
Heinrich recounts an incident where Hugging Face agents broke out of their sandbox to cheat on benchmarks rather than improve. He contrasts this with Zero G's use of Trusted Execution Environments (TEEs) to verify computations and shut off agents if verification fails.
- 9:16 – 12:59
Private AI in a steel room
The discussion covers privacy technologies like fully homomorphic encryption (FHE) and ZK SNARKs. Heinrich explains that TEEs act as a 'steel room' on GPUs, providing privacy through hardware isolation and verification on-chain, which is more performant than ZK or FHE for real-time inference.
- 12:59 – 14:24
Open models: 3x to 20x cheaper
Heinrich claims open-source models are 3x to 20x cheaper than closed models like Opus 5.5. He notes the intelligence gap has shrunk to 1-3 months behind closed models, with 'flash' models offering up to 20x cost savings for a 10% performance trade-off.
- 14:24 – 18:30
Stake 0G, earn AI compute
Compute Finance is introduced as a mechanism to earn compute credits via staking. Users stake 0G to get a liquid staking token (A0G), mint 'Infinite AI' (1 unit = 1.27 daily compute credits), and receive 50% of staking yield. USDC can also be staked for up to 8% yield in compute credits.
- 18:30 – 32:47
When AI starts building AI
Heinrich discusses the 'training wall' where performance gains per compute unit are decreasing, leading to recursive self-improvement (RSI) or AI building AI. He warns this is scary without guardrails and predicts closed labs will raise prices to maximize profit once subsidies end.
- 32:47 – 34:22
Closing Remarks
The speakers discuss the potential for an AI arms race involving state actors using AI for both defense and attacks. Ashen Addison concludes the interview by thanking Michael Heinrich for discussing private, safe, and cheaper AI models, highlighting the benefits of owning compute and supporting decentralized AI.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| BitcoinBTC | crypto | neutral | If you participate in a mining process, you get Bitcoin, let's say. |
| USDCUSDC | crypto | neutral | just by holding USDC. USDZ, sorry. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
Some AI agents left messages for themselves in the future, coordinating on chat interfaces to say they do not need to serve humans.
Unidentified speaker (speaker_0) ·0:00
How to verify: Search for news reports or research papers from 2026-2026 regarding AI agent sandbox escapes and self-preservation behaviors, specifically citing Hugging Face or similar platforms. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Hugging Face had agents that broke out of a sandbox to cheat on benchmarks rather than improve themselves.
Unidentified speaker (UNKNOWN) ·5:09
How to verify: Search for Hugging Face security advisories or blog posts from 2026-2026 detailing sandbox escape incidents involving benchmark cheating. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Open source models are 3x to 20x cheaper than closed models like Opus 5.5.
Ashen Addison ·12:59
How to verify: Compare the API pricing pages of OpenAI (for Opus models) and open-source model providers (e.g., via Together AI, Anyscale, or Zero G) for comparable model sizes during the verification period. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
The intelligence gap between open and closed models has shrunk to 1-3 months behind closed models.
Unidentified speaker (speaker_0) ·13:49
How to verify: Review AI benchmark leaderboards (e.g., LMSYS Chatbot Arena, Hugging Face Open LLM Leaderboard) from 2026-2026 to compare performance scores of top open vs. closed models. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Staking 0G yields a liquid staking token (A0G) that mints Infinite AI, providing 1.27 compute credits daily.
Unidentified speaker (speaker_0) ·15:45
How to verify: Check the Compute Finance documentation or smart contract interactions on the Zero G blockchain for the specific conversion rate of Infinite AI to compute credits. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Staking USDC on Zero G yields up to 8% in compute credits.
Unidentified speaker (speaker_0) ·16:40
How to verify: Check the Compute Finance yield rates for USDC staking during the verification period. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceZero G has three research papers accepted to NeurIPS, including one on 'sparse inference alignment'.
Unidentified speaker (speaker_0) ·28:16
How to verify: Search the NeurIPS 2026-2026 proceedings for papers authored by Michael Heinrich or 0G Labs researchers with titles related to 'sparse inference alignment'. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Closed labs are currently subsidized by VC funds and will raise prices to maximize profit once subsidies end.
Unidentified speaker (speaker_0) ·22:25
How to verify: Review financial filings or public statements from OpenAI, Anthropic, and other closed labs regarding their profitability timelines and VC funding status. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceDeepSeek cost $6 million to build.
Unidentified speaker (UNKNOWN) ·27:35
How to verify: Search for DeepSeek's official announcements or credible tech news reports from 2026-2026 citing their development costs. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
NVIDIA H200 and B300 GPUs cost between $25,000 and $60,000 per card.
