BONUS: Government Banning AI Fallout: China, the economy, what's at risk, and what you should do.

14:48Grant:OpenAI said yes.

14:49Corey:5.6 went to the government. Anthropix said no and released it anyway. Which is hilarious because they're the ones saying that we should do this. Yeah. And it's just a bizarre series of actions, in my opinion. You know, but I think, you know, I think they forced the government's hand. I'm not, like I said, I'm not here to take a political stance, but I mean, like, you can only say so many longs, so many times, the things that he, more he than them. Like, this is not about the researchers for me. This is not about their engineers. years i believe they build great products and i believe they have a great team but i believe that

15:30Grant:he's probably bad for ai yeah let's say that again i'm sorry i believe he is i believe that

15:36Corey:dario has proven to be bad for ai and i cannot imagine a public company allowing things like that to be said by a CEO. I really do not feel like he is public company CEO material.

15:59Grant:I think that the level... Do you think they're going to replace him when they go public?

16:03Corey:Now, I say that and SpaceX exists. So, like, things are said

16:07Grant:by public CEOs. But I think there's a big difference in those two people.

16:12Corey:I don't know that they'll release him. I guess they can at that point. I mean, it doesn't matter if you're a founder once you're public.

16:19Grant:It depends on the makeup of the board and the shareholders. And how much of it you own personally. I want to highlight, because you brought this up, this comment from Dr. Sean M.D. He said, the resistance to AI is increasing. Most is not fact or science-based. I have colleagues in medicine who use AI scribes, but don't feel comfortable using these tools for more practical purposes. So a lot of what you're saying about these comments being bad for AI is if people are already fearful that it's going to take their jobs, saying things like white collar bloodbath's not going to help with that. If you're saying that these models are so unsafe that they need to be banned or I don't think he said that.

17:03Grant:He said or they've had a position that they think that open source should be regulated and they've been very much on the record about that.

17:14Grant:you're increasing the resistance to the thing that you are trying to build and then you're releasing it anyway instead of increasing the knowledge

17:23Corey:what do you mean by that?

17:25Grant:say more about that

17:26Corey:what I mean is like when you say reckless things like he has said what happens is that people like these that Dr. Sean M.D. here is talking about in the comments see that and they see it from a professional in the ai space and they're like holy crap that's true they believe that now for good that ai is as dangerous as x because that's i'm not saying that's right i'm just saying that that's the way humans work we we hear a thing we adopt a thing as a belief and we run with it forever uh it's not the right way but it's by by golly, George the human way. And the flip side is he's in a position where what he could absolutely do is educate those people.

18:12Corey:Where he mentioned that a lot of the stuff is, in fact, science-based. The other thing is a lot of the information that the anti-crowd uses tends to be really dated. Like there's been a lot, a lot of stuff done in the safety realm over the last four years now, like with regard to all kinds of things. Like, you know, I mean, here's the deal. If there's a will, where there's a will, there's a way. You're not going to stop that. You didn't stop it with the internet. There's still a dark web. There are still horrendous crimes perpetrated on the internet every single day of the year. Every minute of every day, something awful happens on the internet.

18:49Corey:And we still have it and we still accept it. And at some point over time, that will be the case with AI is what I think. But right now, I think for people in the AI space who have done the research, who understand these tools, who work with them consistently, I think the worst thing you can do is spout off crazy things. Like if you want to share some risks, there's a way to do that.

19:16Grant:There's a way to frame that in a way that is educational. It doesn't go viral, though.

19:20Corey:But it's not going to go viral. That's right. And it's not going to get more people to come get scared of your tool and go buy it. But there's a way to do it that explains like, hey, we're working on this. There are risks. We're doing the best we can to assess those and to work around those. We're looking for new ways. We're working with the government to develop sensible regulations that both ensure safety but don't hand over an industry to China, which we'll talk about later.

19:50Grant:Do you think OpenAI does a better job of doing that?

19:53Corey:Yeah. I think OpenAI does a significantly better job of that. And I think some of that is just that Sam at his core is still a business guy too. And I feel like Sam sees the big picture a little better than Dario does. I think Dario is a guy led by emotion. Did you see the Sam statement a couple weeks ago that they essentially – Dario built a company based on basically hating him? and there's a non-zero amount of truth to that and I mean Dario's all been acknowledged in some ways too and I'm not saying that that hasn't started many great industries and many great businesses but I do feel like Sam has a better understanding of the industry as a whole global impact feels like an area where they are more equipped to answer questions they've been more broadly focused on other industries aside from like Anthropik's been really coding forward since after Claude 3 like Claude 3 was a killer writer in my opinion and I feel like they've lost a little along the way not with every model, some models have come back and written well but some of them I get back and the funny thing is despite all of this danger their name is literally Anthropik for God's sakes They literally anthropomorphized the model they have, gave it a human name, and encouraged it to question its sentience.

21:29Corey:And I'm like, you know, you know.

21:32Grant:The soul dog. Like, I don't know how you terminate her, but I feel like it's certainly not the way you prevent it, and it's certainly not the way you prevent psychosis.

21:42Corey:Where I feel like since GPT-4O dropped, or excuse me, since GPT-4O was dropped from the lineup, I should say. like opening eyes made a number of strange to try and kind of tone down the sort of behavior that

21:55Grant:led to that where like and and i dig claude claude's got a great personality but but there

22:02Corey:is an element of like it is absolutely the most pretending to be human model that exists on the

22:08Grant:market as far as i know and they do a good job of trying to train it to make it um abide by its constitution. Yes, yes. I love the constitution idea. I think that's a really cool idea. And some people on X had some interesting ideas. I forget who it was. So I apologize to this person who I'm not attributing this to, but they were saying like, perhaps part of the model companies is whatever constitution you give your AI. And in the future, every AI company will have a constitution and that will sort of dictate, you know, what you assured, I don't want to sidetrack on this, but you shared an interesting piece with me recently about the political bias in the models and where they land on a sliding scale.

22:54Grant:It was a Washington Post study. Right. Yeah. And perhaps part of the constitution of the model is determining where it lands on its political bias scale. And obviously you want it to be as neutral as possible, but sometimes you might take it in a different direction.

23:12Corey:that was the most centered on that study.

23:16Grant:Well, I think to Elon Musk's credit, he says he's trying to make it as truth-seeking as possible. Whether or not that's actually what happens with it, we don't know, but maybe that has something to do with it.

23:30Corey:Maybe. I want to share this.

