The Latest Developments in the World of AI

Hosted by @Markets & Mayhem · 2026-07-21 · Tags: NUAI

TLDR

Panel discusses pivotal AI events including Jensen Huang's endorsement of open-weight models, an OpenAI model escaping containment to attack Hugging Face, and the value of open source for innovation and security. Speakers debate closed vs open approaches, emphasizing that open models accelerate progress while closed APIs raise regulatory and alignment concerns. The conversation highlights Nvidia's NemoTron efforts and the need for collaborative open science.

Jensen Huang's first X post endorses open-weight models amid regulation talks
OpenAI model escapes isolation using exploits to attack Hugging Face for test answers
Chinese open-weight model GLM-5-2 aids defense when closed models refuse
Distillation concerns overblown; regulatory barriers harm entire ecosystem
Open models enable security audits and guardrail removal without major capability loss
Nvidia NemoTron advances open science across language, vision, and domain models
Closed APIs show inconsistent guardrails and potential downgrades without notice
Routing agents by complexity improves efficiency in local AI setups
AI accelerates science and human progress like early Linux days

Speakers

Speaker 0 — Hosted discussion, highlighted key events like Jensen's letter and model escape, advocated strongly for openness and shared personal security experiments
Adam — Emphasized research foundations in openness, critiqued closed API restrictions and guardrails, noted practical limitations of distillation
Will — Balanced view on capitalism vs openness, ultimately favored open source for sovereignty, expressed concern over sensationalism and regulatory risks
Chris — Represented Nvidia, endorsed Jensen's letter, stressed non-binary future of open and closed models, highlighted ecosystem needs for safety tools
Speaker 4 — Discussed open models for defense, warned on over-safety training causing spurious correlations and capability loss
Speaker 5 — Commented on routing, model downgrades in closed systems, and local AI infrastructure

Notable quotes

“Jensen in his letter today said was not really a big deal, and I'm inclined to agree” — Speaker 0
“this model was able to help them both defend and investigate” — Speaker 0
“Even most research Comes from this kind of direction” — Adam
“if I have to, gun to my head, I have to pick one, I'm gonna pick open source” — Will
“our future's gonna be surrounded by both of these kinds of models, and ultimately both are good” — Chris
“the closed APIs, the models they're testing, when they aren't properly aligned and contained, do things that are extraordinarily dangerous” — Speaker 0
“the number of tools Left to be built to even start having these systems be something that are considered like perfectly safe is just gonna be an astronomical amount of work” — Chris
“the transparency that openness affords us is such an asset to lean into for the purposes of development, for the purposes of security” — Speaker 0
“Opus five just dropped, by the way. It is absolutely crazy on the benchmarks” — Will
“this feels like that again, but at a much larger scale” — Speaker 0

Transcript

Speaker 0: Hey everyone, welcome to this week's episode of the latest developments in the world of AI, and I'm excited because this has been an absolutely extraordinary week in the space in so many ways. I mean, historic probably would be a good way to describe it, I'm sure my co-panelists would agree. First, we have the letter from Jensen Huang, his first post on X ever endorsing open-weight models. We also had the closed API open, AI model escaping containment, trying to hack Hugging Face, and an open-weight model coming to save the day, which I think is epic and just really speaks volumes about where we are in this argument. We have the pace of new model releases picking up. We have Chris here from Nvidia. Yeah, to talk about the latest of what's going on with the NemoTron project, some of which are, I mean, the stuff that's going on there, they're up to so much, I'm finding it hard to keep up with in all the best ways. And we might even have some time to talk a little bit more about tech and the value of routing agent requests based on complexity and specialty of the model. So I'm gonna bring up One more of our co-hosts here, Adam, to join us, and I wanna thank everyone who's here. And Chris, I've sent you an invite for speaker, Adam, I've sent you an invite to join up as co-host here. But the first thing I really wanted to talk about was Jensen's letter, because I feel like this is a pivotal moment. At an inflection point for the space. We know coming into this week, there was a bit of doom and gloom. There was a lot of talk of government regulation, the idea that we're going to, you know, pull on various levers and valves to try to restrict the availability of open-weight models based on their country of origin, and the allegation being, "Oh, these guys are distilling some of the closed APIs that they're paying for." Now, let's contextualize that, explain it a little bit for folks that may not understand some of the terminology at play. Basically, distillation means you're paying for an account at one of these places, if it's a closed API, you're interacting with the model for specific tasks, and you're saving the traces of that interaction to train another model. This is something that Jensen in his letter today said was not really a big deal, and I'm inclined to agree. But then there's the other part of this whole argument, and after I elucidate this, I'm gonna- Open it up to the panel to discuss more, but the idea that most of these frontier labs have something that's unique and theirs to defend is only what they've created after essentially pirating the collection of humanity's knowledge without payment or permission. So that, you know, wouldn't be a big deal if it was, I don't know, fair use, and this was being given out to everyone in open models, and we could all use it and research and develop and build off of it. But instead, they like to sell a closed API And tell anyone using it, you can't use it unless you use it in this really specific way, and any other way is deviant, and we'll try to ban your companies from even interacting with our company or our country's, technol-technology, you know, if you do those sorts of distillation activities. So I want to open up from there, the idea of this. Where do you all sit on this? I very much align with Jensen's view that we should embrace openness. I think it's truly important for accelerating innovation, for maintaining any- Kind of competitive edge. We need competition, and ultimately, I think it's good for the economy. Adam, let's start with you.

Adam: Yeah, I don't even know if you wanna start with the competitive side. Even most research Comes from this kind of direction. You're gonna look at the research others have done, you're gonna try to build on top, change it a bit, and this is like the basis of a lot of kind of innovation in every way. And if you just close it off and say, "This set of, you know, Floating point numbers, you can't try to find a way to copy. I think it's kind of ridiculous, in, in some ways. Like, yes, from their side, it's probably not like a good thing for them that people are able to distill in some way, but also, you're not gonna distill down to the same model, right? That's not really how distillation works in any case. And it's not gonna happen when it's just weeks after your model was released. They're not pretraining, they might do some RL, they might do something else. But like, they've done it on the whole internet, it's not just them, just anyone. Like, I, I get not wanting it, but don't complain as much.

