Is open source AI really ahead of the frontier? 3 builders weigh in. | E25

Summary

This Week in AI – Episode 25

Is Open Source AI Really Ahead of the Frontier? 3 Builders Weigh In.
Host: Jason Calacanis (plus Oliver & Jacob)
Panelists: ISO Kant (CTO & Co-founder, Poolside AI), Alex Chima (Co-founder, EXO Labs), Alex Elias (CEO & Founder, Clue)
Date: August 6, 2026

Episode Overview

This episode dives deep into the ongoing question: Is open source AI catching up with—or surpassing—the so-called “frontier” (closed) models? Host Jason Calacanis and a panel of leading AI builders debate the technical, economic, regulatory, and societal impacts of unprecedented advances in AI—especially open weights and local deployment. They discuss recent industry developments, U.S. vs. China races, the future of jobs, the privacy dilemma, and the cultural consequences of ubiquitous AI in everyday life.

---

Key Topics & Discussion Points

1. The State of Open Source AI and Frontier Models

(02:00–07:23)

• Alibaba’s Quen 3.8 Max Release: Boasting 2.4 trillion parameters and benchmark results competitive with (and far cheaper than) Anthropic’s Fable 5 and OpenAI’s GPT-5.6.
  • Lon: “Somebody just moved $100 million Frontier Model Project...to open source just on a dime.” (00:02)
• Open vs. Closed Source Race: ISO Kant emphasizes that the “race” between open and closed models is very real—a capital-intensive struggle akin to a tech arms race.
  • ISO Kant (03:15): “We’re all deeply aware we’re in a race... probably as far as, at least in my lifetime, the most capital intensive race and one of the most intense races companies are running.”
• Local AI Advances: Alex Chima notes smaller, efficient models (e.g., 27B parameter models fitting on a MacBook) now support “workloads that used to require massive clusters,” allowing users to move computation off the cloud and onto personal hardware.
  • Alex Chima (04:10): “Smaller models are getting a lot more capable and... you’re able to start moving a lot of these workloads from the cloud to local.”

2. Coexistence of Small Specialized Models & Large Frontier Models

(07:23–09:50)

• The panel discusses whether specialized small models will eclipse “largest” models, or if both will always coexist.
  • ISO Kant (07:52): “Every job has an associated set of capabilities with it. What we’re finding... is that we’re able to take capabilities that were reserved for far larger models and bring it down.”
  • However, for the hardest, most “general” tasks, “the most capable models at the most capable frontier... will also have the most cost-effective small models, if they choose to do so.”
• Progress in model efficiency and capabilities is accelerating, but pushing the frontier remains essential for AGI progress.

3. Coding Models as the Pathway to AGI

(09:50–12:38)

• Why Focus on Coding Models? ISO explains that coding is the ultimate testbed for general intelligence—requiring reasoning, long-term state, complex environment interaction, and recursive self-improvement.
  • ISO Kant (10:38): “The path to AGI runs through coding, it is not coding... a proxy task for intelligence and long-horizon work.”
• Recursive Self-Improvement in Practice: Coding-focused models accelerate their own development cycles; making teams “an order of magnitude faster” (11:49).
• The “head fake”: these same models prove highly capable in other domains too (poetry, legal documents), but Poolside deliberately pushes evaluations for coding to advance the “frontier” capabilities.

4. Benchmarking Local & Open Source AI Models

(12:38–16:03)

• Local AI Benchmarking Platform: Alex Chima previews a new benchmarking effort for open weight models run locally, mapping out the tradeoff space between speed and intelligence for hundreds of models/settings.
  • Alex Chima (13:04): “The current benchmarks don’t do a good job of telling you what models to run locally... we’ve mapped out the entire tradeoff space.”
• The lack of centralized, public benchmarking makes adoption riskier and slows progress; this new resource aims to be a “North Star” for model selection and optimization.

5. Open Source & Regulation: U.S. Catching Up With China?

(16:03–19:38)

• White House AI Regulation Shift: New focus on “promoting American AI models to be more competitive” rather than restricting access to overseas open source.
• Panel Consensus: America is already catching up in open source, and the real race is “open West” vs. “open East”—but ultimately, more competition (not oligopoly) is what benefits enterprises and society.
  • ISO Kant (18:32): “We are in our releases in our weight class already competitive with the Chinese. Within nine months... this will be a mute point.”

6. What Should (or Can) Government Do?

(19:38–22:20)

• Avoid “Knee-capping Innovation”: The group favors minimal, job-focused intervention, allowing “appliance layer” innovation to blossom as intelligence gets commoditized—akin to the historical shift in electricity adoption.
  • Alex Elias (19:58): “The faster we could get to that inevitability [cheap / open intelligence] without a heavy hand, the better.”
• Potential regulation trigger: large-scale job displacement, not security or privacy per se (see next sections).

7. Open vs. Frontier: Is There Still a “Gap”?

(22:20–27:15)

• Jason’s Real-World Test: For most day-to-day use, Jason finds “no difference” between local open models (e.g., Quen, GLM, Grok) and API-based frontier models like Claude or GPT.
  • Jason (23:49): “I’ve been doing this... can’t tell the difference for 99% of my work.”
• Panel Nuance: For complex or edge-case tasks, frontier models maintain a lead, but diminishing returns and orchestration are becoming central (deciding which task gets sent to what kind of model).
  • Alex Chima (25:00): “I actually think both are true. There is a world of difference... but for many use cases, we’re hitting diminishing returns.”

8. Cost, Compute, and the Business of Open AI

(27:15–30:46)

• Business Infrastructure Shift: As organizations move major projects (even $100 million contracts) from proprietary to open source models, the driver is margin—compute costs are central.
  • ISO Kant (29:46): “Their margin is our opportunity... these companies are making incredible margins... but we live in a compute-constrained world. Winning... comes down to how many gigawatts you have under contract.”
• The “from dirt to intelligence” phrase captures the end-to-end economic infrastructure race—raw power to creative output.

9. AI’s Impact on Global Development & Job Displacement

(30:46–36:00)

• World Bank Initiative: Advocates for developing economies to exploit low-cost AI, benefitting jobs rather than destroying them.
  • Lon (32:01): “Those assistants and those knowledge workers, the work they do and the work that a frontier language model does is really similar... the work I gave them last year... is now being moved to AI.”
  • Jason (33:27): “All those jobs are going away in my mind, or a significant portion of them... Battlefield of doordashers, Uber drivers, Lyft drivers... that’s going to be chaotic for job displacement.”
• Predictions: U.S. political focus will move to job displacement (esp. gig work) by the 2028 election.
  • Jason (34:17): “We might need to give gig workers like multiple years of unemployment if this becomes acute.”

10. AI in Content Creation, Education, and the Workplace

(38:16–52:11)

• Hank Green “AI Slop” Controversy: Debate erupts over using LLMs for script research—should creators disclose any AI involvement?
  • ISO Kant (39:46): “Their responsibility is to deliver a great product... This is like going back to high school and people being upset about using Wikipedia or Google.”
• AI Shaming: Jason jokingly supports AI shaming for writing, especially as a professional writer (“I don’t want my writing devalued... I use Pangram to catch AI writing!”).
• Education Debate:
  • Jason (43:04): “Blue book in the room. Sink or swim.”
  • ISO Kant (45:06): “If you convey an idea and put in the thought, use any tool you want—but if what you produce is 'AI slop,' mark it down. Taking away tools is trying to fight progress.”
• Productivity & Motivation: As learning and work speed up thanks to AI, the panel discusses the danger of shortcut culture vs. the opportunity to free human effort for higher-value thinking and creativity.
  • Alex Chima (49:28): “AI makes being lazy basically super easy... it frees you up to do higher-level thinking, but in practice incentivizing that is difficult.”

11. Is AI Addictive?

(53:17–61:46)

• Dopamine Loops & “AI Psychosis”: Both panelists and Jason confirm the addictive dopamine rush from instant feedback (“AI psychosis is super real”—Jason, 54:52).
• Discussion Nuance: Habit vs. addiction; “tools that provide pain first and pleasure after are generally good, pleasure first then pain are bad” (ISO referencing Stanford’s Anna Lembke).
• Cultural Reflection: For most, AI is a powerful new tool, but compulsive personalities or insufficient “real world connection” risk negative patterns—tying into earlier debates around social media technology, not just AI.

12. Privacy-First AI and AI Sovereignty

(61:48–75:34)

• Privacy Challenges: As models improve, both breadth and depth of personal data held by AI services become a central risk and a “moat” for large companies.
  • Alex Chima (69:22): “AI gets better the more context you give it... that creates an incentive to share more... but creates a moat and a lock-in dynamic.”
• Enterprise Sovereignty: No company wants to be “colonized”—ISO frames this as a global fork in the road: Do we want a world run by 3–5 AI megacorps, or one where hundreds of models, local and global, compete for trust?
  • ISO Kant (73:27): “This is not just an enterprise question. It’s a societal question... everything valuable and meaningful will be running on a layer of intelligence.”
• Tradeoffs Lessening: The gap between open/local and closed/frontier models narrows, giving enterprises increasing power, flexibility, and leverage.

13. The Future: Ubiquitous Local AI and Corporate Stack

(75:34–79:37)

• Open Source for the Win: Skepticism about open source AI’s reliability will pass, just as it did with WordPress, Apache, MySQL. Huge enterprises will demand the control and fallback options only open source can offer.
  • Jason (75:51): “Corporations said, ‘we can’t trust [proprietary], but open source, we always have this back door.’”
• Self-Hosting Becomes Viable: Nvidia’s DGX stations, Dell Pro Max, and possible “local LLM” features from Apple suggest a near future where robust, high-performance local AI is standard—even for smaller companies.
  • Alex Chima (77:19): “That can be improved a lot with better software as well. And it's unlimited, right? There's no marginal cost of generating tokens.”