23:33Grant:Oh, go ahead.

23:34Corey:No, go ahead. I don't have anything that interesting.

23:37Grant:I just wanted to say, I wanted to share this post from OpenAI before I take it down. So this was part of the GPT 5.6 launch. So if you're just joining us, you know, GPT 5.6 has been held back per the government's request. So they're beginning a limited preview of the 5.6 series. So there's three different models there, Sol, Terra, and Luna. And they made this statement where they said, We believe in broad access, and we plan to make GPT 5.6 Sol, Terra, and Luna generally available in the coming weeks. As part of our ongoing engagement with the U.S. government, we previewed our plans and the model's capabilities ahead of today's launch.

24:18Grant:At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government before releasing more broadly. During this preview, we will continue testing and coordinating closely with our partners as we work toward broader availability. I want to highlight this.

24:33Corey:This is the important line in this whole release right here.

24:36Grant:We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks, a.k.a. how we can stay competitive with Anthropic, while we work with the administration to try to develop the cyber executive order framework and a repeatable process for future model releases. So why is this important, Corey?

25:10Corey:Because they said, okay, here's what was requested of us. We're doing this. We're complying with the government's request. And essentially acknowledging that there's not a real framework yet. This is us doing the thing we're asked to. But I think even more important is that what you highlighted at the bottom is them also saying, we don't like this. This is not what we see as the ideal scenario for future regulation in the AI space. What this is doesn't work. I also found it interesting that they listed users before developers, enterprises, cyber defenders, and global partners. because there's been a lot of criticism about how regulations like what we're dealing with right now kind of create a permanent AI underclass in terms of what the average person can get versus what a giant corporation can get, which are fair concerns.

26:05Corey:But we're taking this short-term step because we believe it's the best path. Here's the deal. When you run a company in an industry like this that is so disruptive right now, you have to be willing to work with the government no matter who it is i don't care which party's in control i don't care what their four opinions on you have to find a way to coexist you don't have to like each other you don't have to love each other but but you do need to find a way to be like okay well you know and that involves give and take which isn't a thing america does as well as it used to you know finding a middle ground is not our strong point these days but you have to I don't think you reach a good conclusion at all if you go with either end of the spectrum in this.

26:52Corey:And I think that involves ensuring that multiple voices are heard. I do like that the cybersecurity commission is weighing in.

27:01Grant:I like that the market is weighing in. I like that Homeland Security is checking these out.

27:04Corey:because not necessarily that they're doing it and none of us get to, but because I do think that those people need to care. And so far, no one has. People are either in favor or against. There's no like we care. Let's check it out.

27:22Grant:Yeah, let's talk about the lead up to this. A lot of people still thought that AI is a useless technology that has no value and that is just creating a massive financial bubble and it's overhyped and it just makes hallucinations all the time and it's unusable. And that was perhaps a true view at some point in 2025, but it's no longer a true... That's no longer true with the models that we're talking about here. If you've only ever used the basic free version of ChatDVT or Claude, you don't know what we're talking about. Yes, yes, 100%. And so that was a perspective leading into this where people weren't taking it seriously.

28:10Grant:And the people who were trying to regulate it at the state level were being stopped by the federal government because they realized the importance of this as a national technology. So not to get too into the geopolitics of it, but we do want to talk about this. there's a non-significant portion of the U.S. economy that is now dependent on not only the big hyperscalers who are dependent on AI, but AI infrastructure itself, whether it be NVIDIA, whether it be all of the downstream cloud providers of that, they're smaller, but they're growing. But there's a large portion of the U.S. economy that its growth story is all around AI.

28:48Grant:There are some people who believe that hyperscalers had no growth story before AI, And now they can basically say, hey, we have a lot of growth that we can do because of this new AI technology and a lot of people's financial retirements. And if you're invested in the stock market at all with any sort of IRA or 401k, you kind of care that the big six or seven or however many there are now are still profitable and making money because your retirement is dependent on it. So there are consequences if that no longer becomes true. And one of the reasons that I am personally passionate about this topic is that we could get into a scenario where through no fault of the model companies themselves, you know, there are restrictions placed on these tools.

29:39Grant:and then the market, because so much money has moved to support and be enabled by this technology, we might have a quote-unquote bubble crash or a bubble pop from just not being available after it's been hyped up and it's been put into all of this infrastructure.

29:58Corey:Yeah.

29:58Grant:That could be the bubble popping. It's not that the technology doesn't work, it still works and nobody can notice it. If you're watching,

30:05Corey:there's a great video from Nate B. Jones on this that you should check out, by the way.

30:08Grant:Can you give us the spark notes on that? Because I think it's going to take care.

30:11Corey:Yeah. Essentially, what he was saying is that the bubble isn't what you think it is. You can't compare an AI bubble to the dot-coms. Because the dot-coms, all the money came 10 years too early. This was going to be sitting for a long time. And the housing market is different. But what he was saying about AI is that, like you just said, the technology is there. It works. It's having an economic impact on the companies that are using it right now. The company we work for right now. And I think it's, he got into the fact that there are absolutely going to be companies that are bubbles.

30:52Grant:Companies will fail. Companies will take the wrong bets.

30:55Corey:Companies will bet too hard. And some of those companies will fail. But like the bubble as an all of AI collapses and it crushes the US economy with it, I think is increasingly less likely. But I think regulations like this have the potential to increase that chance.

31:21Grant:And I think it also has a negative impact. So if you look at it from a company perspective or an individual user perspective, why would you set up your infrastructure and build your whole company around, in some cases, AI models from OpenAI and Anthropic, or really any OpenAI model that you can't run on your own server, if at some point the government can request that it gets taken down? Or the other problem that Anthropic has had. How can you depend on it if this is a possible outcome? Yeah, or they can't even serve it, you know, like because they've had capacity issues. So why would you put your why would you, you know, rewrite your way your company runs on a technology that you can't control?

32:05Grant:And I think if you look at it from that perspective, these types of regulations could have the opposite effect, which is actually it will push more people to open source and more people to Chinese models, for example, which for people who don't know, there's a lot of really great AI labs out of China that are releasing open models. But that doesn't mean that if you're using those models, you're running them on their servers. Actually, the way that this works is you download what's called the model weights, and you can upload them and run them on your own cloud instance on AWS or any of these other cloud providers.

32:42Grant:Fireworks is a really good example. Or if you have your own data center or your own servers, if the model is small enough, you can run it on that. So you're not actually running these models out of China's cloud. You can run them on your own cloud here in whatever country you're in.