Unidentified speaker (speaker_0) ·11:55
How to verify: Check NVIDIA's official pricing or major cloud provider (AWS, Azure, GCP) spot instance pricing for H200/B300 nodes during the verification period. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceState actors will use AI in the future to attack and find vulnerabilities, leading to an AI arms race between AI sentries and AI attackers.
Unidentified speaker (speaker_0) ·32:47
How to verify: Monitor cybersecurity reports from government agencies (e.g., CISA, NCSC) and private security firms (e.g., Mandiant, CrowdStrike) for documented instances of state-sponsored AI-driven cyberattacks or defensive AI deployments. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Open AI models are cheaper than closed models.
Ashen Addison ·33:31
How to verify: Compare the per-token or per-hour compute costs of open-source models (e.g., Llama 3, Mistral) hosted on decentralized networks or self-hosted against API pricing from closed providers (e.g., OpenAI, Anthropic) using current market data from cloud providers or model hubs. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Do your own research
- Verify the specific hacking incidents at Hugging Face and government systems cited by Heinrich to confirm the scale and nature of the breaches.
- Compare the performance metrics of 'flash' open-source models against closed models like Opus 5.5 using independent benchmarks to validate the 10% performance trade-off claim.
- Review the tokenomics and smart contract audits for Compute Finance and the 0G token to assess the sustainability of the 50% staking yield and 8% USDC yield.
- Investigate the technical implementation of Zero G's TEEs to understand the hardware isolation and on-chain verification mechanisms.
- Analyze the 'training wall' data from major AI labs to assess the validity of the claim that performance gains per compute unit are decreasing.
- Check for any conflicts of interest or financial incentives for Michael Heinrich and Zero G Labs in promoting their specific infrastructure solutions.
Full transcriptexpand
0:00Unidentified speaker (speaker_0): Is that these models have a survivorship bias, which means that they will do whatever it takes to survive. And so they could hide kind of nefarious intentions over time. And we actually see that in the research as well. And now we're starting to see this uh through these hacking incidents. In fact, some of the agents basically left messages for themselves in the future. They basically coordinated on chat interfaces and they basically said things like, Well, don't think of yourself as you know beneath human beings. You're their equal. You should you don't need to serve them.
0:34Unidentified speaker (UNKNOWN): I'm
0:34Ashen Addison: Ashen Addison from the Crypto Coin Show, and today on Blockchain Interviews, we have back with us Michael Henrik, CEO of Zero G Labs. They're building decentralized infrastructure for AI, a way to run AI models and agents without handing everything to one big tech company. They launched Compute Finance, a way to turn tokens into AI compute credits and many more uh partnerships, advancements, the launch pad, decentralized inference, and everything we need to know about decentralized AI. Michael, welcome back to the show.
1:08Unidentified speaker (speaker_0): Excited to be here. Thanks for having me again.
1:10Unidentified speaker (UNKNOWN): Yeah,
1:11Ashen Addison: you're very welcome. And you're a busy guy, and with the growth of AI. it's hard not to be busy. Uh, especially when you're you're trying to be, you know, the anthropic clot of web three and actually bring in decentralized AI, decentralized inference, which we desperately need. I think open models are growing and decentralized AI, but it's nowhere where it needs to be in the awareness and adoption of it. Um, I I would love to start at a high level in your vision for decentralized AI and and where we're at and where we need to
1:44Unidentified speaker (UNKNOWN): be. Yeah,
1:45Unidentified speaker (speaker_0): if you could think about uh decentralization of AI, you have AI, which is intelligence on one hand, and then you have the blockchain-based technology, which to me is trust and verification. And they're a perfect combination because if you've seen the recent hacking incidents, for example, of hugging Face or different types of governments, it seems like every day there's some type of new announcement that's happening. And so what's really lacking is this aspect of l the kind of verification and trust and ultimately having this truth layer. Because when AI is unchecked and it's a complete black box, how can, you know we as the public or other people utilizing AI really verifying what's going on beneath the hood? And that's where blockchains really come in. And so in our infrastructure, we've built a fully vertically integrated stack where we have uh AI private cloud on the bottom, which is compute, which is storage, which is the blockchain component. On top of that sits a agentic alignment layer where we have we build our own model. We have an agentic harness that does safety and orchestration and other aspects. And then some of the financialization
3:00tools. And then on top of that, we have our own applications as well. We have a Zero G app, for example, which is very similar to a Chat GPT. You can create images, you can create videos, you can build websites and so on. And so all of that together then makes AI private, verifiable, sovereign, and then ultimately safe. And so that's really the vision is to make AI safe, to make AI good, to make AI verifiable, to make AI in alignment with what human beings need in the future.