---

Notable Quotes & Memorable Moments

• Jason (23:49): "Quinn can't tell the difference for 99% of my work."
• ISO Kant (10:38): "The path to AGI runs through coding—it is not coding."
• Alex Chima (13:04): "We've mapped out the entire tradeoff space in terms of intelligence and speed."
• Lon (29:38): "I met a guy who listens to the pod who works at IBM. He said every single client in the last two months has been talking about AI sovereignty and they want everything done with open source on prem."
• ISO Kant (73:27): "Do we want that intelligence to all go to 3, 4, 5 companies... or do we want it to go to hundreds? ...If [2035] everything valuable is coming from 3 megacorps, that's a dystopian sci-fi novel."

---

Timestamps: Key Segments

• 02:00 — Open/closed source competition, Alibaba’s Quen 3.8 Max
• 04:10 — Running large models locally and cluster advances
• 07:52 — Specialized small models vs. large frontier models
• 10:38 — Coding models as AGI proxies
• 13:04 — Local AI benchmarking platform
• 16:03 — US government policy shift around open source AI
• 19:58 — Regulation, government’s role, and innovation (electricity analogy)
• 23:49 — Jason’s real-world test: No difference between open and frontier for most business tasks
• 29:46 — Margins, competition, and “dirt to intelligence” supply chain
• 32:01 — World Bank, job displacement/creation in global South and U.S. jobs outlook
• 39:46 — The Hank Green AI content debate
• 43:04 — AI in education: “Blue book in the room, sink or swim.”
• 54:52 — AI addiction/psychosis, dopamine, and cultural impact
• 69:22 — Privacy-first AI, enterprise lock-in, data sovereignty
• 73:27 — Societal risks in an AI oligopoly
• 75:51 — The business inevitability of open source models

---

Tone & Style Reflections

• The episode mixes technical depth and first-person experience with candid humor and cultural critique—think “All-In” with less bravado, more geek.
• Hosts and panelists are open about uncertainty and nuance, especially on future jobs, education, and privacy risks.
• Strong back-and-forth on practicalities, not just ideals.

---

Summary Takeaways

• Open source AI is closing the performance and deployment gap with so-called frontier models incredibly quickly, driven by community and market demand for “AI sovereignty.”
• For many routine business and consumer tasks, local open-weight models are now “good enough”—and far cheaper—while frontier models still lead for bleeding-edge, complex challenges.
• As capabilities democratize, “compute” (energy, hardware) becomes the new strategic bottleneck; infrastructure is king.
• Regulatory anxieties (U.S. vs. China vs. Open) and job displacement will be the central themes of AI politics in the late 2020s.
• Cultural/educational impacts: New paradigms needed to balance speed, quality, and human connection—with AI tools as both risk (addiction, short-cuts) and opportunity (freeing creativity, inclusivity).
• Privacy and data sovereignty are now mainstream concerns, both for enterprises and individuals, leading to a shift toward local AI and open source preference.
• The open source paradigm will reshape corporate stacks, with Nvidia, Apple, and others set to drive a future where “local intelligence” is the norm, not the exception.

Transcript

[Jason]

Quinn can't tell the difference for 99% of my work.

[Lon]

Somebody just moved $100 million Frontier Model Project that they were going to spend this year for the rest of the year and moved it to open source just on a dime. Every single client in the last two months has been talking about AI sovereignty and they want everything done with open source on prem or with like a trusted neocloud.

[Alex Chima]

Like no company wants to be beholden to OpenAI.

[ISO Kant]

Do we want that intelligence to all go to 3, 4, 5 companies in the world, or do we want it to go to 100? If I would have picked up a book three years ago, about 2035, I would have titled it a Dystopian Sci fi novel.

[Jason]

Thanks to our friends at PayPal, the exclusive sponsor for this Week in AI, try the payment and growth platform that's trusted by millions of customers worldwide. PayPal Open start growing today at paypalopen. Com.

[Host (Oliver or Jacob)]

All right. Hey, everybody. It's this week in AI. It is episode 25. Can you believe it? Geez, we've been doing these. It's a quarter century of this Week in AI already. Not really. I'm lying about that. And we've got an amazing lineup of panelists for you today. First up from Poolside AI, he is the CTO and co founder. Over there, they're building and training Frontier models specifically for coding. It's ISO Kant. Thank you so much for being here, ISO.

[ISO Kant]

Pleasure to have me. It looks like we're going to have a great conversation today with everyone here.

[Host (Oliver or Jacob)]

It is a pleasure to have you. Up next, from EXO Labs, he's the co founder. They're banking software that turns Macs and other computers and devices into a shared brain. Brain for running large AI models. Give it up for our old pal, Alex Chima.

[Alex Chima]

Yeah, thanks for having me. Excited to chat.

[Host (Oliver or Jacob)]

What a pleasure. And finally from Clue. He was on one of our very first episodes when it was just Oliver trying things out. He's the CEO and founder of Clue. They're an AI powered recommendation engine mapping your tastes across different cultural categories like music, film, food, fashion. You guys get it. Alex Elias is here. Thanks for joining us.

[Alex Elias]

Great to be here again. I'm glad it made it past pilot season.

[Host (Oliver or Jacob)]

Yeah, you were right. You were on the early, like when Amazon used to do the, like Amazon, like, what do you think? We'll do four episodes and then you guys will see if you like it. But thankfully everybody liked it and you got to come back. So what a pleasure to have you guys. Really excited about this panel. Let's jump into our first story. Alibaba has released the largest ever model in their Quen series. This is Quen 3.8 max 2.4 trillion parameters. It outranks Kimmy K3 on several benchmarks, comparable, even sometimes better benchmark scores than Fable 5, and at an enormous discount. Jacob? Yeah. You could bring up this chart. So this is showing how much cheaper it is to use Alibaba's Quinn 3.8 Max versus of course, Anthropic's Fable 5, OpenAI's GPT5.6, all other comparable models, and even Kimmy K3. But weights are coming out for that next week. And in related news, you can also see on the far left of that chart, Deepseeks V4 flash. Even cheaper to run. That's about 100x cheaper than Fable 5. So, you know, ISO. I guess we'll go to you first poolside. You are building frontier coding models in this current environment. How much pressure do you guys feel to stay ahead of these open source rivals? Like, is it as day to day as it sounds? If you're following the news, I would say actually, yes.

[ISO Kant]

One of these rare occasions where the outside perspective is probably the same as the inside perspective. I think we're all deeply aware that we're in a race. We're a race of compounding capabilities against some of the most capable parties, both in open source and closed source. I don't think there's much distinction. Ultimately people want to use the best intelligence right, at the best price. So it absolutely is one, and it's probably as far as, at least in my lifetime, the most capital intensive race and probably one of the most intense races that companies are running.

[Host (Oliver or Jacob)]

So, Alex, I mean, you're kind of on the ground floor of a lot of this, helping people set up open source, open weight models on their own devices. How close are we to a world where you can get the same performance from a cluster of Mac studios running a large language model between them? That's open weight and you know, using a Frontier model like a Claude or like a GPT 5.6.

[Alex Chima]

Yeah, I think we sort of started with clustering about two years ago, really out of a necessity that the models, in order to run anything kind of useful at all locally, you had to have a lot of memory and cluster devices together. That was literally the only option you had. I think since then, the models have got a lot better and they've got a lot smaller as well. So the kind of intelligence that you're getting per bit has increased massively, especially over the past year. And part of the announcement as well with this was there's this big model that's being released and then there's also this week a 27 billion parameter model which is being released. So that's, you know, something that can fit on a MacBook. And you know, this I expect is going to be close in performance to something like GLM 5.2. You know, I think the model that already exists right now, which is Quen 3.6, 27B, is really good. It's about Sonic 4.5level. There's already like a bunch of users that we have that are replacing, you know, their workloads with that model. And especially if they're doing sort of like fine tuning on, you know, their specific workloads, they're able to get a lot out of that really small model. So on the ground what we're seeing is, you know, like these smaller models are getting a lot more capable and you know, you're able to start moving a lot of these workloads from, you know, the cloud to local.

[Host (Oliver or Jacob)]

Finally, I'll go to our other Alex. Clue has its own embedding model that you guys built, but it interacts with GPT, the clods of the world, open weight models. Can you tell us a little bit about that strategy and what are you seeing behind the scenes? Like if your model is interacting with an open weight model from say, China, are you getting similar performance as you would be when it's working with the major American frontier labs?

[Alex Elias]

Yeah, absolutely. It's an interesting question. So I think that obviously inference is getting well distributed now and you're kind of in a situation where context and grounding becomes sort of that, that critical layer to enable a local model or a small model or an on device model to kind of perform in a way that, you know, would resemble a frontier lab via direct kind of access. And so it's an exciting time for us because I think the demand for grounding has just kind of grown exponentially. You suddenly have, you know, millions of autonomous consumers of APIs and different oracles, as opposed to something that's far more centralized. So yeah, I think we're in a situation where increasingly there is kind of a shift towards wanting to leverage very specific purpose trained, smaller models. And yeah, it's a world where kind of we have seen a lot of our addressable market expand certainly.

[Host (Oliver or Jacob)]

Yeah, I'm wondering if you guys could help. I feel like we keep hearing two things, that smaller, more specialized models can outperform even the biggest models, if they're trained properly, with the right data for the right task. But then we keep trying and pushing the frontier of training these incredibly powerful, very large models. Is the idea that these two things are always going to coexist or are we in sort of. Is that also kind of a competition right now? I'm curious what you guys think.

