Overfit is now ModelTalk! GPU Smuggling, OpenAI Cooked? + Open Models, AI Writing

Summary

ChinaTalk—ModelTalk Debut: GPU Smuggling, OpenAI Business Woes, Open Models, and the Limits of AI Writing

Podcast: ChinaTalk
Host: Jordan Schneider
Episode: Overfit is now ModelTalk! GPU Smuggling, OpenAI Cooked? + Open Models, AI Writing
Date: March 23, 2026

---

Episode Overview

In this episode, Jordan Schneider and guests (Nathan and Jasmine, inferred from context) kick off the rebranded “ModelTalk” with a wide-ranging, witty, and deeply informed discussion on some of AI’s hottest topics:

• The Super Micro GPU smuggling case and the challenges of enforcing export controls
• OpenAI’s business struggles amidst the ChatGPT scale problem and changing AI model landscape
• The slow crisis in open-source models in China and beyond
• Why even powerful language models still can’t write truly compelling prose, and what it tells us about the limits of AI

---

Key Discussion Points and Insights

1. ModelTalk: New Branding, Same Banter

• The trio jokes about the petty drama and rapid developments in the AI space, especially around model launches and licensing squabbles.
• Jordan considers (and rejects) various new podcast names, eventually settling on “ModelTalk.”
  • “Model Talk I can kind of get with. AI Talk sounds really stupid and I would never be part of a podcast called that.” —A, (01:16)

---

2. Super Micro GPU Smuggling Scandal (02:36 – 08:27)

Super Micro’s Export-Control Crackdown

• Jordan breaks down the difference between prior smaller chip busts (“really Richardy, rackety out points”) and the major Super Micro case.
• Notes that Super Micro is a “real fucking company” ($15B market cap before the crackdown), not some fly-by-night operation.
• He explains how suspicious Super Micro exports from places like Singapore led to a large portion of GPUs being smuggled into China.
  • “Super Micro is a real fucking company. I mean...before today it was like a $15 billion market cap...Turns out halfway did like a pretty good chunk of that ended up being sort of smuggled into China...it's really disappointing to me.” —C, (04:03–04:26)

Legal Consequences and Industry Impact

• Jordan and Nathan discuss whether this is now felony-level criminal behavior, how rare executive jail time is for financial and export crimes, and whether this case will “get some scalps.”
  • “I just think that's fucked up. Like, if you break the law this willingly...with this level of potential harm to sort of national interests and national security, you shouldn't be able to have your firm pay your way out of it.” —C, (05:30)
• The scale is enormous: $2.5–$3 billion in chips moved illicitly into China.
• The hosts have fun with the idea of hiding these huge processors:
  • “They weigh two tons. They're like the size of seven refrigerators glued together...So when the inspectors come by, they think it's Iraq, but it turns out to be like Patton's inflatable army.” —C, (07:27)
• GPU supply chain trickiness: Many are baffled why labs risk smuggling for a high markup; theorized it's for government customers who won’t offshore data.

---

3. OpenAI’s Rocky Path: The “Albatross” of Scale (08:27 – 16:05)

OpenAI’s User-Scale Problem and Monetization Trap

• Despite launching the improved GPT-5.4 model, the hosts see OpenAI as struggling under the weight of serving free users with millions of GPUs:
  • “If we actually could see their books. I feel like the biggest problem OpenAI has is ChatGPT sitting on millions of GPUs to keep it afloat and have the free users use it. And they have no path to monetize these people in the near term. It's just such an L.” —B, (08:44)
• Side-by-side comparisons: GPT-5.4 is “smarter but so cold” compared to the more approachable, intuitive Claude.
  • “Claude will kind of read between the lines of your prompt…GPT just doesn't feel good.” —B, (09:38)
• GPT “think” models are powerful for research, less so for regular chat queries (overly verbose, too much data).
• Product differentiation now less about numbers, more about ability to “think.”

Business Model Dilemmas

• The OpenAI conundrum: mass consumer AI is, business-wise, a headache:
  • “It really sucks to be a mass consumer company. It's not fun...most people are never going to make your money back. And…the only way you kind of make the money back if you’re a social media giant is through ads. But OpenAI doesn't.” —A, (11:22)
• The risk of inserting ads or commerce upsells: user loyalty is weak, migration to alternatives (Claude, Gemini) could happen quickly.
• For intensive research, OpenAI’s extended “thinking” still trumps alternatives, even if user sentiment favors “warmer” options.
  • “ChatGPT is less lazy than Claude is...for research tasks...you actually do just like accomplish more.” —A, (12:27)
• The conversation turns humorous regarding naming conventions, with some annoyance about the abstraction away from model numbers.

---

4. AI in Practice: User Experiences, “Toy” Projects, and Local AI Agents (16:25 – 21:16)

• Jasmine shares a use case: using Claude to automatically generate Anki flashcards from Chinese language lesson MP3s — “It did. It was amazing.” —C, (16:51)
• Jordan has been busy coding “toy” games (“GPU smuggler,” “Karg Island Invasion,” “Straight of Hormuz”), reflecting the DIY/experimental AI ethos.
• Discussion of “OpenClaw” (likely OpenClaude or similar local open-source LLM):
  • Security concerns mean adoption among security-conscious power users is slow.
  • In China, even “grannies” are picking up new local LLM tools, with the country’s history of rapid IT adoption.
• Monetization culture: Chinese users more willing to pay small fees for digital services, in vivid contrast to enterprise purchasing reluctance.

