Ex-Google Insider Reveals The Future Of AI in 2026... Transcript, AI Summary & Key Points

00:00 You said PhD in your pocket. What did you mean by that? >> You'll [music] have a system in your pocket that you won't pay a subscription for. You'll have the smartest system ever built that represents hundreds, thousands of years of human experience and knowledge in your pocket. It's never going to go offline. It's never going to get stupid. It's going to have a better memory of exactly what you like or don't like.

00:19 Every person's going to own their own intelligence. And as long as you charge your phone, you'll have that PhD in your pocket. >> For the AI models coming out of ChatGPT, OpenAI, Anthropic, what does that mean for them financially? >> I think what what we're going to see is a much more diverse explosion of models and applications that people own and run themselves.

00:37 [music] And right now it's really just hackers that are doing this. But in the near future, we're going to see everyone's going to have the ability [music] to have something downloaded that they own and run. You can see Anthropic actually making like this desperate [music] attempt to cover everything. They do design, chat, coding. Now they're doing private equity stuff.

00:53 They're doing marketing. They're doing legal. Can they win that all? There's actually just so much surface area. I find it difficult to believe. >> Apple and Amazon seem to have just gone into a different path. You don't really see, at least here, of a lot of action there. What was your thoughts on that? >> We're super excited about what they're doing.

01:11 And they have been a slow mover and said, "You know, we're not going to build our own model. It's not really what we do." But they see this wave coming too of why shouldn't we have a PhD in your pocket [music] for every Apple user? Why shouldn't one live on your laptop and live on your iPhone and live on your Mac mini, which is already happening, right?

01:27 They have really committed to on-device AI that their customers will own, which really fits into their mantra of privacy and ownership and fully vertical infrastructure. And I [music] think the story by the end of the year will be Apple has emerged as one of the clear AI winners because of that. >> Jack, we are approaching a world where most people no longer need OpenAI-level models for the things they actually do.

01:53 Would you say that's true? >> Yeah, absolutely. It didn't seem like we would get there till the end of 2026, but we are essentially already there. >> What does that mean? Because I think for a lot of our viewers and our listeners, uh majority being founders, investors, engineers, but for the non-technical folks, what does that mean for them? >> Yeah, I think if you look at Anthropic's last few releases, Opus 4 6, Opus 4 7, they're incredible models.

02:19 They're they're really, really powerful and they keep emphasizing these are really effective at the most difficult challenges. They're really great at the most difficult challenges. They're pushing the frontier on there's this extremely difficult thing and we can actually do something that was never been done before. But what's been catching up behind them is the open-source models that are really good at yes, less difficult problems, but that's becoming a bigger and bigger problem space of what they can solve.

02:46 So, kind of big open-source models are already really good at the things that most people use chatbots for today. Definitely holding a conversation, doing research. They're really good at coding now, almost frontier level with Opus and from Claude and Open AI. Most of the valuable work that you do using language models on a day-to-day basis can be done by open models already today.

03:08 And then the thing that's amazing to us, these newest releases of smaller language models, specifically some of the Quen models that have come out recently, are as good as the frontier big models think from from Open AI and and Claude and the big open-source models. Now you have tiny models that can fit on your laptop, potentially even on your phone, that can do that kind of frontier level work.

03:30 We already have that already. They might not be good at the most difficult challenges, but the reality is most of the work you do on a day-to-day basis isn't the most difficult work. >> For someone who's not in technical, who's not in the industry like you and I am, and perhaps is a teacher, what would be a example of an open-source model and one that they could have on their phone or their laptop, and what can they do with it?

03:54 >> Yeah, absolutely. My My sister's a a fifth grade teacher, so I I've talked to her about this a little bit. If you use ChatGPT to Let's say you are a teacher, you're you're create lesson plans, you ask it for questions, help revise, you give feedback to to students or their parents. Any of that work can be done for much cheaper using an open source model.

04:12 And there's really two ways to use it. [music] You know, number one, these all have kind of competing chatbots to OpenAI or Claude. [music] And what's getting better and better is you can download essentially a really big file onto your computer that contains the actual bottle that you can then run and chat with yourself. And then instead of paying a a monthly subscription, you really you own your own intelligence.