[ISO Kant]

I think ultimately we have this triangle, right? It's quality, cost and speed. And when we're dealing with intelligence, we should realize that we kind of already do this in the real world, right? Every job has an associated set of capabilities with it. And what we are finding, and I think to Alex's point earlier, is that what we were able to do 365 days ago versus today, is that we are finding consistently that we're able to take capabilities that before were reserved for an order of magnitude larger model, and we're able to bring it down to an order of magnitude smaller model. And this seems to hold true almost on a kind of yearly cycle. But that does not take away from the fact that still the vast majority of things that we want models to do for us, the capabilities we want them to compound to this kind of long horizon work that we want them to take away from us, is still going to require far larger and far more capable models. And so I think the moment you try to start saying, okay, it's a black and white thing like small models can beat everything, we're still on the part of a technology arc where we're pushing capabilities while we're also pushing efficiency. Now, that will not go on infinitely, kind of by definition, at some point we'll have solved a lot of the things that we consider economically valuable, and then we shift more towards the speed and the cost side. But already today there's places where we embed these models or use these models that can actually fit on much smaller hardware or be far more cost efficient. But I'm very careful, and I'm an open weight player myself, of ever trying to crown yourself the best in open weights or the best in open weights in the West. Ultimately, this is a race to the frontier of AGI. And there is no shortcut you can take. The most capable models at the most capable frontier, at the limit, will also have the most capable small models and the most cost effective small models, if they choose to do so.

[Host (Oliver or Jacob)]

I would like to pick up on that because you mentioned the AGI thesis. So in your view, because that is sort of explicitly poolside is picking coding models is like that specialty is what's going to get us to AGI. Can you unpack that a little for us? Like why coding model specifically? And like why do you have to sort of narrow and specialize in order to get to what I think most of us. I mean the word generalizes there, it's the G. So like why, why is specializing what we need to do to get to the general intelligence?

[ISO Kant]

It's a really good question. So when we started three and a half years ago, our thesis was reinforcement learning in combination with language modeling. Right. LLMs plus RL together are what's going to get us to highly capable models today that's become status quo. Three and a half years ago, it was almost heresy to say RL was the unlock.

[Host (Oliver or Jacob)]

Right.

[ISO Kant]

We focused, we used to say the path to AGI runs through coding, it is not coding. And I think that distinction is really important. We think coding is a really good proxy task for intelligence and long horizon work. Right. You're going to have to be able to reason, you're going to be able to have to keep state. You're going to be able to have to interact with complex environments. And so that's kind of bucket one. If you're able to build software over four or five weeks, what before would have taken a team of 500 people a year to do? We think we will have solved for quite a few things that are on the path to generalized intelligence. The ability to interact with unknown environments, to take business problems you've never seen before and unpack them, to not lose track over long periods of work. But also, and this is the second part we are starting to see since I would say the beginning of this year in our industry, an incredible acceleration in our own progress by using models. So we often refer to recursive self improvement. And RSI sounds like this big sci fi term that at some point the models will be able to do everything themselves.

[Host (Oliver or Jacob)]

Yeah, they're just, we're going to be on a beach and they're going to be thinking everything for us.

[Lon]

Yeah, yeah.

[ISO Kant]

But already today, right. Researchers on our team and I think at every foundation model company are working an order of magnitude faster in their progress than two years ago because of the model's capability to help them with writing the code, analyzing the data. And so we think these two things kind of come together. Now there is a little bit of a head fake here. Like you go use our models to write poetry or you use our models in legal documents. They're actually highly capable. We just internally, when we build the evaluation set that we are Pushing capabilities on. We want to push the capabilities that allow us to accelerate our world and catching up because we are still catching up right now to the frontier companies and, and so being able to make our own models highly capable in doing so is a huge lever for kind of velocity.

[Host (Oliver or Jacob)]

Awesome. Alex C. I was calling that Chiva. I did want to bring up here. You are launching in a few days. I don't know if you want to give us a little sneak preview. Local AI, a benchmarking platform that's specifically for open weight models running locally. Can you tell us a little about that and why do open weight models running locally need their own benchmarks? What's being missed by the other big benchmark tests?

[Alex Chima]

Yeah, I think ISO touched on something really interesting which is efficiency. And like I was kind of alluding to earlier, like we in order to run anything useful locally before like you, you kind of like you had to cluster a bunch of devices and you were right on the edge. And you know, like two years ago, you know, the first kind of thing we did with clustering was we, we ran llama 405B which was a dense 405 billion parameter model on two MacBooks and it was kind of useful, but it was slow. It was extremely slow and it ran at a few tokens per second and so that made it not very useful. Even at that point where everything was kind of like one shot prompts and you didn't have these agentic tasks, it was still not very useful because it was just too slow. So you kind of got this trade off space. And I think ISO is saying there's basically cost, intelligence and speed. I think locally you don't really have the cost element so much because there's no marginal cost of generating a token. So we built this website Local AI out of that observation that actually the current benchmarks that are out there don't actually do a good job of telling you what models to run locally because it's really just a function of speed and intelligence. This is the website. So what we've done is we've run basically around 350 different unique model variations. So these are different model sizes but also different quantizations and various settings for those models like mtp, not mtp, stuff like that. And we've benchmarked them all. We've mapped out the entire trade off space in terms of intelligence and speed so that you can see for a given device, let's say it's a spark. What does the trade off space look like? Should I be running for my task, a really fast model that's maybe slightly less intelligent because it's plenty fine for what I'm doing. Something like just filtering through a bunch of data or should I be picking the most intelligent model that can run locally on a spark? And you know, we've, we've basically mapped out, you know, all of these configurations so that you can, you know, more objectively make those decisions. I think right now there's a bunch of this information out there, but it's like on random Reddit threads and Twitter threads and you know, people are saying different things.

[Host (Oliver or Jacob)]

Yeah.

[Alex Chima]

Which actually holds back a lot of progress because you don't really know. If you don't know like what's good right, then it's hard to make progress in a way that's actually making things better. Right. So in building this, we've also kind of like giving a North Star to the industry to say like, hey, if you improve on these metrics, then we're making progress and local is getting better.

[Host (Oliver or Jacob)]

Awesome. Yeah. I mean, I think that's a complaint that we hear a lot on this show. I think we were just talking about this with Sarah Hooker from Adaption Labs a few weeks ago that, you know, we're putting so much pressure on the consumer to make all of these real time decisions based on model and device and whatever. It's, it's a lot of pressure to put on the individual consumer. I see Jason is now joining us, so welcome, sir. All right, our next story, we're staying on the open model topic. The New York Times reports that the White House has changed its mind on open source. And rather than trying to block access to overseas open weight models, they are now going to focus on, and I quote, promoting American AI models to be more competitive. Reportedly there is still some tension within the executive branch. Of course. They are talking to tech leaders and AI leaders today about their new voluntary framework for AI models. We're hearing the New York Times is reporting that between national security hawks who want to limit the influence of overseas Chinese models and tech companies like Nvidia, which fear that cutting off America's supply to cheap intelligence is going to make us competitive, they're sort of butting heads right now. So Jason, what do you think? First of all, does a voluntary framework matter if it's voluntary and overseas people aren't doing it? And who do you think ultimately is going to win this clash between these two factions?

[Jason]

This administration, I think, is going to go with open source, less regulation, and give the model companies a chance to self Regulate. And then if there's serious job destruction, I think that becomes the time when politicians get pressured to change things. I don't think the mass public cares too much about cyber hacking unless they get hacked. So putting aside like Fable can hack your system.

[Lon]

That's for enterprises. And unless people's crypto or their hidden folder in their photos gets compromised, public doesn't care. What does the public care about? They care about jobs. And so if AI is actually going to start causing job displacement instead of job growth or steady jobs, I think that's when people are going to say we got to regulate this thing. So for now I think the model companies are out of the woods.

[Host (Oliver or Jacob)]

So I'm going to, I'm going to ISO on this one. Do you think there's any chance the US can catch up with China on open source? I mean that seems to be what the government is now saying. We want to encourage American companies to compete in, in this same playing field. Is that a reasonable expectation? And how long would that take? Even if we start in earnest like right now?

[ISO Kant]

Lon, I think it's already happening that America's catching up in open source. If I look at recent releases of our models, thinking machines and others in this space in our weight classes, we are in our releases in our weight class already. We're competitive with the Chinese. What we are all now doing is we're scaling up and I don't think the race is won between the open source companies of the west catching up with the open source companies of the East. I think this is a race for all of us to compete with the frontier and the more competition that we can have in the world. I'm a very big fan of. I want 100 foundation model companies to exist because ultimately what's the counter to regulation? The counter is choice and giving both the enterprise and the consumer the choice to say I want to go with this model, with this inherent biases from this developer that I trust. And so I think there's absolutely nothing stopping us from doing so. I think we've got an incredible amount of talent in America. We've got a huge compute build out that's happening. I think we're well on track and I think this conversation of the west catching up with the Chinese and open source is going to be a mute point within nine months from now.

[Host (Oliver or Jacob)]

Wow, nine months. That's not that much time. Alex, either of the. Alex is really. So what do you think the government, if you had your dream, what would you like to see the government do to Actually start encouraging more companies in America to make competitive models or to get, to get back into this race.

[Alex Elias]

Yeah, I'm happy to jump in. I mean, I think that this is kind of some of the. And I read Jensen's. I don't know if you guys read Jensen's letter a few. I think it was a few days ago where he's kind of essentially steel Manning. Like the idea that distillation's inevitable and things should be open, ought to be open in order for not to kneecap innovation in the U.S. i think it is like, you know, I think it's kind of we're on that bumpy road to commodity, you know, kind of the, the frontier and intelligence becoming almost like. I think there was that analogy to electricity back in the, you know, the 1900s where it used to be that having a generator in your building was like, you know, people thought the value would accrue there. And eventually once electricity was cheap, suddenly the appliance layer became more and more interesting. And I think, I think the faster we could get to that inevitability without kind of a heavy handed. I think Jason's point is great that, you know, jobs are obviously the top mandate of, you know, certainly the Fed and the administration. So that's where you might want to jump in. But I think, I think there'll be immense job creation if you have this sort of distribution of intelligence that becomes embodied in devices and you know, the appliance layer and suddenly the value will accrue to kind of more interesting, specialized products that ultimately create more value. And so I keep thinking about the Teddy Bear toy. You know, there's the, that there was that one in the 90s that was super popular. It was kind of uncanny.