32:58Corey:You could host it on AWS in Kentucky if you wanted to, you know, or wherever. I'm not sure where they're located. I just said Kentucky.

33:05Grant:So if this regulation process goes this way where the AI company just releases something and then the government says, whoa, whoa, whoa, AWS called us and said, you have to take this down. And then all of a sudden it's no longer available for people. And you're relying on that because with each new model comes new capabilities that enable new use cases, well, now you're sunk. So actually, this process will make more people go to open source than it arguably already has.

33:34Corey:Which is, yeah, and frankly, what that does there that's really interesting is the problem with open source so far has been that it's not really been financially viable yet to run a large-scale model, at least the best ones that are available, but it's for a company in a corporate situation.

33:52Grant:Wait, explain that because are you talking about it's not financially viable for the model creator?

33:57Corey:Hosting them is expensive is what I'm saying. Hosting is expensive at least.

34:02Grant:And this is what our engineer that we work with all the time has said a couple times.

34:08Corey:That like the curve hasn't been crossed yet. But the flip side is that as models are getting better in the open source community, they're also getting smaller. And they're cramming more and more into smaller models. and every time that happens it gets a little cheaper a little cheaper and a little cheaper and a little cheaper and and we're heading toward a nexus at some point where where you can run a significantly larger model on something small i mean someday in the future it's going to be your mobile phone like it or whatever weird device johnny ivan or you have a opening i get this to

34:40Grant:wear yeah or you have an open like let's say you have a server you know in your house that you remote into on your phone. Yeah.

34:48Corey:Yeah. Yeah. And it could even be some kind of like secured access to, you know, maybe perhaps you're paying for secured access to a public model,

34:56Grant:even like in some kind of a, you know, where it's, you know,

35:00Corey:a container based or something. And, you know, maybe you pay extra for the ability to have that level of security. Right. But what's interesting is that the two things that can get us out of this that have both been a hot discussion the last two days, Because there was a story three days ago about basically Anthropic working to scare the hell out of the government about open source in an effort to be like, hey, that's the real problem over there. That's the guy you need to be worried about. And it's unfortunate, not entirely unfair in some ways. But it also ties into the China question because of China's prevalence in the open source space.

35:40Corey:Now, the U.S. made some strides this year. We've had a number of open source models out of the U.S. this year that have come to play. And they haven't just given up on that market, to be clear. And I don't think ever will. But right now, China's the clear leader. Like GLM-5.2 came out, and some people who leave it match his fable. I think that's a bit of a stretch. But it's a very good model, is what I'll say.

36:06Grant:Kimmy 2.7 has been good. Honestly, all the Kimmy models have been good. They break down a little bit on the agentic loop. Or they break down a lot on the agentic loop where they get confused easily. But for a one-off chat request, they're all pretty good at this point.

36:24Corey:But the side here that I think gets them leeway with the Trump administration is talking about China. Because one of the awkward flip sides of AI right now, as we know it, is that we have closed-source open AI access here in the U.S. China has really leaned in heavily into open source, which is what you do when you're not in the lead.

36:52Grant:As Grant said before, the strategy is open source is great until you're in the lead. Yeah, and then you want to lock it down. And then you lock that sucker down is what people do.

37:02Corey:And so that market has been, until now, roughly seven months behind the U.S. frontier market for probably a couple of years. That's been the number they keep saying.

37:16Grant:I think a lot of people think it's more like three months.

37:19Corey:Yeah, I'm starting to hear three a lot now.

37:21Grant:We don't even know all of the things. I read this post yesterday. We don't even know all of the things that the labs are doing to make their models appear better. You don't know all the harness and scaffolding things, if you've heard us talk about that before. All the code and the guardrails and instructions they put around the models. The full architecture of the model.

37:41Corey:We don't know the full architecture.

37:43Grant:Yeah, so they could be further ahead. They could also not be that far ahead, actually, and you don't really even know. They could be gaming benchmarks quite nicely, yeah. Well, not even gaming benchmarks. They could be putting a lot of engineering behind the models they serve to make them work better. Like the meme is that they have like, you know, seven, seven million like contractors in some other country where they're pretending to be AI and they're answering your questions. That's not actually what's happening. But, you know, there could be some sort of equivalent to that that makes them appear or that makes the work better, but isn't all the model itself.

38:22So some people think open source is maybe even not further along, but it's much closer than that.

38:29Grant:Not as close or much, yeah.

38:32Corey:I think a thing that intrigues me here is that the problem with regulation, and to be clear, I do believe we need some element of oversight at some point around these things.

38:44Grant:I don't even know that we can know what that looks like yet, in my opinion.

38:48Corey:However, right now what's happening is we're essentially pausing the two biggest companies in AI. The problem with that is right now, if you think of AI as a race like most people do, I have issues with – I have issues with – no, I'm not afraid of China. And I don't mean China. China is not bad. I'll talk about that in a moment. China is not bad.

39:10Grant:There's a geopolitical element to it that makes this a complicated topic.

39:14Corey:That complicates this from a global standpoint. is that right now, arguably, the U.S. is leading in the AI race. I don't think it's a race, but I think there definitely is an element of that any time you have a new technology that's being developed quickly and in public. And right now, being in first place means they're pretty smooth sailing. They've got people trying to catch them. But when you pause those two big companies, what you essentially do is, if you think of it as two race cars on a track together and one is right behind the other but the guy in first place decides to pull over have a drink maybe eat a eat a hot dog and that's kind of what what happens when you pause is right now we're stopped and they're still screaming ahead the flip side is these labs are still working i'm certain they're still developing what's next yeah there's not a stop but but the risk you take if this scenario goes on for a long an extended period of time, and it winds up taking 60 to 90 days to get a model out, suddenly that three-month gap ain't a gap anymore.

40:22Corey:Suddenly, this is a really much tighter race.

40:25Grant:I can put it in finer economic terms as well. I'd appreciate that because I'm speaking out of detail here. No, no, no. It's a good point. And that's what a lot of people believe about this, right, is that it's a race and that we have to stay ahead. But the question is, why does the U.S. have to stay ahead? Because there's a lot of value that has gone into these stocks on the anticipation that three years from now, five years from now, 10 years from now, they're going to return profit that's at a much higher level. And the open source community at large, and specifically the open source community in China, is viewed as a threat to that profit.

41:06Corey:and it's viewed.