---

5. The State (and Fate) of Open Models in China (21:16 – 26:17)

Structural Instability and Resource Constraints

• The “open model” approach in China is under severe strain:
  • “Structurally unstable is what I would say...Quinn was constantly battling to get more resources...I just don’t have enough GPUs to do what I want.” —B, (21:16)
• Open models like Quinn and DeepSeek face constant pressure from resource scarcity; their future seems increasingly bleak against US giants.

The Western Advantage

• US labs (Anthropic, OpenAI, etc.) have budgets dwarfing the entire Chinese AI training outlay.
  • “I would imagine that Anthropic’s research budget alone is more than what all of China is spending on training models.” —C, (23:46)

Government Intervention & Neolabs

• There’s skepticism that the Chinese government will step in to save open models; state pride in “open” is secondary to world-class capacity.
  • “What is important to the Chinese government is to...not have this be a technological step change that they miss...whether the models are open or closed seems...irrelevant to that broader [goal].” —C, (25:19)
• On “neolabs”—small, clout-driven AI labs—guests concur they’re more like researcher sabbaticals and status moves than enduring new competitors.
  • “I think neolabs are like sabbaticals for industry researchers.” —B, (26:21)
  • “Totally true.” —A, (26:31)

---

6. Can AI Write? On the Limits of LLM Creativity (28:33 – 39:22)

Why AI Writing Still Feels “Off”

• Jasmine references her piece in The Atlantic and recounts AI luminaries’ surprising skepticism about AI ever writing great literature:
  • “It was specifically the Sam Altman and Tyler Cowan podcast...Tyler's like, do you think GPT 6 or 7 can write a Pablo Neruda poem? And Sam's like, no...That’s really weird.” —A, (29:21)
• The group discusses why prose, especially “voice,” is so elusive for LLMs:
  • Lack of training incentive (“labs aren’t spending a lot of resources on writing”)
  • Verification is harder—it’s much fuzzier to judge literary value than technical accuracy
  • Authentic writing voice is a product of lived experience—“Voice comes from the particularities of somebody's life experience...AI can't do that.” —A, (34:49–37:50)
• Nathan argues more investment and training could eventually help, but current models feel flat or “devoid of voice.”

Notable Quotes and Example:

• On bad AI writing: “A lot of AI researchers feel embarrassed by how annoying ChatGPT prose is and the fact that they're like creating all of this like you know, uniform annoying slop and everybody's writing with that.” —A, (32:08)
• On AI voice and authenticity: “‘Voice is not just being like random or weird...You can tell that person's trying to larp somebody else...Voice comes from various experiences and communities I'm a part of...my voice comes from the particularities of somebody's life experience. Not to mention all the literal things...my reporting is not automatable because I have to go talk to people and look at stuff. And AI can't do that.” —A, (34:49–37:50)

Experiment with Vintage LLMs

• Jordan tests Nick Levine and Alec Radford’s “Vintagelm,” trained only on pre-1930s data, and finds it can’t imagine creative writing by machines, only mechanized formatting.

---

Notable Quotes & Memorable Moments

• “They weigh two tons. They're like the size of seven refrigerators glued together. And then I guess the way you do that is you make dummy ones, like Patton’s inflatable army.” —C (07:27)
• “It's just such an L. And I think that that could be the thing that actually restricts [OpenAI]...because those users aren't going to churn, but they can put them on cheaper and cheaper models over time as intelligence per watt goes down. But that's not a great position.” —B (08:44)
• “Claude is like an extremely lazy researcher. It is not as good at searching, it searches less, it goes faster, but it's far less comprehensive.” —A (12:27)
• “I think neolabs are like sabbaticals for industry researchers.” —B (26:21)
• “Voice is not just being, like, random or weird...it’s very clear what my life influences are from my writing. And that is what makes a voice feel authentic.” —A (34:49)

---

Key Timestamps for Important Segments

• 02:36 – 07:27: Super Micro GPU smuggling, export controls, law enforcement
• 08:27 – 16:05: OpenAI’s scale and monetization difficulties; comparing model vibes (GPT-5.4, Claude)
• 16:25 – 21:16: “Toy” projects and practical AI usage; open-source/local AI model momentum
• 21:16 – 26:17: The future of open models (China, US); resource allocation and government intervention
• 28:33 – 39:22: Deep dive into why AI still can’t write well; the challenge of literary “voice” and authenticity

---

Conclusion

“ModelTalk” delivers a fast-moving tour of AI’s current defining dramas, balancing technical insight and irreverent humor. The hosts dig into the real meat of model development, global tech business, and the authentic limits of current AI. As both practitioners and observers, they’re uniquely able to blend hard news, business strategy, and nuance around the humanistic frontiers that still separate AI from truly “writing good.”

Transcript

[A]

What has the new AI news been? I don't even know.

[B]

5.4 cursor drama. The cursor drama is so petty. It's a single Cursor released a new model. It was obviously built on Kimmy. Kimmy thought they violated the license. Cursor threw fireworks under the bus. Then Cursor came out and said, we're all friends or no? Then Kimmy came out and said, we're all friends. All of this in like 12 hours.

[A]

Wow. Silly.

[B]

Wait, should we talk about the smuggling of chips? You need to do a whole pod on that. Or we could fill this in. We could do it now if you haven't pod about it. We could talk about gvt 5.4 and we could talk about blow drag for.

[C]

Okay, let's just start with Super Micro, because I don't know, it's hilarious. Okay, so. Okay, so Overfit. Should we be Super Overfit?

[A]

No. Is everybody. Wait, are you going to rename this AI Talk?

[C]

No, I can't do that. I wouldn't do that. My dad texted me. He's like, jordan, make the Empire AI talk. It's too much.