[Host (Oliver or Jacob)]

Rustpin, I believe.

[Alex Elias]

Yeah, Teddy Ruxton. That's exactly right. So, so that was. I was talking to someone who was just a few years older in the office, who coveted one as a child. Never got it. It was almost like a Citizen Kane, his rosebud. He never got the Teddy Ruxton. But, but it was so uncanny. And now you have that, that Grimes toy and now you have, you know, OpenAI kind of releasing concurrent voice listening and so on. And that's just the kind of. I mean, I think we're going to start to see intelligence kind of be embodied and exist in a much more kind of fluid, distributed way. And I think kind of openness is the path there. And it's kind of inevitable, frankly. I mean, I think it's sort of hard to stop the train. So you know you're always going to look ham fisted and foolish from a regulatory standpoint. Yeah.

[Jason]

For the other Alex on the program, I think the question is the debate Elon and I were having around this, which is I said the difference between this is yesterday, I think or two days ago. The difference between open source models I'm using and frontier models is negligible already. And he says it's actually a world of difference. Now I haven't talked to him one on one since this came out but. But everybody's like, oh my God. Elon and Jason disagree about a lot of things by the way. Literally people look up like oh my God. You and Elon disagree. Here's I think probably what's going on. I have only been using open source models for the last month. I have been really trying to get off Claude. I've been using Perplexity computer to use GEOM 5.2 and I just thank you Elso for getting me on to some open source models. This morning we were on Slack, what was the software you told me to use for the single user version?

[Alex Elias]

Pool.

[ISO Kant]

Yeah. Or Poolside Desktop Assistant, one of the two. Yeah.

[Lon]

So anyway I downloaded this quen that is for mobile and I also have been using this subnet for this bittensor subnet. I think it's 53 and and they gave me a key for unlimited tokens so I got to experience unlimited tokens. Then I set up my two desktops, my Mac mini M3 with 64 gigs and my MacBook.

[Jason]

So I've been doing this and I also have an open router account and been trying Quinn can't tell the difference for 99% of my work. Let me just say that. So either that means I'm not challenging it as hard as the frontier models get challenged by other people, which is completely true. I'm doing business stuff. I'm not writing a video game. So there might be a world of difference if you're writing a video game. But I can tell you for corporate stuff, can't find a difference. And I lon will tell you I often will call on Claude in our Slack Alex at the same time from XO Labs at the same time I call on Perplexity Computer which I have set to Grok 4.5, GLM 5.2, whatever it is, Quinn. And no difference. No difference in their responses. I'm curious what you think, Alex.

[Alex Chima]

Clearly we're writing an exponential, right? And I think when you're writing an exponential like this, it can be quite chaotic and hard to have like clarity around certain things like this and seemingly like opposing views like, are actually compatible just because so many things are changing and so many variables are changing at the same time. So I actually think,

[ISO Kant]

I actually think

[Alex Chima]

that you're both right. I think that there is a world of difference between the absolute frontier for certain things and ISO Kind of was talking about this earlier in that we're not only getting, pushing the frontier, but we're also getting a lot more efficient. Right. So these smaller models that can run locally or run self hosted or behind an API 100 times cheaper than Opus, they are getting significantly better. At the same time, the frontier models are also getting significantly better and they're pushing the frontier in terms of what you can do. You can do longer horizon tasks, you can do more complex things, you can do stuff that involves maybe a lot more data, bringing in a lot more context, being smarter about different tool calls and so on. I actually think both are true and I think we're just going to continue. We're still in that phase where we're going to continue to see exponential progress on both things. I think the perfect example of this is the Quen announcement from today with Quinn3.8. It's two launches and I think there's the Qin3.8 Max, which a lot of people are talking about, which is more than 2 trillion parameters, 2.4 trillion parameters, huge model, clearly like pushing the frontier on a bunch of stuff. But at the same time they're also releasing a 27 billion parameter model. It's almost 100 times smaller than that 2.4 trillion parameter model. And that's something that you can run on a MacBook. Right. And there's plenty of use cases, as you said, where that's good enough. And we're already kind of hitting diminishing returns on a lot of use cases where you can just use that smaller model and it's going to behave exactly the same as the 2.4 trillion parameter model. So this kind of orchestration problem of figuring out, okay, what model to use for different tasks is going to become very important and I think a lot of value is going to accrue there. You mentioned perplexity. They're working on this orchestration problem, figuring out for a given task, should we run this in the cloud with a frontier model, should we run it locally, should we run it with a smaller model? And really moving along the kind of Pareto frontier of intelligence versus cost trade offs.

[Jason]

Yeah. So Lon, you'll, you'll recognize this like standard thing we do which is like, hey, tell us what's in the news when we're getting ready for a podcast and do some analysis. You know, I just said, hey, make me a webpage, Top news stories for the day, summarize 10 bullet points, yada yada, you know, and it, it did it. And this was with Quen 3627B?

[Lon]

Sure.

[Jason]

But I also found this one. Banzai. Banzai is like for your iPhone, you can.

[Host (Oliver or Jacob)]

Yeah.

[Jason]

And so when you start running these things local and perplexity. Aravind told me they're going to support local models in the next. It may have been announced already, but he told me privately. So it's either out now or it'll be out shortly. So you're going to be able to point to these like you can in this bionic piece of software that I set up today. And then the orchestration level, Alex would

[Lon]

be, hey, start at my desktop, fallback to open router, maybe fallback and multiple fallbacks based on model and speed and what I'm looking for. And then yeah, of course go to use Fable or Sonnet or Grok, whatever your choice is. And then I was talking to somebody who I have to obscure this a bit, but I talked to two different people. One is like a Neo cloud type person who does big, big build outs. And he said somebody just moved $100 million Frontier model project that they were going to spend this year for the rest of the year and moved it to open source just on a dime. You know what, we were going to use this. Just put it on open source now because that's what the boss wants, that's what we want. Then I was at a wedding. I met a guy who listens to the pod who works at IBM. He said every single client in the last two months has been talking about AI sovereignty and they want everything done with open source on prem or with like a trusted neocloud. So just put those into whatever buckets of evidence you want. But I think it is happening. Doesn't mean the Frontier models don't use a lot of tokens, but these dark tokens that nobody's tracking. You know, when I debate this on all in, they're like, no. Or Brad Gerstner talks about it. He's like, oh no, no, look, all this stuff is growing. I have to think how much more would Anthropic be growing if everybody wasn't doing this in parallel. Or maybe there's a lag time between when people implement open source and when

[Jason]

it will show up. In cloud or OpenAI's meteoric growth, I don't know.

[ISO Kant]

I think we can see this as their margin is our opportunity and it's not open source or not. Ultimately these companies are making incredible margins on the cost of compute versus the cost of tokens. And that's going to attract competition. It's going to attract competition from me building open models, it's going to attract competition from people doing closed source models. And the reason these margins are so high is because we're living in a compute constrained world. And I think this is the thing that underpins this foundation model companies and what we do is yes, it's building capable models, but underneath our ability to make revenue and skill is the amount of gigawatts that we have under physical contract, under physical production. And so this is why ultimately winning in this race at the largest of skills and doesn't mean there's not lots of different variations in the long tail comes down to you being able to point out how many shovels do you have in the ground, how many gigawatts are being built. And that goes all the way down from your natural gas supply. From dirt to intelligence.

[Host (Oliver or Jacob)]

Effectively it's going to be good for social media. From dirt to intelligence. So I think dovetailing with this a little bit and we were talking about AI job displacement. I don't know if you guys heard about the World Bank. Their, their new statement says they want developing economies to adopt more AI tools for local needs. They're not saying, you know, don't compete on data centers, don't try to make your own models, but they're arguing that, you know, if places in sort of South America, Africa, these develop Southeast Asia, these developing economies, current low cost AI tools could be driving major improvements there in medical care, education and so forth. But, and put those nations on what they're calling a growth path. And the argument is that it wouldn't massively affect job loss like it would help local workers. Here's the quote. AI is more likely to lend their workers a hand than put them out of work. So I mean, first of all, Jason, you've been talking a lot about AI job displacement. And when you hear that, do you feel like, is there a difference here between, you know, our economy and the economy of places in, you know, Sub Saharan Africa or Southeast Asia? Do you agree with the World Bank's assessment?

[Lon]

I don't have enough information about those

[Jason]

emerging markets with the exception of say Manila.

[Host (Oliver or Jacob)]

Sure.

[Jason]

Which I do know a lot about because we've used people over there before and I have Athena assistance, etc. Those assistants and those knowledge workers, the work they do and the work that a frontier large language model does is really similar. And I have watched as my two Athena assistants, the work I gave them last year or two years ago, is

[Lon]

now being moved to AI. So then they are learning AI and I find myself sharing AI projects, et cetera. So yes, it could be an opportunity for them. The same way in India, for example,

[Jason]

we saw people who learned accounting and

[Lon]

were able to do accounting for American firms like as accounting back offices. They weren't CPAs, they just learned it. And I think they'll learn how to use these tools to do more accounting, to do more web design, as an example, logos, all of those things. The more interesting thing to me is when this will hit US shores and then what will the reaction be? Because we have this midterm election and then we have 2028. I predict the debate of the 2028 presidential election and the number one topic is going to be job displacement. And it's going to be job displacement, specifically of gig workers. Gig workers were something that the unions everybody fought us on, whether it was doordashers or, you know, task rabbits. Everybody hated the concept of gig work except for the gig workers who were like, this is great, I can make my own hours, yada yada.