04:33 It's never going to go offline. It's never going to get stupider. It's going to have a better memory of exactly what you like or don't like, and it's going to live on your laptop as long as that laptop has power. >> That's incredible. You know, you think about technology and how it's evolved. From the internet, it was very expensive. It was dial-up, and then broadband came along, and now it's getting cheaper and cheaper.

04:52 And you almost internet is becoming an essential asset you need to have. Cloud, you know, with storage was expensive with on-prem, then cloud comes along, and then you look at mobiles, they're getting cheaper and cheaper and cheaper. I don't even upgrade my phone anymore. I think I have an old iPhone. The question is, is that what you are basically stating here, that AI is going to play in a very similar field where maybe free models, completely open source models, will I think you mentioned this to me in our calls,

05:18 you said it the small models can do 95% of economically valuable AI work. >> Yeah, I think that's right. I think we're already there. And the model is probably the most important part of this entire ecosystem, what's actually what powers the intelligence at the other end of the line when you're you're doing a chat or you're doing some sort of work. The small models are really smart, and they're only going to get better from here.

05:39 And the question is, okay, so as the frontier models push out, they continue to do things that no one else can do. They do frontier AI research. They do research in physical sciences or or biology that's never been done. That's going to be extremely valuable for this frontier labs. But the open source models are going to do what most people care about.

05:58 Think software engineers, think teachers, doctors, people who are, you know, working with computers every day, doing research, doing sales. That kind of work the models are already smart enough, ones that can fit on your laptop. And I think, you know, there there's basically two directions then where the frontier labs are going to continue to to push the envelope.

06:16 They're going [music] to do that work that no one's ever done before. What you probably seen in the headlines is they're really starting to build out apps to use their models. So they've said, "Okay, we actually realized that this model revolution is coming. People can own their own intelligence. They won't need to come through us." So what have they done?

06:33 They've built these frankly incredible number of products that you can interact with them. Think Claude Code and Claude Design and Claude Co-work and and Claude Chat. And then you have Codex and and Chat GPT and those things. And the question will be, you know, do you want to pay a monthly subscription to use one of those products or do you want to own your own intelligence on your laptop or wherever it is?

06:54 >> Is it like AI is going to become like electricity? Don't really care where it comes from. It just generates and we use it. >> In some ways, yes. In some ways, no. I think the threshold that we've been going after is like can it do a given task? And once you reach the point of, "Okay, it can do this task." then you start to care about other things.

07:11 How well does it know me? How well can I rely on it as it completes a number of tasks in succession? You know, what's its personality? How quick is it? How much does it cost? Once you pass that accuracy threshold, those other questions come into play a lot more. What we really focus on specifically is is a couple of those. We want your AI to be really fast.

07:30 We want it to be really accurate even as it encounters a lot of data over a really long period of time. That's really what, you know, my company Subconscious does specifically on top of open models. And you know, I've talked to some companies who really think that personality is going to be the differentiator. That think that you know it's ability to understand certain types of data.

07:49 Think a model for doctors versus bio pharma versus software engineers. But they were there will probably be some distinctions between those. >> One of the I guess realizations I've seen in the market now is AI has in many ways removed this protective layer for a lot of industries. Like management consultancies, even doctors, lawyers where they're charging extremely high fees or maybe inadequate in their services, AI is becoming that adequate solution.

08:18 >> We're taking things even further talking about AI becoming effectively free with open source. What does that mean for the professions that have notoriously always been very expensive thinking lawyers, doctors, architects. What what happens there? >> Yeah, it's it's a difficult question to to figure out [music] cuz in any time frame that's reasonable over the next couple years, as smart as the models will be, there's not enough trust built up to put all of your faith into these models.

08:43 Do I really want to trust a model versus talking to an actual doctor? If I'm building a house for the first time, do I really want an AI system to do that instead of actually hiring an architect who's done that tons of times. If I have a very important, you know, I'm on trial or I'm raising money for my company, do I really want to not talk to an actual lawyer at any point?

09:03 Probably not. But they are smart enough to do a lot of that work. We'll probably see that shift but not in the immediate short term. >> I think what I'm noticing at least and I agree with you. I don't think it's something where you're completely not going to use a lawyer or doctor or an architect. But I I feel there's maybe some elements of some industries and sectors where sort of the cost, the unit cost of that service is so high pre AI, now with AI they can really focus on the higher level work where perhaps it's

09:34 worth paying for a lawyer if that makes sense. >> I think so, and I I think like it it makes it much easier on their end to do a lot more of meaningful work, potentially billing less hours. And then on our end, you know, we can go into those conversations with a lawyer, you know, with an architect, with a doctor, and use their time better, also. So, I think it'll come from both sides, but we'll still need human expertise, at least in any time scale that will probably matters for our lives.