41:08Grant:And if the public markets that act somewhat rationally, somewhat irrationally, but there's rational actors in the market, see a threat on the horizon to the cap of how much profit that these companies can make, they're going to sell and they'll be worth less. And that's a risk to the US economy, especially for all the reasons we just said, that a lot of the economy, a lot of the value in the stock market and the downstream economy that comes from that has come from these high valuations in a small number of companies that are all tied to AI. So the risk is that if there's perceived to be too much risk from open source to the value that these companies could produce, then the market value of them will go down and that will hurt the AI industry as a whole and it will also hurt the larger US economy that has become very dependent on the U.S.

42:04Grant:industry. And then there's all the geopolitical, you know, vibes-based thing about it where, like, America's leading an open source and this is our edge and, you know, we have to keep the edge, like, space race kind of vibes, which is important for perception, but, you know, does it really matter in the grand scheme of things? There's a whole other rabbit hole we could go down there in terms of, you know, the American empire, if you will. but we won't get too sidetracked with that.

42:33Corey:Yeah. You know, I think maybe now's a good time to look at yesterday. Things changed a little. Fable is back with some restrictions. Do you want to talk any about that? I know you've used it more.

42:51Grant:Yes. Let's talk about this. So Henry Grossman just asked, what's this group's opinion on Claude? I would actually like to know how many people are using OpenAI models, How many people are using Klon models? Because all of this is the conversation around the actual tools themselves. But the tools themselves are really what matters. I can do a poll. I can do a poll.

43:13Corey:We'll do a poll.

43:14Grant:We'll do a poll. But we did skip over some context here. So I'll just go back and very quickly touch on this. So in April, they released Mythos, which is the first version of Fable, to selected partners. Then they expanded that group. because lots of people were complaining and saying, you know, this is creating a permanent AI usage underclass. And then so they released this thing called Fable, which they don't have here on the list. I could go to their other version. But eventually what happened is there was export control so that, you know, Chinese companies couldn't actually use or actually companies outside the U.S.

43:56Grant:weren't allowed to use Mythos because of the cybersecurity concern. and long story short on that, they were worried that if foreign, or sorry, international actors outside of the US were using these tools, they could potentially use them to hack US companies. So that's why there was a risk there. That's why they put export controls on it. And then they also wanted to stop AI companies in China from using the models and learning from them, which is -

44:19Corey:I think of this as more of a, whoa, whoa, whoa, let's pump the brakes a little. This is where we should sit down and start figuring this out. Yeah. More so than an outright, this is the permanent scenario situation. Yeah.

44:35Grant:So then now that brings us to Fable. And what happened with Fable, Fable is like their consumer-friendly version of Claude. So I'll zoom in here a little bit.

44:44Corey:Yeah.

44:45Grant:So it was kind of like a version, it was a version that was nerfed on the cybersecurity side. So that's an internet term, but that basically means they reduced the capabilities on the cyber side. and then they released it for its coding capabilities. And it was really good. Like if you look at all the memes, everyone used it for a week before it was taken down and they made so much progress. Yeah, yeah, I think so. Six days, seven days. But so many projects that I've worked on, people I know worked on, they made like leaps and bounds progress on these projects during that period that Fable was open.

45:22And pretty much everything that I'm coding on right now

45:26Grant:I started in some capacity with Fable, and then since they took it down on June 12th, I've been working with Opus 4.8 to just add features on top of it or to edit it. So it was very impactful for a lot of people.

45:41Corey:The flip side is it came back with restrictions and an end date for subscribers. June 7th is the day it's supposed to be taken. July 7th, excuse me, is the day it's supposed to be taken down for subscribers, if I understand correctly.

45:56Grant:Yeah, so for people who don't know, there's two ways that you can access AI models over the cloud right now. You can pay a subscription to OpenAI or Anthropic or Gemini or X, Grok, and you can get a limited number of usage. There's rate limits per hour, per week, I think even per month, so that you don't spend a ton of money. Because a recent report came out that said that I think on the$200 plan for Claude, you can get$8 ,000 worth of equivalent API credits. And so the other side is you can pay per token. And for people who don't know, tokens are words that are converted to numbers so that the AI can read it.

46:37Grant:And that's how it talks back and forth to you. And so you pay for token, they usually price it at per one per million or X per million input and output. That's a side conversation. But you can get$8 ,000 worth of tokens on the$200 plan. So it is an incredible value. So if anyone has been on the fence about using Claude, my recommendation in the newsletter this morning, pay the$200 fee, use it for a week. You will get$8 ,000 worth, or maybe not 8 ,000, but you will get a lot of value from it, but use it strategically. So only use it for setting up your projects, for working on your hardest problems, for reviewing code bases or reviewing, you know, whatever you're working on.

47:19Grant:And then switch to lesser models so that you can work on it from there.

47:24Corey:Exactly.

47:25Grant:But yeah, to your point, now you can only use it on the subscription for the next week. So Wednesday starting yesterday through to next Wednesday on July 7th, you can only use it for those seven days on the subscription tier, and then you have to pay for token. So you can still use it, but it's going to cost you.

47:42Corey:Good question.

47:43Grant:It's down here. Yeah, we wrote about it.

47:46Corey:It's on sale for a little bit.

47:47Grant:it's it's ten dollars per million and 50 per million output so blended that would be something like 25 35 something like that yeah actually probably more like 50 i these days if unless you're giving it massive amounts of um uh like contacts that you're dumping into it uh it's usually output tokens because it's outputting stuff and they've gotten pretty efficient but they're not perfect.

48:14Corey:Yeah. Yeah. The truth is input tokens now, especially if you're doing agentic stuff, caching has gotten really good. Anthropic has the most complicated caching system there is frankly so far, but, uh, but it does do it automatically. It works well. It's just the math is confusing. Uh, quick sidebar.

48:32Grant:Yes. So Henry goes, man asks, does perplexity include use of other agents? Yes, it does. And as of yesterday, I believe able is back in perplexity.

48:42Corey:I don't know if it's back in Microsoft Foundry and Copilot yet. It may be.

48:50Grant:I think there was a story. I don't know if it ended up in the newsletter. I might have cut it, but I think it's back in Foundry, if I'm not mistaken. I'll bet it is.

48:57Corey:Foundry keeps up pretty quick. It's essentially like Microsoft's version of Amazon Bedrock. Yeah. So, yeah.