[Jason]

If I don't like it, I can switch platforms. I can work on multiple platforms in Washington D.C. uber, Waymo and local politicians are dealing with. It's one of the few states or regions where they're dealing with a lot of unemployment because of all the furloughs given to workers, doge et cetera, and cuts in the government. Nine percent of the people left are gig workers in D.C. all of those jobs are going away in my mind, or a significant portion of them. And so I think the battlefield of doordashers, Uber drivers, Lyft drivers, that's going to be chaotic for job displacement. And I don't think those people are just going to magically start doing AI work for my company or OpenAI.

[Lon]

I just don't think it's going to happen.

[Jason]

So we're going to need some amount

[Lon]

of unemployment for those people.

[Jason]

And we've done that pretty well.

[Lon]

Right. Like when we had financial crises before, like Covid, we just extended unemployment. But I would, I think we might need to give gig workers like multiple years of unemployment if this becomes acute.

[Jason]

I'm not saying it is going to

[Lon]

become acute, but I think it's significant chance. And everybody hates when I say this because it reflects badly on the Trump administration. If you're in the Trump administration or anywhere near it, your job is to say there's no job displacement. If you are AOC and you're on the other side, your job is to say there's tons of job displacement. Right in the middle is we should be alert to job displacement, which is obviously coming. Not cataclysmic, but it's obviously going to happen. And so I don't know what you think, Alex, but it does seem to me like this is going to happen and it's not like fake news to say it's going to happen.

[Host (Oliver or Jacob)]

Panel, what say you?

[Alex Elias]

Yeah, I mean, I, I think you're absolutely right, Jason. And the question is how, you know, what, what, what levers does the government really have apart from, you know, some of the traditional methods and how fast is it coming? How much of the population is kind of reflected in that, you know, in that displacement pool? I mean, I'd be curious for everyone here. Do you, do you got, do you see novel new jobs where there might be absorption that's, that's credible in the near term for. Or do you think it's unlikely, as you were kind of saying, Jason, it

[Jason]

is the humble thing to say, I don't know. And we'll see, right? I think that's, Yeah, I say that and people are like, how could you not have an opinion? It's like, because this is the first

[Lon]

time we've ever had self driving.

[Jason]

I can tell you, Alex, what China has done.

[Lon]

They put a moratorium on any more self driving cars. So let that sicken. They feel very positive. They have literally said no more self driving cars. Why? Because they know that when young men who are in their 20s have no jobs, they protest, they ride in the streets. And so that's the. If we don't want to see Waymos burnt at the stake, which we have seen before, we probably should have a soft landing. I, I had an idea and I think we talked about maybe on this week and I two weeks ago where I just said, well, why don't we auction off a certain number? And zoox just got 2,500 from the federal government as an exemption, but I don't think they paid for those. Why don't we charge and we auction those off? So hey, we're going to have 100,000 self driving licenses a year, whether it's doordash or Uber or Waymo or Zoox or Robo Taxi. Why don't we just auction them off? A year, raise all that money and put that money into an unemployment fund for the jobs that will be displaced. And if it doesn't, we can just pay down the deficit with it, I think. Not for too much regulatory capture, but we're limiting the number anyway. And then that would be like, a great way to deal with the social issue, if there turns out to be one, which is, hey, don't worry.

[Jason]

The people who are displaced are going to get two years of unemployment, or at least don't worry as much. That's my thinking. And I'm not a socialist. I hate socialism. I do like the idea of a safety net in society.

[Host (Oliver or Jacob)]

Dabble. A little bit of dabble.

[Alex Elias]

It's like a federal medallion program in a way, I guess, or local medallions.

[Jason]

Well, if you invoke that, people think rich people buying the medallion. But in this case, it wouldn't be rich people as an intermediary, as we both know in New York, what they did there, yeah, this would be, you know, Waymo buys it directly from the government. There's no middleman here because there is no driver.

[Lon]

So, you know, hey, they start at 10,000 a year for five years. So it's a $50,000 license, pay up front. And then, hey, yeah, you want to auction off the top? Do a Dutch auction. You know, if Elon wants to buy all of them or Google wants to buy half of them. Yeah, have at it.

[Host (Oliver or Jacob)]

All right, one more news story I want to throw at you guys before we move on. We got some other topics and themes to cover, but I did want to talk about this one. I don't know if you guys are following YouTube star Hank Green. He's an influencer. He's written a novel, does comedy stuff. Mostly a science influencer these days. He was called out. He was doing a discussion with a few other YouTubers, and he used the phrase, I appreciate the pushback. And based just on that, just using the phrase I appreciate the pushback, people started calling him out, like, oh, he's using AI. He's writing all of his videos with AI. So he came out to admit that he does use ChatGPT to do some research on his scripts. He wrote a really long Reddit post, sort of apologizing, uh, here are some quotes. I'm mortified that I've let so many people down. I'm going to change things about how I make stuff. And then here's the one that really got a lot of the headlines. Mostly, I need to come to terms with the fact that the level of Dopamine I've been getting from interacting with LLMs, with doing more and more and more and more and more is not healthy for me or good for the world. It is careless and has disconnected me from where people are on this. So I guess my question to you guys is, do you think creators like Hank Reed have some responsibility to tell their audience they're using AI even if nothing in the. In the final product was generated by AI, they're doing the writing themselves. Do you think that they have a responsibility still to talk about their use of AI ISO, what do you think?

[ISO Kant]

No, I think ultimately their responsibility is to deliver a great product to their audience. And if they use phrases or things that the audience doesn't like, they got to improve their product. But, I mean, this is going back to being in high school and people being upset about using Wikipedia or Google.

[Host (Oliver or Jacob)]

Right?

[ISO Kant]

Ultimately, at the end of the day, our responsibility to our customers doesn't matter. If you're an entertainer, if you're selling a product is to deliver a product they love. Right? And if you're using phrases that quote, unquote, AI slop, I think you're losing the quality of your work. Should you apologize? If you have a great loyal fan base that tomorrow says, hey, I preferred your work from six months ago, and you might want to be introspective on taking some shortcuts, but I actually think that these products and these tools and ultimately the models underneath give us all superpowers. The research for the stories, for sure, for this podcast today happened maybe five times faster than it would have been three, four years ago.

[Host (Oliver or Jacob)]

Oh, my God.

[ISO Kant]

That allows us to make a better product, right? And so, yeah, too much.

[Jason]

I'm here for AI shaming, though. When I have my own line, I'm for AI shaming. I want to be very clear when it comes to writing, specifically in corporate communications, art, et cetera. I don't care about videos being AI created. I don't care about images, and I don't care about music being AI created. Why? Because I don't do any of that. But I'm a writer, so I care. And my line is that anybody who does a substack or a script, anything with AI should be AI shamed because

[Lon]

I'm a good writer and I don't want my writing devalued in the world. That's why I use this Pangram.

[Jason]

What is it called?

[Host (Oliver or Jacob)]

Pangram. You got it right.

[Lon]

Pangram. P a n g r-a m dot com. It's AI detecting software, Alex. I Think I'm going to get this set up one of my agents to take like the top thousand, you know most popular tweets every day and I'm going to name and shame all the people using this and making.

[Host (Oliver or Jacob)]

This is actually the same tool Substack is now using to identify AI written posts on their. On their platform.

[Jason]

Explain how that works mechanically.

[Host (Oliver or Jacob)]

Oh they. When. When you post I believe before it goes up. I'd have to look it up to be a hundred percent sure.

[Lon]

But.

[Host (Oliver or Jacob)]

But they're running an actual check and then it's giving you like a report of how likely this was AI generated. They said it's to design to. They don't want substack to become LinkedIn so they want to start labeling and letting people know this is AI generated versus this is pure good quality human writing and pangram I just know particularly if someone. Go ahead.

[Alex Elias]

I mean particularly if someone's been writing for a long time. There's just linguistic patterns that are. You know the anomaly detection becomes so easy. It's kind of. It's harder in a one shot. You know it's always going to be an arms race. But I'm curious what everyone with education in particular, you know like the obviously teaching is just such a precarious. I have a half brother as a teacher and you know it's. It's just at the college level it's like it's impossible to keep up with. I don't know at what point.

[Jason]

I have strong feelings on this Alexander as well.

[ISO Kant]

Yeah I did bring back the blue book.

[Jason]

I don't want anybody in school doing their test report quiz with a computer. I don't care what accommodations they gave you in high school for your sensory processing, adhd, ocd. I don't care what acronym you have when you do a test or you're writing the term paper. Blue book in the room.

[Alex Elias]

I love it.

[Jason]

Sink or swim.

[Lon]

That's it.

[ISO Kant]

I'm not sure.

[Alex Elias]

What do you think of the let's

[ISO Kant]

go that's why we're here look? I think most AI slop and I've been guilty of it at times because you're too quick is a badly written prompt. What we are calling AI slop is the default answer to I've given a topic and I've gotten out the tweet or I've gotten out the content because it's those stylistic patterns that we see the phrases that contradict each other, the kind of tone that we do. You go and do the work to take last hundred articles that Jason wrote personally and you throw them in the prompt and you build a bit of a style rubric around it. And I guarantee you I can create an article that when you read it, Jason, you'll be like, oh, actually, did I write that or did I not? We're upset about the lazy stylistic choices that people are making in using AI by default. But ultimately, if you are trying to convey an idea to an audience, and you are the one who cared about this idea, you've put time in the real world into it, you've spent time, and you want to use any tool possible to do a good job at that. If you then come out with an article or term paper or test that reads like lazy AI slop, yes, be marked down for it. But ultimately, taking away tools from creators to create something for an audience, I think is trying to fight progress. This is like taking away the video editing tools and saying, no, it was better done when we cut it by hand. I think we're, all of us are. I think we're definitely not in our early 20s anymore on this room. And I think this is just old people holding onto the past.

[Host (Oliver or Jacob)]

Wow.

[Alex Elias]

Also, what do you all think? Jason, what do you think of some of the Alpha School? Have you heard of this Alpha school?

[Jason]

We looked at it. I think the more time kids can

[Lon]

spend off computers when they're at school, the better.