09:58 >> I think, you know, the the mobile and the internet was very transformational piece of technology that has transcended everywhere. I think for a lot of people they don't quite realize what cloud does, and unless you know what cloud does because you use it, but, you know, if you're in the industry. But, for mobile and for the internet, you don't have to be in tech to to have used it.

10:14 I think with AI, I just think about how fast it's changing on a weekly or on a bi-weekly basis, the technology is exponential. >> Yeah. >> The question I have for you, Jack, is when you look back at the last 6 months, just in the last 6 months, maybe the last 3 months, has there been a moment for you where you thought, okay, I did not expect AI to be where it is now at this rate?

10:36 >> Yeah. Yeah, absolutely. There was a moment when really the Qwen 3.5 open-source models came out. And and Qwen is a series of open-source models from Alibaba, a Chinese company, where I would say a step-wise change in what open models can actually do. And then a couple weeks later they upgraded and said, one of these we're actually going to upgrade to version 3.6, and that's this version that is a 27 billion parameter model.

11:00 To put into your mind, 27 billion, that's a pretty big model that can fit on your laptop if you use the right techniques. Something that runs in the cloud that you use behind ChatGPT or Claude is trillions of parameters, so so 100 times, you know, bigger, something somewhere in that range. And we have a model that is 100 times smaller and is as good as those that live in the cloud.

11:20 And I've been talking to my my co-founder about this, and we said that'll probably happen sometime, maybe December of this year. That's roughly what the trend line looks like. It happened in March, and I think we're going to keep seeing acceleration that's as quick in that direction. It's It's pretty impossible to not believe that these things are going to get small enough to to really live on your computers, to live on your laptops, potentially even your phones um in the very near future.

11:43 So, we were just surprised about how quickly intelligence got compressed. >> That's impressive, Jack. You used the phrase on our previous call, you said "PhD in your pocket." What did you mean by that? >> Yeah, you think what is really the benefit of >> [music] >> these models getting so small so fast? And it really comes down to you'll have a system in your pocket that you won't pay a subscription for it that as long as you charge your phone, you'll have the smartest system ever built that represents hundreds,

12:10 thousands of years of of human experience and knowledge in your pocket to do whatever you need and whether you need to ask it questions, you know, help plan or coordinate events, do research, maybe you want it to do coding, whatever you want it to do. You'll have that capability. You know, today it's already in your pocket. You can download the ChatGPT app and use it.

12:29 What I'm talking about is something that will really you'll own that someone else can't turn off, that's not going to get stupider because ChatGPT is is has too many people using it and they're going to kick you off for the day. You have something that's that's really yours and lives in your pocket and understands you [music] and as long as you charge your phone, you'll have that, you know, PhD in your pocket.

12:49 >> From a I guess evaluation perspective or the value of these companies, even the ones that are providing the GPUs like Nvidia, maybe less [music] them, but for the AI models coming out of ChatGPT, OpenAI, Anthropic, what does that mean for them financially? >> I honestly think their valuations are justified. I think that the narrative though isn't quite right where there won't be these two dominant forces that that battle for Anthropic wins the enterprise and ChatGPT wins all of the consumer usage.

13:16 I think what we're going to see is a much more diverse explosion of of models and applications that people own and and run themselves. And right now it's really just hackers that are doing this, people who are are tinkering enough and have the the strong enough computers to to get that frontier intelligence on their own devices. But in the near future, we're going to see everyone's going to have the ability to have something downloaded that they own um and run.

13:41 I think it's going to have a lot of implications for what consumer AI looks like, what an individual asking questions, planning their life, asking for advice, doing work, both school work and and kind of business employed work means. But for businesses, I think there'll always be a need for some sort of cloud offering. There's certain things that you want to run for really long periods of time or going to need that kind of frontier power.

