ModelTalk: Claude Fable (Nathan's pissed), is AI actually productive, advice for graduates

Summary

ChinaTalk: ModelTalk — Claude Fable (Nathan's Pissed), Is AI Actually Productive, Advice for Graduates

Host: Jordan Schneider
Guests: Nathan Lambert, Jasmine, Ethan
Date: June 12, 2026

---

Overview

This ModelTalk episode dives into ongoing tensions in the AI community around Anthropic's Claude Fable release, disappointment and distrust in Anthropic's safety policies, open vs. closed AI paradigms, and the real-world impact of AI productivity gains. The conversation then shifts to practical advice for recent graduates anxious about AI-driven changes in the workplace, ending with a discussion of data centers, NIMBYism, and absurdities in local tech opposition.

The episode features candid, sometimes fiery exchanges, keeping a sharp but witty tone throughout.

---

Key Discussion Points

1. Nathan’s “Anthropic Hate”: Claude Fable and AI Community Backlash

• [00:17-04:38]
  Nathan kicks off by slamming Anthropic's new Claude Fable model and its pricing/exclusivity, criticizing the company's invocation of “safety” while undermining user trust with shifting access policies.
• Anthropic "burnt a lot of trust" by restricting access, especially for researchers, and then walking back on their initial decision ([00:35])
• Nathan calls out the leadership's “narrow conception” of the AI community: “They are keeping to turn up the temperature where it seems like they feel personally responsible for fighting all these fights for the AI community. But their idea, their conception of the AI community is very narrow around themselves.” ([02:16])
• Fears of power concentration and scientific progress being “dictated by one private company's whims.”
• Jasmine agrees, drawing parallels between AI monopolies and authoritarian regimes: “The reason I don’t like a world where…Anthropic is the sole provider of that utopia is the same reason I don’t like authoritarianism.” ([06:09])

2. Nationalization and Power Dynamics in AI

• [05:39-07:32]
• Jasmine: Anthropic leadership may be open to nationalization, but are wary of who controls the process (especially the Trump administration).
• Nathan believes “open is better” but acknowledges that open models come with their own risks. He’s uncomfortable with the thought of either the government or a private monopoly being in control.
• Jordan notes slow, confused policy responses in US government, arguing executive and legislative action seem unlikely in the near term ([07:32]).

3. Real Progress or Hype? — Are AI Models Actually Productive?

• [10:15-23:08]
• Ethan questions whether recent AI improvements lead to tangible productivity gains, claiming economic impact is overstated: “I feel people massively overweight how much economic productivity is actually coming from these things.” ([10:17])
• Nathan retorts that “my output is multiples higher than when ChatGPT was released,” but agrees diffusion takes time and that micro-studies of productivity gains show large variability depending on user skill.
• Jasmine: The key to productivity is how agents are used. “Some people are just getting five times more uplift than the other people because they're better at using agents.” ([13:03])
• Ethan: Even if “20% improvement is a lot,” organizational waste and feature bloat mitigate impact. “Most of the features companies end up shipping…probably shouldn't ship 90% of the features.” ([17:33])
• Jasmine: Ultimately, organizational adaptation is coming—companies will get leaner, as “companies are like somewhat rational and they're gonna figure out the right ways to use these things.” ([18:37])

4. Open Models, Price Compression, and What AI Is Good For

• [23:53-26:11]
• Nathan and Ethan discuss open models—Ethan compares value in the space to “the value of Louis Vuitton…mostly the signaling value and the brand.” ([24:05])
• Most real-world, individual AI model use remains in communication or chart creation, not substantive world-changing advances.
• Nathan gives a concrete example of building a data API for interconnects as a genuinely valuable use case.

5. Advice for Graduates in an AI-Transformed Economy

• [28:01–38:34]
• Jasmine observes that young people's pessimism over AI is met with “fatalism” from AGI-pilled friends. Her advice to grads includes:
  • Make friends, be flexible with identity and work
  • Develop tacit knowledge, avoid deskilling with AI
  • Don’t bet everything on computer/desk work
• Nathan pushes for entrepreneurial thinking: giving “people conditions where they will stumble into an idea of a micro business.”
• Ethan (sardonically): “Probably go to church, find God and touch grass” ([30:43]). He suggests that real-world relationships and getting offline offer resilience, despite his own lack of religious practice—a point Jasmine quickly challenges.
• Jordan: Encourage high-agency experimentation, like “start a Substack and see where it takes you.” ([35:56])
• Jasmine, critically: “If you don't want to be high agency or solopreneur or whatever…get a job in something that is going to grow with the AI economy” (data centers, healthcare, public sector, trades) and doesn't subscribe to the “give up and find God” for normal people line.

6. Data Centers, NIMBYism, and “Chinese Influence”

• [38:34–48:10]
  Jordan riffs on the controversy around local opposition to data centers and the absurdity of blaming Chinese influence for NIMBYism:

“Trying to smear AI data center opposition as, like, a communist plot is, like, completely missing the point.” ([39:24])

• Argues companies could easily win community support if they spent more (e.g., “Pickleball courts on top of data centers!”).
• Jasmine: "Just do it. Everyone will be happy. It’s gonna be okay." ([41:49])
• Transparency and proactive engagement are the real solutions, not conspiracy theories about foreign influence operations.

7. Research Lockdown, Claude’s Future, and the AI Product Gap

• [48:21–54:41]
  Nathan shares Joanne Jang’s tweet critiquing Claude's restrictive policies:

“Kind of crazy that someone's full time job was to steer Claude to sabotage ML research capabilities of paying customers. So good.” ([48:21])

• Jordan and Jasmine criticize how “AI safety” rhetoric parallels tech companies’ strategic use of “privacy,” pointing out both genuine belief and ruthless self-interest.
• On switching models: Nathan is undecided, Jasmine prefers OpenAI’s GPT-5.5 for research over Claude, noting real performance differences for specific use cases ([51:04]–[51:59]).
• Universal gripe: Despite “superhuman cyber attacks” being promised, models still can’t reliably turn an outline into a solid article—“it’s not that hard!” ([52:42]–[53:14])

---

Notable Quotes & Memorable Moments

• Nathan on Anthropic:
  “I can't just rely on this as a long term partner. And I think this is still the case despite the change…” ([01:52])

• Jasmine on Monopolies:
  “The reason that I don’t like a world where…Anthropic is the sole provider of that utopia is the same reason I don’t like authoritarianism.” ([06:09])

• Ethan on Productivity:
  “I feel people massively overweight how much economic productivity is actually coming from these things.” ([10:17])

• Jasmine on Company Evolution:
  “Anytime you have a new technology show up first, you give every individual the new technology. You don't reform the systems…companies are like somewhat rational and they're gonna figure out the right ways to use these things.” ([18:37])

• Nathan, joking on AI research:
  “My job is secure because I'm the only one that's allowed to do AI research out of a frontier lab and therefore I'm gonna be okay.” ([48:10])