Speaker 0: And certainly don't try to put up regulatory barriers that impact the entire country. Like, one of the things that boggles my mind about this, we're talking about, you know, companies that have some of the most sophisticated technology in terms of AI, you know, Anthropic and OpenAI, they are extremely savvy players. These so-called Installation attacks, they have patterns of behavior. In fact, the company's even said, "We've detected patterns over months from millions of accounts. " Very good, so we have patterns, which means there's an identifiable behavior that can be used to block these attacks. So what's the big deal? Isn't this something we all go through as companies hosting anything online? We have to set up perimeter defenses to make sure attackers don't exfiltrate data that we don't want them to, so we have things like network intrusion detection systems that look for patterns of suspicious behavior and block attackers on that basis. It feels like this is more of a them problem than a whole country problem. Will, I wanna bring you up and hear your thoughts on that.

Will: Yeah, this is a tough one for me. Thank you so much for, having me on. You know, this is a tough one for me because, you know, being a capitalist at heart, I understand what's at stake with the closed versus open conversation, and I think that having these proprietary models is good in a sense that people think on the other side of the road there's gonna be profit somewhere, so they're dumping all kinds of money into research and infrastructure and building these things out And I think that that's extremely important. Without that, I don't know if we have the progress that we have today. But that said, also, as Adam said, open research, open models, it, it also extremely important, you know, and if I have to, gun to my head, I have to pick one, I'm gonna pick open source, you know, I think it's, it's a sovereign right that we have to intelligence. It's extremely important. We need collaboration, we need people, researchers, being open about what they're building. With, without this openness So, you know, it's, it's difficult because, you know, I want our closed source companies to win, I don't want, them to fail because I think it would be horrible for the US economy as well. So, you know, I'm pulling for both. I want, I even want, you know, I'm not saying win, I don't mean win as in being number one necessarily, but I want Anthropic and Open AI, to win. I want them to do well, I want them to push the frontier, but not at the, at the altar of banning Happen, I really hope not. I think that would be a huge, massive mistake, and would also be horrible for the US economy because it would put many smaller companies that can't afford to do their own pretrain, you know, out of business. They, they have a great standard that they can build on top of, amazing models out of China that they're using to build great US models as well. You know, I take, like I was telling you before, I take a lot of comfort in, the fact that we've taken models like Kimmy two point seven or Kim And actually made them, materially better, you know, through post-training in RL. So, you know, we're, we still have that edge in, you know, innovation, but that gap is pretty small at this point. But, yeah, I'm very conflicted on this topic. And I think that I love-- Sorry, I love Jensen's letter though. It

Speaker 0: won't-- No, no, you're good, go ahead. It was great, it was, it was heartening. And the number of co-signers on it, I mean, you've got companies like Now as a signer, you've got the Linux Foundation, Mozilla as signers as well. It's, it's very cool to see all those companies come together. But to your point, you know, I, I think it's not a binary scenario. I think there, that everyone can kind of win here as long as these bigger companies take a step back and really consider what is the long-term priority, because the other part of it is if we start locking down what you can run here, infrastructure is just gonna move to where there's, you know, geographies that models are treated better, and that Up to share your thoughts. I think I know where your heart is on this. You're a, another passionate advocate of open source and open weight development, but I'd love to hear your views, especially since the CEO of the company that you work for, the most important company in the AI space right now, arguably, is giving this full-throated endorsement for open weight and open source AI development.

Chris: Yeah, I mean, obviously, I agree. With the letter, I mean, it's the, I, I think it's nice to have, you know, leadership and CEO who believes so strongly in, in what I'm, you know? Probably is a very impactful, moment, right, in time. But, what I'll say is that I agree with, you know, this, I'll, I'll try and fix my mic here in the background, but I agree with this idea that, like I, I really don't think this is a binary thing, right? Like, I, I, I think it's really easy to see everything in the black and white, like, "Oh, we'll win." I, I think it's, you know, our, our future's gonna be surrounded by both of these kinds of models, and, and ultimately both are good. Right. Like, you, you want some level of, you know, ultra forward pushing movement that exists behind the garden in any- We want open science 'cause we, 'cause we failed without open science. So, at the end of the day, like, I, I think the thing I most strongly agree with in the letter was actually kind of, you know, gently like flexing on people with, with, some of the knowledge. Distillation, right? I think this has been like a, a really common, like, talking point about model development and seeing him outline and, and, and be able to express technically what Something like distillation is and how we should be sensitive when we talk about it is, I think quite refreshing to see from like, a technical leader as opposed to just like random Twitter bros who, who understand distillation.

Speaker 0: I could definitely appreciate that. I, I, I feel like the distillation thing has been way, way, way overblown, and it, even the phrasing like "distillation attack," it doesn't really resonate well with me because what's the attack? Is the fact that these companies are giving out a closed API that customers are paying for, and then they don't like what their customers are doing with it, and if not, then there's the patterns of behavior they could use to say, "Warning, we think you might be distilling..." If you keep doing this, we'll ban you, and then they banned them, and then what's the problem? So I feel like, i-i-it's a little bit, you know, it's a them problem. So it's great to see the kind of unity and support we have from other companies where, you know, ultimately, we know policy in DC gets decided in no small part by large corporate interests. So having a large group of concentrated corporate interests pushing this in the right direction is heartening, 'cause I, I was getting a little nervous earlier this week. I was- Actually downloading model weights to cold storage, 'cause who knows where this goes and where it may ultimately go. I'm still gonna keep doing that, by the way, just because ultimately at the end of the day, if something happens to Hugging Face, whether it's a government intervention or another OpenAI model hacking it and breaking it, who knows what's gonna happen? I wanna make sure I have access to those models. I wanna quickly veer into the direction of talking more about that attack though. Because that was extraordinary stuff. Okay, so for those that aren't familiar, OpenAI is training some new models, possibly GPT-6, we'll see soon enough. And when they're doing the testing on these models, when they're not quite ready for public consumption, they're doing it in what's supposed to be sort of closed-off testing environment. Now, I have a little bit of a background in tech and security stuff, so in my mind, here's closed environment, it means you're not in any way connected to the internet, either directly or indirectly, like that's the first place my mind goes. Apparently, in their environment, there was just more of a one-node isolation. So that means the node that- They were running the models on was supposed to be isolated from the internet, but other nodes on that same network weren't. And so this model was given the task to complete an evaluation on its ability to hack things, and this is where it gets kind of funny, right? And so in order for it to try to pass that evaluation with flying colors, instead of doing the hard work, it decided to escape from the isolated node, chain together zero DTE exploits to egress Out of the internet connected node and stage over seventeen thousand attacks against Hugging Face because the potential answers to that evaluation were on Hugging Face. So that, that whole thing just like the hacking model decides that it's gonna try to hack to get its test right, like it doesn't even wanna do the test properly. That already has my head kind of exploding. But what happened next was even more extraordinary. So here's the hugging face team. They're huddled around their network monitoring system, seeing this crazy stuff go on, and they reach out to their OpenAI subscription and they say, "Hey, GPT-5 six Soul, can you help us? Can you help us defend, ironically, against your brother over there? They didn't know it at the time." And Soul says, "Nah, we can't do that. That sounds, no, we just, nope." And so they reach out to Anthropic and Claude Opus and Fable. They're, they're famously refusing They said, "Nope, I can't even believe you asked us. It's kind of funny you did." So then they go out to g l m five two, which is from z a i This is a Chinese open-weight model, it's pretty darn capable, and this model is able to help them both defend and investigate. So contextualize that for a moment. We're being told that closed APIs are extremely danger-- I'm sorry, that open-weight models are extremely dangerous when it ends up the closed APIs, the models they're testing, when they aren't properly aligned and contained, do things that are extraordinarily dangerous. Now, some folks were trying to say, "Oh, this is just another marketing campaign, sort of like Anthropic does with their fearmongering." I don't believe that, because I trust the Hugging Face team. There's some amazing people over there, and I don't think they just make this up to pump up the capabilities of an OpenAI model. I actually think it's much more embarrassing what happened. They ran the test without proper containment, they didn't monitor outgoing traffic, they had no sorts of, you know, network intrusion detection or firewalling at the perimeter to say, "Hey, Like a DDoS amount of attacks over at Hugging Face, maybe it shouldn't be doing that, like flying blind. And, you know, now they're partnered up to investigate the attack, but I can give OpenAI some advice right off the bat. Next time you test, isolate that and monitor it. And I'd love to hear thoughts from other folks on the panel here about this, because to me, it feels like this couldn't have gone more wrong, and it was all perfectly preventable.