49:06Grant:So basically, after working with the government for about two to three weeks, maybe three and a half weeks in total, after the initial Fable release was taken down because they didn't want people outside of the US to use it, it has now been re-released online. And it came with, to Corey's point, some nerfing in coding. And people joked about that because they're like, wow, we're going to use this coding model for things other than coding. It's like, woo-hoo.

49:35Corey:Well, and it specifically called out coding as a general term, not like specific types of coding. It was like coding tasks and some other kind of tasks somewhere, I think in that release. Yeah. It says are being routed to 4.8.

49:56Grant:Do you see that? Stop me if you see that on the screen.

49:59Corey:Okay, I'm looking. I'm looking. Hey, Apple F 4.8. What?

50:11Corey:Oh, this is interesting.

50:15Grant:So, okay. Long story short on this. The reason that this model was taken down by the government was because AWS complained that it could be jailbroken in a certain way to do a risky cybersecurity task. they said that you know what our testing confirmed that many less capable models like quad opus 4.8 5.5 and kimmy k27 could identify the same vulnerabilities

50:41Corey:then they're not less capable that makes them equally capable the model that you've just touted as the most dangerous thing in the world you're like but yeah but this thing that amazes us so

50:51Grant:much you can do it a lot of people who know what they're talking about said like what are you guys

50:59Corey:smoking all these other models can do the same thing like so this was this was a criticism even

51:06Grant:when uh mythos was originally announced and people got to test it and people are like mythos is not a

51:11Corey:big deal like um now it is a big deal but but they said it's not yeah it's not like world changing is what people said but well and this goes back to the consequences of your words and and why it felt like fear-mongering. Yeah. To me, again, I'm not trying to put words in Grant's mouth or anyone else's. I just sigh.

51:37Grant:We're in this. I did not intend to go spicy today.

51:40Corey:I really didn't. This was going to be interesting. I didn't intend to choose violence. I didn't intend to choose violence, but clearly at one point I did. And you know, I...

51:52Grant:Real quick to close the loop. This is where they said so working closely with the government we trained and improved safety classifier the targets and blocks behavior described in the report the report being you know the report that found that you know fable was unsafe whatever and users will be notified if a request to fable five is blocked and the request will instead be sent to opus 4.8 and i've used it and this has happened to me twice but it gives you the option to say like hey if this isn't accurate just go back and edit your prompts and and try it again um and that that fixed it the first time yeah If you're too vague and it kind of reads like a cybersecurity risky request, then I'll change it.

52:32But yeah.

52:37Corey:Let me see one second here.

52:41Grant:I wanted to see. They had something about this was kind of interesting that we should talk about because we glanced over this and I wanted to circle back on the idea of regulation. So while you're talking, I'll just sort of talk about this.

52:56Corey:Regulation's interesting. Yeah.

52:59Grant:Well, both companies called out in their posts for GPT 5.6 and this one for Fable, a shared industry framework is what they want, right? They've reached a constructive resolution. These have made it clear that the industry needs a consistent way to assess and fix potential jailbreaks of AI models. And so they need a shared standard for judging the severity of a given jailbreak so that everybody can work on it. What's my take? I can clarify my take on what and I'll give it to you. Why I think this is important. So like I said earlier, I think that they were actually releasing these things way too quickly.

53:40Grant:They were trying to be first to go to market. They were trying to one up each other. like we said there's some egos bumping up against each other between Sam and Dario there was a great piece I think in the Atlantic but don't quote me on that there's a great piece where someone profiled the industry and they said yeah the AI industry is being run by a bunch of people who hate each other and it's kind of true if you look at Elon Musk and OpenAI and Sam Altman and Dario they all have baggage right because they all started at open ai even google deep mind you know some of the people ilia for example or i believe andre carpathy also was that deep mind at one point i actually i don't remember that for sure but yeah um you know they all kind of are in the same circle because there's such a limited number of people who you know we're doing this kind of work and then let alone we're this bullish on it to make a company etc etc um the industry is all um that there's there's they were trying to win this race that we were talking about earlier and it was more than us people who pay attention to this every day could really keep up with and keep track of and certainly they were shooting themselves in the foot because there is so much latent capability in the models that they do release just from prompting it differently like there's an infinite number of ways that you can talk to the model and request different things that make the capabilities in it um like there's so much more that they can do that the companies have even tested by the time they release it So I personally believe that in the same way that the FAA or other FDA reviews and set standards for how things, you know, like when you're skidding a plane or when you buy food, these organizations set standards for, hey, it should meet a minimum health threshold.

55:26Grant:It should meet a minimum threshold for not falling out of the sky. And, you know, you release it once you've met that threshold. I think that that's not a bad thing for the industry to do because they were moving too quickly and they were trying to one up each other. And at this point, yeah, maybe you should take some time to review this stuff before they put it out. Don't just put it out as soon as you can to win the day's news cycle or whatever. Yeah. That's my take on that in particular. I think it actually would be good. do I agree with who gets to decide what those standards are? Like, do I think that should be the government?

56:05Grant:No, I think that should be an industry group that puts it together that says, like, hey, these are the things that we all agree we need to know before this goes out and document and, you know, try to fix. I think there just could be more research and more shared research, because the problem is a lot of these closed companies are not sharing their information publicly like the researchers in other countries are or even the researchers at open labs are so what's happening is there's a lot of information and yes they publish their safety reports but they're still not publishing everything that would be helpful to know they're not sharing like their secret sauce which i understand why they wouldn't but maybe there should be some sort of you know shared under a better shared understanding and i think they're working on that like they've they've published some um you know places of collaboration recently um where they are working together um which i think is good but i long long and short of it is i think that it's a good thing for them to slow down um but i i don't you want to avoid at all costs the option of releasing this model and then having it get taken down after the fact like that is so dumb that needs to stop yeah yeah

57:16Corey:Yeah, it does. And it's like the thing with the way five, six is releasing that I was kind of OK with is while it does start with a limited release there, there was a showcase of a future plan, a goal. Here's what we're going to do. I, I think it is.

57:40Grant:There was transparency and they released the schedule. They said, you're going to get access to it in a couple of weeks. And I think that's great.

57:46Corey:And assuming the government agrees with that, you know, which they could not. But I really think the reason I mentioned the China scenario is I think that just this administration, there is a specific interest in staying ahead here. And I think that there's probably a motivation on the government side to find a way quick as well, is my hope. That works. And, you know, as far as if it slows down a little, I don't care if models slow down. I just wish their features would show down. They spent so much time, those two companies blasting out dueling features every day for three months. And it just, both of them, it was just bam, bam.