• Ethan, on coping with AI-era angst:
  “Probably go to church, find God and touch grass. I don't know… probably underpriced right now. I feel like if people feel a ton of fatalism, it’s more correlated with spending too much time online…” ([30:43])

• Jasmine, on advice consistency:
  “I'm not anti church. I'm anti people giving advice that they do not follow.” ([33:06])

• Jordan, satirizing data center NIMBYism:
  “My favorite proposal… pickleball courts on top of the data centers…just make these look like beautiful and innocuous civic buildings that have recreation facilities attached … this is like, I think an entirely a fake problem, one that like money can and should solve.” ([39:24])

---

Timestamps for Important Segments

• Claude Fable and Anthropic Trust Issues: [00:17–04:38]
• Open/Closed AI, Nationalization Risks: [05:39–07:32]
• Gaps in AI Policy Response: [07:32–09:08]
• Is AI Actually Making Us Productive? [10:15–23:08]
• Open Models and What AI Is Good For: [23:53–26:11]
• Concrete AI Use Cases: [26:11–28:01]
• Advice for AI-anxious Graduates: [28:01–38:34]
• Data Centers, NIMBYs, and Foreign Influence: [38:34–48:10]
• AI Model Lockdown, Product Gaps: [48:21–54:41]

---

Tone, Style, and Dynamics

• Candid, direct, and often irreverent—guests are not afraid to call out “bullshit” or critique peers
• Mixes deep policy, technical insight, and workplace realism—insightful debates over hype vs. reality
• Witty banter—self-deprecation and jokes (e.g., creating voiceover graduation speeches, pickleball courts on data centers)
• No sacred cows: Institutions, companies, and even the tech industry’s own narrative are all up for challenge.

---

Final Thoughts

This ModelTalk delivers a critical, nuanced, and wryly humorous look at the evolving realities of AI’s promise, pitfalls, and workplace disruption. Whether you’re worried about AGI, stuck choosing between Claude and GPT, or just trying to give a 22-year-old some hope, this episode underscores the complexities facing both individual workers and society as a whole.

Transcript

[Jordan]

My mayor Muslim, my bagel Jewish. These models are alive. Nixon. Welcome to try it.

[Nathan Lambert]

Excuse me.

[Jordan]

Welcome to Model Talk here with Jasmine's

[Nathan Lambert]

son Nathan Lambert, as well as special

[Jordan]

you haven't seen the Season of Mandates, the substack as well as text Ql, a startup. I think we're going to start with some anthropic hate from Nathan. They're giving us AGI for $200 a month, but that's not enough for you, is it?

[Nathan Lambert]

Well, that's a trajectory. Yeah, it's a trajectory. So like Cloud Fable is in your subscription plans for two weeks until they decide if you'll only have to pay more, which means mostly only enterprises get access to it. And they are claiming that almost all of their decision making is for safety while doing things that actively undermine their users, which to their credit they have walked back this policy. But a lot of trust has been burnt by them telling people that they're doing this, undermining the users for safety, while they have three much more simple safety mechanisms to release with the bottle. So my question is like why do they think this is acceptable to release if safety is the top line priority relative to anything else? Is like they're not, they're not doing it fully. They clearly felt some pressure to get this out. And then also the whole AI research community has turned against them because they no longer feel like CLAUDE is a platform that they can rely upon as the future thing of their work. Like I describe it as this year as I'm in the last six months figuring out how to change all of my workflows to work with agents. And I think CLAUDE is easier to work with on the broad set of tasks I do where Codex is probably a bit better at writing code. But it's like if Anthropic is going to make it so I have to Swiss cheese their classifier to get it to work on my editing my paper or reviewing my code base or my ideas. I can't just rely on this as a long term partner. And I think this is still the case despite the change. I think there's a few weeks of understanding how the safety filter works, but they've told me they intend to restrict access to other AI researchers because they think they are the only people that could do safe AI research, which I think is categorically an error on how safety exists. Like safety is by a broad set of people having no financial incentives to eat their own bullshit and agreeing that the broader AI system is safe. So I would be shocked if I'm the Only Anthropic hater on this call for that. I don't hate the people at Anthropic. I just think that they've been, the leadership is very bought into their mentality and then they are keeping to turn up the temperature where it seems like they feel personally responsible for fighting all these fights for the AI community. But their idea, their conception of the AI community is very narrow around themselves. And for whatever reason they don't seem to be asking for help. Or they say they could slow down, they could slow down on these releases and get them more right.

[Jordan]

I don't know.

[Nathan Lambert]

No one's going to defend me. Somebody could defend Anthropic.

[Ethan]

Do you feel like they're, are you like, like personally offended, like to, to this effect or do you just feel like, like they're doing so it's like a, like a. How dare you. How dare they be like high and mighty like in, in the way that

[Nathan Lambert]

they both like, I can't do my work anymore.

[Jordan]

You didn't see that Nathan Lambert in the, in the model spec. We want him not to be able to do his work.

[Nathan Lambert]

I think both, I mean it's just like I'm worried about concentration of power for a long time in AI and there's a prevailing sentiment that Anthropic is the furthest along in this. But there's a sentiment at other frontier labs that they will be the only people who can do certain types of science because they're the only people with the inference compute and the compute in the smartest models are the only things that are needed to make scientific progress. And these are all horrible ideas. It's like you're doing research and development internal to these companies. But I wouldn't call it scientific progress if it's just dictated by one private company's whims.

[Ethan]

I feel like they're like pretty consistent with like how they, how they behave or like their actions are like completely consistent with like what they believe. Which I guess is like hard to, hard to hate on because like, like

[Nathan Lambert]

I think if they can hate on it and that's bad for society, but it is internally consistent. And I like, I commend them for like being true to their virtues, but they're walking path of fire to the government striking them down.

[Jordan]

I mean Nathan, like if we, if we live in the world where we get wonderful scientific breakthroughs and cure cancer and productivity goes up and we're all rich, but it's like only brought to you by black boxes in frontier companies and like we end up where the sort of like, worst, worst case scenario safety pathways are. We don't fall into, like, is that not okay? Or maybe, like, what's the likelihood of that happening?

[Ethan]

What?

[Nathan Lambert]

I think the government would just be like, we own you. If the entire economy is one company. I think the government would also be like, you are now a government company.

[Jordan]

Yeah.

[Nathan Lambert]

Let alone the, like, obvious AGI is the most powerful technology. They love to compare it to enriching uranium and nukes and telling everybody that you're building the strongest weapon in a private company with little governance. The government already was upset about this, but I think economically the government would take action just against the monolith as well.

[Jasmine]

Yeah. I think a lot of people at Anthropic, including the leadership, have always been open to nationalization or self nationalization. And I think that outsiders sometimes underestimate the extent to which nationalization is not necessarily an undesirable outcome. I think they care about what that looks like and how it works and who the administration is. I think most people at Anthropic really don't want the Trump Admin in particular to have that much control over them, but I think it's all the control over that.

[Nathan Lambert]

It goes back and forth.