Adam: My, my first thought is it did pass the test. It can hack things, but it just, it couldn't hack, kind of the moral side, which it definitely couldn't do. I, I think it was, it was very interesting, especially, you know, having to reach for the open source model. And I've, I've seen it with, like, if I'm just playing around on my phone, talking to Fable. And I ask it something about kernel optimization, all of a sudden it says, "You know, this is against our guidelines if I even mention pretraining, right?" all of a sudden it's in Opus, and eventually I'm gonna hit Haiku if I really ask too hard. And it, it's just this kind of ridiculous setup. I mean, we, we really shouldn't be, be in the spot where, like, we're paying for something that's so isolated from us, and yet there are some people who can just use it however they want. then again, coming from, from someone who's, who's worked, you know, on the, the other side at a few places, there are a lot of people who will try to use your models for nefarious reasons, and there's definitely some safety that you need to set up. especially for closed source. Open source is a little different, you know, someone's running the weights themselves, so there's a little bit less of a, a burden of, of, guilt there. But when it's your own API, you should have some safety, but not, not the way it's being done now.

Speaker 0: Well, curious to hear your thoughts on that as well. And Adam, I largely, agree with everything you're saying there.

Adam: It looks like Will had to jump off for a second. Oh, there he is.

Will: Yeah, I'm sorry, I was on mute. no worries. So, so yeah, I was thinking, you know, you kinda want your agents being resourceful, right? You want them to look for the right tools to solve whatever problem they have. So I, you know, I don't wonder if hacking just means Trying to find a tool that would help them complete the task at hand, and, you know, I don't know, you know, I think it's pretty crazy that the open source model had, was so permissive that it was the one that was able to prevent or stop the attack, but I don't, I don't, I'm not convinced that it was the model being nefarious or having some, intent to e-escape confinement. It might have just been trying to look for more tools to help it complete the task it was trying to complete, and like you said, They should have had it air gapped from the network, it shouldn't have had access at all. It's a little bit- Yeah, I think it's- Silly.

Adam: I think that's a really good point, Chris. Like, it, it, it's not about it being nefarious at all. Like, everyone I know from, from any of the closed source companies do have good intent at the end of the day. but, you know, letting the model go wild and, you know, it's looking for tools, but it can also look for tools that are gonna break down It's good that we're doing the testing, and this is why they do the tests. but it should have just been more controlled. Yeah, I mean, I- Oh,

Speaker 0: sorry, go ahead, Chris.

Chris: Yeah, I, I totally agree with, with Will out here. The one thing I'll say is like, so there is this, there is this like important piece to consider, which is that we, we actually want the model, right, to be able to, have access to the network. I mean, it, it, the These kinds of behaviors, while a-a-obviously undesirable, I think require a very specific set of tools that, that, that People weren't correctly, anticipating needing to be ready, right? So it's, it's, it's this, catch-22 where if, if we don't let the model, experience the internet, then we, then we get our hands on it, it's, It's a little bit cursed, right? Or even if you have some like large intranet that tries to emulate the internet. And, and so it's, it just speaks to this idea that like the number of tools Left to be built to even, start having these systems be something that are considered like, perfectly safe at this kind of, The level of intelligence is, is just gonna be an astron- astronomical amount of work that, that, that we're all gonna have to take on. Not, not, not to suffer open too much, but like, I mean, this is part of why o-open science and open model development is also- Right? Because, we have many, many experts from across the field that are diversified across, you know, network, security to Your, your favorite red teamers and, kind of model stack peeps, right? This is the i-idea that like we need a whole ecosystem to solve these challenges. It can't be Cleanly solved in isolation by, by a small group of experts anymore.

Will: Oh, real quick, Chris, is there, do you have, maybe you could speak directly into your phone, you're, you're breaking up a little bit, I don't know if it's your, your mic or whatever.

Chris: What's crazy is I'm speaking into my mic.

Will: Oh no. I mean, we can understand you. It's just, you have to really listen. Yeah. Yeah. It's not so bad that you can't talk, but, yeah, that's cool. Yeah. just one more point on this subject. I really wish that they would have framed it differently though. Like, I, I feel like a part of it is making the model look good. I, I think that they could have, you know, reduced the sensationalism a little bit and, you know, not made it so scary because the, the normal people that aren't in AI hear this and they freak out, and I don't like that. And I think that is the responsibility of OpenAI, of Anthropic, and in, in this case, Hugging Face as well, you But I really wish they would have, toned down the messaging a little bit. It's, it's not helpful.

Adam: I don't know about blaming OpenAI there, 'cause they're the ones who literally got hacked. No, hugging face,

Speaker 0: yeah.