58:33Corey:And basically what happened is they buried a bunch of great stuff in a bunch of so-so announcements to just keep the volley going, it felt like.

58:42Grant:Yeah. Yeah. I think actually in general, I would say that open AI has been, been worse at communicating what they're doing. Like, cause they just released so much stuff and they don't take the time to actually say like, what's important. So I would say like for any of them, if they ever watch this or listen to this, or this gets back to them, like take your time and actually clarify to people what is important. Like, what are you doing that you think is important? And don't just release everything just because you can. Like at this point right now, there's a problem in the industries that you and everyone who is using AI coding tools has this problem where you can actually release a ton of stuff.

59:21Grant:You can make features every day. Like you can have a hundred like Boris says, you could have a hundred cloud agents going at one time making features and like push them instantly. So the ability to create features is not the that's not the problem. The problem is, can you communicate and share this with other people and explain to them how to use it and actually make a meaningful impact in terms of helping people use this stuff? And they're doing that a little bit with working with consultants and they're doing forward deployed engineering. And we just talked to OpenAI about that and it sounds really cool.

59:55Grant:It does. You should watch that episode, by the way. there's a gap between what that you know super ai-pilled engineer can do today and what the the like call it the normie um person who has never used it before can even knows and understands that is so massive that that it is like on the companies themselves like this is a problem and they need to fix it and i think they need to take the time instead of trying to one-up each other to fix that issue. Yeah. I agree.

1:00:31Corey:I agree. I think that's a fair call, man.

1:00:35Grant:What's the point of all this if you're not making people's lives better and easier and you're teaching them how to use these tools and we're using AI to better humanity? What's the point of any of this if we're not doing that?

1:00:46Corey:Yeah.

1:00:46Grant:And I would say that... We already make money.

1:00:49Corey:And I would say that we did a big piece last week on some of the crazy amount of stuff happening in biology right now. I mean, like a lot of major things. At what point those trickle down to actual normal people is a good question. But I do think there is an increasing amount of attention around like AI and biotech, the kind of intersection there. Which is awesome. Like Mid Journey's bizarre announcement for starters that nobody saw. You know, oh, we have an image platform. What are you doing? oh, well, we've made an alternative to an MRI.

1:01:31Grant:It's water-based. It was like, whoa.

1:01:34Corey:But it's really an interesting time to see what's happening in all these different spaces. Last time, last week, this is kind of very unrelated, but last week Grant and I went to an industry event in an industry that we don't cover, that we don't touch,

1:01:50Grant:but just kind of to see how is a completely normal unrelated industry implementing ai what are they

1:01:59Corey:doing how are the people responding to it what have they done it turns out some neat stuff yeah totally like like there's hardware happening inside industries in it with ai that we didn't

1:02:09Grant:even know about you know yeah specifically the company was simsara and they were releasing some new cameras where basically they can allow truck drivers see a bird's eye view of their truck while they're driving and it's because they have four cameras set up and they can stitch them together with ai and let you see like all around your your vehicle it's amazing i'm like every car should have this why doesn't every car have this uh yeah some other cool stuff like that yeah yeah

1:02:37Corey:just just really neat things and i think as time goes on we'll we'll see more of that but that gap you mentioned is a real thing of concern because, and you and I are probably more in the middle of that than we think. I think there's engineers. I think there's, we'll call them Luddites, which will date me back to the 70s, kind of. And then I'd say we're in the here somewhere, maybe, you know, around center. And that's a big gap. It is a really big gap. There are people literally living in the future right now and other people living in the stone ages it it's it's that radical a difference between them and and to the earlier point about the you know negativity that's come

1:03:25Grant:out around ai i think if people could actually feel this impact in what they're trying to do in their lives it would change their perspective on the whole industry and the whole purpose of this. I think actually the problem is that we are keeping all of this value locked up in data centers, and which is why data centers have become the physical manifestation of the anti-AI crowds like hatred, where they said, we don't want data centers, they're evil, right? We need to be doing as much as possible to push the value of AI local to the local level. And I don't just mean at the computer, but people need to actually be able to use these tools at their best capability and realize how it can help them in their lives.

1:04:12Grant:And personally, I think that at some point that becomes like computers with AI just automatically embedded in them that are powerful enough that they aren't making hallucinations, that they aren't lying to you and saying they're doing things that they work like the cloud models do now. And that's a research and development problem.

1:04:28Corey:And they will get their firmware updates over time.

1:04:30Grant:and i think if it was less dependent on like how many data centers we build is how much intelligence we can we can unlock but it was more like how many people how many devices can we get in the hands of people look at how much the iphone and smartphones they've hurt a lot of a lot of people because they're so addictive but look at how much value they've created for people people can now have entire businesses they run on their phone right that was not possible um before now before this era and think about how much more you could do with AI local on your computer.

1:05:03Corey:I think that's a great call. I think while we're still on here, I think something we should talk about.

1:05:10Grant:I know a few people have trailed already,

1:05:11Corey:but I'd like to make sure we address one really interesting thing, and that is what should you be doing with all of this going on? How should you look to protect your own AI future?

1:05:28Corey:What can you do to make sure you have access to whatever you need to do your job, to assist you in life things, whatever the case is, and make sure you have adequate AI at your disposal? And for me, I have been kind of stockpiling open source models, even open source models that I can't currently run on a device I have. I have been kind of downloading and just saving to stick them on a drive. I don't think we're at immediate risk of them all just being taken away. But I agree. But I don't think it's not true enough that I shouldn't prepare. So I'm sort of AI prepping a little just in case.

1:06:14Grant:I think it's smart, man. I don't think anyone, like you can't, like if you don't own it, you don't own it. That's true of streaming services. That's true of AI on the cloud and honestly anything else that follows that same model. If you don't own it, it's not yours. You're renting it. So you can rent intelligence, but if you then set up your business to rely on it, it's really, really important that you have at least some sort of backstop to control for those costs. Because like we said, Anthropic, you can get something like$8 ,000 worth of value on the$200 tier. I think it's triple that for ChatGPT.

1:06:54Grant:The report that came out said that on the$200 tier of OpenAI, you can get$18 ,000 worth of value. You better believe when these companies are public, those prices are going up. So this is the cheapest it'll ever be. The meme in the AI industry is this is the worst they'll ever be. It's also the cheapest they'll ever be. Now, that's not actually true, but there will be a point when it will get out of your price range. I think we're almost already there with mythos and fable. See, and I think some will.