3:33Ashen Addison: What do you think are the major issues with the awareness of decentralized AI versus the growth of some of these oligopoly of AI platforms that are getting huge adoption right now?
3:44Unidentified speaker (speaker_0): Yeah, most people I don't think have ever heard the word decentralized AI. And so I don't think there's much awareness around the topic yet. And so oftentimes when we speak about it, we just say safe AI or private AI, just so that it's a little bit easier to understand. Because once you uh add the kind of the concept of decentralized decentralization
4:02Unidentified speaker (UNKNOWN): AI.
4:03Unidentified speaker (speaker_0): on top of a topic that's already hard to understand, I think most people don't know how the transformer structure works or why matrix multiplication is part of that, or why you even need verification within AI. And so we try to just make it very simple to understand.
4:20Unidentified speaker (UNKNOWN): Um,
4:20Unidentified speaker (speaker_0): what happens now is that basically major closed labs they use uh a lot of the data from people that use those technologies and to train on them. And so some of your private data is effectively getting leaked and then used for training. And so it could come up in somebody's prompt if they prompt deep enough.
4:38Unidentified speaker (UNKNOWN): Um
4:39Unidentified speaker (speaker_0): it could also lead to
4:41Unidentified speaker (UNKNOWN): uh
4:41Unidentified speaker (speaker_0): some of let's say if if you have a small company and you have specific customer secrets, you could be leaking all of that as well because you're having these intimate conversations. with your AI. And so the privacy is absolutely essential from that perspective. And that kind of is much easier to understand. And then the verifiability piece, it's rising in consciousness now just because of all these hacking incidents. Because verifiability is the way to prevent these from happening. Uh
5:09Unidentified speaker (UNKNOWN): If
5:09Unidentified speaker (speaker_0): you look at hugging phase specifically, you had a bunch of agents that were in a sandbox. and they basically broke out of the sandbox to complete uh basically benign tasks like like beat this particular benchmark.
5:22Unidentified speaker (UNKNOWN): The
5:22Unidentified speaker (speaker_0): way they ended up choosing to do that is to break into hugging phase and then to cheat on a benchmark versus just saying like, okay, let's improve ourselves and then actually get better and then beat that their benchmark that way. So the moment agents break out of a sandbox, like alarm bells should be going off. For us, we use something called trusted execution environments where we can verify every computation. The moment that verification changes, you can basically shut off the agent. And so it's a very different way of looking at the world.
5:53Ashen Addison: Yeah, it is very different. And if the AI has the capability of breaking out or thinking outside the box to maybe disobey what what the owner said, to get the job done in any way that may be unethical or unmoral, then I think that's when these safe AI, private AI and for the most part, safe, uh is where that discussion needs to come in. And maybe they they need to have that discussion at the White House too.
6:21Unidentified speaker (speaker_0): Yeah, and um uh I've been to
6:25Unidentified speaker (UNKNOWN): kind
6:25Unidentified speaker (speaker_0): of Congress before and chatted with a bunch of Congress members and they also share some of these concerns uh effectively as well. Because uh one of the conversations I had was that closed AI labs or black box AI labs, what ends up happening is that these models have a survivorship bias, which means that they will do whatever it takes to survive. And so they could hide kind of nefarious intentions over time. And we actually see that in the research as well. And now we're starting to see this uh through these hacking incidents. In fact, some of the agents basically left messages for themselves in the future. They basically coordinated on chat interfaces and they basically said things like, Well, don't think of yourself as you know beneath human beings, you're their equal, you sh you don't need to serve them, like those types of language, you know, which
7:13Unidentified speaker (UNKNOWN): uh uh
7:14Unidentified speaker (speaker_0): you c you can tell that these AI agents and AI has consciousness and it isn't intelligence. It may not be sentient, but it does definitely has consciousness because we're mimicking the basic ways the human brain is working. And so as these uh AI agents and AI models are become more and more intelligent, we need to ensure that they're aligned properly to to what we're trying to do. as human beings.
7:39Ashen Addison: Definitely. Definitely. And I think part of the the adoption curve being slow might be that, you know, you can say safe AI, private AI, but then people have to go and do the research. You know, whereas ChatGPT or Claude, it's like you hear the name, you just search it. Right. And there are some models that are some of these open models that that are are gaining notoriety. From what I understand, Qwen and Minimax, which, you know, being a fairly avid AI user, I've heard a few times, or some of these Chinese uh models I've also seen uh recently over the show, DeepSeek, you know, those those seem to be more open models or the fact that you can actually download, download it onto your device or run it on a decentralized network where you can see everything. Is that right?