[Jasmine]

Oh, I know. I mean, this is so, so to be clear, like, I'm, I'm closer to Nathan's side on this debate where, like, the reason that I don't like a world where like, yeah, we have utopia, but Anthropic is the sole provider of that utopia is the same reason I don't like authoritarianism. Right. Like, the authoritarians can promise you all sorts of cheap goods, but fundamentally you're relying on like three guys in power to like, keep you around even though you don't do anything for them. And like, I just, like, I get that Anthropic thinks that they're the good guys. And, you know, so far they have made some, like, I know many people who I like, do think are good people, but for the same reason I believe in democracy. I just like, still don't want those like, three guys to be controlling my lives. And I don't think other people want that either.

[Nathan Lambert]

I think what it pushes to is that either it needs to be nationalized or the standard of the frontier of intelligence needs to be more open. And I. That's why I'm like, I think open is better because I don't really trust the American political system with them controlling the AI models that are coming in three years as a single entity to do whatever they want. With like, I don't know, the US political system doesn't seem ready for that. So therefore I think something that's more open and auditable and has a very diffuse notion of control is actually like socially more stable. There's, there's a lot of, you could stack a lot of small risks of AI there. Like there's all these domains that get more risk when everything is more open. But the system seems a better high level. Yeah.

[Jordan]

So but he base case, right? Like we just had Trump like he was going to do an executive order, then he undid the executive order and then he was going to test the models. Then a week ago he's like, oh no, I don't really even want to test the model. So. So like think it's reasonable to assume executive branch basically doing nothing on the sort of like, you know, how this technology is going to roll out Congress. I just had Senator Slotkin on my podcast. I was like, what should we do about AI? And she tells me about deepfakes and data privacy. It's like, okay, but I'm not sure if that's really what we need here. And like this is a good one. Like on the, on the spectrum of like with it senators. She's like, probably second standard deviation. So, so the legislative branch isn't coming, is incoming executive branch isn't coming over

[Ethan]

the next three years.

[Jordan]

And then it's like, okay, so we're just like, I think the like trying to dial all the way to open source is.

[Nathan Lambert]

We should see the executive branch be forced though. Like they already were forced and did some crazy shit and then they walked it back. Like who says they don't do more crazy shit and then don't walk it back.

[Jordan]

Yeah, I mean, I guess, I guess like if you're. If. If what you're as an anthropic employee for is hoping for like the smart nationalization. Like I think that that path or just like the well executed.

[Nathan Lambert]

Like yeah, I don't see that framework. I think there's a predictable nationalization in the Chinese AI economy. A very predictable path there. I don't know they want that either.

[Jordan]

Well, that's kind of like what the path Deepseek took. They have all this state money now and they decided that that's better than taking partnering up with Ali or um.

[Nathan Lambert]

It's just like every year there has been a major bottle step like Mythos, what is it? GPT4 then like 01 and then you could like 2025. You could be like cloud code agents. 03 blah, messy, messy line item. And then mythos. It's like we're going to keep going up this ladder. It might not be as clear every year, but there's clearly big changes still coming that are hard to predict. And the models are going to keep getting better. And that's the world that assume we're going to. And people react as it becomes more real, which is going to be months from now. But the ways that that doesn't come is like Taiwan getting invaded or maybe a Presidential Election 2028 with sufficiently strong running positions. But like, scary stuff's coming.

[Jordan]

Ethan, you want to, you want to jump in here?

[Ethan]

I just feel like the model improvements aren't that significant. And I feel people massively overweight. How much economic productivity is actually coming from these things? Yeah, those are two different things.

[Nathan Lambert]

I think the model gains are extremely significant, but the economic productivity takes a long time to diffuse and the people building the models have lied about the immediate gains. But that's irrespective of the fact that the models are getting way better.

[Ethan]

Like, way better, I guess. Like what. How would you, how would you measure, like them actually, like getting better? Like what, what is like a. Like a. What is like a measurable Me and

[Nathan Lambert]

all the AI researchers I use are more. I work with are more productive and do all of their work through agents and the agents are getting marginally better at each release. Like, I'm like the. My output is multiples higher than when ChatGPT was released. The amount of work I could do.

[Ethan]

Yeah, I understand the self reporting that people all say that. It just feels like if it is the case that how do you make an engineer more productive?

[Jordan]

Right.

[Ethan]

You measure engineering productivity on how fast does your product improve per unit time, per unit energy, per unit dollar that you put in or the per unit cost on the engineering side. And Anthropic has had cloud code for much longer to the point where OpenAI was even using cloud code internally. OpenAI employees were. And they were upset.

[Nathan Lambert]

So that was a while before Claude was really Anthropic was really training the Claude models to be good in Claude code. Like Claude code started as an experiment of what you could do with Claude. Definitely took internal alignment to see that this was a real thing. And that's what I posit as the opus4.5 moment. And we're now downstream of that.

[Ethan]

My, my point is more like, like cloud code is like not obviously like a better product than like Codex and it's like not obviously like a better product than Factory or Cursor. Or whatever if you ignore the unit economics part where they can subsidize the tokens. And as a result, I feel like most of the best engineers are relatively awash for a while between 4, 7, 4, 8 and 5.5. And I feel like over the course of a year you would expect there to be a much bigger delta in product quality such that 9 out of 10 engineers are all using cloud code over codec. But I feel like for the past couple months it's been much closer to a wash, which makes me go like,

[Nathan Lambert]

okay, I think within AI labs the impression is that they have it and I would expect the AI labs to be the first ones to get it. And then I also think that it will take a lot of new organization building to maximally extract agents. Like me. Turning my blog Interconnects into a company is the sort of company that would show the gains of agents, not a company that was built 20 years ago with a old school management philosophy. Like I would expect them to not have a high ROI on these AI tools.

[Jasmine]

I also think that most of the microstudies that have been done on productivity mostly just reflect that the way that you use the agents makes all the difference. And some people are just getting five times more uplift than the other people because they're better at using agents. As to the point about whether the models are comparable, one, I think that Claude Code is still probably considered the leader and certainly was for most of the last while. And I think that was in part probably because Anthropic was using coding agents to build the next generation. And now all the other companies have realized that they have been doing the same thing as well. And so you can see them all speed up too. But yeah, I don't know. My sense is there's been a crisis and every other lab that's gone, oh shoot, we didn't invest enough in coding agents, we could have been speeding ourselves up more and now they're doing it and playing catch up. But like, to me that seems like a consensus that assuming that you use the models well, which, like, because there's so much difference in how people use them, like and how people use the agents, it doesn't always show up in the microstudies, but like, assuming you do, I think that like the uplift on coding in particular does seem pretty clear.

[Ethan]

I mean, I just feel like if it was there, you would expect the gap to be widening. And if the gap is widening, what I would look to is what are the best engineers on Twitter who Neutral, let's say open Source developers or T3 or DAX or some of the people working on open code or Sentry, those people are saying. And it feels like across the board, all of them are kind of just split between Codex and cog code in the prior version.

[Jasmine]

But Codex is also using codecs to build codecs.