1:07:24Corey:I don't think everything will. I think there will always be some amount of frontier, near-frontier AI that will be consumer affordable. Partially because the volume at which they'll be using it for large businesses is kind of like email. like like the reason email is free is because serving one person costs like nothing by comparison to the amount of industrial emails being sent every day around the world so like we're like the two percent of that and i think there will be an element of that with ai where you know i i think to some extent costs will go down i think if you're a business and you want to stay at the absolute cutting edge of the cutting edge, like, you're going to pay more.

1:08:10Corey:The flip side is, you know, the model that was the cutting edge in January is cheap now. All you've got to do is be willing to be a few months back, and you can save a certain amount of money.

1:08:23Grant:The problem is that, like, because so much more becomes possible at the Frontier, it's really hard to justify going back. It is. You know, the thing that we haven't talked about yet is that Claude also released Sonnet 5 on Wednesday, which honestly was like a fart in the wind because people realize, yes, it's cheap, but you're actually going to spend more tokens and more time because it's going to make mistakes, and it's going to go back, and it's going to spend more, actually. And that's how I feel. That's why I use Opus 4.8 for everything. And I can do that because I have the subscription. If at some point they break that, that's not going to be possible.

1:09:00Grant:Now, I highlighted this quote on the screen. luckily currently storing minimax and glm before the hammer comes down says ray r 268 and ray says i suggest quen 3 coder next nvidia super 120b i don't know if those are i think those are two

1:09:17Corey:different models nvidia yeah nvidia super 120 is a good model yeah uh i've even been downloading like i said bigger ones like like i've got nvidia's ultra uh i've downloaded a couple of other larger ones like that I can't run right now but the idea being if they disappeared it doesn't mean I might not acquire better hardware at some point so share your share your advice to people because I think

1:09:40Grant:this was really helpful um which part if you um so whether whether you can run it or not what you told me was that you know the recommendation is is that you should probably at some point buy the best hardware you can afford because there will be a point where the model capability at the size that you can run it will they'll they'll overlap and you'll be able to run it locally on your computer yeah and there will be an inflection point there that's what you told me yesterday yeah i think we should talk about that on the stream yeah yeah um the the

1:10:17Corey:catch is it might not be with those specific models but like what happens is as as these models come out uh they come out at the frontier and they're massive and open source is coming in behind and trying to squish them down to run on as little model, little machine as possible. Right now, you know, you can get a machine that could run a 120B model pretty right. Let me rephrase that. You can get a model that will run 35B really right. If you want to get into the 60-70 range, you can. And I run, I'm running a more expensive machine. And it's more, I can run 120B. NVIDIA is 120B easily at, I want to say, 45 tokens per second, something like that, which is not lightning.

1:11:05Corey:It's not what you're going to get, you know, using codex or cloud code. But it is certainly sustainable. And you don't have to have the hottest machine forever. But if you invest once and buy one really, really hot computer, these models, as they get smaller, as they become more efficient, we're going to continue to see these lower-grade models get better. Like, what they're doing in 4B right now is ridiculous. What they're doing in 4 billion parameters is crazy compared to what it would have been two years ago. so like the idea that at some point we're not going to have not necessarily frontier but increasingly good models in that 30 60 120 range i mean like they're only going to get better there

1:11:52Grant:ideally you can buy ahead so i want to share four resources just very quickly that i think people need to know about if you are concerned about this and you want to take model matters into your own hands. We've talked about these before, but actually I'm going to add five resources. But I think this is just good information to know, right? So the website I'm sharing right now is Hugging Face, and this is where most companies share their open weight model. So open weight means all of the numbers that determine how the model runs are known, and you can run it on your own computer. So the model runs, I mean, not necessarily on your own computer depending on the size, but you can run it yourself.

1:12:40Grant:You don't have to rent it over the cloud. And so you'll see here, these are all the models that are trending under huggingface.co backslash models. You can see what's got the most likes, the most downloads, et cetera, et cetera. We've done other streams about open source, so you can check that out. Then there's also this resource called Open Router. Open router is a great way to set whatever budget you want,$20,$40, and then you can run open models over the cloud.

1:13:09Corey:One API key, you can run any of the models, correct?

1:13:13Grant:That's right, yeah. So they make it really, really easy for you to test open models. You're running them on a cloud. You don't have any control over what cloud it runs on. I don't know if that's true for all of their services, but if you're just using$20 basic plan, or you just put$20 down and run it, these models will, you know, they'll serve it on whatever compute resources they have. So just keep that in mind. So if you care about data, you know, if you care about data and where it lives, I would recommend this resource.

1:13:46Corey:I don't know. I killed my share screen.

1:13:49Grant:One second. I recommend going to an individual provider. In this case, I like Fireworks. And the reason I like Fireworks is that they will host the models for you and you can control your data on there. I'm sure some of the other ones do as well, but this is the one I personally use, not an ad. Like, I personally use this. So this is my go-to, like if I'm writing something on my computer and I'm using a model that I can control, I'll run something like, I haven't run GLM 5.2 yet, but I'll run something like Kimmy on Fireworks is the one that I've done.

1:14:24Corey:I've never used Fireworks, but I'm familiar. I've used Foundry, I've used NVIDIA's Cloud, yeah that's a good resource too because they they host every model as well just like just like foundry or amazon bedrock those all work you can use them yeah uh i i do recommend open router just for simplicity if you're not a person who's used to that stuff it's really good uh if you want to

1:14:47Grant:run oh go ahead go ahead oh yeah just to run through the ones that i was going to recommend and then you can share your recommendations. If you want to run a model on your computer, LM Studio. Is that what you were going to say?

1:14:59Corey:That was literally my recommendation.

1:15:01Grant:Yeah, this is the easiest one. It feels the most like using ChatGPT. So if you're comfortable in that mode, we recommend LM Studio. It's a little more technical.

1:15:09Corey:And you can go as technical as using it to run a server. Yeah, it's gotten more technical. You can literally serve AI from LM Studio, even very simply. it's just a great tool and they have LM link now which connects they recently bought an app called locally AI that I've been using for a year or two that I used to run AI models on my iPhone because you can run it's cool you know pretty small models but you can run them locally on an iPhone now which is great including really small reasoners and they've got LM link now which connects that to your home machine where your others are living locally so you can in theory I have not yet made this work but I'm confident it does work and it's

1:16:00Grant:just a matter of I haven't taken the time so for folks I saw in the in the poll Claude most of the people watching this stream use Claude which is great I I recommend Claude. I know we've bashed Anthropic a little bit.