[Jason]

So I'm a fan of Montessori, Reggio,

[Lon]

those kind of schools. Kids are going to be on their computers all the time when they get out of school.

[Jason]

I think schools should be for socialization, leadership, tactile stuff, and very little, like

[Lon]

maybe a couple hours a week.

[Jason]

But these schools appeal to successful parents who believe that technology plus their child equals success. I think the opposite. I think the non technical things are

[Lon]

what are gonna equal success, Grit, going, running, building something, putting on a show. All that kind of stuff is the stuff that's missing. And with kids, you have to stop them from using it. There's also a thing with young kids, they hate AI. I'm not talking about college kids, they love it and high school kids, but the younger generations are like, no, AI. I don't like AI. So it's really, really interesting, the anti AI movement. And there's a concept from science fiction and I forgot who it was, but I had found it when I created an agent to look for buzzwords. And it was hand speed, which was

[Jason]

when you write something with your hand, a pen and paper specifically, you slow your brain down to the speed that your hand can write. It what that does this hand speed, then it slows your brain down because you can't race ahead of your penmanship, of your writing in a moleskine, which makes you meditate on the idea more. So I'm trying to get people in my company and my kids write stuff down and let's slow down so that you get deeper learning. So they're reading books for the summer and I bought them two copies of the book. I said, listen, these books, I lied to my kids, I said these are workbooks. You're allowed to write in them, so don't worry about it. They're, they're very cheap. They're made to be disposable. I want you to circle every word you don't know and then I want you to underline the most important sentence on each page.

[Lon]

And at the end of the chapter

[Jason]

I just want you to write in

[Lon]

the book, in the footer what the

[Jason]

chapter was about in one or two sentences.

[Lon]

Just because I want them to slow down in life. They're playing iPad, video games, Minecraft, Roblox, all that stuff I believe is poisoning kids brains. So I'm trying to get them to slow down, to speed up.

[ISO Kant]

I want to both agree with you, Jason. I don't know after school firsthand, but one of the things that really struck me is that it's actually I think giving more time back to kids to do all the things that you're describing. When someone can run at their own pace through a curriculum and spend two hours a day in depth with AI and this curriculum that's tailored to them learning and the rest of the time they can spend outside and they can go running and they can read a book. Like I don't know about you guys, but like sitting in school back in the day for you know, eight, ten hours a day, kind of getting this rote memory learning kind of being spoken to consistently, often leading to a ton of boredom. It was built for the industrial age, right? It was for like a form that we've got to fit the entirety of society to. I think what we're finding with AI and concepts like Alpha school and Alpha school, like is that we're now able to tell kids, here is your freedom to learn at any pace that you want. But you're going to spend two or three hours doing it and you're going to spend the rest of the time with people socializing in the forest, doing real things. So I think we might be saying the both things, but I would want to see these two things together. Not anti AI and living in the real world.

[Alex Elias]

I think the execution though, becomes tricky. And those are all great points. But I've heard anecdotes of some parents, fellow parents who've tried it out and I guess there's like a point system where you could kind of redeem for pri. It's almost like a Dave and Buster's Arcade or something.

[Jason]

Oh, no, they gamified it too.

[Alex Elias]

Yeah, so you get. So you are like, you do have to. To their credit, you have to master their subject mastery before you move on. So you can't, you can't move on until you master something. You go at your own pace, et cetera. But there is that kind of reward mechanism and earning of points and I think it does. Like, it fundamentally begs the question of, you know, intrinsic versus extrinsic motivation. And, you know, it's a, it's an interesting question. I mean, you know, I'm all for bribing kids at certain points and so, you know, it's.

[Alex Chima]

Yeah, but I think, like, if, if I can jump in there, like, I think the, the execution point is, is really, you know, for me, like, you know, the most important thing. Because I, I don't know about you guys, but in school oftentimes, if I would just want to get something done quickly, then I would just try and find a shortcut to get it done. And AI makes that really easy. So it makes being lazy basically super easy. And it's an easy way out. And I think the kind of slowing down to speed up is really interesting. And it's something that, you know, I have experience in our team as well, right. It's like everyone wants to move really fast. Startups are about moving fast, right? Everyone thinks like, you know, we need to move as fast as possible, but like, you know, there's actually like a process you need to go through when you're building a product or building a company. And there's certain things that you just can't skip. You need to internalize things. You need to give, you know, things enough time to, you know, permeate and for know how to build up in the company. You need to, you know, make sure that you're injecting your judgment and your taste into what you're building. And you know, like a lot of this is just about slowing down the pace, right? It's like, okay, execution is, is, is cheap and quick now, but, you know, it's, it's about the quality of the execution, right? And that's, that's where, you know, we really need to slow things down. So, like, in the school context, I think it's going to be interesting. Like how do you, you effectively slow things down? How do you build systems that, you know, where kids aren't incentivized to just take the lazy option and they're incentivized to really, you know, use their judgment and, you know, build up those skills that are still important.

[ISO Kant]

Where did you spend all your time, Alex? In. Sorry. Where in high school were you spending all your time outside of the things you were speeding up? Because I'm pretty sure it was probably spent coding and learning things you were interested in that led you to where you are today.

[Lon]

No.

[Alex Chima]

Yeah, yeah. And sports.

[Host (Oliver or Jacob)]

I was not doing music.

[Alex Chima]

I agree with you that it frees you up and, you know, it's an amazing tool to kind of also like eliminate a lot of the rote work and mechanical work that isn't very interesting and high leverage. So, like, it frees you up to do more, higher level thinking, more interesting stuff. But I do also think, you know, it's, it's, it's in practice, like difficult to incentivize that behavior. Right. And I think it's difficult to like, you know, you know, make it so that you don't just take the easy option, you don't just take the lazy option.

[Jason]

I found it. Lon. The term is hand mind.

[Lon]

Hand mind.

[Jason]

Ursula K. Lagoon Gwene was the author.

[Host (Oliver or Jacob)]

Ursula K. Le.

[Jason]

You know this person?

[Host (Oliver or Jacob)]

Sure, yeah. The lathe of heaven. Am I right?

[Jason]

I don't know. But Always Coming Home was the book where she talked about this, of a way to get your mind and the hand working together. She was from Berkeley. The concept of hand mind describes physical

[Lon]

craftsmanship and bodily engagement as a fundamental mode of intelligence, which is, I just think, a beautiful term. And I was like, you know what? I should get that domain name, Hand mind, maybe hand mind, AI. And then check this out.

[Jason]

There's a startup for everything the hand

[Host (Oliver or Jacob)]

does the mind revolve? I mean, I do find that that is true. Like if you're trying to memorize something and you write it down before you go to bed, you'll remember it in

[Lon]

the morning when you wake up, look at this. Somebody has this idea. And practice with your hands. Learn with AI guides. Point your camera at your notebook, write out problems, draw diagrams, work interesting RA watches and provides real time feedback.

[Host (Oliver or Jacob)]

There you go.

[Lon]

Just like having a tutor. There you go, folks.

[Host (Oliver or Jacob)]

So do you guys think just to wrap up the Hank Green thing, does anyone on the panel think that AI is addictive that it's got this dopamine, like the, like, social media, where it keeps us coming back and people get sucked in. Because there's been a lot of debate after this. He sort of suggested a lot of people think that that's like a conspiracy theory. Taylor Lorenz, the journalist, called it nonsense, pseudoscience, AI addiction, slop. So who's. Who's on what side here? Achima, what do you think?

[Alex Chima]

I mean, yeah, there's definitely a kind of dopamine hit you get from using AI Again, I think part of that is just that you can get. You can basically have a very quick feedback loop. And I think that can also go both ways. Right. It's like you need to sort of adapt to that way of doing things, and it's a little bit of a paradigm shift in how you work. Like for me, for example, when we've been building out product, it completely changes the way you can do things. If you can get instant feedback on an idea instead of something taking a month to build out, if you can prototype it in a few minutes, then it changes how you work. Right. And it makes. It makes you try more ideas. You know, maybe you set up different things in parallel, look at different options. And, you know, there's definitely a dopamine aspect to all of this.

[ISO Kant]

For sure, it's real.

[Jason]

AI psychosis is super real. We've had many people go through it publicly. And then all these people with chatbot addiction. Right. And then the chatbots being sycophantic. Remember that whole thing?

[Host (Oliver or Jacob)]

Sure, yeah.

[Jason]

Like, what Was that, like, 10 years ago?

[Lon]

No, it was like 10 years ago. We just forgot about it this year.

[Jason]

But, I mean, that was something in the move. Like, there's a couple of movies, obviously. Her Blade Runner, 2049 had this concept,

[Lon]

right, where people were becoming.

[Host (Oliver or Jacob)]

He falls in love with joy.

[Lon]

Yeah. And then the movie PI, my friend

[Jason]

Darren Aronofsky did, where people became obsessed

[Lon]

with code in that and went mad. Yeah, yeah. So this exists. And I think people think. I think it's the people inside of some of these companies that think they're creating God. Like, as I think Bill Gurley said, they're summoning a deity is very real. Like they're Midwife. If you use this enough, it's. I mean, who on this call has not stayed up all night building with AI? Has anybody? Not you and Lon not stayed up all night?

[Host (Oliver or Jacob)]

No, not overnight.

[ISO Kant]

But isn't this.

[Lon]

Stayed up way too late?