14:02 But they might need smaller compute footprints than we think. But at the end of the day, my my overall take is we're just scratching the surface of how useful these systems can be. They're only getting smarter, you know, the trend lines aren't showing any signs of stopping. Personally, it's just made my work life really productive and there's a couple things, you know, here and there that I use in my personal life that I think is is really fun that that AI's brought to the world.

14:27 >> There was a thing that I kept asking about 2 years ago, who's going to own the application layer? Who's going to be the big players in AI? We've got some incumbents here. We've got some companies here that are doing exceptional work like Anthropic, like OpenAI. But in your opinion, who do you think, if any, is the jury still out on who the winners are in AI?

14:44 >> You know, I think it's it's pretty clear that Anthropic, OpenAI, probably Google also, they'll have models that do things that no one else can do. And they have the compute footprint to do that. They have the research scientists in order to to train the models. They're going to push the bounds of what we thought was possible with AI. When it comes to, like I said, 95% of consumer AI use cases, there's going to be this explosion of what what models uh can do.

15:09 And there's just a number of open source models that are out there, so the the cat's out of the bag. And what's powering any given app is is totally up for grabs. Um I think what the surface looks like, like how does someone actually interact with that, is also pretty up for grabs. Uh you could see Anthropic actually making like this desperate attempt to cover everything.

15:28 I mean, they do literally like design, chat, coding. Now they're doing private equity stuff. They're doing like marketing. They're doing legal. Can they win that all? Maybe, but there's actually just so much surface area. I find it difficult to believe. And then when you come down to individuals who are using chatbots, there's some value to having built up a lot of context within a given app that the first one you use for a long period of time, you're probably going to use for a long time.

15:55 But I think we're we're really just starting to see these things like really penetrate into people's lives, and it feels like the playing field is still open. >> Somewhat of a sticky piece of software, but let's say you're using it personally and it's on your phone and as you said, you you can use it personally for a number of different use cases. Yeah.

16:12 You're teaching it constantly who you are, what you like, what you had for dieting or for exercising or health or hobbies. And it's interesting to to think how sticky these technologies are because I know you can migrate it over to a different model, but you kind of become attached to it because it knows you. And it's interesting also that you say Anthropic, OpenAI, and Google because I I recently had uh Andrew Dye, who is a founder of a new company called Alloyent that just raised 50 million.

16:38 And Andrew was in the research lab at Google Brain. He worked with the founders of OpenAI, Anthropic. And the entire topic that we really touched on during the podcast was all about the culture of that particular team and how they became I think at the end I said it was the Google Mafia. Someone said the PayPal Mafia. It's really fascinating to see this new shift happening.

16:58 But if you were to pick a player and I think about this as far as like let's use the cloud era with AWS and GCP and and Microsoft Azure. Where do you think Anthropic is, OpenAI, and Google if you were to sort of look at them in comparison to the cloud? As we are right now, you know, we're talking we're recording this in May 26th, so this will probably go out towards the end of this month, but as of this moment right now, who do you say is the biggest player?

17:24 Who do you say is the biggest out of the three? >> I'll I'll go from first of all, our preferences as a team. We use a lot of the Anthropic models ourselves, like as our models have gotten better. Um we're trying to use more of our own, but but we really like Anthropic and for a number of reasons. I think number one, their models seem to outperform the others.

17:41 Number two, they've really just completely dominated the mindshare among developers. Um everything from Claude code to they developed the MCP, um to they developed all of these tools, uh they developed skills, um different things that developers all over the world use now, whether you're using Anthropic or not. And then they just seems like they continue to dominate the new cycle with with constant releases.

18:03 So, I would say they are likely number one. Number two, I would actually say right now is Google. I'm an ex-Googler myself. I worked at Google for a number of years, um on their on their search team. I just think they've done they've done a really excellent job. Their their Gemini models are really strong. They're very friendly to the working with developers and and new companies.

18:19 Their notebook LM product is is pretty great and they've continued to come across as like a very good, you know, respected player in the space. Um and then third, I'd say is is ChatGPT and and Anthropic and OpenAI. You know, they obviously came in first, um but our take has been that their their models are less reliable for these agentic tasks uh that we're really building our systems around.

18:42 It was just less reliable infrastructure for us and it seems like they are they are dropping in their mindshare. I think we have a a big trial ending in about 2 weeks that's going to really reveal where they sit, so I'll leave it there. We just internally just shifted away from >> OpenAI over to Anthropic and we use both Anthropic and um and Perplexity as well, which we find Perplexity computer is quite good.