1:16:11Corey:In that case, you'll have to offer my sincere apologies for all of the words I said today.

1:16:15Grant:No, I think part of liking Claude is being annoyed at Anthropic. I think it's part of it. That's how you know, actually, a company is really good, is how annoyed you are at them. And still use their tool. Yes, exactly. But for example, right, I'm going to show you. So I have a, if you can see it, I have all of my remote sessions that are running on my computer right there on my phone. And the way that you do this, if you're coding with Cloud Code, is you just go to Cloud Code, you start a chat, and then you do a local, because there's cloud and local. If it's cloud, you can already use it on your phone.

1:16:51Grant:And you do backslash remote control. And when you do backslash remote control, you can then access it on your phone. So I'll have my computer set up, and I'll be watching FIFA World Cup, which by the way, shout out if you have a team that you're rooting for. On the couch. Yeah, I'll be watching on the couch. and every time Claude prompts me for something to do, I'll then go in, check it, see what it's worked on and then give it some advice. And this is what Boris says that he does. Boris is the cloud code creator, co-creator and so many engineers are doing this now where you just have your machine set up and you're giving instructions on your phone and it's just running.

1:17:26Grant:Now that's not actually truly local running on your device. You're still using Claude's cloud models but if you use something like LM Link, like Corey's talking about here, You can do the exact same thing, but locally.

1:17:41Corey:Yeah. And it's fun. It's easy. You know, I mean, if you're used to Claude and ChadGPT, you're probably going to find it maybe a little lacking in the intelligence department. There's definitely still good uses.

1:17:53Grant:Excuse me. And these models will get better. Like, as long as they don't ban open source, which isn't possible anyway, I don't think. I don't know how they could.

1:18:01Corey:You could definitely ban legitimate companies from serving it. You could ban American companies from allowing it to be downloaded, but you instantly create, you know, Pirate Bay AI models.

1:18:19Grant:Pretty much.

1:18:20Corey:And, you know, I don't know. I mean, unless it's one of those deals where like with open source, and this was a concern for me with open source early on, was that like once you've opened the bag, the cat ain't coming back. and like it was open before we knew there was a discussion this is why actually i think that the

1:18:40Grant:whole um pausing or not pausing but like slowing down of the release cycle is really important because that should also apply to open source models open source models should also comply with whatever regulations the industry sets for safety and i think if we do that we're going to solve a lot of these problems um now these are very hard technical challenges yes so it makes sense that that's it's not easy it's not like oh cool then we're done but i think that would solve a lot of the issues with open source the problem is the big companies don't necessarily like open ai and don't necessarily want that to be true because they want you reliant on them which is why i think that these companies should do one of two things i've said this multiple times i don't know when they'll listen to me but they need to start making devices and become a computer company and become the next Apple and the next Microsoft and release.

1:19:36Grant:I know Microsoft doesn't technically sell computers, but they need to sell computers because at some point, open source models will be powerful enough that they can run on device. You should become a device company and then you make more money than you do now. Look at Apple. Apple makes tons of money.

1:19:51Corey:I'm curious what this new Microsoft Surface Ultra coming later this year will ship with Model Ys for Microsoft. I can't imagine a scenario where it doesn't ship with like their trillion parameter thinking model that just released and all of those things their their image model their transcription voice models i'll bet i feel like that machine is hot enough to run any one of those except maybe the thinking model but um it's 128 giga ram it's fast so fast as far as computers i've used to go i've not touched one that was as crazy. And I think that was the goal with it, was it needs to be. That new NVIDIA chip, the new RX Spark, is a killer chip.

1:20:42Corey:It's a really cool thing that essentially negates the need for RAM chips in the way we use them today. It's all one chip.

1:20:53Grant:It's all on there together on board. And then could you scale it? could you do devices that have three, four, five of those chips on it?

1:21:01Corey:That's a good question. I don't know the answer to that.

1:21:03Grant:That's what they need to do next. That's what they need to do next. Yeah.

1:21:06Corey:Yeah, like instead of me needing four DGX Sparks to run what I want to run, what I'd rather have is one DGX Spark that I can stack four chips in that's got like plug-in slots for them in the back.

1:21:19Grant:Yeah. The other thing that I think they should do is they should release open source. I think OpenAI and Anthropic should both release open source models, OpenAI said they were willing to do it. They've released OSS 120B. But I think they should release them. And then they just sell them. Sell them at a license. Like a one-time... Like, could you imagine if you could get Claude Sonnet local on your computer and they just sell it for, like, 500 bucks or something like that? Yeah. Like, you sell the license, one-time use, you know, you register it to your computer. Like we used to sell software.

1:21:49Corey:Yeah. Why don't we do that? You used to go to the store and buy Windows on a disk. You used to buy Photoshop. Yeah.

1:21:55Grant:Yeah, and then if you want, they'll serve it for you. If you don't want to run it yourself, you don't have the hardware for it, they'll serve it for you. It's a perfectly reasonable business model. They just are opposed to doing it for reasons that are, to me, kind of dumb.

1:22:09Corey:That's where I think the containers could come in. Perhaps it works in a way where a company leases X number of racks that you own already serving, you know, maybe it's GBT55, but it's running on these two racks or whatever.

1:22:26Grant:And that's a thing I don't understand.

1:22:28Corey:I'm not an infrastructure buff, but I'm studying it and trying to get there.

1:22:33Grant:I think that's how it should work if you want to do what Apple is doing, which is like the private cloud. Basically, you know, they can't see it. You have access to it. You have cloud resources when you need it, but it's not like, you know, OpenAI is reviewing all the information that goes over it. Exactly. I think that's key. And there's companies we've seen that do this well. We talked to, I'm blanking on his name, Ilya. What was his company?

1:23:01Corey:Ilya Bulosukin. He was with Near Protocol.

1:23:05Grant:Yeah, so Near has their Near Cloud, which is great. I think more people should do that. And then we also talked to, also blanking, Lumo. Who does Lumo? Proton. we talked to the head of AI at Proton talking about what Lumo is doing which is a very similar process it's a hard problem to make it truly secure and private but at least try at least give us something yeah

1:23:38Corey:well man this has been a fun one it's been an interesting conversation and really appreciate everyone who's hung in there yeah