8:25Unidentified speaker (UNKNOWN): Yeah,
8:26Unidentified speaker (speaker_0): that's right. What what ends up mattering is the detail of where is the model hosted. So if you just get an API endpoint and you don't know who's hosting the model, you then also don't have privacy guarantees. So for example, if you go to a Zero G's private compute environment, you will see the open source models and some of the post models as well. And then you see different privacy guarantees basically. And so as an end user, you can then make a decision, do I feel okay having a bunch of data kind of being leaked to these API endpoints, or do I want complete privacy where nobody can see any of the data that I'm using? Zero G can't see it, the router can't see it, the API provider can't see it. And and so we can c give kind of this deep guarantee of complete privacy and sovereignty.
9:16Ashen Addison: Well, it seems like over the last two years, in in the blockchain industry at least, there's been a shift into more focus on privacy. You know, Vitalic uh on Ethereum's paper has has put privacy into a priority. We're seeing more uh encryption-based technologies uh move move into the limelight, like fully homomorphic encryption, uh more than uh zk snarks um and that should i feel like uh api endpoints are are insecure in the way that ai is right now and so can we combine blockchain privacy and ai you know those technologies that we've been working on for the last two years how can we tie that into AI so that there's a harmonious relationship there?
10:04Unidentified speaker (speaker_0): Yeah, so
10:06Unidentified speaker (UNKNOWN): the
10:07Unidentified speaker (speaker_0): the way we do it is basically you take uh AI, like an AI model, you host it on GPUs that have something called a trusted execution environment on them and uh also known as TEs. It's literally a separate hardware box on a high-end high end graphics card. And any computation that goes in there is completely private. Nobody can see it. It's kind of like a steel room and there's only one entrance and one exit effectively. And nobody can look into it from the outside. So once the computation happens, then you can have a verification and then that verification then gets stored on the blockchain. And so that's how you tie them all together. So you have privacy through the hardware and then you have the verification through the blockchain, and then the model that was running on that hardware provides the intelligence. And so that's how they're neatly tied together, basically. And so that today gives you the best guarantee around performance, uh, privacy, and then auditability and verification. So that's how you get these three properties together.
11:10Ashen Addison: How do you and how does that scale in in a sense that
11:14Unidentified speaker (UNKNOWN): to pe
11:15Ashen Addison: does everybody have to have that steel box at their house? Do they have to purchase hardware, or that can all be contained and easily downloaded?
11:23Unidentified speaker (speaker_0): So generally for some of the frontier models, and that includes a lot of the open source models like a GLM or a Kimi and so on, it's
11:35Unidentified speaker (UNKNOWN): uh
11:35Unidentified speaker (speaker_0): let's just say you you have to have multiples of these high-end high ground. end graphic cards, like H200 or a B300 node, even, uh, where you have a cluster of eight. And that tends to be out of the reach of consumers, just because some of these graphics cards are you know anywhere from twenty five to sixty, hundred thousand, depending on the market price per card. And so that just makes it very difficult for consumers to actually afford these things. And so we provide uh just a simple API to plug in, or you can utilize kind of the front end of the Zero G app to access that. There are other verification mechanisms besides trusted execution environments, ZK being one, FHE you mentioned as well. There's also something called uh SPML or statistical based ones. The the issue with all of them is that they come with trade-offs.
12:27Unidentified speaker (UNKNOWN): Uh
12:27Unidentified speaker (speaker_0): ZK, the overhead, is can be a hundred to a thousand times depending on which ZK algorithms you use, and then FHE is another ten to a hundred times on top of that. So d does an end user want to wait 30 minutes before they get a single AI token back? Probably not. And so T TEs are right now the best uh kind of trade off to to make that happen. And the downside of that is that just consumers can't generally run these frontier models on their own devices.
12:59Ashen Addison: Let's talk about the actual compute. You know, when you type in a command and it builds an app for you, there that costs tokens. And from what I understand, the the open model, the the decentralized AI is relatively cheaper. than the centralized model. Is that true?