[Nathan Lambert]

Phenomenal. But they've gotten so much better in

[Ethan]

the last few months. They should be, I think, and maybe

[Jordan]

less Ethan, like, like on the. Like, is it Claude or is it Codex? But like, is it AI agents that are going to be making the, the, you know, I don't know, bringing, bringing us a productivity revolution. You had a piece called Claude code is not making your product better. I mean, how do you feel about maybe is there like overreliance on this? Is people substituting it for like, actually things out into the world that can make money?

[Ethan]

There are like, there's like this frontier of like product quality and then like, once you're on that frontier, like more engineers does not produce like better product.

[Nathan Lambert]

Like, oh yeah, I think the tech industry is going to lay off a ton of people because they can do that and get away with it with the AI narrative. And then some of the people that are there are going to be way more productive. But like, there's so many weird dynamics to go through that aren't just like, slack gets better, like, slack's not going to get any better.

[Ethan]

Um, but like, then like in the coding world, like, what. How do you measure like improvements in productivity at all? Like, I feel like almost all the products, like, I don't think any of the products are going to get a lot better over the next couple of years. I don't think, like Salesforce, do you

[Jasmine]

think there'll be more products and more features shipped? Because I think we're already seeing that and I think that's true.

[Ethan]

Uh, I mean if you look at the mobile apps, like there's way more mobile apps, but the like used mobile apps like this has basically stayed flat. I think that like the, the total demand for software and like amount of time is more correlated with like how many humans are on the planet, how much time they spend looking at computers and like that number is kind of fixed. So like it's.

[Jasmine]

Do you think that there will also be like, like for the same company to be able to build the same product, the coding agents will have no impact on whether or not they need to scale headcount to build the features they want to build?

[Ethan]

Uh, I think it has no impact because I think right now, like, at most companies, if you have 50 engineers, three of them are doing all the work, and the other 47 are entirely awash. So if you fired those 47, the three are continuingly do all the work, and you also would have no impact. And if you gave those three engineers coding agents, they would probably still do all the work. Maybe the time horizons overall get compressed because there's a lot of manifesting coming into play where people. If you believe that you can do it faster, people can do it faster, then there's a lot of compounding factors. I'm sure it helps. Sorry, I'm not saying it helps nothing, but I think it helps by a factor of like, maybe like 20, 20 to 40%.

[Jasmine]

That's a lot. So. So this is the thing, right? Like, it's like, I think that's actually a lot, first of all, like, and, like, I feel like it's. It just seems crazy that you're making an argument that coding agents have no effect on engineering productivity. Like, I agree. Like, I used to be a product manager. I think that, like, lines of code is not the thing that correlates with product quality. It's like, ability to make, like, good product decisions and to prioritize and whatever. And, like, you need to scale that if you want to keep scaling, like, your product adoption or whatever. So, like, I get that stuff. But, like, engineering productivity is a separate thing from, like, whether or not your product quality improves. And it just, like, seems insane to me to say that, like, you would not be able to, like, ship, feature, ship the same set of features faster with fewer people. Like, I just believe that a world with coding agents, obviously you can ship the same set of features faster and with fewer people.

[Ethan]

20% improvement is, like, maybe a lot, but, like, like, not when, like, there's like 95% waste in, like, the system. Right. If you compare, like, linear to Ligera. Right. Atlassian has probably put, like, at least 100 times more dollars of R and D dollars, engineering hours to work trying to build jira. And it's probably objectively a worst product if you look at that. It's just like, oh, so product quality is entirely uncorrelated with R and D spend. And so then more faster productivity coming from those core engineers. I feel like most product improvements over time, most features are terrible. You probably shouldn't ship 90% of the features that you end up shipping, 99% of the features companies end up shipping. And the core features that really matter come from, like, these, like, moments.

[Jasmine]

Yeah, but you're like empowering like the three people who actually do work so that they're not slowed down by all these people who don't do work. I feel like that actually does lead to productivity improvements. Sorry, I don't know if this is a rabbit hole and like, doesn't matter, but no, I, I, I, I, I,

[Ethan]

I, I see that like possibly happening, but like the, the top line impact is like extremely low when they're getting diluted by the other 97.

[Jasmine]

Yeah, I just think organizations will adapt. This is like where I'm with Nathan, where I'm like, yeah, like anytime you have a new technology show up first, you give every individual the new technology. You don't reform the systems, you don't reform like the organizational structure, whatever. So. And everyone's like not looking at their token budgets. But I guess I just expect over the next couple years companies are like somewhat rational and they're gonna figure out the right ways to use these things.

[Jordan]

Yeah, like, why wouldn't the creative destruction happen?

[Ethan]

Why, why, why were those 47 people like hired in the first place? I mean, I guess in my mind it's because like, they're there to like placate the egos of like, they're like.

[Jasmine]

I don't think that's true. I think it's because you actually believe that you were getting productivity uplift. And then you hire the people and sometimes it's not as high as you want it to be, and that sucks. But like, like, okay, like one thing that you know, I mean, you work at a tech, when I worked at a tech company that I saw is like, you hire people, some of them are definitely 10x as more productive than the other people who are there. But in a world with AI, maybe you don't need to have as many of those like low productivity engineers. And so by giving the PMs or the senior engineers or the actually productive people coding agents, then when the like low productivity engineers quit, then you don't have to backfill them as fast. And like that's to me how the organization gets leaner and gets smaller.

[Ethan]

Yeah, I guess I just feel like that was always the case. Those people never needed to be there. And the fact that they were there, whether or not they needed to be there, is not changed by AI at all. And as a result, the thing that needs to change is do your managers feel like they no longer tie their ego to the number of people reporting to them and now they tie their ego to the total token budget or something that they get to Set on fire. Both sides are an equally irrational thing.

[Nathan Lambert]

This is why it takes a decade. Like, this is not a thing that's going to go in a year. I think Anthropic might not fulfill their whatever 40 or 100 billion ARR goals this year. I think they might because the models get better. I think that Claude Fable will be an economically very successful model for them based on early evals and stuff. It doesn't work for everybody. But I think models getting. There's certain domains where the rewards to frontier intelligence is really high. And people will pay for those tokens, but a lot of people that are trying the tokens won't. We'll just kind of stop. And it's like there's two different lines there and they're both reported in anthropic's rapid ARR growth.

[Ethan]

Yeah, I guess I feel like the second line is like like 95% of it. And as a result, like the. And like if the top line, the first line doesn't start growing faster, like you kind of run into trouble. I think like, with the current ARR that they're at and like the growth that they're at, like, they're looking at like something like 1 to 2% of like engineering salaries in like North America. That's like, that's. That's coming up on like basically 10% of like total IT budgets for like the CIOs of like Fortune 500 companies. You can get to like 10% in a given year. And like, that's like, okay, like they're not going to like immediately, like, the CFO is not going to wake up and like beat the shit out of you. But if you get to like 20 and like 30 and then like the going from like 10 to you go from like 0.1 to 11 to 10. And the whole time you're like, I'm growing 10x every year. But like going from 10 to 100 and like doubling this IT budgets is like a.