19:04 One more question I wanted to ask before we dive into talking about where you are right now, what you're building as a founder of a company. I'd love to tap into that. But the last question I want to really dive into as far and unpack as far as this topic of the big players is there are two other big tech companies that seem to have just lost their way when it comes to this new evolutions, new wave of AI.

19:24 One being and I could be completely wrong. I'm looking at Microsoft and thinking Microsoft made a massive investment to open AI. That was kind of their play. But Apple and Amazon seem to have just gone into a different path. You don't really see or at least hear of a lot of action there. What was your thoughts on that? >> Oh man, I I think that's going to be an outdated take pretty soon.

19:46 Amazon, first of all, they own a lot of the GPUs. They have this, you know, big deals with Anthropic and a lot of teams that we talk to are using Anthropic through their Amazon Bedrock accounts. They're running them on, you know, Amazon GPUs. I think their massive cloud footprint isn't going anywhere. >> Well, that's the thing. I think I think it's a lot of this is around branding and marketing and how what is known in the market, right?

20:11 >> But but they've become, you know, Switzerland in a way because of that and they're just benefiting from you might have the best model in the world, but you got to have the GPUs and the infrastructure to run it. And then on top of that, too, as it gets cheaper and cheaper to run a language model and do this agentic work, the the computes surrounding it gets more important.

20:28 How do you call tools? How do you access data? How do you do the things that AWS does really well? On the other side, I'll say Apple, too. We're we're super excited about what they're doing. Um and they have been a slow mover and said, you know, we're not going to build our own model. It's not really what we do. But they see this wave coming, too, of why shouldn't we have a PhD in your pocket for every Apple user?

20:48 Why shouldn't one live on your laptop and live on your iPhone and live on your Mac mini, which is already happening, right? >> Interesting. >> And they have really committed to on-device AI that their customers will own, which really fits into their mantra of privacy and ownership and [music] fully uh, you know, vertical infrastructure for whatever end experience they're giving to their users.

21:13 And I think the story by the end of the year will be Apple has emerged as one of the clear AI winners because of that. >> Wow. Okay. Well, we'll do a second episode in possibly and talk about that. >> [laughter] >> Yeah. >> Jack, let's move on. This is really fascinating. Let's move on to where you are today. I'd love to get to know your business a little bit more and understand what is it you're building today?

21:33 What is the problem statement? Start off there. What is the problem statement that you're solving? >> We basically ourselves heard a lot of the hype around building AI agents, but really struggled to actually build agents ourselves and we think it was because the tooling that was built around AI was really built for chatbots and not specifically for agentic systems.

21:56 So, what we do as a company is we take open-source models, we retrain them to think in terms of these longer running workloads, and then we run them on top of our own infrastructure that's that's very, very efficient. What that amounts to is we offer a series of language models to customers that are more accurate, especially on long-running complex tasks, and much, much cheaper to run and then can run at really any scale.

22:22 Anything from data center scale all the way down to, you know, the devices that we're talking on right now. So, this wave has been really exciting for us cuz what we can do is we can take a language an open-source language model that's already really great. You know, maybe it can live on your computer and be that replace your chat GPT subscription cuz it's as good a chatbot as anything else in the world, but we can post-train it, run it on top of our own way that we would serve up the model to you in a way that it can

22:47 solve long context problems. Things like, [music] you know, deep research and coding and interacting with your browser. Things that today only the best, most frontier models can do. We allow smaller language models to do that uh any scale. >> It's very fascinating. I'm just wondering when you thought about this being a area of interest, what was that epiphany or a moment where you realized this is an interesting area to go after?

23:12 >> We really started the company around how do we build agents that can reason over long context, not lose the thread, and solve these these longer, more challenging tasks. And our process to do that as a small company with limited resources is let's actually prove this out with small language models, and then we'll show what we've done, we'll go raise more money, and then we'll be able to do it with big language models, which take a lot of time and resources to to retrain and run.

23:35 And through that process, we started working on these smaller language models, and we hit some of our stretch benchmarks for much later uh trainings of much more bigger bigger and powerful models. And we thought like, "Oh my god, the tech is already here. What we've just done is actually, you know, proven this at a small scale, but proven it in such a way that it actually makes these small language models act like the biggest and baddest AI models that are out there."