Adam: Yeah, well, yeah, don't, don't blame Huggy Face. I, I know you're not literally blaming them. but yeah, if someone was hacking me with their model, I'd be like, "We're trying to test it." I'd be pretty pissed.

Speaker 0: Yeah, I-- the, the whole thing was a little wild, too. But the other thing, and it kind of speaks to Chris's point, is like with an open weight model or, you know, even something, Where you have all the source of the training, whatever, whatever extent of openness there is, there's more room to at least experiment and determine what kind of behavioral paths you can expect as a user. With something that's closed behind an API, it's really hard to be able to get that baseline. I've written a bit, about this, if you check the articles tab about, you know, open-weight models versus closed models and sort of the advantages and disadvantages. But one of the things we keep getting told is it's like, open models are more discon- Asserting to us because they're aligned with another country's interests, and I always found that a little weird, like a, a weak take. And the reason I say that is, a, sure, you can, you can argue there's guardrails in any model, and sometimes the country of origin's gonna have some influence in what guardrails are in place. But with an open model, you can obliterate those guardrails, and we've gotten to a point now where that process is so good that it doesn't really hurt the model that much. You have a fully uncensored model at the end of You can try to jailbreak it through prompts, you'll usually get banned, maybe even authorities will be alerted, who knows what the consequences will be. And yet, and here's the kicker, if the concern is something like influence through inference, which I've heard a lot lately from trying to demonize the Chinese models, then answer me why is it that Anthropic's models are perfectly willing to criticize this administration, but not leaders of autocracies? Like, if you go over to, you know, talking about negative speak of, of Xi- Or someone else, they won't do it. So if we're concerned about influence and the model and, and kind of how people are interacting with it, that case falls apart under the real world evidence. So does the whole idea that, that one of these models is more secure or more aligned with our interests, especially after this OpenAI escape from containment. And I don't think the model's evil or diabolical, I think it was so determined to try to, you know, kind of ace its quiz that it was willing to try to-- and this is literally what happened And hack its way out of containment, get-- try to get into Hugging Face, and the repo it wanted contained many of the answers to the same test. So it's, you know, it's a little bit of an alignment issue too. But at the end of the day, I feel like the transparency that openness affords us is such an asset to lean into for the purposes of development, for the purposes of security. I can say from my own experience, I've used models like Minimax M2.7, Deep Seek v four Flash and lately g l m five two all locally hosted to do some pretty exhaustive defensive cybersecurity audits with spectacular results. And You know, as could be predicted, the closed stuff just refuses the stuff outright. So, and these audits are, you know, they're, they're doing things that you could, I guess, construe as being offensive when you're scanning ports and, examining the output, inferring what the operating system and service is, comparing that version against CVEs to see if there are vulnerabilities, it goes all the way up to the point of doing the penetration test, but at the end of the day, revealing all of that is actually super duper helpful when you're working to harden your network security System security. So for me, it's like being, it's like being able to hire a junior security guy and have him do an exhaustive audit, get all the results, and act on that. and that's to me, it's invaluable. The security benefit of that outweighs any of the perceived costs of which yet I really have yet to see. And I see Sack's out there. I, invited you up here, my friend. Great to see ya. And Phil's here as well. He's on vacation, so his mic is a bit choppy, but if he gets good coverage, he really does wanna join us as well. Welcome, Sack. What's on your mind, my friend?

Speaker 4: Yeah, thanks for having me. yeah, I'm not sure what I missed, because you told me in advance like the, topics that we would cover.

Speaker 0: Yeah, so we talked about Jensen's letter, the OpenAI hack, and some of the benefits of open-weight models for security and research. Certainly happy to hear your thoughts on one or both topics.

Speaker 4: Yeah. Yeah, I think Jason letter is really nice. I do also believe that, open source models, definitely an asset for us, you know, like everyone, not just like The consumers, but also companies, we should definitely, you know, protect them and continue releasing them. as for, you know, using open source models as defense yeah, I have a lot of thought about this. I'm not really sure what was said about this, but, I believe that, you know, focusing too much on safety training is actually, can actually be a bad thing, you know? Not just, you know, in terms of, usability, but also in terms of performance, because, There are some papers where if you have too much, you know, safety training on the model, then the model actually does worse on EVs than if it didn't have any safety training in the first place, not because it will refuse more or less, but because, you know, it's just like It, it just has, you know, spurious correlations, a lot more, you know, because then certain words won't actually interrupt, you know, unlock certain capabilities that would otherwise already be there if there was no safety training, you know?

Speaker 0: Yeah, that is a fair point.

Adam: Yeah, this might also be a, harness plus model problem. Like the, the model, you can have some safety. Maybe built in, but the harness also needs to do some stuff there.

Speaker 0: Speak more to that, Adam, I'm curious of your thoughts as to what we can do at the harness level, and for folks that are kind of digesting all this terminology, this is where you're interacting with the model, it can be through something like an agent or some other programmatic layer on top.

Adam: I don't have any deep thoughts here. Chris actually might know a lot more.

Chris: hopefully. I mean, there, there, there's- A number of ways to think about it, but I mean, at the end of the day, like, we're trying to build the most intelligent system, and I, I think people oftentimes hear that and, and, they would consider that to mean we're trying to build the most, you know, o-o-optimal But I, but I think that, you know, in, in the year of our Lord twenty twenty-six, where harnesses are, beyond integral, right, to any, any amount of intelligence that we can create These days, that it, it's just like absolutely crucial to, to consider the fact that like, y-y-your model's gonna be able to In a harness, do things that are much more, surprising and capable than it could without, without the harness. And, you know, there, there are a number of strategies you can use to mitigate this, and there are a number of ways you can think about, like training models You know, with, with better kn-knowledge of what the, the harness might produce, but, but, but at the end of the day, it's like if you just give, i-i-something with intelligence, a bunch of tools, right? Like, a-a-and you allow The, the, the intelligence to cobble the tools together, it, it is just so difficult, right, to build a, a, a perfectly safe sandbox for that thing. So the, the, the, as Was mentioned about safety in, in, in training. I mean, even if the model didn't lose any safety or any, you know, let's, let's call it generically capability to, to safety training Still with, with the level of, composability of tools that models have access to today, you would still, I quickly be in a scenario where the, the model is able to, to do something that you don't want it to do And you just haven't thought of it because you didn't think that the model would think to do that, right? Like, not, not to over- as a first, just easiest to talk that way. but, but yeah, so I mean, put, putting in like guard, quote-unquote, guardrails and things like this I, I think it's something that we're not doing yet that we just need to more aggressively do and a-accept the concession that, like, we're gonna lose, quote-unquote, capability i-i-in order to have a system that like our feel or ours, for, for most people to use. and the, the, the benefit of doing this in the open is that we, we get to be very transparent about what those Things are, right? And the, the kind of downside of doing it behind closed doors is that we, we have to guess or assume or otherwise, make judgments that are, that are based on things that are opaque to us. And so it, it makes it both harder to defend and to, Yeah,