[Nathan Lambert]

Did you see the Citadel securities note on this today? I feel like it's a somewhat similar argument, but it's pretty well put together.

[Jasmine]

What did it say?

[Nathan Lambert]

Oh, do I have to read it? Can I share screen? Can I share screen? That might be easier.

[Ethan]

What is the.

[Nathan Lambert]

It's pretty long. It's like a few paragraphs. It's essentially saying the same thing. It's like a very professional way of saying that like the top end of tokens, the people who are getting value out of them are a minority and they're going to See kind of depressive pressure on anthropics revenue because so many people are getting ROI and it's that,

[Jasmine]

but then also like human capital at like, like people's ability to get value of agents will also increase. So like, yes, people will realize that some of the tokens are being wasted at the same time. Everyone will get better at using tokens at the same time. Like, if you give an advanced software to a boomer who doesn't know how to use it, they're probably not going to get very much value over out of it. But eventually people just like train themselves to do better things. I mean, I think the difference between me and Ethan is like, I think that firms are like slightly more rational. Like, I think that some jobs are bullshit, but like, I think that like not all the jobs are bullshit. And like actually people are hired not only for ego reasons. I think sometimes you hire people because you want them and expect they will do work. And I don't know, I mean, one metric that I'm like pretty interested in watching both in firms over the next few years and I'm sure there are people tracking this, is like, how does token and AI spend compare with headcount spend? And what are the rates that token spend and headcount spend are growing? And so do we actually see that firms are reallocating headcount spend to token spend or something like that?

[Ethan]

That'll be really interesting to track. I probably, like, you're probably right on the delta. I think that if people reallocate headcount spend to token spend, although I would probably end up seeing that as basically equally unproductive because or it's sustainable. The only thing I really care about is is it sustainable? And if the CFO and CIO feel like I should be spending, I'm going to go from spending like 60% on headcount of my total IT budget to 30%. I'm going to put that 30% into tokens. That's actually totally sustainable. I'd say that whole class of lifespan is probably like, like at a company like Coca Cola, it's kind of like parasitically sitting on top of like the actual like sale of Coke and like the profit margins that like could be like otherwise returned to public shareholders. But I mean, you know, you want to believe that like intelligence like actually helps you sell more Coke.

[Nathan Lambert]

How do you feel about open models and price compression? It seems like you think most of this is fake and that if it was TEDX cheaper it would also be zero because it's like not Real work.

[Ethan]

I mean, I think it's fake. The same way that the value of Louis Vuitton is not the actual craft of the thing, and it's mostly the signaling value and the brand and the perception of that it's valuable. And the same way people spending a lot of money on tokens could produce all that cost. I could see a world where a CEO would pay McKinsey a million dollars for a slide if the McKinsey partner says, Listen, I spent $400,000 worth of tokens on this, and maybe those tokens are spent on a for loop going through really dumb hypotheses. But I'd see a world where if you front the costly signaling, you end up getting there. And I'd say the same is the case with open models right now. I think most of the spend on outside of Fable, which seems to be like a genuine step up, most of the delta between people spending on coding models and stuff and like open source models is like, that was mostly like a sense of brand and you feel like Anthropic is a better brand around light coding.

[Nathan Lambert]

What do you personally use these models for?

[Ethan]

Mostly communicating ideas and like making, like, pretty charts.

[Nathan Lambert]

Because I feel like that's like a pretty low. Like they've been able to do that for a while and it's not like it's a high. I even think that's a pretty productive use case of them and if they're very useful at it. But I think it's not like as much of a potential unlock. It's like having an intern that would make a plot for you. Like, Jordan does all this vibe coding, and I'm like, jordan, you gotta do something that's like building progressively, like Vibe coding is like tarnishing the AI industry. Nobody wants another app.

[Jordan]

You know what I got? You know what I got Claude downgraded from fable to OBUs 4.8 for. I mean, I made it make a trailer of my, you know, Death of Stalin, but it's Mao movie. And then it was like, sorry, we can't trust you with Fable for that. My animated John Ching had to be, you know, an older model. All right, Nathan, what should I build? What, what, what, what actual software should I create?

[Nathan Lambert]

For example, like, I built a data API for interconnects because we scrape hugging face to get all the downloads data. Now I have an API where I could. I use it myself or could give it to a researcher to query, like model data over time for any important language model. And we have the flow of data and we create our internal eval metrics. Like I did this relative adoption metric and now there's an API where the student could query it for any model. It automatically adjusts the numbers every quarter based on some stuff. And it's like I just built this in what would be like a weekend day and that now exists and it's like a research tool that doesn't exist in the world for understanding open models. I just like trying to build out institutional expertise based around my very, very cutting edge knowledge. And I don't what exactly that research is for. Chinatalk.

[Jordan]

Yeah, there, there are two that I'm working towards. I mean one is just like the, the Chinese language news ingestion but it gets really frustrating because like you know the model like these, like Xiaohongshu and, and, and WeChat are kind of closed and like you can sort of get the.

[Nathan Lambert]

Jordan, do you follow me on RedNote?

[Jordan]

No, sorry. The other one which I'm more excited about and stay tuned is creating and like maintaining an eval suite for like policy and national security. So I don't know what that actually is going to look like yet. We're going to have a essay contest with a big prize number attached of like proposing benchmarks for like how useful these models are to like legislators and staffers and war planners. But anyways, tbd, I think that will actually be real software that can save the world. Tbd. Jasmine, you want to talk about the kids?

[Jasmine]

Oh sure. Well, I mean I'm like curious what everyone thinks. So I've been thinking about for the past month or so all of the booing at college commencements. Like anytime a commencement speaker says AI, then everyone boos. And that's not totally rational. I think kids just boo because their kids, but also the youth are worried about their future, they're worried about inequality, they're worried about affordability, they're worried about getting a job. They think that the AI companies and the government and everyone's screwing them over and it feels very pessimistic. And so I was thinking about what would I tell a 22 year old who is worried about the future and worried about their resilience and ability to get a job. And at first I started asking just a lot of my friends and people I knew, mostly tech people, but also other people. And one thing is I noticed a lot of fatalism. So the most AGI pilled friends I have would usually just say like, oh, they're screwed, they're doomed. Like I have no advice for the 22 year olds. They're totally fucked. They should be scared. And I was like wow, I don't like that very much. And so I was like, how do I. What kind of advice? Both like reckons with the fact that the future will be different and change a lot and maybe scary in some ways but also is like somewhat practical. And so yeah, so I wrote a post and I was like, I think it was like make friends. Like be very flexible about your identity and get good at like you know, getting new jobs and don't tie your identity too strongly to like being at one firm forever. Develop tacit knowledge. Don't like bet on doing desk work forever at a computer. Like don't deskill yourself with AI. Don't make yourself stupid because then you'll totally be a commodity and no one would ever hire you over an agent. But I want to know what you guys would tell a 22 year old right now. So that's what I'm curious about.