24:01 And so, that's really led us down this path of, "Okay, it's very clear to us every person's going to own their own intelligence. Companies, businesses around the world are going to have the opportunity to have, you know, their entire workforce have these agents on their computers, live in workstations in their office so that data doesn't have to leave the floor, let alone the the company.

24:19 And then there's a bunch of implications for consumers, too. You know, as can is this something that you can reasonably download with an iPhone app and and live on your phone? That's what we're doing. We're We're more selling to businesses now, but we just see the the world opening up in front of us. >> What would you say the time is for your company for a market like this?

24:37 >> There's kind of two markets that we sell into right now. Number one is we sell inference uh to companies via some cloud providers, and we're able to offer frontier performance at a much, much lower cost. Uh we're doing that with a number of customers now. We think that market is pretty pretty massive. Hard to put a number on it, but there's there's been tens of billions of dollars spent on AI inference in the past year.

24:59 Probably more spent in the last 5 months than in all of 2025. That is a clear, large, and and and growing market. The other thing that we've been doing is working directly with some hardware companies to ship our models on device. So, if you buy not yet disclosed workstation, it will have our our subconscious models embedded on the system. That is a small market today, but we think it's going to grow to be something that's that's very meaningful.

25:24 >> Why did you call it subconscious? >> We wanted it to represent doing work on your behalf in the background. And we think that it really just captures the essence of what we do. And even as we've refined how our system works under the hood, it reflects it even better. Because really what we're doing is over long long reasoning chains, we are as the model is deciding what words what tokens to share next, we're doing some compression of of the previous information that's encountered.

25:51 Specifically, information that's no longer relevant. And what that information does is it compresses it, and it still sits in its memory, but in kind of a latent space that doesn't have full context of all of the data around it. And so, in the model's perspective, you know, we we we stuff that in its subconscious. So, it's still aware of everything that it's encountered, but maybe not every exact detail, but enough to complete whatever task over these really long periods of time.

26:14 Kind of number one, models working in the background, number two, it actually really reflects what we do under the hood. >> Love it. It's been a pleasure having you on the show. You and I talked about doing this, and the thing that stuck out for me is this whole idea, as you said, of open source and AI and smaller models. The number 95% really struck out.

26:33 The fact that you said 95% of what is effectively we need to do could be done by a smaller models. It really is interesting. Jake, just want to wrap up with one final question, which is a book recommendation. Yeah, I'd I'd love to hear yours, you know, what what would be a book that you'd recommend to our audience. >> God, my favorite book of all time is uh in the Three-Body Problem series, the second book, which is called uh The Dark Forest.

26:59 So, so good. It's a uh kind of fate of the world. Um I don't know how to really explain it without explaining the book, but I would say the first book in the series is seven out of 10, the second book, which I'm recommending, is a 10 out of 10. And the third is a nine out of 10. Uh my co-founder's from China. The book was originally actually written in Chinese translated to English.

27:19 And we'd work together for a year and a half, and we finally put that together a couple weeks ago that we both love these books. >> Wow. >> really, really incredible. Definitely worth, you know, yeah. >> Yeah, well, nice timing for this uh for this podcast. Thank you so much for the recommendation. Thank you for your time. Yeah, wish you all the best of luck.

27:37 It's it's really interesting the area that you're tackling. It's it's a very important topic, which is why I wanted to do this. And when you talk about such large numbers in terms of the impact on what this really looks like going forward, AI in your pocket. I love the phrase you used. It's like having a PhD in your pocket. Uh >> Yeah, I I think it's something that it's percolated through the hackers and the people really on the edge really like oh my god, this is amazing.

28:02 And it will be, you know, top-tier news in the next couple months. And I think there'll be talks of, okay, what does this mean for OpenAI and Anthropic, like what you're asking. I think they will still have a business, but I think it's going to open up a whole new world of possibilities, and that's what we're really excited to to play in that arena.

28:18 >> Amazing. Jack, it's a pleasure. Thank you so much for being on the show. >> Yeah, thanks a lot. >> Thanks for tuning in to another episode of Inside the Silicon Mind. This podcast is powered by Harrison Clark. For more episodes, don't forget to subscribe and hit that notification bell. As always, stay curious, stay consistent, and stay inside the Silicon Mind.