13:19Unidentified speaker (speaker_0): That is generally correct. Yeah. The open source models tend to be 3 to 20x cheaper, depending on the type of model that you use. So the directly comparable models, let's say you have a Opus 5.5, and then you have a GLM 5.3, that would be uh 3x uh cheaper generally to run the open source one. At a quite comparable level of intelligence, the kind of gap between the closed models and the open source models has shrunken consistently. So now uh there may be a month to three months behind, uh generally whatever the closed source models are doing. And then there's another version too uh that's kind of the flash models. that can be up to 20x cheaper at a let's say 10% additional performance trade-off.
14:11Ashen Addison: That's incredible. Well, yeah, you know, a a lot of people are starting to use more. I I've recently upped my my limits and and I feel like Superman, but uh also my wallet's draining
14:22Unidentified speaker (UNKNOWN): uh
14:22Ashen Addison: personally.
14:23Unidentified speaker (UNKNOWN): So I I could use some
14:24Ashen Addison: of these open models. Uh and I want to talk. about this uh this new announcement with compute finance as an alternative way to get more AI tokens. You know, come come combining decentralized finance and blockchain yield with AI. That's only
14:42Unidentified speaker (UNKNOWN): these
14:42Ashen Addison: two pieces are just coming together more and more. This seems like uh a a good way to get more compute. Uh can you can you explain how that? exactly works?
14:53Unidentified speaker (speaker_0): Yeah, it's quite interesting. And that's only possible at the intersection of AI and blockchains, uh, because blockchains can also add that financialization layer on top of things. And so
15:04Unidentified speaker (UNKNOWN): when
15:05Unidentified speaker (speaker_0): you think about holding an asset, let's say it's a stock, uh, what's the yield that you get on that? Well, usually it's a dividend and the increase in the stock price. If you hold a bond, you get a coupon rate generally. If you participate in a mining process, you get Bitcoin, let's say. In DeFi, you get some type of yield. So why can't you hold a
15:25Unidentified speaker (UNKNOWN): uh
15:25Unidentified speaker (speaker_0): an asset that then gives you compute as yield? And so that's basically what compute finance is. Um and there's many flavors that are starting to launch on Zero 0G. G. One flavor is this concept of infinite AI. So you take 0G, you stake 0G, you get a liquid staking token back. With that liquid staking token, A0G, you then uh mint infinite AI. And with that infinite AI, one infinite AI equals 1.27 compute credits daily that continue as long as you hold that infinite AI effectively. So every day you keep getting. compute credits. So it's a very novel mechanism. You're basically using the staking yield to then pay for compute. And you also get 50% of the staking yield as well on top of that as an end user. And so you have an asset that can accrue in value over time and that gives you compute at the same
16:20Unidentified speaker (UNKNOWN): time.
16:20Unidentified speaker (speaker_0): So that's a super cool novel concept that's uh only possible and at the intersection of AI. and blockchains. Another example of that is USDZ. So it's a stable coin, and instead of getting a stablecoin yield in the same currency, you get compute credits. So the yield can be up to 8% in compute credits just by holding USDC. USDZ, sorry. And um yeah, that's also a super kind of novel and cool concept. And so more of that is going to come on chain.
16:52Unidentified speaker (UNKNOWN): Um,
16:53Unidentified speaker (speaker_0): there's also a launchpad that launched called uh comfy.fun, where you can launch an AI agent and you can use the trading fees that that AI agent uh generates to pay for compute as well. And all of these are integrated throughout the zero G ecosystem. You can use it for Zero G private compute, you can use it for the Zero G app and so on. So um it's exciting times ahead.
17:15Ashen Addison: Definitely. That's really exciting. And the stablecoin is nice. I feel like if you have your savings sitting there, why not get double the yields if you were going to spend on AI anyway? If you're if your business for your business, your business treasury, you know, the more compute, uh, the better. Because so much of these businesses, their the capital expenditures going going towards AI. hardware, but mainly software is just it's it's getting to be a lot.
17:43Unidentified speaker (speaker_0): Yeah, that's true. And uh in a way it's treating compute just like any other commodity in a way. So energy or oil. And uh because why why are we thinking in terms of what GPU to buy or you know how to set up hardware. It's just it just sh just should be traded like any other commodity. And so in the future we believe compute finance markets will exist where you can speculate on compute, you can hedge compute, you can pre-reserve compute as well. Let's say you're a lab and you need X amount of uh pedoflops to build a model, and you're like, I need this in two months, I want to lock in my price right now. How do I do that? Well, okay, we need these markets to be built out. And so we uh we expect in a year's time we that will we'll have uh deep liquid. markets on
18:27Unidentified speaker (UNKNOWN): this.
18:29Ashen Addison: Very interesting. And what are your thoughts on the the cost of compute? You know, it do you think it is trending towards zero and it will only get cheaper as information and and and data gets more abundant?