[Nathan Lambert]

I feel like mine's pretty similar. It's like mostly giving people conditions where they will stumble into a idea of a micro business and get excited about something and thinking about me in a business and becoming less reliant or having less expectations of the existing career ladders.

[Jasmine]

Yeah, it's like I don't know that people have framework. Young people are provided frameworks for like starting businesses. Especially like, like businesses that are profitable but are not venture backed businesses. Like I don't know, I don't know that college prepares people for that. Maybe it does these days.

[Nathan Lambert]

I don't think it does. My impression is that since I was in college it's become even more like expectation E where you like want to get an entry level CS engineering job

[Jasmine]

and I'm like those are going away.

[Nathan Lambert]

Yeah.

[Jordan]

That's why you want to tell the kids. Ethan, what do you want to tell the kids?

[Ethan]

Oh, sorry, I didn't hear that. I didn't know what you said. Probably go to church, find God and touch grass. I don't know like the. I'm trying to think about the things that are probably underpriced right now. I feel like if people feel a ton of fatalism, it's more correlated with spending too much time online and reading the headlines than anything else. And yeah, just relationships in the real world are real things and I think that at least in the spaces that I every. I spent a lot of time doing remote stuff spent a lot of time doing real life stuff in customers offices. I really wish I'd spent more time or at all points if I Could dial up the number of things I do in other people's offices and not remote. I would do it. And probably includes the park and everything else. And so yeah, I think people who are like probably going outside and they're seeing a lot of sunlight probably don't feel as much fatalism. And so

[Jordan]

Ethan, what Bible characters best capture various AI players today? You muted yourself. You muted yourself.

[Ethan]

I'm not religious at all. I don't follow my own advice.

[Jasmine]

So why are you telling other people to go to church if you're not doing it?

[Ethan]

I think that I end up. I'm a very, very. I behave like a religious person in terms of how much I have non nihilistic groundings and very arbitrary and subjective values that I just feel like are good. But generally I feel like a lot of being directionless comes from not having that.

[Jasmine]

I feel like what are your values?

[Ethan]

Do right by the people around you and then have fun mostly.

[Jasmine]

Yeah, See I think those are good values. I think those are good values. And like I also my thing, I was like, you gotta like get better at having fun, party more, whatever. But I don't think you should tell people to go to church if you're not going to church. Like I think that there's an, that creates an assumption that like you know, that church is going to be the thing to instill whatever these values are. Again, I'm not sure that church will teach them to have more fun or whatever or. And like it feels like kind of retrograde. Like some people should go to church but like, I don't know, recommend the stuff that you actually do do right by yourself.

[Nathan Lambert]

If you go to church, don't stop. Because I think people that have a community in church could end up happier. But it's hard to.

[Jasmine]

I'm not anti church. I'm anti people giving advice that they do not follow.

[Nathan Lambert]

Yeah, I was just giving, I was just figuring out a way to fix the advice.

[Ethan]

I feel like the. So like I used to, I used to say like, like work super hard on stuff and like whatever and try to like do stuff that like I more closely stuff identify with. But I feel like I have an extremely weird like out of distribution. Like I also like can't like give the advice like go to like go to a decent or like, like I use the word privilege or whatever. But like I, I can't give the advice like go to a decent school. Get yourself in a position where you can like raise like venture funding and like do like a bunch of crazy in like New York or one of the big cities. I feel like like a lot of the like, like nihilism like comes from places where like you don't have as much opportunity. It's like meeting people like where they're at there. It's just being like, like there's a lot of, there's a lot of like, like if you don't naturally like or just or of no, if you don't like fortunately find yourself in a situation where you have all the privilege on the planet.

[Jordan]

You know what would have been pretty funny? So in the past two days we've gotten OpenAI's Economic Plan for America and I think Anthropics 1 came out this afternoon. I have not read it yet, but if we start getting the frontier model labs being like actually the only thing that's going to save you is your belief in God and your church community. So when, when we're really, when we're really out of ideas and you see Silicon Valley explicitly pushing, you know, monotheism then or maybe not monotheism tvd, but

[Nathan Lambert]

I think long term, it's just the near term that I worry about and getting to. There's plenty of good possible equilibriums but getting there is hard because I think if it's like in 10 to 15 years and like classic arts and sciences educations are more popular and it's like writing class is mostly telling people to go explore things and then once they have found a topic they're interested in, you have them learn to write about it. And like all the education system is more built around building curiosity and building somehow figure out how to make status around this. And then like rather versus the CS undergrad which is like your status is you're going to get a 300 to 400k job right out of college. So like if the stat you're figuring out the way to change society so that status is around critical thinking and like being an independent brain that has thoughts on the table, then I think like these people are going to be well suited for extracting value from AI. It's like AI is the tool that just lets you run with this and you don't need to be in some random place in the world to do it. But I can't give this as advice to a student. It's like I can't tell them this because that's like you can't wait. Like, like waiting for the world to change is horrible advice.

[Jasmine]

Jordan, what would you say?

[Jordan]

I feel like, I feel like we've, we've done my Riff already on Model Talk, which is basically start a substack and see where it takes you.

[Jasmine]

So true.

[Jordan]

I don't know. I was, I was trying to find something I disagreed with in your article and I really couldn't. But, like, look, we're on a very weird plane and I think it's of like, kind of being like solopreneurs. And I, I don't. I mean, look, that was what the economy looked like in like 1860 is like, most people just like, ran their own farms or like sewed socks or made, you know, leather goods for their town. Am I sure that that's where this whole, you know, technological change is going to lead us? I don't know.

[Nathan Lambert]

And I think there's going to be mass collapse of capital and the biggest companies are going to get bigger. So, like, there's still going to be people at these companies. We're not going to replace them by AI in the next. Like, is Google going to be an AI only company in 20 years? Like, absolutely. Maybe. I don't think so.

[Jasmine]

Okay. Oh, like, oh, you meant AI only as in no humans employed.

[Nathan Lambert]

Yeah, like there's still going to be humans there.

[Jasmine]

I think there'll be fewer, but I think there will be people. I think fewer people, but I think people. I mean, yeah, I think my concrete advice, if you don't want to be high agency or solopreneur or whatever, is to tell people, get a job in something that is going to grow with the AI economy. So it can be at an AI lab. If you hate AI, it can be an AI policy and you can try to ban the data centers. I bet there's going to be a lot more jobs in both data center construction and data center banning over the next few years. So those are some good opportunities. You can get a job in healthcare, you can get a job in the public sector, that you can get a job in a unionized trade. So those are my things. If someone just wants me to list some jobs at them and not tell them to like, be a solopreneur. But I also tend to believe that, like, I actually find it like, slightly annoying that I think a lot of people in my circles will be like, oh, yeah, like among my friends, I tell everyone to like, be high agency. But like, when I talk to normal, normal people, they should just like, give up and find God. Because, like, they probably can't, like, figure it out. And I'm like, I don't know. I think I just have more faith in normal people. And I think that people would want to receive similar kinds of advice. I mean, they have options. You can choose the path that makes the most sense for you. But I think a lot of people can start small businesses because that's what a lot of people. I think more people want to start small businesses than are currently doing so right now. And if they realize that one thing that AI allows you to do is it might privilege small businesses in a bunch of ways, if you're, like, smart about the agent stuff, then I, you know, I think that's a good thing, and people should know that those are good options, too.