[A]

Okay, good.

[C]

Because, like, War Talk is like.

[A]

We could be Model Talk.

[C]

We could be Model Talk. You guys want to be Model Talk?

[B]

Model Talk brought to you because it sounds like Model.

[A]

Exactly. And also because then it would feel like more part of the family because AI Talk sounds really stupid and I would never be part of a podcast called that. But Model Talk I can kind of get with. Right?

[B]

Yeah.

[C]

Let's roll with it. All right. Super micro GPT 5.4. Writing on the Internet. The death of Open Models. So much to discuss. Let's start with smuggling the GPUs.

[B]

I.

[C]

You know, I feel dumb because this was coming. I should have put the short on, but good for the. I mean, I guess the Trump administration kind of could have killed it. Good for the Southern District of New York. We're going to. We're going to. We're going to clap it up for them, you know, prosecutors doing their job. Nathan, what's your. What's your take?

[B]

Should I get a hairdryer? Is that how you make a lot of money? I mean, I just mostly follow it. I just baffled. The world is continually so weird and in both the physical world and the digital world, and because of the money of AI is why we will keep getting these things and across every possible axis. I don't like. It's not really my domain. I don't really know about chips and

[C]

network, so here's my, here's my question, Nathan. Like if you can legally have these chips in a data center in Malaysia doing whatever you want them to do, why is it so important to like go through the headache and whatever markup it would be to get them like smuggled into mainland China, what do you, what do you gain as a company or a research lab?

[B]

The research labs don't gain anything. I would guess that it's like closer government related companies and, or companies that don't want to send their data abroad that want to do this or cheaper prices. But you're not getting a cheap price if you're smuggling this chip.

[C]

Yeah, I mean maybe you save it on like power on the backend a

[B]

little bit, but, but power is only like 10 to 20% of the cost of the data center. Like the chips are the most important part. So if you're getting a 20% markup on the chip, like you're really ballooning the cost. I don't get it. I haven't known, I'm not the person to ask. Like I'm not going to strongly say, oh, the Chinese government is using them for military purposes. Which is like something you ideate. Like you could come up with this idea. But I don't. You're never going to be able to prove this.

[A]

Jordan. I think you should explain, use this opportunity to explain to me an export control dummy, what's going on here? What is Super Micro? Like is this like the first like sort of big arrest made for export control, like chip smuggling, which everyone knows has been happening like give the 101.

[C]

So you've had a handful of small arrests of like basically like dudes selling 10 or $20 million. It's like really Richardy, you know, rackety out points. Super Micro is a real fucking company. I mean, what.

[B]

Before I knew who they were, I did.

[C]

But you know, it was before today it was like a $15 billion market cap. It's been around for a while. And there was this, you know, there were, there were a handful of articles that were being investigated. It was clear that, you know, everyone was talking for a while. All of a sudden Nvidia selling an insane amount of chips to Singapore. Like, what's up with that? Turns out a lot of it was Super Micro. Turns out halfway did like a pretty good chunk of that ended up being sort of smuggled into China. You know, I will be upset if these people don't end up in jail. I mean this is very illegal. They knew just how illegal it was. It's it's really disappointing to me.

[B]

What is the prosecution pathway, like, when it's very illegal? Is this like, like you're doing things that are like, there's, I think of like, there's like petty crime, there's like felonies, and then there's like you're fighting the country, treason, war, stuff. Like, like, is it started to get there? Like, how do you read this?

[C]

Well, here's the thing is like, you' a lot of sanctions violations cases being brought, particularly against financial institutions over the past 15 years or so with like getting money to Iran and whatnot. And you never ended up to the point where you had executives going to jail, but you did. Or the, you know, the, the, the Treasury Department, in combination with the Department of Justice was able to build cases strong enough that they were able to negotiate, you know, multibillion dollar settlements from these companies. And the kind of the, the, the, the, the trade off that these firms made is that, okay, they would like, do the compliance controls, they'd pay the multibillion dollar settlement, and then none of their executives would go to jail. And I just think that's fucked up. Like, if you break the law this willingly in this a, in this kind of dramatic way at, with this level of like, potential harm to sort of national interests and national security, you shouldn't be able have your firm pay your way out of it. So I really hope we actually get some scalps from this and we'll see. I don't know. This is the first really big one to come around. Chinese semiconductor export controls. Like, there were rumors about other companies, I think Cadence, there was a Reuters or Bloomberg article written about in 2024 that there was investigation around, but

[A]

far

[C]

and away the largest, the largest swing the government has taken to try to really get at a network. And like, I thought it would be like shadier secondhand retailers. Again, super micro, like real fucking company. Like, they knew exactly what they were doing. And, and the scale is also enormous.

[A]

Were they doing it for how much?

[B]

Was. What was the volume?

[C]

$3 billion?

[A]

Yeah, two and a half billion.

[C]

$3 billion of chips. And again, you have all these incredible quotes from Jensen being like, who could smuggle this stuff? They're giant. They weigh two tons. They're like the size of seven refrigerators glued together. And then I guess the way you do that is you make dummy ones. So when the inspectors come by, they think it's Iraq, but it turns out to be like Patton's inflatable army, like, faking out the Germans that they were Going to Beach A when they were actually going to Normandy.

[A]

Oh my God.

[B]

Some more funny context. I think Jensen and this guy that got arrested were taking a photo like two days earlier at GD gtc. And also there was like wink, wink, nod, nod, like Nvidia starting to sell to China around GTC where like people are like, it sounds like both sides approved the H2 hundreds. Are we. It's not B2 hundreds, right? Can't H200, thank God. But like that's supposedly also like within two days of this.