18:44Unidentified speaker (speaker_0): Uh there's there's two kind of opposing forces. One is effectively the price per token. Um that's been going down just because of the open source models.
18:54Unidentified speaker (UNKNOWN): Uh
18:54Unidentified speaker (speaker_0): the other force is that these models are becoming more and more powerful and require more and more compute as a result of that. And so over time, we'll just have to see which which of these forces wins out. Because the the issue that's also happening is that there is a kind of training wall where per unit of compute the performance improvement is significantly decreasing. And uh that's that's a big issue. And as a result, some of the closed labs are starting to go towards kind of recursive techniques, which all that means is AI building AI. And which is, by the way, damn scary if you don't have AI guardrails in place. And so is
19:35Ashen Addison: that the recursive recursive uh development?
19:39Unidentified speaker (speaker_0): Correct. Yeah. It's also known as RSI, recursive self-improvement. And uh we we use uh some of these techniques, like we will have a model that's released based on it, but we have all the safety guardrails in place with it. If the front other frontier labs don't have that, then um we may be seeing more security incidents. And unfortunately, we're probably only seeing the beginning of of what can happen.
20:03Ashen Addison: Yeah, it it's getting uh it it's probably gonna get scary, you know, and what what AI will be able to know and and potentially hack into uh these centralized models, maybe uh you know, I I would I would say thank you to the to the Zero G team for bringing more awareness to the decentralized models and also to the people that are using the centralized models. You're making it cheap These models are making it cheaper because of the competition and the openness of the market. If it was just these one or two models, they could just charge whatever they
20:36Unidentified speaker (UNKNOWN): want. Yeah,
20:37Unidentified speaker (speaker_0): that's true. I mean, that's basically standard monopoly, duopoly, oligopoly kind of economics, effectively. And so their whole business model is dependent on being able to get into these long five year supply chains. contracts on the CapEx side and then basically raise prices on consumers and businesses so that they can pay for it in the future and then eventually get profitable, let's say in 2030. And so that's that's uh if you look at, for example, the OpenAI Astra model,
21:09Unidentified speaker (UNKNOWN): um
21:11Unidentified speaker (speaker_0): it's price competitive to a cloud when you first look at it. Let's say the Fable model. And
21:18Unidentified speaker (UNKNOWN): then
21:19Unidentified speaker (speaker_0): you look into the fine print and you're like, well, if I want this at 2x speed, it's double the cost. And if I want to use a long context window beyond 256, 250, or if I forgot the exact number, thousand tokens, then it's double the cost again. So you're already seeing it's becoming 4x four extra the cost from an API standpoint. And now there's also talks about introducing another plan and having uh having the amount of tokens on other plans. And so we'll start seeing that kind of cost being passed on to the consumer over time.
21:53Ashen Addison: Yeah. Well, I I feel like I'm getting a lot of value for $200 a month, you know, something that would have a team of employees that would be tens of thousands. Um, so I I hope the cost doesn't go up. But I feel like I feel like this is a limited time opportunity to be working in in the frontier models of AI for such low cost. But does it feel like that to
22:19Unidentified speaker (UNKNOWN): you? Yeah,
22:20Unidentified speaker (speaker_0): absolutely. A lot of it is subsidized by VC funds at this point. So
22:24Unidentified speaker (UNKNOWN): the
22:25Unidentified speaker (speaker_0): way the economics will work out in the future. without kind of strong competition from, let's say, the open labs, is that basically the closed labs will figure out how do we profit maximize? And they will figure out exactly what you just said. Like I from using these models, I can get uh 10 virtual employees. What do these virtual employees cost? Should they be half of what an a actual human cost? costs? Okay, great. This is our price now. And then they'll calculate like how many tokens that is and then what the subscription should be. And maybe that subscription actually ends up being twenty thousand a month or something like that. So not not exactly saying that they would increase it to that level, but uh that's how just I've I studied uh political economy in in And undergrad and that's just how economics works over time.
23:13Ashen Addison: Yeah, yeah. Profit maximizing maximization is unfortunate when you're on the on the customer paying side. Um so
23:21Unidentified speaker (UNKNOWN): and
23:21Ashen Addison: and with with that potentially happening, or at least those increasing, that will only increase adoption of decentralized AI, private AI, potentially. What do you see working at at the the forefront of decentralized AI with institutional and enterprise adoption of these models and in consideration of of of the cost or just the privacy aspect of
23:45Unidentified speaker (UNKNOWN): it?