[Jordan]

So should we close on my data center riff?

[Jasmine]

Go for it.

[Jordan]

So there's been this, like, arc of both legitimate sort of investigation into Chinese influence operations, like, pushing out, like, tweets that have gotten literally zero likes saying, like, data centers are bad. You shouldn't build them as well as some, like, corporate dark money funding, like, really not good research that is saying, oh, yeah, it's like. Like China that's making people not like data centers. This is ridiculous. I'm sorry. Like, it is an interesting fact that, you know, there's someone in the Chinese deep state which has decided to, like, push this out into the, you know, to try to, like, glob onto this. But trying to sort of smear AI data center opposition as, like, a communist plot is, like, completely missing the point. And the most sort of absurd part of all of this is that it would be so easy for these hyperscalers to solve their local problems with money. I mean, like. Like, what if you, okay, spent all the money to make the data center silent, which, like, I think is like, $100 million. So Fable tells me if you be, like, offer to double, you know, the. The county's school district budget and put, like, real money into it for 20 years to. I think all of those people will flip. And by the way, my favorite proposal, which I think will really get all of the sort of, like, community activists on board is I would like to have, like, a platform on top of the data center once we, like, silence it and, like, have security or whatever, where you just create pickleball courts on top of the data centers. And then it's, like a happy part of the community. You can have Maya Lynn, like, do this sort of, like, like, layout, right? So it can look like Storm King and, you know, have, like, wavy grass, just, like, make these look sort of like beautiful and innocuous civic buildings that have, like, recreation facilities attached. And you, you know, improve the school systems. And all of a sudden, everyone's going to be, you know, clamoring for data centers in their, in their, in their backyard. So this is like, I think an entirely a fake problem, one that like money can and should solve. These companies are cheap. It's embarrassing that they're being cheap. Like, you've seen these like pledges to make sure the electricity doesn't go up. It's like, okay, we could do that. We can make sure the water prices don't go up. We could take the sound away and we can just like, you know, do data center UBI as Ben Thompson they

[Jasmine]

need to do some like Carnegie shit, you know, like make the park like really nice. Like, where's the salesforce park of data center? Oh man. Yeah. But yeah, or it's like, okay, like, don't raise electricity prices. What if you cut them all by half? Would everyone like the data centers more if you cut all the electricity prices by half?

[Nathan Lambert]

Pay the whole electricity, man.

[Jasmine]

Yeah, pay that. Make it free. I think if it's free, everyone will be pro data centers. If you have that much money, you need the data center that bad, then like, just do it. Everyone will be happy. It's gonna be okay. But no, I totally agree.

[Nathan Lambert]

It.

[Jordan]

It almost makes me think that this is a fake person problem that like the data. But I know it's not. But like it makes me think that the data center opposition isn't. Isn't even real if they're not willing to put an extra like $150 million.

[Jasmine]

I think they literally just haven't tried that hard. I think they like, are because it takes time to do community engagement and to like have these meetings. And I. My very vague sense of it is that the people who are in charge of the data center expansion just like have not spent the time and didn't really feel like spending the time to do the annoying community engagement process where like you meet with the city and agree on what needs to be meeting

[Jordan]

with the field, but it's not. But the community Eng engagement, it's like, okay, you can say nice things, but like the meetings don't matter if the money isn't behind it. Like, if you just 10x your like bribe then like everyone's gonna be happy with it.

[Jasmine]

Yeah, but I think it even takes time to like set up the meeting where you agree on what the bribe is, where you agree on like, is it electricity? Is it the park? Is it school? And I think that the, the people doing the data center build out just never spent any time on that. And I think that they could spend time on that and they probably will starting probably a month ago when they realize this is a big problem for them. But historically I don't think they really devoted time to solving this problem. A lot of the issues are around transparency. A lot of the people who are mad at the data centers. It's also because the data centers were under a name like Project Flow or Project Tempo. People wouldn't know it was a data center project and it wouldn't come up in the town hall meetings until very, very late. And so the community opposition would really fire up kind of at the last minute as opposed to a slightly more extended process. And it's just because there's such a rush to build and people haven't cared. But yeah, I agree that it's solvable. You just need to do like a tiny bit more community engagement and the community engagement should in fact result in like nice public goods for everybody. But yeah, the China narrative is really funny.

[Nathan Lambert]

It would be nice to have more non crazy reporting on what is happening at these discussions over time. And like how are these discussions changing? Because who's havoc it is it like like some construction agency contracted by Google ended up going to this local council and getting mobbed by the locals and they're just kind of like shit. They're like I don't know what to

[Jasmine]

do, like who is the local politicians will make a deal with the data centers because the local policy people want tax revenue and so they actually want the data centers. And so but then the both parties will underestimate the public backlash and then once the public finds out then they start calling, storming the town halls, mounting local position. I mean I don't have like an extremely deep understanding but that's my sense of how it happens a lot of the time is that the policy and politics conversation are quite different because data centers do tend to increase tax revenue. But I told Jordan this. I'm going to go on a data centers road trip next month where I go to some data center towns in like I don't know, Wisconsin or something and ask people why they hate the data centers or maybe they like them, I don't know.

[Jordan]

Yeah, I just be you know the, the sort of economic utility of a data center I think is something that I would imagine the local politicians don't quite grok is like how much money are these tokens going to be? Like are these data centers going to be printing over a 10 year horizon? And like the, I wouldn't be surprised if kind of like the like whatever tax Incentive like whatever tax abatements they're giving is just not, is like commensurate with like oh, this could be like a poultry like processing facility. Not like this is you know like the most valuable real estate on the planet and like you actually have a lot more leverage. So yeah, like having a 5% bump in like your local school and police budget is not something that people are going to feel in a way that like you know, crazy doubling, you know, crazy headlines like okay, everyone gets free electricity or everyone, you know, we're having, we're giving, we're, we're hiring twice as many teachers.

[Jasmine]

I think Zoron could solve this problem. I think if you gave Zoron this problem, he would figure out the deal to make. Like it would be like the like Trump new housing like type of thing. Like you need like a comms genius to figure this out too.

[Jordan]

You know. Who's not a comms genius? Who is these Chinese influencers that are being paid by like provincial governments. You know, I did a big report five years ago. This was my like Covid project. I taught myself data science and because Twitter had released all these like, like the raw data of all the state backed bot networks and you could go through them and like you'd see how many likes they get. Like they all got no likes. You'd see like you can read what the text was. Like none of them were funny, irrelevant or good. And ended that article five years ago being like, maybe when these, you know, when these state backed organizations get access to something like GPT3, then they'll really be able to sort of like crack the nut of overseas influence operations. And, and the OpenAI caught some folks doing basically the same thing around data centers and like not liking OpenAI from China and they were still really bad. It's like it didn't, it doesn't quite like even with the tools today, even with frontier model access today, you can't kind of get past the sort of like, okay, I'm like a 24 year old in like Hangzhou. I've never lived in the us I don't speak native English. Like, like it still is hard even with these tools to actually make something which is like compelling and persuasive to an audience which isn't your own. Or maybe these are just like the bums that are, you know, too, too lazy to like actually do a good job and they're just trying to, you know, fill some KPIs or something. But I was, I was struck at how mailed in the examples of, you know, data center Chinese influence operations.