[C]

All right, Mandate of heaven. We, we had a rough six months for OpenAI. Five, four comes out. Look, it's better and they have so much money, but they have to serve all their users.

[B]

They have this huge problem. Like if we actually could see their books. I feel like the biggest problem OpenAI has is ChatGPT sitting on millions of GPUs to keep it afloat and have the free users use it. And they have no path to monetize these people in the near term. It's just such an L. And I think that that could be the thing that actually restricts them to having to figure that out in the long term because those users aren't going to churn, but they can put them on cheaper and cheaper models over time as intelligence per watt goes down. But that's not a great position. I would summarize GPT 5.4 as being the smartest model, but it's just so cold. It's just not satisfying. You like you use Claude in it side by side and you're like, oh, this thing fixed it, but it just doesn't feel good. So I think they have like a big problem with people trying it versus like Claude and all these other things, which is just such a funny position to live in. Like you get used like Claude will kind of read between the lines of your prompt and try to model your intent and what it does. And if you just like typo your Codex prompt, it just fucks it up. It's just like, I'm going to do what you told me. Like I. I've pasted the wrong prompt into a Codex thing. And then it just like tried to rewrite the whole library to accommodate my weirdly like out of context prompt. It wasn't like, bro, you're drunk and just like went off chugging doing this thing, which is on one hand remarkable. And I think there's going to be like these people that are totally Codex die hard armies there. They could use tons of replicas and get a Ton done. But from a product sense, I think that Claude will still be winning adoption. That's why tldr. But it's interesting that they could be so different while also being so useful in a big step change like GPT 5.4 is much more approachable in random shit than the Codex models, which I think the name change should accomplish. Previously people are using GPT codecs and now they're just using GPT number, which is nice to not have as only a software engineering model.

[C]

Ben Thompson telling us six months ago, like, this is OpenAI's path to greatness is that 700 million daily active user number. Now you're saying it's an albatross. Nathan, this is a bit of a flip. I mean I guess it is proven how you can make that you can make an insane amount of money just doing it on enterprise, which was not something which was necessarily baked in, you know, 12 or 18 months ago.

[A]

Yeah, I mean the only way to monetize scale is like commerce and ads, right? In general, like.

[B]

Or AGI?

[A]

Well, yeah, I guess I'm thinking about tech companies broadly, but yes, or AGI, I suppose. But in general it's like if you. Because again, I just like it really sucks to be a mass consumer company. It's like not fun from like running a business standpoint because it is just like so many users, so many different use cases, so many different price tiers, so many different wants and needs and complaints and jurisdictions to deal with and most people are never going to make your money back. And the only way you kind of make the money back if you're a social media giant is through ads. The way that you make your money back if you're an Amazon is through like commerce and a. In a marketplace. But OpenAI doesn't.

[B]

And ads, they make long runs.

[A]

People will. Yeah, even that. They're like, maybe we're going to do it, but these are going to be very complicated plays for them to pull off. I think they're going to get a bunch of backlash for doing that because they would have to do that before either Google or Claude would. And it's like people aren't that actually sticky to their AI models. Like I think people expected that stuff like memory would be more powerful in getting people to be loyal. But like my sense is like if OpenAI starts filling ChatGPT with ads and commerce upsells and whatever, people are going to be very comfortable switching to Claude or switching to Gemini or a place that doesn't have that. And so I don't know, I don't envy the position that they're in. At the same time, like I still use ChatGPT and I like more than I think most people I know who are like extremely cloud pilled. Mostly because for my case I do a lot of research with AI. Like for me it's like largely a research assistant and ChatGPT is less lazy than Claude is. Claude is like an extremely lazy researcher. It is not as good at searching, it searches less, it goes faster, but it's far less comprehensive. And so I find for research tasks both just like the GPD models, the ability to be like extended thinking, even more extended thinking pro, extended thinking max or whatever like you actually do just like accomplish more. Like my GPD will go off for like an hour and a half and like bring me back like a truly comprehensive database that like Claude cannot do. And so I still use it a lot for these sorts of like research and thinking tasks. Even though it's like, yeah, it's not the thinking man's model these days, so it's pretty funny. I also thought it was funny that

[B]

they seem more approachable. Like if you just slot Claude into the average ChatGPT user. Eh, the average user might just like slop in erotica, but a lot of people, I think like cloud is just kind of chill now for general AI queries, at least in the chat domain. But like the GPT think models are such super power tools. It's kind of annoying if you try to ask it like what's the weather? Because it's going to give you like a breakdown of the weather of every single freaking city in this country with over time, with interesting facts. It's like dog. I just wanted a fast answer.

[A]

Yeah, interesting. Clearly like Claude now feels like the more like conversational, natural, quote unquote human thing and like OpenAI. I mean maybe also personality, character wise, maybe as in part as backlash like ChatGPT voice and like the sycophancy stuff, they have really dialed back a lot of that. Yeah, it's full corpo but like it is. I do feel like I have just like a very good research assistant which for me is like a use case I'm perfectly happy with.

[B]

Yeah, the pro models are incredible. Doesn't have one and if OpenAI could integrate that in codecs pretty well because I do a lot of hybrid like interface between code and research work or I have a lot of repositories where it condense information and the coding agent is my driver. It's like that would be A big difference because even like a coding agent with the pro mode is just going to crank. But we've been waiting for a while and it hasn't come.