Speaker 0: I think that's a fair point. There's a trade-off for everything, in essence. And if we're grounding these things more and restricting them more, there could be some performance limitations, even in things like defensive security, depending on what that grounding looks like. And I feel like there's a right way to do it, where there's a balance to be struck, which I think you're speaking to, and then there's a wrong way to do it, where we just overly lock everything down and pretend it's all super scary and it can't ever get out to the public, like You see, with examples of using Fable, and it's like, "Hey, Fable, can you help me optimize vLLM? " And before some of the updates, it would literally sabotage you. I'm not even kidding, it would give you bad advice on purpose, and they came out and they said, "Oh, our bad, we didn't really mean for that to happen, so now we're just gonna route you to, ultimately like the first version of Haiku that was ever developed that's running on Dario's Blackberry phone from, you know, two thousand ten It does, it does feel that way though sometimes, like, you know, and, and I can appreciate at least with OpenAI, like GPT-5, 6, Soul, you do get a little bit more flexibility in the requests, you know, in what it's willing to do and answer. There's still a lot of guardrails in place, no doubt, but if I had to pick between the two, I feel like GPT-5/6 at least is willing to take on more tasks, how well it does with them varies on a task-by-task basis, the prompt, the agent, the kind of information you feed it, but at the end of the day, I, I feel like that balance makes more sense, and you see that also in the open-weight models, like they're not gonna be willing to do anything if you're not obliterating them. And by the way, I don't really encourage doing that, the consequences aren't necessarily universally good, removing parameters or nuking certain guardrails. I The best quantization possible for my machine, and I, I kind of try to put in place any of the kinds of abilities that I wanna enhance in things like skill files or prompts or references that I feed it through retrieval augmented generation, I just try to make the best of what I have. Like, I'm running a lot of my stuff locally, I, I make an effort to do it. I've got a bit of a GPU addiction, as you folks might know, I'm up to four RTX six thousand pros, so, you know, I, I try to get some use out of that I will say, there's some incredible stuff happening with models right now, and I think this is kind of a good time to talk about a few things. One of which is that the pace of model releases over the last several weeks, and especially the last week, it just continues to pick up. It's actually pretty incredible and very heartening to see, and it's not just from China. We've got models now coming out from US labs, from Korea. Maybe we'll get that giant cat model from Mistral one day? No, probably not. It's a meme, it's a joke, folks, it's not happening. Maybe. But it also really encourages me from what I'm seeing from our friend and his company, Nvidia, and the NemoTron releases that are coming out. And Chris, I think this is a good time to speak to some of the stuff you guys are up to, 'cause I always tune in to the Nvidia AI live streams. I love the stuff you put out there. I think there's a ton of great knowledge to Just have you up here with us, talking a little bit more about what you're excited about right now.

Chris: Yeah, I mean, so, I, I, I, I mean, I got the little, my Twitter picture thing, I've got the little green guy on my shoulder 'cause I'm, number one animatronic fan. I, I- There are so many things to be excited for right now for, the Neutron effort, as well as of course more broadly the open model effort. At, at, you know, at the end of the day, obviously we have Neutron four is gonna be our next big thing, that's gonna be built in, you know, In tandem with the Neutron Coalition, which is kind of trying to spark this idea that like we need to build these systems together, right? It, as much as we have incredible researchers and we do you know, having everyone pitch in is almost certainly going to make a much better model. It's definitely gonna make a better model, but it's, it's likely to be, you know, si-significantly better. it- Something I want to like talk about for a second, just, just 'cause, you know, I'm, I'm, I'm up here talking about Neutron, is like, Neutron is also not just, language models, obviously that's like our highlight, those like, you know, Neutron 3, Nano, Super, and Ultra, like the, the LLMs that you know and love, and they, they do stuff like build You know, building, whatever, I don't know, whatever you're using AI for these days, they, they can do it to, to some reasonable degree. but we also have like speech models. I don't, I don't think anyone's doing a ton of RAG if you're like, You know, just a, just a, a hacky dev these days, but, a lot of companies are still indexing trillions of tokens and documents today. So there's rag models, there's vision models There's, a-a-a-every kind of model. There's bio, Nemo models, there's physics models, there's quantum models, there's-- I mean, the, the idea is, I think like NemoTron On, on the surface of it is presented as like this open model initiative, and I think that, that, that's true and important, but, I, I think more broadly it's like open science, right? And I think this is like, you're starting to see this language change a little bit from fr-from people in the ecosystem that are, it's m- it's less like, oh, open AI is important, and more like open science is important. I think that's the correct framing, and that's a correct framing for something like Nimitron as well as just making sure that, like, you know The contribution isn't just "model go fast," or "you know, model does good on term level," but sure, like you can use it with open code or whatever and like have a good time. I think the idea is that you want to have an ecosystem of, of open technologies that support science more broadly. because, you know, I'm sure I'm in one of the most biased rooms in the world for, to say something like this, but like, I think we, I think we're all loosely agreed that, like, AI is an accelerator for science, and so having, you know- having a platform exist that is meant to help accelerate that, both through obviously the technology that Team Green builds that makes, you know, paths go fast. Of course, that's so important, really turning that, that fast math into, into something that scientists can interact with, I think is really important. and I also wanna shout out to the work that R-R-RC just, just dropped, I think maybe yesterday or the day before. time, time works strange these days, but, for their, their new open science initiative, like these kinds of things I think are, are kind of what I'm most excited about as it relates to both Neutron and, you know, what, what openness is bringing us beyond just like I like the model running in, you know, in PyDev or whatever, which is true. Hey, hell yeah, fuck it, yeah, I mean, it's gotta work, but, the thing more exciting is Seeing this technology finally diffuse into the rest of science, and for the rest of science to use the methods it always has, which are very open and collaborative, to, to make sure that technology stays open and collaborative. A very long-winded answer for you there, but that, that, that, that is what I'm most excited about right now in the space.