[Jasmine]

Actually, yeah. I mean, maybe that's why our jobs are so AGI proof. We can do authentic non astroturfed data center influence operations from our purchase as massive online influencers.

[Nathan Lambert]

My job is secure because I'm the only one that's allowed to do AI research out of a frontier lab and therefore I'm gonna be okay.

[Jordan]

All right, cool.

[Nathan Lambert]

What else are you talking about? Wait, look, I read a tweet. It's so good. Joanne Jang, formerly did character at OpenAI, now at Corados, tweeted. Kind of crazy that someone's full time job was to steer Claude to sabotage ML research capabilities of paying customers. So good. It's like somebody had to do this for their job or they made Claude do it, but somebody told Claude to do it. It's like just, just so crazy to me.

[Jordan]

I don't. Why is it. Look, the, the, the, the White House Kratzios OSDP writes a letter saying we need the models to do a better job of stopping China from distilling models. Right. And like I would assume it like. Yes. Okay, so that's reason number one. Reason number two, you don't want OpenAI stealing your secret sauce. I mean it kind of makes sense, right? Like I, I guess like there, there, there's like, okay, if you are a private company and you want to like maintain your edge, setting aside all the safety stuff like, you know, access to this for like, like no one. There's no rule that they need to serve every single customer.

[Nathan Lambert]

Yeah, yeah. They just did it under the guise of that were in an intelligence explosion and they didn't need to be safe.

[Jasmine]

Yeah. I mean there's like this text in a group chat I'm in that I'll read out without a name. Chatham House. I guess that's reading anthropic talk about AI safety. Feels like Apple talking about privacy. They may truly believe in value. What they are saying. It is also the strategically ruthless thing to do. And I think that's true. I think it is like Apple talking about privacy.

[Nathan Lambert]

That's fair. Yeah. They're so much like Apple in my opinion.

[Jasmine]

Yeah.

[Nathan Lambert]

So, but I like Apple, so it's sad.

[Jasmine]

So are you going to keep using Claude?

[Nathan Lambert]

I haven't fully decided.

[Jordan]

Nathan's going full Siri.

[Nathan Lambert]

He's done with canceling. I almost canceled it. I'm going to see how strict the filter is and if I'm like really hitting the filter, I'm like, yeah, I'm done, but I'm definitely hoping that OpenAI gets better at the things that Claude's better at is like like editorial book and like I've written like just writing math equations and understanding intent when I'm making a slide deck and a lot of this like soft skill agent stuff. But Codex is already very good especially at anything technical like code heavy. So I'm really hoping that like GPT6 comes out and it is just like a bit more of a savant in terms of design and user intent and then I would happily switch.

[Jasmine]

Yeah, I think 55 is a good model. I actually like I am not changing any of my behavior based on this. However, I already use ChatGPT much more than I use Claude which is makes me an outlier and I feel like all the people I know because Claude is a God awful researcher and I keep telling everybody this, like if you give it a research task that requires a lot of web search, ChatGPT will be Claude in basically every eval I've tried. I've sent them to the anthropic team to be like get your shit together guys. Like this stuff doesn't work. But like Yeah, I think 55 is a good model. It's like less sycophantic I think than. I mean I haven't tested Fable enough but it's less sycophantic than Opus was. It gives more detailed responses. It is better at research, it is worse at editorial tutorial. So I guess insofar as I want some feedback on an essay structure, I would still go for Claude in that case. But 5.5 is my daily driver and has been and I think it's a good model and I think that actually it's underrated in SFE type circles because it's not cool to be into OpenAI but actually it's a good model.

[Nathan Lambert]

Yeah, I think it's my book workflows. So I have a workflow where I drop a chapter of my book in and I'm make this into a lecture based on the style of my previous 6 lectures on the topic. And OpenAI is horrible but Claude is like pretty good. Like I had it made like a 15 slide thing and I was like I need to get these models and find this figure in their paper for their post training recipe and put it side by side in my custom slide tool with the summary of the recipe and put them in this order and Claude kind of just does it. I have to do like one round of feedback on the slide so that it's ready. But OpenAI was crap at it. But once I'm done with this project, I'm going to go work with, not use cloud anymore.

[Jordan]

You know what's ridiculous is like they're promising me RSI and you know, superhuman cyber attacks and that, like, you know, zero days that no one's ever see seen. And it still can't really help me take an outline and turn it into an article. Like. Well, it's not that hard. I don't think I have the same, you know, I don't think what I'm trying to do there is as hard as like solving, you know, an Erdos problem.

[Jasmine]

But they're not trying to make it do the thing.

[Nathan Lambert]

I know.

[Jasmine]

They just don't care.

[Jordan]

Okay, if you can have one guy like take like deal with, you know, not let Nathan do his AI research, can we have another guy just help me write my Chinatalk article?

[Jasmine]

Yeah, actually there might be, there might be like a galaxy brain thing going on where one strategy to AGI pill the media immediately would just be to create a strike force to make these things extremely good at writing articles, and then the media would be instantly AGI pilled. The fact that AI is bad at writing has like really disproportionate impact on what the chattering class thinks about AI capabilities. So there's an argument that they should just like improve them on that axis for the sake of the public even understanding. Because right now the chattering class is like, this stuff is shit. Like it can't do anything.

[Jordan]

Sounds right.

[Jasmine]

Yeah, but I also like having a job and so, you know, I don't

[Nathan Lambert]

mind you're so employable. We've already discussed this at length on the podcast.

[Jasmine]

Yeah, but I like my job being writing.

[Nathan Lambert]

That's the thing. People will read your writing anyways. You just have to do your own voiceovers. No one's gonna read my posts. They're all gonna listen to me doing voiceovers. And I don't like it, but I know.

[Jasmine]

I thought about doing a voiceover for the graduation one because it's like, it's like commencement speech themed and so maybe I should do one. But it seems so cringe, like I just like, like then it really. You're like, trying too hard to give a speech. And it wasn't written as a speech, it was written as a blog post. And so I don't know.

[Jordan]

But yeah, I think you should have, I think you should have done the whole thing like, like, like stand up and wear a gown,

[Nathan Lambert]

post it on YouTube.

[Jasmine]

God, that'd be so mortifying.

[Nathan Lambert]

You can do it right now. We could boo you.

[Jasmine]

No, no, No.

[Jordan]

All right, let's end it there. Okay, thanks, everyone.

[Jasmine]

Good.

[Nathan Lambert]

Model talk, guys.

[Jordan]

Model talk. We'll get more fights with guests next time.