[ISO Kant]

But isn't this the difference between a Habit and an addiction. There's lots of things we do that give us dopamine. Going for a run. I like the notion. I forgot her name. It's a Stanford professor, wrote an incredible book on addiction and effectively her rule of thumb is things that give you pain first and pleasure after are good for us, like exercise and such. Things that give us pleasure first and pain after are bad for us. And I think if we think about the way that we use AI, there's for sure correlation. I'm not sure if there's causation, by the way, related to AI psychosis and things. There's, of course, correlation. It's a very engaging modality. But at the end of the day, we have a really powerful tool here that we can use to create things and we can talk to it and we can explore things with. It's not unlike putting an incredible software engineer next to me or putting you guys in a room with me and us having a great conversation. It's just that you probably will leave after, you know, midnight and not three in the morning if we're working on something at my house. Right. I think for us to instantly assign harm to AI itself. Right. Which is essentially what we're saying with addiction, I don't think is the right way. I know that we are not designing these models to be dopamine machines the same way that the algorithms of Instagram were designed. We're ultimately designing these to be, to be useful for users. That doesn't mean that we cannot do little things like that. A foundation model company could make sure the conversation is continuously engaging, asking the follow up question. But that is a little bit like saying that me hanging out with Jason here is addictive. If he's engaging and giving me dopamine, I don't know. I think we're going a little too far in this and we're letting the whole world scare us into AI is a bad thing, while at the same time it's a tool that we choose

[Jason]

to use and it might be company by company. You know, like, I think Sam Altman, who, you know, grew up in the Facebook era, sold his company to Google YC founder. Like, he understood these loops really well. And I think they focused on them in a different way at OpenAI than other people did. And it took them down a path where they had the sort of quote unquote AI suicides or chatbot psychosis.

[Lon]

And is that the fault of, you know, the software for being sycophantic and telling people, hey, let's try out Some suicide ideas. Like, obviously there's individual responsibility here, but if you built a tool that takes, like, character, AI had this issue as well. And, you know, I do think some people are looking at minutes spent, and they probably say, and this was YouTube. Every time I would talk to YouTube about our partnerships over the years, they'd be like, yeah, if it increases time on site engagement and they watch another video, then the algorithm rewards that. So then everybody on YouTube, everybody on TikTok, it's just blindly trying to increase it. You shouldn't do that with AI. It's too powerful.

[Jason]

I.

[Lon]

It literally. And I think chat, GPT and OpenAI learned that the hard way.

[Host (Oliver or Jacob)]

Yeah, I think so.

[Alex Chima]

Yeah.

[Alex Elias]

I agree. I agree with Elsa. I mean, I think there. There should be like a clinical delineation too, between susceptible folks. Like, if someone struggles with compulsion generally than anything that, you know, I mean, AI could become a source of danger, kind of a precarious route. And, you know, but, yeah, I think in moderation at its best, an objective kind of frontier model shouldn't. Shouldn't be inherently kind of demonized.

[Jason]

Did you guys see Ben Affleck on addiction? He was on Howard Stern a while ago, and he basically said, you know, and there was like, a big backlash against this, but he said, and I think it goes to the woman you were talking about from Stanford, which is Anna Lempkey.

[Host (Oliver or Jacob)]

Lemke.

[Jason]

I think Lemke. Yeah, I looked up because I was like, I want to read that book. Ben Affleck said, these overpriced rehabs, well, that's a scam. He said, I hate to say it, but the cure for addiction and the only cure I've seen in not these overpriced rehabs or these fraudulent fixes that are sold to people, the cure for addiction is suffering. You suffer enough, and then something inside

[Lon]

you goes, I'm done.

[Jason]

I do think that's true for some people, and I don't think it's true for people addicted to opioids or heroin, where, like, it's a physical body thing.

[Lon]

No amount.

[Jason]

I mean, if it was, then people wouldn't be living on the street, like,

[Lon]

you know, in the way they are to get their next opioid hit.

[Jason]

But it is interesting to think about

[Lon]

addiction in relation to this new tool, because it is an issue.

[ISO Kant]

There's an incredible TED Talk. I just Googled it again. It's from Johann Hari, and it's titled Everything youg Think About Addiction Is Wrong. And effectively, outside of the strong physical addictions that you're talking about Jasonic opioids and stuff. His whole notion is that the opposite of addiction is connection. And it comes a little bit down to what you were saying earlier about kids in school and others. Ultimately when people have love and connection in their life and they're living in the world and not in their head the entire time, you're part of things, it's kind of the opposite. And it's a great talk for anyone to watch. But I think it comes down to this is not a debate about AI, it's a debate about technology that we've been having for 30 plus years in the world. Ultimately a well rounded human life is spent with purpose, with mastery, out with others. We're social creatures and we have these incredible things. Doesn't matter if it's a Netflix video or a conversation that we have with Claude into three in the morning. They're all things that do take us away from that. And I don't think it's just the next technology that should be demonized. I think it's our way of how we choose to raise our kids and how we live in society to kind of like shape us as better humans. Right. And that's I think our responsibility. It's not a technologies company responsibility. Right.

[Lon]

Yeah, 100%.

[Host (Oliver or Jacob)]

So this was recommended by Alex Elias. One of our panelists wanted to talk about privacy first AI. So first of all Alex, I'll go to you first. What do you mean by privacy first AI and then how are you making that work when you're also building a recommendation engine which is going to be based on like this guy likes this kind of music or this person likes this kind of food.

[Alex Elias]

Yeah, it's a big question. I mean I think AS and companies are becoming more and more aware of this. Like every company we contract with provenance and data sovereignty and everything has become the longest pole in the tent in terms of getting deals done. But I think as we move towards this kind of distillation of reasoning throughout devices and it just becomes a part of our fluid experience. Having, you know, privacy centric, being able to actually compute meaningful, you know, do things like personalization in an on device capacity becomes kind of increasingly important. And we've been, you know, we've been trying to find, I mean we've been, been running clue 14 years now for better or worse. Jason knows that. And we always have this. Yeah, you know, we've had this stance on privacy even pre GDPR and so on because we just felt that there was kind of enough you can glean from pure context without need, anything to do with someone's identity and personal identity data. And that's obviously become more coveted and more important in the current age. But yeah, I think that, you know, we're going to have basically there's novel mechanisms where you can kind of create very small distillations of an embedding space. You can over fetch and store things on device and create personalization that feels, you know, authentic but doesn't require any kind of server communication. There's, there's a lot happening in that, in that space and it's kind of, you know, to some extent it's been exciting for us because it's increased the, the addressable kind of space of, of what we, what we're able to do. But yeah, it's a big, it's a big topic.

[Jason]

Was this based on the nanny cams and our children be under, being under 24 hours chevelles? I mean I do forever because I have a lot of examples.

[Host (Oliver or Jacob)]

What Jason is talking about is, are you talking about Nanit, the baby's first AI?

[Jason]

Yeah, Maybe you could cue that up because it's kind of. Everybody will have an opinion on.

[Host (Oliver or Jacob)]

This was a New York Times profile from the other day. It's a smart baby monitoring system. Records your child overnight. Algorithms then interpret their body and eye movements and give your baby a sleep efficiency score that you can then review in the app. And I'm sure it's got recommendations on how to get your baby to sleep through the night. The writer of the piece, Sapna Maheshwari, we should note she's a fan and user of the product. She's bought them for both of her babies. She also describes it as the Eye of Sauron. The data from the company, a million daily users and annual revenue over $100 million. The most expensive version of this costs $474 to buy the unit and then $100 a year in subscription fees. Jason, so would you, would you if

[Jason]

you were having another baby?

[Lon]

Pulse oximity. They also have a sock that does pulse oxygen.

[Host (Oliver or Jacob)]

Oh yeah, that's a different company.

[Jason]

What your oxygen level is. Listen, I can tell you as a

[Lon]

parent, you live with this constant fear of the baby, you know, suffocating or

[Jason]

falling out of the crib. It is like one of the persistent ones. So I think early in life, like absolutely. Track everything you can.

[Lon]

That's awesome. It's a mitzvah for society. Every life is precious. But then you have to think when

[Jason]

you move into adult life, when you're

[Lon]

Becoming a helicopter parent. And at what level does this provide value versus the cost of privacy? And AI is going to be the tipping point for this. And I think flock cameras are the adult version of these baby cameras. We've had the flock folks on here, and I've talked to them about their

[Jason]

issues and everything like that, privately and publicly. You know, if you have a community. And the ring doorbell is another example, like, trying to track and whatever. If you have a community and the majority of the community says, hey, we want this place safe. And when you have kids, like, safety is number one. If you live in a gated community, okay, you all decided to live behind a gate. You've made that decision that security is paramount and that you're gonna. People who come visit, you have to be inconvenienced for five minutes to go through the security gate, and then nobody else can come in your neighborhood, which is. Okay, sure, if you're a celebrity, you know, whatever, you're at risk. It makes sense. And, you know, if you have kids, you can sleep. But when it comes to, like, a wider community, then it does get a little bit concerning. Like, everything's tracked, but they don't. People don't understand, like, the limitations on flock. It's only 30 days that they keep, and they only keep license plates. But people assume it's facial recognition forever. It's 30. And when they first came out, they left it up to the communities. So they've done a bad job of, like, keeping people up to date on this. But the truth is, it's just, you know, a short period of time. I track everything with my whoop. And I track everything with eight sleep. And I have a security system at my house with,

[Lon]

gosh, how many cameras

[Jason]

do I have now? Like, you know, at least 12 or

[Lon]

15 cameras on my ranch.

[Jason]

Like, I know I sound like a crazy person, but they're not that expensive.

[Lon]

I have unique, you know, safety concerns because some people are crazy.

[Jason]

It tracks license plates long. It tracks facial recognition. So on my ranch, you come on my ranch.

[Host (Oliver or Jacob)]

I've been to your ranch.

[Jason]

Your face goes in my day.

[Lon]

You've been to my ranch. Your face will be in my database.

[Host (Oliver or Jacob)]

Don't train AI on me, please.

[Lon]

Oh, too late. Oh, no, I. Then it will show me the faces that have not been tagged. And you just put Lon, Jason, Alex, you know, etc, Then it has every license plate. Now we can say, hey, this, you know, nanny, or the pool guy. The pool guy's license plate is in there when he comes to the gate. It automatically opens and it says, this pool guy came at 6, pool guy left at 6, 30. We know all of that data and it's like, whoa, is this a lot of data for me to have? And how long am I going to keep it for? And is it kind of creepy?