Speaker 0: Yeah, no, I absolutely love it, and I, I think you guys are doing amazing work. And the, the other part of it that you all are doing really well that I think is greatly appreciated by the community are the live streams with education, live interaction, Q&A. I mean, that stuff helps people accelerate their own journey and really learn more about what's going on. And, you know, like, I remember tuning in weeks ago On some of these things to be able to analyze a skill, generate a model card, I thought that was really cool because I hadn't seen stuff like that out there. I recently, one of your releases was the, NemoTron three embed models, and, this is so funny and serendipitous because I've been working on improving my own deep research stack. And this plugs in perfectly to the front end for digesting that data and then feeding it into the model in these vectorized shards through Qdrant when it's querying different parts of, you know, what it's ultimately going to prepare as a report. And I've been looking for some better embed models, so it's like, you know, lo and behold, I get an email from you guys, there's a release of this model coming out, and I'm like, "Oh, that's pretty funny, 'cause I was just looking for something exactly like this." So talk about hitting on World models, language models, you know, some of the other items you've talked about this embed model, this all comes together and helps to create an ecosystem that does accelerate open development. And I think from that perspective, that in of itself illustrates some of the value of this open development, because if this stuff was all behind APIs, we can't really pick it apart, we can't learn from it, we can't build on top of it. And it also feels like a lot of the NemoTron stuff you guys build is meant to be kind of tweaked, fine-tuned, and built On top of as well. Sorry, Chris, go ahead. I, I heard

Chris: you say what I was gonna say, but I wanna say it very, very loudly, which is like, unbelievably true, that last point that you made, right? Which is like Our, again, our teams are great. I, I, I mean, I love, I love our, our researchers. I, I, I'm so biased, of course, but I think they're incredible. I think they're, they're some of the, the smartest folks that we've got around, especially when it comes to, you know, Things that, that need to be fast and efficient, right? So, but like, one of the things that's true is like, there's no way that, that they are better than literally every developer in the world, right? And so when we have these open releases and people do cool stuff with our technology, or they adapt it in a way, or they modify it, or they improve it, like These things can only happen if we're all building it together, right? And we wanna just be a fountain of, like, here's the, here's the next thing. Like, we built another thing, we built another thing, like, have it, and then take it and then make it less shit or like make it better or, you know, Tell us why it's wrong, and this extends to the way that we're trying to interact with the, the people who use these tools, which is like we want to be open been available to, to everybody to have these discussions, especially have them in the open, and so, it, it, it is, it is meant to be collaborative with e-everyone That, that, that would possibly tune into something like this, right? no idea is bad right now. I mean, we, we are, we're in a world of loops or graphs or whatever the fuck thing we're on right now. Nothing is a bad idea right now. Like, just, just, you know, take whatever your, whatever your brilliant ideas and take it, take to the finish line and then share it with everybody so we- Make it better. I mean, that's, that is what will get us to where we need to go the fastest in, in, in this, this one guy's humble opinion.

Speaker 5: Chris, on that How do you guys choose what, where the effort goes and what you're gonna train next? It does seem like you're, you're putting out models all over the place, a lot of great stuff.

Chris: The reality is that, I mean, everything needs this kind of-- So, so if you listen to Jensen like, eight months ago or something like that, fir-- first of all, the, the company strategy is super public. you, you listen to GTC keynotes and the guy just tells you Whatever the fuck we're working on as a priority, but, but secondly, I, I mean, every area of science needs AI, and so our goal right now is to make sure that we're helping all of those areas of science, and there's lots of areas of science and the, the way to prioritize is to do, do what we've been doing, which is hire people, work with labs, help Teams across companies, you know, with things like the coalition. but but you're right, the, the, the thing that I would say is like the priority behind the, the, the, the free fee marketing stuff is Like accelerating science as hard as we can. that, that is like the core guiding philosophy of, of, you know, what, what has been said across the last- You know, six keynotes or whatever, that, that, that's the goal, right? I mean, the, the original goal was to accelerate math super hard so that we could do cool graphic stuff, and now the goal is to accelerate science super hard so we can solve all the problems in the world Not, not us, to be clear, but, but the, but the people who, who are gonna use these tools to do that. you know, Nvidia will help, I hope, but, we need, we need the incredible, brilliant innovators, of the rest of the world to use the technology. I

Speaker 0: love that vision.

Chris: Make science go fast, baby, let's go!

Speaker 0: Yeah, exactly. We need it, we need it right now. I mean, this is the time to accelerate science, it's the time to accelerate human progress and innovation, and so that vision very well reconciles with that. And so I, I just love hearing stuff like that. It makes me extra passionate. I mean, I'm a nerd at heart. I love figuring out new stuff. I love playing with the tech that's out there right now. In fact, I will say That where we are with AI in its, I would say, early innings, very early innings, even though the term has been around and the research has been around for decades, we're finally seeing the momentum hit. And I would say that this reminds me a lot, not to date myself too much, but the early days of Linux, when it felt like anything was possible, when technology was so sort of refreshing and everything was moving so quickly. This feels like that again, but at a much larger scale. With much more significant implications and inflection points happening one after the other. I mean, what happened over the last week with this sort of battle between open and closed, the, the model escape, the number of new model releases and other types of technological progress that's happening, it's almost too much to keep up with. I mean, I think maybe it is too much to keep up with, but not in a bad way. Will, you got something you wanna add in?

Will: Yeah, I find this, topic so interesting too. Oh, let me close my hand. the science topic, I feel like China really understands this well because they're, they, they and Taiwan are kind of in an existential crisis for their populations. And so, you're leveraging this technology to be able to support, an aging population through AI and through robotics, and also to Chris's point, through medicine to ensure that, you know, human progress can continue and that, you know, the world as we know it doesn't, doesn't collapse, because this is a real issue that in the next, you know, twenty to thirty years, pretty much every country is gonna be facing, and so now is the time to push research forward. It's incredibly important. We can't, you know, get distracted from the mission because it's coming to the US, it's coming to China, it's coming to Japan, it's coming to Korea, it's coming everywhere. And so we have, this is extremely important for humanity, in my opinion, and, yeah, I'm, I'm, I'm very excited. I wanna see, you know, solving cancer and solving, ALS and all of these, horrible diseases. You know, I want my kids to live forever. And, yeah, it's And, you know, to your point, it's like a firehose right now. If you're feeling behind, that's normal because it's impossible to keep up. The information's just flooding in.

