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

**Jason Calacanis** (0:00)
Qwen, can't tell the difference for 99% of my work. Somebody just moved a hundred million dollar 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 Neo cloud.

**Alex Cheema** (0:19)
Like no company wants to be beholden to OpenAI.

**Eiso Kant** (0:23)
Do we want that intelligence to all go to three, four, five companies in the world, or do we want it to go to a hundred? If I would have picked up a book three years ago about 2035, I would have titled it a dystopian sci-fi novel.

**Jason Calacanis** (0:34)
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.

**Lon** (0:47)
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. 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 Eiso Kant. Thank you so much for being here, Eiso.

**Eiso Kant** (1:12)
Pleasure to have me. It looks like we're going to have a great conversation today with everyone here.

**Lon** (1:16)
It is a pleasure to have you. Up next, from EXO Labs, he's the co-founder. They're making software that turns Macs and other computers and devices into a shared brain for running large AI models. Give it up for our old pal, Alex Cheema.

**Alex Cheema** (1:30)
Yeah. Thanks for having me. Excited to chat.

**Lon** (1:33)
What a pleasure. Finally, from Qloo, he was on one of our very first episodes when it was just Oliver trying things out. He's the CEO and founder of Qloo. 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** (1:51)
Great to be here again. I'm glad I've made it past pilot season.

**Lon** (1:55)
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 Qwen series. This is Qwen3.8-Max, 2.4 trillion parameters. It outranks KimiK3 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 Qwen3.8-Max versus, of course, Anthropic's Fable 5, OpenAI's GPT-56 all other comparable models, and even KimiK3.
But weights are coming out for that next week. And in related news, you can also see on the far left of that chart, DeepSeek's V4 Flash, even cheaper to run. That's about 100x cheaper than Fable 5 So Eiso, 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? Is it as day-to-day as it sounds if you're following the news?

**Eiso Kant** (3:14)
I would say, actually, yes. 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.

**Lon** (3:46)
So, Alex, 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 using a frontier model like a Claude or like a GPT 5.6?