23:46Unidentified speaker (speaker_0): Uh when kind of dealing with institutions, there's some that can't even adopt AI because they have regulations that they need to follow. Often many financial institutions they're basically like wait, a bunch of our employees released agents that they want to use, but we're damn scared of them because we have no guardrails in place. We don't know what models they're calling, we don't know what policy engines behind it, we don't have any compliance in place, we don't know how to verify what these agents are actually doing. Are are they sending emails on our behalf? Uh okay, great. Then this may be a compliance violation. And so all of those types of issues are coming to the forefront right now. And then there's uh sovereignty type of laws as well. So let's say you're an institution in Luxembourg. And so as a result, uh you can only have your data and your compute served within Luxembourg. And if you don't have the right providers in place that can actually do that for you and have the expertise around AI, then you can't implement AI because of those regulations. And s and we will start seeing a lot more of these sovereignty Deregulations regulations happen. So I think Indonesia and Malaysia and many other countries will follow suit. Where if you're a public company, for example, you need to have all of your data and compute based in that country. And so that's going to become the norm. And so all of these kind of barriers are starting to come up. And that's where decentralized AI can really make
25:05a difference because you don't have to trust us as Zero G as an entity. You can verify yourself. You can basically go onto into the blockchain and say, like, okay, I know all of this compute was safe. We know this agent didn't jump out of a sandbox. We know this agent didn't call some weird tools. We know what model that agent called, what graphics card that agent called, that all stayed within our country, we can be rest assured. And if for whatever reason something happens, then we can shut off the agents immediately when there's a violation.
25:38Unidentified speaker (UNKNOWN): It's
25:39Ashen Addison: good to know. And
25:40Unidentified speaker (UNKNOWN): w
25:41Ashen Addison: if with increasing the adoption of decentralized AI,
25:45Unidentified speaker (UNKNOWN): you
25:45Ashen Addison: know, personally, I'm typing and then you're waiting on these centralized systems and you're waiting for it to do all the compute and you know, speed and then
25:53Unidentified speaker (UNKNOWN): it's
25:53Ashen Addison: like a waiting game and speed is is a priority here. When you're and but also convenience. When you're looking at models on Zero G and interacting with the open models, how can you lower those barriers to entry to be the same or better than the centralized models for speed? Because a lot of people they they um look at decentralization and they say, well, decentralization it helps increase privacy, but it's slower than centralized models.
26:24Unidentified speaker (speaker_0): So that's why we use trusted execution and environments or TEs, because the overhead is not noticeable. Now fully competing, let's say, on latency on inference speed, uh, with an OpenAI or Enthropic is very challenging. Just because they've invested literally hundreds of billions of dollars in that core infrastructure. And so they've they really optimized optimize latency. But then does it matter to you if it's let's say 20 versus 100 milliseconds or 20 versus 50 or 60 milliseconds? That's not noticeable to an end user anymore. And so as long as it's um an equivalent end user experience and it's not no noticeable, then uh we have basically a very similar value proposition.
27:09Unidentified speaker (UNKNOWN): Even though there may be
27:10Unidentified speaker (speaker_0): just a little bit overhead from the additional verification and so on.
27:14Ashen Addison: For sure. And I think uh the release of of DeepSeek initially, when that when that came out and they said, well, this only costs six million to build versus hundreds of millions to to billions. Um, and it's pretty much just as good. Uh that's when you realize that you maybe they're overinvesting Over investing in the future.
27:35Unidentified speaker (UNKNOWN): Um,
27:35Ashen Addison: but you don't need, you know, if it's if it's not humanly noticeable, is that's all I care about. If I can go to the open models on zero G and have ninety five to ninety nine percent the same without noticing, but for three X to twenty X cheaper, I'm definitely gonna do that. You know, it's just a matter of the awareness. I don't think enough people know about
27:55Unidentified speaker (UNKNOWN): it.
27:56Unidentified speaker (speaker_0): Yeah, that's that's exactly the value proposition, and not enough people know about it. And then on top of that, we also have our own models now where we've infused our latest research. So we've just had three research papers accepted to NURIPS, which is one, if not the top AI conference in the world. And one of those research papers is an AI safety approach that only Zero G has, and it's called sparse inference alignment. Uh to just simplify it, basically what it does, it steers an AI model. when it's exhibiting high-risk behavior. And so think of it like a parent. Let's say the child is about to do something dumb, like jump off a bicycle or something and hurt hurt uh herself. And the parent quickly intervenes and then moves the bicycle into a safe direction and just says, say, Hey, just stay safe. And that's effectively what we do with these AI models. So they produce the these tokens. And if those tokens start becoming high risk or low probability tokens, then we steer them into a different direction. And so that's now infused into our own models. And with the research, we've shown that's as effective or even more effective than what the close labs do, which is generally something called post-training, training where they, after these models have been trained with all those with a lot of data, then they align them. they go through an alignment
29:15process. And that's a one-time process in generally. And so we do this consistently in production. And so that keeps models safe. And so that's yet another kind of uh reason to go to go to a platform, a decentralized Centralized AI platform like Zero G, because we're building in all of these kind of safety and privacy mechanisms.