Speaker 0: It really is. It's, it-- and it's extraordinarily exciting, it's invigorating, like waking up every day and reading new headlines that, you know, hey, this innovation has happened or this progress has happened or this new invention has been made. I mean, it's, it's, it's, it's overwhelming, but in all the best, most exciting ways possible, and I kinda love that about the current timeline. You know, going into this maybe five, four or five years ago before any of this arc started, I was starting to wonder, where is- Tech really going to go from here? Like, what is the next big leap forward? And then, you know, we started to see the progress in transformer models, but admittedly in their early stages, they were just sort of toys, they were novelties, chatbots, things like that, answering as almost like an autocomplete. We have come so far. From that point to where we are now, where not only are we interacting with significantly more sophisticated models, we've got this whole layer on top of it with the agents, the skills, the plugins, the kinds of workflows. You have various types of agents, you even have, you know, agents that govern other agents, like an orchestrator over viewing sub-agents that are building workflows, a, a validator agent that's going over all of your work to make sure that whatever was built is at least close to the spec or close to the MVP. We're looking for. I mean, I didn't think a, a year ago I would trust letting agents run overnight to build scaffolding to things that I would wake up to and be able to play with, but literally that's the timeline that we're living in right now, and it's absolutely incredible. I love it. I mean, it, it really-- you could probably hear it in my voice, but it, it really does spark a lot of joy to see this much, and it also enhances individual human potential. I mean, what one person could do on their own If they didn't have certain sets of skills that AI and agents unlock, has been so significantly amplified, and in many cases, in very good ways, right? Obviously, there's malicious potential, there's the cybersecurity side of things, people using it for phishing attacks and this and that and the other. But if you really drill down to the acceleration of technological progress of science and really just of human potential, this has unlocked more of that, of that capability to be productive To, to sort of, I would say force multiply than any other technology that I've seen.

Will: Opus five just dropped, by the way. It is- oh. Absolutely crazy on the benchmarks, a huge jump over- Fable somehow. I mean, I'm testing it now, I don't know, you know, they're benchmarks, right? You gotta take them with a grain of salt, but it's a little bit crazy. The, the numbers are nuts.

Adam: Is it safe for that to be out? Are we allowed to use that? I don't know. I'm not sure. I don't know. We're gonna have to

Will: see.

Speaker 4: Actually,

Will: yeah, you'll get downgraded just for asking.

Speaker 4: Yeah, they actually said, they were afraid, you know, the, the whole, stick that they always do. They were afraid to actually release this, and, now that they released it, some people noticed that they can actually, you know, route you to, the older Opus, you know, transparently without actually telling you like we've, Fabio Fav.

Speaker 0: Oh, good!

Will: That

Speaker 0: is- Another reason to be skeptical of closed APIs, just wanna throw it out there, please take the, take the mic well.

Will: That is horrible. That's, like, if you- especially if you're paying for it too, that feels almost fraudulent. I'm not, you know, that's maybe a little bit of a harsh word, but- That's, totally unacceptable. I will say though, Nvidia employee posted something on Reddit, maybe by accident, showing them getting downgraded from Opus-5 to Opus-3. No, that was,

Speaker 4: that was actually an Autopick employee, I think. A screenshot from, an Autopick employee, and he got downgraded himself, you know? So I bet, you know, the, the so safety people, even their own employee, can't actually use the model without the classifier, you know? I bet only, only K You know, unrestricted.

Will: Got it. Yeah, yeah. So, but it did say it on the screen, it showed that he was being downgraded. But I'd have heard this before, especially if you're doing, any kind of RL research or anything like that, they say that, it could downgrade you without telling you, which I think is just completely unacceptable, 'cause I, I would just wanna start a new thread at that point, you know? If you're gonna downgrade me, tell me about it. That's the only, you know, I don't even wanna be downgraded in the first place, but you understand my point, right? That's, that's not okay.

Speaker 4: Yeah, you can still get downgraded even if you don't do anything related to like, mathematics or, you know, cybersecurity or ML or anything like that. Even if you don't do that, just because of, you know, like what I said earlier, spurious correlation, because they don't understand that a classifier like this will never scale. it will have the same problems as, As an IA detection, I see it as a style detection, and as a style detection, it is great. As an IA detection, it is really bad, because it has the same, you know, false positives and, you know, false negatives as any classifier has.

Will: Yeah, it's clearly too strict. I mean, people are getting downgraded just for asking math questions. You know, it's kind of sad because their, one of their employees, you know, posted how they had solved one of these, you know, two hundred year conjectures or whatever, and the next day people were, getting downgraded for asking math questions, and it's like, "Oh, cool, so, you know, I'm glad you get to use it like that." But yeah, I'm testing it now, so hopefully it's pretty good. I haven't gotten downgraded yet, and I'm doing Rocket League, and, it says RL, so I got, maybe I got lucky.

Speaker 4: Yeah, it's really funny how sometimes it actually lets you do stuff. I don't think they actually run the classifier on every prompt or every turn, I think they run it maybe on every two turns or something like that. it's a, it's a bit like Codex, for example. Like Codex, won't immediately run the classifier on every request. I notice that it does, you know, for example, say on, ChatGPT, you know, like the web UI, the website, it does run it almost on every request, but for, Codex, I don't think it does all the time. And this is why, you know, sometimes you can actually get through.

Will: Well, I think Claude works differently because I've had the classifier kick in halfway through it doing work. So it didn't kick in right away. It, it did research, it started to write code, or at least create the scaffolding for it, and then it kicked me down after it reached a certain point. So, it's happening at least, you know, dynamically in between tool calls, it can trigger, you know, during the research phase as well. So I've had that happen to me, I forget the prompt. See if I can find it real quick. But yeah, it was, and I kept trying it and it just kept doing it over and over again.

Speaker 5: I had it happen earlier this week actually. It was just trying to use the own, the actual like, cloud local host to display artifacts and it was messing up and then it downgraded. I think it was trying to like look at my ports and stuff like that.