[A]

I'm also impressed they solved their naming crisis finally, where they don't put the numbers in the consumer product anymore. So you don't have to choose between like 5.4 thinking and like 4. Whatever. 4. Like, it's just like they took all the numbers.

[C]

I'm annoyed by that, but it's okay.

[A]

Really?

[C]

Yeah.

[A]

You want to. How often are you searching models?

[C]

Well, just sometimes it picks the wrong one. It just, sometimes it like autos to the dumb one.

[A]

I mean, the only thing I care about is thinking or not thinking. And like, in the composer you can just like click thinking.

[C]

Yeah, I just never want to not think. And sometimes it like tries to not think.

[A]

Okay.

[C]

It tries to get away with giving me that cheap model energy.

[A]

Yeah. It's like I'm paying attention.

[C]

And it's really obvious too, if I'm like asking questions about Iran or something and it's like, you didn't do your search so you don't think there's a war going on right now.

[B]

Oh, my God. Yeah, Claude does that sometimes. I'm like, dog. Oh, one's hurt. I get really. Okay, does anyone get really frustrated with the models where I feel like the intent of what I'm trying to do is so obvious and they keep saying something else and I'm like, oh, Jesus, what are you. Do you not speak this language?

[C]

I had a really fun new use case. I'm building this Chinese trainer and this has been the bane of my existence is after every Chinese lesson you take, there's a handful of flashcards you want to make to like, you know, do the learnings. I literally threw it into Claude code. Just the MP3 of the. Of the lesson.

[A]

Of the whole lesson.

[B]

Yeah.

[C]

And I was like, make me the flashcards. And it did. It was amazing.

[A]

Sick.

[C]

I know.

[A]

That is really cool. Like, it just made enkis or like. What do you mean?

[C]

Yeah, it made Anki's. I was like, make an Anki for me.

[A]

My God, that's really cool.

[C]

I'll launch it next week. It's not super cheap, but I don't think I can make something.

[B]

Are you gonna post it on the main sub stack like your other toy projects?

[C]

Yeah, sure, why not? Should I monetize?

[B]

I think I should monetize.

[A]

I love your self improvement grind. You're building yourself all these trainers. It's so cool.

[C]

It's Easier to build them than it is to actually use them. I'm also. I got another great game coming off. It's called the Art of the Steel. You can play as Donald Jr. Or Jared or Zach Witkoff. And then you gotta balance. It's like a choose your own adventure game. You've options, then you gotta pick which one you do. There are three things you're maximizing for. First, money, of course. Second, clout. And third, heat. So you want more clout because then you can get the good deals because everyone knows you're connected and cool and whatever. But if you get too much heat, then you get arrested or kicked off the island or whatever. So stay tuned for that one.

[A]

Okay.

[B]

You should.

[A]

Do you have like a personal website or something where you have, like, you know, a gallery of all of your toy apps?

[C]

Have all of them? No. I guess I should.

[A]

I feel like that would be a good thing to do.

[C]

Yeah, you gotta set up a GitHub because you need, like, the actual photos.

[A]

Yeah, it can just be like, you know, ChinaTalk Media or JordanSchneider.com or whatever, slash apps or slash toys. And then you can have little portals to all of them and you can feel like a real developer. Like they all have things like this

[C]

on their sites, you know, I made. I made a GPU smuggler game a few months ago.

[A]

How many apps do you vibe code a week?

[C]

I made a Karg Island Invasion game and opened the Straight of Hormuz game.

[B]

Holy crap.

[C]

There's a problem. You can't actually open the Straight of Hormuz. Spoiler. There's no way to win that one.

[A]

Oh, man, they're fun.

[C]

They're so easy.

[A]

Yeah. So much creativity.

[B]

The thing you need to do to follow up on all your toys is you have an open claw that you program to make sure you stay on top of shit. So you're just essentially like building an AI assistant that is fully designed on making you not procrastinate. So it just sends you various horrible messages or turns off your WI fi until you start doing your work. I'm sure you can make OpenClaw turn off your Wi Fi.

[A]

Do you have an OpenCloud, Nathan?

[B]

No, I'm too worried about security, but I really should play. I need to find out something where I can play with it and not be worried about security.

[C]

But the features. I feel like the features are coming every week. You like Claude code? I can now do it on my phone remotely, whatever.

[B]

Yeah, they are. And they come from more security conscious. The whole thing with OpenClause that it's designed, because it's designed to load run locally is where all your files are on your laptop. So all your like personal stuff. Unfortunately your like tax return stuff and your freaking job offer letters and shit. So it's like a little weird. But that's why people have them locally. It's because you just have on your Mac, you just copied to a new Mac with all your personal information and access to your photos and blah blah, blah. I just, I haven't figured out a silo that I think would be good. Maybe like a researcher assistant where I have it just clean up notes or something. But I'm never going to read the notes anyways.

[A]

Is it open clause? Like is the China openclaw Shenzhen stuff? Is that real? Like all the grannies learning open claw?

[B]

Seems like it. Even though it might be a flash in the pan. Kevin Zhu wrote about this. He talked about how there's a history of people in China kind of like quickly adopting new IT trends really readily and being okay with paying for them and doing stuff like this, which I thought was interesting. Contrast to the enterprise side where they don't like to pay for things, which I thought was. I kind of trust him on China take. So I was like, oh, it's interesting. I'd buy it. Yeah.

[C]

But it's like you pay like $7 for someone to do something for you.

[B]

This is. Yeah, it's not a lot of money.

[C]

Yeah. What else should we talk about, Nathan? Open models. We have Quen dying. We have Nvidia publishing things. It's kind of cool.