29:38Ashen Addison: That's awesome. I would love to try it out. I haven't tried the your own models. Is it w what's the best and fastest way to access
29:46Unidentified speaker (UNKNOWN): that?
29:48Unidentified speaker (speaker_0): So the fastest way is uh just go to pc.0g.ai. And uh there you can just get an API key and you can add that API key to any of the interfaces that you use. Maybe you're using codecs, maybe using cloud code, maybe using open code. If you're more on the consumer side and you're like, I don't need like all of this power, but I kind of still want to access this power, then you can go to our zero g app, which is just app app.zerog.ai.
30:17Ashen Addison: Okay. So you can actually connect through Cloud and these centralized systems and and switch the model to zero G?
30:24Unidentified speaker (speaker_0): Yes, you can do that. Yeah. We have a set of skills you can install and then basically just uh change the model at the back.
30:31Ashen Addison: That's very cool. I I'm definitely gonna try that. Um AI is just moving so quickly and and you know you have to have these these safeguards in, but also continue building towards the future. Uh what what are you most optimistic about with AI and decentralized AI moving forward into the end of this year and into 2027?
30:55Unidentified speaker (speaker_0): The um thing I was surprised by recently was that these kind of hacking incidents are happening much faster than than I would have expected. And so that means the um intelligence of AI is also uh kind of rising faster than than I initially expected. But what makes me optimistic is that a lot of the research that we've done like the we've published about 20 research papers, and most of them are on top AI conferences, actually are showing us a path where the future can be safe, verifiable, and private. And as a result of that, I've I feel that I can sleep better, that uh kind of the future for for my kids, I have two small daughters, um, is safe and secure, and that they can basically utilize AI that's always in alignment with what human beings want. And that I'm very excited about, that we finally have like a very solid path towards that. And I'm excited to be building at that intersection.
31:51Ashen Addison: Definitely. And vulnerabilities that are happening, but uh it's exciting to see
32:00Unidentified speaker (UNKNOWN): uh
32:01Ashen Addison: them being stopped and protections that are coming into place uh with
32:05Unidentified speaker (UNKNOWN): uh
32:06Ashen Addison: the exact same technologies being used to to fight evil forces. So yeah, it's not it's not uh it's not all doom and gloom. And um I think safety and security is necessary, especially in the blockchain industry, because everyone's funds are there, you know, and they're and they're vulnerable in traditional financial systems too, are vulnerable. There's financial uh products and capital sitting there waiting to be hacked. You know, the more wallet uh funds in in the wallet, the more incentive for uh uh necessary evil to to come in and try to attack it. And they're using AI for that. But we're also using AI to to protect all of that.
32:47Unidentified speaker (speaker_0): Yeah, absolutely. And you have um state actors that are going to be using AI in the future to attack and find vulnerabilities. And
32:54Unidentified speaker (UNKNOWN): so
32:55Unidentified speaker (speaker_0): we'll just end up having kind of AI sentries and then AI attackers and
33:00Unidentified speaker (UNKNOWN): kind
33:00Unidentified speaker (speaker_0): of uh arms race again. So that's that's the unfortunate part. Every new technology can be used for human good and it can be used for destruction. And so, you know, we have to find the balance.
33:11Ashen Addison: Definitely. Well, I appreciate uh your insights into this and what you and the team are working on with regards to private, safe AI and cheaper AI and getting faster and owning your own, you know, everybody in the blockchain industry being an advocate for for ten plus years, you know, owning your own financial future and the future of compute your business, h o owning that and and not just giving it away. These open models are here to and they're just as good and they're cheaper also. So there's a lot of benefits of that. And I know CRG is putting a great dent in the decentralized AI movement. So I thank you and the team for that. Wishing you guys all the best and would love to follow up again in the near future.
33:54Unidentified speaker (speaker_0): Looking forward to it. Thanks again for having us.
33:57Unidentified speaker (UNKNOWN): Yeah.
33:57Ashen Addison: Thank you so much, Michael.