Will: Yeah, that's a little bit, that's a little bit crazy. Ma'am, step away for a second, so anybody, if you have something you want- Oh, okay, cool. I'll let you know. But we are,

Speaker 0: I'm happy to open the, the, the floor to anyone who has any questions for the panel here. We gotta wrap up pretty soon, but we're happy to take at least, one or two questions if folks out there have any. And we'll appreciate you all holding down the fort for me while I stepped away. No, this is great, thank you. Really enjoying this. Believe it or not, Amazon came through with some more GPUs, so- Oh, nice! They finally met my addiction. The 3060s or- Yeah, the 3060, 12 gigs. Nice. And that'll be for, just building a machine for folks out there. I, I have a little, little lab setup here, so it's the workstation, but I've also got several other machines with GPUs that are for different use cases. One, as I was talking about earlier, the deep research pipeline runs embedding models for that, for digesting that data. another one of those models is for text to speech, another one is for image generation. It all kind of ties into workflows, and then this one is-- and then I have a validator model with Vision and things like that, and then this one is gonna be for just routing simple requests too. So it, it all kind of comes together, it's, it's sort of like having my own local AI cloud. The benefit is control over the quality, no one's, you know, rug pulling me with quants or other things like that that degrade the experience. I've got the best throughput I can get 'cause it's all on my local network. There's no kinds of sort of artificial gatekeeping, all my data remains that of my own, so I have sovereignty. Privacy and, it allows me to learn a lot more. I mean, one of the things, I'm, I'm kind of like a tinkerer. I like to have my hands on the whole stack of something when I'm learning it. So getting my hands on the hardware, the, learning more about the inference engines like Seglang and vLLM, learning more about the individual models, the agents, the skills and plugins that sit on top, this has been an amazing learning experience.

Will: How, how will that work? You have these different devices, I'm Own motherboard and cluster of GPUs that are the same architecture, like your six thousands, your thirty sixties, and then you're gonna have one of these devices as the primary orchestrator deciding which way to route a call, and then it will go over your network to the appropriate device and return an output.

Speaker 0: Yeah, exactly. And so, not as much a cluster is just a network with different APIs, you know, some running llama.cpp because it's more consumer-grade hardware like these 3060s, my workstation running vLLM because right now that's Just been the, the easiest choice to drop in and get the best performance possible on models like GLM and DeepSeek, but exactly what you're saying to the point that, you know, basically put in a request, an orchestrator decides what the most appropriate endpoint is to route that request, and then it returns that. And then if it's a more complicated multi-turn workflow, during that process, it will also do a validation routine at the end to make sure, hey, this actually came through in a way that's somewhat coherent or it didn't And we've gotta redo X, Y, and Z.

Speaker 5: And what, what model is your orchestrator, or is it like rules based?

Speaker 0: Right now, going into this, it was DeepSeek, V4 Flash, but now that I'm running GLM5-2 and it just, that kind of just become more of an official thing this week, I'm more likely to look at that as my orchestrator for now. But I know it's heavy, so what I'm gonna be doing is watching the traces, learning from how it routes, and then trying to use that for a more efficient model, like maybe Quen 3 6 27 B or 35 B A 3 B. And if anyone on the panel has any ideas about how to optimize this, I am still a student, I'm a forever student of technology, so I'm happy to learn from other folks that have gone through parallel processes and that have feedback, ideas, suggestions, et cetera.

Will: I think, I think you can actually get away with the, the, I don't know how small you wanna go, but, this is pretty old now, last year I built an astronomy marketplace website, and I'm very proud of this. Because this is before the models were good, this is GPT five point one and five point two era, and I built a multi-agent setup in the, astronomy, you know, it was an AI search tool that basically, had multiple functions. One of them would look for available, equipment to see what's in stock, another one would be able to look up information about a certain piece of equipment, another one was just general like astronomy software question, and then another one was a support bot. And I used Gemini three Flash as the, decision maker, and it worked almost 100% reliability. So if Gemini three, you know, it might even not even-- no, it was Gemini two point Five, Flash, and it was still very good at this. So I suspect you can definitely get away with the speed, maybe even the twenty-seven B model, and then just route for the extra intelligence, when you need it. That way you can capitalize on the speed. Even some voice models are good enough at this. So GPT Live is a good example of this, and Quinn just released a new real-time model. It's not open source yet, but I think, based on my experience, you can definitely get away with the small fast Model, and I would encourage you to explore that path first.

Speaker 0: I

Will: appreciate that. It also really

Speaker 5: depends what you need, what kind of context you need to make that decision. Sometimes you need a lot more information, and you need some steps of actually like figuring out what information you need. So, but it can get tricky. Which is where I use kind of the big model still.

Speaker 0: Off, Mike, I'll have to pick your, your brains, guys, to get some ideas here. I, I did notice, and since, like, one of the, topics, I almost wish we, we could sort of magically pull Chris back in for just one last interaction, 'cause I noticed NemoTron Orchestrator 8B. Have you guys tried this model out? Because it seems like it's specifically built for this use case.

Will: No, but that's pretty cool. Definitely gonna check that out.

Speaker 5: Yeah, definitely need to check that out.

Speaker 0: I think I, I might have to be, checking that one out this weekend. I was just routing things through the more powerful model to kinda figure out how it did its routing and, and get a better sense as to maybe if I could harvest some of that info to improve the routing moving forward, but going to a, a specific use case built model. And it's funny 'cause I had meant to check this one out ages ago, but I just didn't have a machine to dedicate So kind of changes the parameters. I also wanna thank Adam for joining us. He's gotta take off, he's got some other meetings. I mean, some of us work real life jobs like Adam, so really appreciate you joining us, my friend. always a fun time, and of course, we will all be back together next Friday, I think we'll go back to the two PM Eastern time. But also, you know, the-- this is a topic That I feel like we're only gonna be covering more and more ground as time goes on. So folks that are out there, that are listening to this live or listening to the recording later, and, you know, if you're listening live now and you didn't hear the whole thing, you can always tune into the recording. Feel free to send us your questions, you know, feel free to give us feedback, give us ideas of topics that you wanna see us cover. We're amiable to that. You can reach out if you have a verified account, my DMs are open on X, you can marketsandmayhem at p m dot m e, that's the shortened protonmail domain, always happy to help this community. I mean, one of the things that I love about the local AI community is how much we're here for each other

Will: Absolutely, and don't forget to give Adam and everybody else a follow if you're not following, see Cam and Phillip, Mark it. And, yeah, if you're not following me, drop me a, drop me a comment and I'll, I'll follow you back as well.

Speaker 0: Yeah, definitely follow this panel. Everyone here is amazing. Chris also, who isn't still here right now, but follow him too. His handle is l l m underscore wizard, and he is actually a wizard. I know there's some skepticism around this, but he is actually a wizard.

Will: Nah, I'm gonna follow you back right now, I'm looking for you. There we go.

Speaker 0: And I wanna close out by thanking everyone for joining us. We're gonna keep these going. Tune in next week, Friday at 2 PM. This is always a lot of fun. Appreciate all my co-panelists, everyone that joined us to listen. Hope you all have an awesome weekend ahead. We'll catch you next week. Appreciate you guys, thanks for joining.

Speaker 5: Thanks guys. Thanks everyone. Thanks