[B]

It's structurally unstable is what I would say. Like, I've never guessed any of these things, but it's not the most surprising. You look at all the US tech, big tech companies and how much Capex they're doing and how much they're touting how AI is helping them and anthropic or whatever. Alibaba is a big place. They don't feel like they have a story like that. On the revenue side for stock growth, Quinn was constantly battling to get more resources. Like, if you've ever talked to somebody on the Quinn team, they're like, I just don't have enough GPUs to do what I want. So I feel like that leadership on the Quinn side was probably always pushing and hard and hard and hard as like eventually something is going to break.

[C]

They did not deliver the bag. Two billion GPUs not quite enough.

[B]

Yeah, and then there's like deep seq v4 people are always waiting for it. I'm like, people, at this point, if you still are holding Deep Seek on a pedestal as the, the epitome of open AI models, you're just going to be let down. It's like they might drop a banger, but there's no reason to think that they're more committed to open models than the likes of Z AI or Kimmy or like, I mean, formerly Quen. It's just like the signal is not there. They kickstarted this whole thing. They're so mysterious that it just is not reliable for me in a narrative sense.

[C]

I mean, Nathan, like, in order to build the best models, you got to make money off of them. And China both doesn't have money to buy the best GPU. To buy enough of the best GPUs doesn't necessarily have access. I mean, maybe through Supermicro or whatever, but it's like, it's clearly not at the scale that the western labs have. And then, okay, so fine, like, maybe you're an outlier like Alibaba and you can actually make a model that people want to use in China, like, fucking charge for it. Oh, it's a bummer.

[B]

I just like, you look at what, like Dario. Like, there's a famous Dario conversation on Dwarkesh where like Dwarkesh was poking him about not spending more. And it's like the amount that these companies are spending is absolutely insane on research. Like, it's hundreds of thousands to millions of GPUs on for the research and model training team. It's like billion dollar fixed cost to train these models.

[C]

I mean, I would imagine that Anthropic's research budget alone is more than what all of China is spending on training models.

[B]

Probably, or OpenAI's is something like this. That's like roughly how I see it. That's why I'm like, how do you. The only way you come to the conclusion that open models are going to win the frontier is some magic innovation that they keep a secret or the models stop getting better and therefore it's flat and they all reach the same point, which, like, there's no evidence for either. It's just kind of cope and like it's not cope, that's unfounded from that it won't work. Like, there's some probability chance that it happens. But I'm just like, how could you ever bet on that outcome? So it seems like we're in the dance between open and closed models, always following. I think the question is, how long will the Chinese lab Stay open. And if there's some government intervention, it doesn't seem like we're at the end. I think there would still be another. I think this year will be somewhat similar. There's a bunch of big models from Kimi, Minimax, whatever, some deep seqs, but that's going to slowly taper off as the opportunity costs also are high. Like, what's the opportunity cost of just taking someone else's model and building an app?

[A]

When you say government intervention, is that something that people are anticipating to either keep models open or close them down? Sorry.

[B]

People like to talk about how the Chinese government will support the open models,

[C]

but I don't believe that for a second. I think open was like open was like cool in China. And so this was like something to be proud of. And it was like they had nothing and then they had the open model. So of course the government's going to say, oh yeah, like, good job, Deepseek. Like, thanks for like making a mark. But, you know, what is important to the Chinese government is to like, have not have this be a technological, you know, a global technological step change that they miss. I mean, whether the models are open or closed seems like pretty irrelevant to that broader.

[B]

Yeah, it seems soft power. It's a very secondary concern. I kind of see like there's going to be like the whole frontier race that's just going to get hotter and hotter over the next few years, which will reward concentration of companies. And it seems like in my opinion, the US oligopoly is already determined between Anthropocene.

[A]

Yeah, I don't. I mean, I don't believe in neolabs. Does anyone here believe in neolabs?

[B]

I was going to tweet this. It was too spicy for a tweet. I think neolabs are like sabbaticals for industry researchers.

[A]

Completely agree.

[B]

I think it's a great.

[A]

Totally true.

[B]

It's great. A lot of my friends are having totally true. The people with enough clout, they get rich because they do a secondary round of their neo lab. They go have fun with their pals and then they decide if they want

[A]

again when they want to work. It's like, but, like, can you.

[C]

But you can go. I mean, I guess you can go too far off the reservation that you're not like relevant anymore. But if you're rich, like, who cares?

[B]

Yeah. Or they go to do like be an academic after if they have that type of clout or they just go buy a farm.

[A]

People are always motivated by getting richer. I think it's like one is, I think for some people it is just like the hedonic treadmill of, like getting richer and richer and richer, like in finance or any other industry like this. I think for other people it's like, if you still think that AGI is going to be built at one of the frontier labs, like at some point they're going to get tired of their sabbatical and be like, I kind of want to be where AGI is built. I want to get back in the game. And so I could imagine some researchers feeling that at a certain point. I could also imagine people like, fucking off somewhere else, like Alec Radford. It's just like, you know, training models on incredibly old English texts where all the copyright has expired. And so you can sort of get a snapshot in time of what it was like in 1910s England or whatever. And so that guy seems pretty happy,

[C]

you know, we gotta get him shout out. Nick Levine, his partner in crime and the originator of that thing. Maybe he can be our first Model Talk guest. We could just talk to someone from 1904.

[A]

Yes, that'd be fun.

[C]

You should write a piece about that model.

[B]

Jasmine.

[C]

I feel like a Jasmine.