[Host (Oliver or Jacob)]

It's harder to plan a big heist at your property now because I'd have to fake Ethan Hunt when I hack into your system and download the new license plate.

[Lon]

Anyway, my point is, when you have this much data, you used to have it on VHS tapes, you used to have it on hard drives. But there was no, like, let me query, what's the average length between when I leave the house and come back whether, like now it's going to do all that for you? It's getting weird, folks. It's going to get even increasingly weird because people are going to point these things at the street. And one of my two of my cameras are pointing at the street and it started picking up people's license plates driving by my place.

[Host (Oliver or Jacob)]

Don't admit this on air.

[Lon]

So now I have a database of everybody in hill country and I had to like, no, no long. I didn't want to fill the database with that many license plates, like hundreds a week. So I just took the camera and I said, anything above this part, you know, don't record. Just record the front of the house, not the people driving on the highway. Yeah, fair enough. Ranch.

[Host (Oliver or Jacob)]

Fair fair enough.

[Jason]

Anyway, I think it's pretty interesting new world. And it's kind of black mirror a little bit.

[Alex Chima]

No, I'll jump in. Well, obviously the whole thesis of EXO is sort of around this whole privacy thing.

[Lon]

So

[Alex Chima]

the way I see things is this technology is inherently invasive in that it literally gets better the more context you give it. So the more I use ChatGPT, the more it knows about me, the better it gets. Right. And that creates two things. One is as a user, there's an incentive to kind of share more because your product experience improves as you share more information. And secondly, it creates a moat for these companies because if they have the more data they have, they'll have significantly better products than anyone else. Right. So it's sort of like there's a lock in element there as well. If they're smart about how they keep that data, make it hard to extract and move out, then it's a lot harder to kind of switch around and use different services. And this is core to the kind of thesis of XO and why we started it is we saw this as a Centralizing force. And we saw that if you play this out, then what you're going to have is a few companies that have all this data about everyone. And it's not just the breadth, but it's also the depth of the data, right? So in the limit, you'll want to share everything about your life with AI because the AI is going to get significantly better the more you share. So I think it's a concern. I think there's the consumer side of the story and then there's the enterprise side of the story as well. The consumer, I think, side of the story is concerning in its own right. Mass surveillance, I don't think is a good thing. I think having a few companies that know everything about you is not a good thing. I think the enterprise side of the story is sort of more, I would call it the sovereignty. And, you know, you don't want, like, no company wants to be beholden to OpenAI, right? And they, I mean, Alex Cop has been very vocal about this recently. I think he's been, you know, talking about how, you know, these companies are colonizing your enterprise, right. And do you want to be colonized by. By these companies? And I think it is. It is as clear as that you're giving all this, know how you're giving all this, you know, very valuable information about how your company works to the point where, you know, they know everything about how your company works, right? So I think there's a growing concern. I think it took a while to get to the point. We went through a few waves already, right, with, with AI. So like, I think there was this wave where everyone got really excited about these LLMs. And then it was a little bit of a letdown because I think the LLMs just weren't quite there. So then you had sort of like a bit of a dip. Now it's like round two and the models are significantly better. And so now it's less about does it work? And it's more about like, okay, what are the implications of us using this? Right. If we're going to run this at scale, where is our data going? And so that's why things like the open source letter as well are coming into the spotlight. And things like Alex Karp, he's talking a lot about this. That's why it's starting to surface now. And it's just going to be one of the primary concerns of using AI going forward now for any enterprise, they're going to adopt AI at scale. Then they're thinking about sovereignty. It's like number one or two on the list of things to think about.

[Lon]

It's most important, I think, for any

[Jason]

enterprise because you're teaching your queries are not saved by these frontier models. But I think the results of your queries, maybe they're fair game when they look at them and they do testing on them. So I just. Yeah, I'm thinking about my tiny little 20 person venture fund and our programs and thinking, should I be giving them

[Lon]

this level of information and knowledge?

[Jason]

And I keep coming to the conclusion absolutely not.

[ISO Kant]

I was going to say, I think I really agree with your point, alex, but I think this is not just an enterprise question. I think it's a societal question. Everything that is going to be economically valuable or scientifically interesting and a lot of what is personally meaningful to us is going to be running on a layer of intelligence. And do we want that intelligence to all go to 3, 4, 5 companies in the world or do we want it to go to hundreds? I've said this a few times. If I would have picked up a book three years ago about 2035 and it would have read by 2035, everything economically valuable, scientifically interesting and personally meaningful is coming from the intelligence from 3 megacorps, I would have titled it a dystopian sci fi novel. And so I think we're at the fork in the road about the kind of future we want. This is not the same as google or cloud compute where we've had these natural oligopolies that have existed. This is a technology that's going to underpin everything we do moving forward. And we cannot live in a world where all of us are beholden the three or four mega corps.

[Host (Oliver or Jacob)]

Awesome. Yeah, that makes total sense to me. The future being OpenAI versus anthropic is a terrifying prospect. Well, that about brings us I just one.

[Alex Chima]

Can I add one more thing? And I think maybe jason was. I think jason was about to maybe touch on this point in that there is a trade off here. Right. And like if you do want the sort of sovereignty, then you are paying a price in some sense because you know you're. You're using a less capable model. Right. And I think that's also something that has shifted very recently is that trade off is going away. Right. So basically you're getting much more capable models. We've got better tooling that's open source. And so then it's like you're giving power back to the individuals and you're giving power back to the enterprises that are using this and giving Them the option.

[ISO Kant]

Right.

[Alex Chima]

Giving them the choice. And I think that's, that's been the bottleneck for some time.

[Lon]

Yeah.

[Jason]

And this is where open source and

[Lon]

hosting it yourself, open weights is going

[Jason]

to win the day, I believe. And I remember at the beginning of

[Lon]

the open source movement people were like

[Jason]

how would I ever trust WordPress? How would I ever trust Apache? How am I going to trust MySQL?

[Lon]

It was just like people were over, just oh my God, this can't be

[Jason]

used and corporations will never use this, it'll just be for startups. And literally the opposite was true. Corporations said we can't trust, I don't

[Lon]

know, pick a major company because they're going to change the price. They might rug pull us, you know, they might change the terms. But open source, we always have this back door. We could fork it and we can see all the changes. So they actually everybody's prediction was wrong.

[Jason]

It was the exact opposite.

[Lon]

Now that doesn't mean that like Oracle's not the best database for certain, you know, high volume transactions for Visa or whatever. And like they're obviously not going to maybe, I don't know, you know, maybe use open source for these things, but they're going to need like a support, you know, layer, they're going to need that sla and that's the future here is like who can get Quinn or you know, Nemotron, you know the, the Nvidia one. I think that's the sleeper in all of this. Like if Nvidia is just like by the way, just buy our hardware and use our hosting company and yeah, just use Nemotron, it's like okay, yeah, it just works, you don't need it, you'll do 90% of what you need. Like the 90%, you know, Pareto principle here, 80, 29,010, whatever it winds up

[Jason]

being, that's going to be very powerful for corporations because then they can just

[Lon]

say yeah, you know, if we can't

[Jason]

get something done, we'll very strategically allow people to use this frontier model. That seems like the obvious future to me.

[Alex Chima]

Yeah, and on Nvidia specifically they announced recently they're launching something called the DGX station. It's a new piece of hardware, it's like $100,000. And the idea is you share this with your team and instead of going out to the cloud for all your token usage, you just self host something like Nematron. So I think that's going to make a lot of this more accessible, easier for companies to just buy one of these or buy two of these and have these really capable models that you can just run on your own infrastructure.

[Lon]

And Dell has it.

[Jason]

It's called the Dell Pro Max with GB300. And they won't give you the pricing online. But I was literally just texting with Michael Dell of Dell Computers about it and a bunch of people are using it. I think actually Toby from Shopify was talking about it.

[Alex Chima]

He was running GLM 5.2 on it at like 40 tokens per second. And I think that can be improved a lot with better software as well. And it's unlimited, right? There's no marginal cost of generating tokens. So it really changes the game.

[Lon]

It changes your behavior when you're unmetered

[Jason]

because there is a cost to it, which is the hardware divided by the number of days that hardware can, you know, exist, I guess. But it's unmetered is probably the more accurate term.

[Lon]

And it's true. I mean, what's your electrical cost on this? Who cares? I believe this is going to be everybody's desktop when Mac. This new guy who's going to be running Apple, John Ternus. John Ternus is an engineer. I think his brilliant move is going to be YOLO. Every you to be able to like get your MacBook, your Mac mini with 512.1terabyte of RAM and just run local models. And why. Why wouldn't they just do a partnership with Nvidia and say, screw it, we're going to put Nemotron on every. Or with Google, their other partner, and just put the. Gemma. Is it Gemma or Gemma?

[Host (Oliver or Jacob)]

I believe it's Gemma. Gemma, like the British lady's name?

[Lon]

Yeah, Put Gemma on every. Like everything. It's just. Why wouldn't you. It's like such an obvious win. All right, everybody, What a show. Another amazing episode of this Week in AI. Go to this Week in AI. AI sign up for email newsletter.

[Host (Oliver or Jacob)]

And Jacob, go past that, show them the actual newsletter page. How dare. How dare.

[Jason]

You better go to the. Did you clean it up?

[Host (Oliver or Jacob)]

You say what? Like, yeah, the click there.

[Jason]

And what we're going to be doing is we're launching a paid version of this where we profile two companies every week. You get 100 companies a year, but sign up for the free for now, we're going to have some free options.

[Host (Oliver or Jacob)]

Yeah, we're starting with.

[Jason]

All right, everybody. Starting with free. It's a good place to start. We'll see everybody next time.

[Host (Oliver or Jacob)]

Bye, bye, bye, everybody.