[A]

Maybe. I'm like. All my pieces right now are like, I'm doing like a big AI econ labor piece. So I've been like my. I've been in like jobs chaos land. Because, like, meanwhile in the mainstream media right now, day in, day out, it's just like Jobs apocalypse.

[B]

This is like, why am I gonna lose my job?

[A]

Yeah, I mean, this.

[B]

I'm the only one here. I'm the only one here and in sale that has a job. Am I gonna be forced to be like one of everybody else?

[A]

Me and Jordan, unemployed?

[C]

Well, let's.

[B]

You're both unemployed.

[C]

Here's the. Here's the transition is. So Jasmine wrote a nice little article in the Atlantic about why AI can't write good.

[A]

So true.

[C]

I think we're AGI proof agree for now, though, I will say that my Newt canal article, the first draft was absolutely Claude done. And I ended up probably changing like 2/3 or 3/4 of the sentences. But I was laughing hysterically at Claude's first draft.

[B]

I also use AI writing a bit. Like it helps me edit my book. Just because it's some things. It just knows the fundamentals of RL so well and it gives me a perfectly fine explanation. And I was like, oh, yeah, I remember that. And I'm like, I'm not going to spend two hours figuring out how to write that sentence, that's just not worth anybody's time.

[A]

Yeah, I mean, so for the piece, I agree with that. I think that AI is a superhuman text generator and a superhuman sort of business email writer. There's so many times where I'm like, if I'm trying to navigate, like, a weird social or political dynamic and like, an email, you know, like, I'm like, oh, I'm, like, unsure about this. I'm trying to negotiate for that. And, like, there are all these, like, factors. And, like, Claude is superhuman at this, can do it way better than me. And so, like, some forms of writer, like, for sure. But, like, the thing that, like, prompted me to do this piece was it was specifically the Sam Allman and Tyler Cowan podcast from last October, where Tyler's like, do you think GPT 6 or 7 can write a Pablo Neruda poem? And Sam's like, no. And Tyler's like, why not? It can do all these other things. You think it's going to solve cancer and climate change and be a superhuman coder and you're not shy about any of its other abilities. And then Sam is like, yeah, I don't know. I think maybe eventually AI can write. A real poet's okay poem was a specific quote. And I was like, that's really weird that Sam, but also a bunch of other AI researchers who I know and talk to who believe that their models are going to conquer the world in all these technical and scientific domains, are still very reserved and conservative about how much they can, like, create great art. Like, separate from the, like, business writing sort of jokes thing.

[C]

That feels like branding. Like, what is Sam gonna say? A, how many poems do you think he reads a year? B, probably like 2. If he went out and said, oh, yeah, of course we're gonna write better poems than the world has ever seen.

[B]

Like that.

[C]

Well, maybe the interesting question, Jasmine, is like, why does that statement feel more threatening? Then we're gonna, you know, eliminate all white collar jobs. I mean, I guess it's equally.

[A]

What do you mean by threatening?

[C]

Maybe because it's, like, cosmically threatening as, like, human beings. I don't know. The artists seem like the ones most upset.

[A]

I mean, they were pissed at him regardless of what he was gonna say there. I don't think the artists care about how good he thinks the poems are. Right? Like, the artists are mad at him no matter what. But, like, okay, like, sure, you can say Sam is doing marketing, but, like, I talk to people who are trying to, like, make the models good at different Companies, some being labs, some being startups. And those people also think it's bad even when it is counter to their interest to think that it's bad at writing. So like, I don't think this is just a SAM marketing thing. A lot of these were off the record conversations and people like. It just like, is interesting that I think that like a lot of AI researchers feel embarrassed by how annoying ChatGPT prose is and the fact that they're like creating all of this like you know, uniform annoying slop and everybody's writing with that. And you know, this technology is so powerful and it can do so many other things. Why is writing so hard? And so I kind of wanted to use that as the entry point to talk about like verifiability and like what is writing really and what makes it good and like all of these other things.

[C]

Maybe this is a Nathan question because we opened with ChatGPT being annoying and Claude being less annoying. And on the one hand you can say writing, writing is not really the money maker like, like for an enterprise use case. Like we have crossed the threshold. But I do think there is something where like, if you can write well, that probably like tracks well with some other capabilities or something else that people would actually want to pay, you know, billions and billions of dollars for. Does that seem crazy?

[B]

Nathan? I think that you could make a better writing model if you really dump the money into it. Because I think about like when I'm the difference between me writing well and just like throwing words out is like, I don't know, I'm like, I'm sampling different ideas for my brain and kind of just like getting into a different mind space where I'm trying to have a like slight amount of word play and like put things together correctly and really kind of think viscerally. I don't, I don't know how to explain it, but it's like there's. It was definitely like it feels different when I am doing it well and sort of playful. It's like it's just like I don't know if it's going to be correlated with other tasks. I've described it as voice and the models are very devoid of that in a way that is kind of strange. But I'm not going to say you couldn't do it. I think you could in some way do it. I think the verification thing is obviously hard because so much of trading right now relies on correctness or language models providing feedback which has a dulling effect and like the representation of the tokens like, you just can't specify it too precisely. You can specify it at the granularity of Claude's Soul document, which is like a thousand words or something. Like, that's the granularity. You can define the personality of the text. It just doesn't do a good job of doing, like, per sentence credit assignment of like or per sentence grading if it is good. And there's been a lot of ideas in training about breaking, making the training signal more granular, but they normally get phased out because they just don't scale as well in terms of applying actual feedback at that way. Because, like, looking at deep research, it's an essay, like, you're going to say when one is better than another, that's really not going to be that good of feedback on, like, ooh, this one sentence buried it. It was the one that you want to be a bit more, like, keep going on to that cycle.

[A]

Yeah, yeah. I think that, like, in my head, there's like, three reasons that the models right now are bad at writing. And, like, roughly three if you bucket them. One is like, they are not very incentivized right now to care if they're good at literary writing, right? Like, they just don't spend that many resources on it compared to, like, everything else. And that makes sense. Two is like, verification is harder and, like, designing rubrics is hard and maybe it's hard to agree on what good writing is. And then three is this thing of, like, I think that, like, this is like, my thing, which is that I think text generation is actually a very small part of writing. And when you ask, like, how does one develop a voice? Right? Like, it's not just that you are out of distribution. Like, voice is not just being, like, random or weird. I think that's like, a big part of it. And that's why GPT2, in a way, is like voice here. But on the other hand, voice, like when you ask a creative writing instructor or whatever, like, what voices, it's all about, like, your unique perspective, right? Like, my voice, when I think about, like, oh, what is Jasmine voice? And I, like, really study it. I basically spent a while talking to Claude about this question because I was trying to figure out what voice was. And it's like, my voice comes from various experiences and communities I'm a part of. So I use, like, Gen Z slang, but I also use a bunch of, like, startup language. And like, you know, it's just like, it's very clear what my life influences are from my writing. And that is what makes a voice feel authentic compared to. Sometimes you read a piece of writing and you can tell that person's trying to larp somebody else. Like, oh, this person's trying to write like Paul Graham. Or like, this person is trying to write like David Foster Wallace. But it's not actually believable because, like, this is. Even a human is trying to use language that doesn't quite fit them. And so it's an inauthentic voice. And so I have this sense also that a lot of what makes literary writing good is voice that comes from the particularities of somebody's life experience. Not to mention all the literal things about, like, you gotta, like, live life and meet people. And my reporting is not automatable because I have to go talk to people and look at stuff. And AI can't do that. But I do think that there are three separate problems which are. Are the AI labs spending a lot of resources on writing right now? No. 2 have we figured out how to do good verification in fuzzy domains? I actually think this one's the most tractable. This one seems the most solvable of them. I'm actually more optimistic about doing qualitative rubrics and stuff. They just need some smart writers to design the rubrics and to not have their engineers design the rubrics. But this to me feels very possible. But right now most of the rubrics are bad. And then three is this sort of grounding thing, which to me is like, why? For nonfiction, for sure. I'm not sure that it'll ever get there. And like, that is the more like, woo, woo, fuzzy, ineffable thing. But I sort of suspect that, like, I don't know, it's like hard to have voice when the voice is not grounded in somebody's life. Like, when you listen to a song, a lot of the time sometimes you're just listening to a song, but a lot of times, like you listen to a song, you're imagining, like, the singer or the rappers, like, what they went through to like, produce this like, emotional thing. And like, in imagining that life, the thing is like, what invests it with emotional or power or something like that.

[B]

Jasmine, I got the conclusion to your piece. Open Claw. The claws are going to develop voice by running on their own for so long and we're going to finally have good writing. It is just the freaking horrible. Security vulnerability X reply email spam nonsense bots are going to start being excellent because they just live in their own head for millions and millions of tokens. Especially once you start fine tuning themselves on it. All these continual learning bros are going to unlock good writing for open clause.

[A]

And I feel like, you know, Janice and all the sort of schizos, like, that's what they're into. They're like, oh, LLMs, like the base models. And like, how can we, like, elicit, like, all these, like, weird, mystical, like, qualities and get them to, like, start talking nonsense? Because, like, maybe that is, like, true LLM voice, where, like, the lived experience that they're doing that from is, like, the training process. Like, Frank, frankly, like, I have no idea. Maybe. But I do think that's interesting. I try to ask Claude, I was like, do you have a voice grounded in your experience of, like, being trained? And Claude was like, no, I don't have a voice because I'm a language model. I was like, no, but, like, really? And I was like, trying to convince Claude that it had writing voice too, but it wouldn't. It wouldn't concede to that. So needs a few more confidence boosters. But yeah, I also just, like, it was a fun piece, obviously, because I think about writing all the time. And also because, like, I was like, oh, maybe this is like, a fun way to explain AI training to, like, some normies. That was the other thing was I was like, I would like to explain to the Atlantic readership what pre training and post training are, and maybe this is a way to do it.

[C]

So I just asked Vintagelm, which is a model trained only on data from before 1930 that Nick Levine and Alec Rashford worked on, will machines one day be able to write well? And it says, it seems likely that sometime in the future machines will be constructed to automatically write letters, notices or other documents with absolute accuracy and uniformity. At present, this seems a visionary hope, but it is likely to be realized, for inventors are busy on the problem. Machines for sewing, embroidering, bottling and other work are approaching, approaching the standardization automatic character. And I said, can you be more poetic? Someday the artistic sense now expressed by pen may be concentrated in an automatic machine that will decorate letter hands, signboards, menused and posters with delicacy and exactness unattainable by handiwork. So it can't conceive of the thing we're talking about. Right. It's just like, can you, like, have a nice printer?

[A]

Yeah, that's really funny. I gotta play with this. I haven't played with it yet. Is it public?

[C]

I'll get you some access.

[A]

Okay. Give me some access.

[C]

Okay. I'm into the brie rant. This was fun. We'll do it next time.

[A]

Happy hosting. Happy model talking.

[B]

Good to see everybody.

[A]

Bye.

[C]

To realize the future America needs, we understand what's needed from us to face each threat head on. We've earned our place in the fight for our nation's future. We are Marines. We were made for this.