In this episode of the All-In Podcast, three long-time friends and seasoned investors—Chamath Palihapitiya, Jason Calacanis, and David Friedberg—take the helm, specially inviting Google co-founder Sergey Brin for a dialogue. They discussed how Gemini, with its infinite context and chain-of-thought reasoning, is upgrading AI from a passive executor to a cognitive collaborator, and analyzed the profound restructuring this paradigm shift brings to technological productivity and the boundaries of human-machine interaction.
The following is a transcript of the podcast content:
Chamath Palihapitiya: You're practically punching a clock. I've heard some news, and we've talked about it, you're working every day.
Jason Calacanis: Honestly, this is one of the happiest times of my life. I retired about a month before the pandemic, and at that time, I founded Hidden Theory. I thought it was pretty good; I wanted to do something else, go to cafes, read physics books. But after about a month, I felt that life wasn't for me. So, as soon as I could go back to the office, I started working. Actually, at that time, an OpenAI friend named Dan, I met him at a party, and he said, this is the greatest moment of change in computer science history, especially for computer scientists like you.
Chamath Palihapitiya: You were a computer science Ph.D. student.
Jason Calacanis: I haven't finished my Ph.D.; strictly speaking, I'm on leave. I thought, he's right, this is truly incredible. Everyone is obviously paying attention to all AI technologies, and as a computer scientist, from a technical perspective, this is definitely the most exciting thing in my life.
Chamath Palihapitiya: And this exponential nature and speed dwarf anything we've seen in our careers. It's almost like everything we've done in the past three or four decades has been leading up to this moment. Google grew from 100 users and 10 employees to now having over 2 billion people using it. I think 6 or 5 products have more than 2 billion users. This isn't even worth calculating, because the vast majority of people on Earth use Google products. Please describe this growth rate.
Jason Calacanis: I remember the excitement of the early internet, using Mosaic, and then Netscape. How many people remember Mosaic? (Looks around) Hmm, not many. Do you remember there was a "What's New" page back then? For example, a small elementary school website, a fish tank hobbyist's website, and Michael Jordan's fan page.
At that time, the entire internet only had these three new websites. Obviously, the internet developed very quickly from then on; it was a very exciting period, and then we had smartphones. But compared to this, the development of AI is simply astonishing. Although the internet became popular, technologically, it didn't change much month by month, year by year. But these AI systems actually change a lot. You know, if you leave for a month and come back, you might exclaim: "Wow, what happened?"
Chamath Palihapitiya: Someone told me you started submitting code, which surprised everyone a bit.
Sergey Brin: What happened?
Jason Calacanis: The code I submitted wasn't particularly exciting. I just needed to add some permissions to access certain things, change a bit here, change a bit there, nothing extra. But you need to do this to do some basic things, run some basic experiments. I tried to do this and got in touch with different parts of the system, which I found interesting.
Secondly, being able to return to the company, without any administrative responsibilities, but still being able to delve into every subtle detail, really feels like a privilege.
Sergey Brin: Which parts of the current AI stack interest you more? Are there any areas that particularly attract you?
Jason Calacanis: A few years ago, or rather, a year ago, I started paying close attention to what we call "pre-training." What most people consider "AI training," whatever they call it, for various historical reasons, we call it "pre-training." This is a super project that requires a lot of computational resources. I also learned a lot, seeing us go from one model to another, and also ran some small experiments, but basically just for fun. Recently, the later stages of training, especially with the emergence of chain-of-thought reasoning, marked another huge step forward for general artificial intelligence. So, we really don't know its upper limit.
Chamath Palihapitiya: So how would you explain the relationship between prompt engineering, deep research, and what's happening to a layman? Because I don't think people are clicking the drop-down arrow to watch the deep research in the Gemini mobile app. You have a great mobile app. By the way, after we talked, I also bought a Fold phone, and the "OK Google" feature is amazing. When you ask it to open an app, it really executes. And the number of threads, queries, and UPS it performs in deep research reaches 200, 300. Can you explain this leap and what you think will happen next?
Jason Calacanis: For me, the most exciting thing about AI, especially today, is that it hasn't reached the complete Artificial General Intelligence (AGI) that people are pursuing, nor is it superhuman intelligence, but it's already quite smart, and it will definitely surprise you. Its superpower is that it can complete tasks at a scale I can't reach. By default, when I use some of our AI systems, it will grab the top 10 search results, or a book, and then extract the information you need. But honestly, I can do that myself, it just might take me more time. But if it grabs the top 1000 results, and then performs follow-up searches on each result and reads deeply, that would be a week's worth of work for me, and I can't do it.
Chamath Palihapitiya: That's what I think people who haven't used the deep research project haven't fully realized. Before, we had an F1 driver on stage, and I was a novice and knew nothing about it. I asked, what's the average annual number of fatal accidents in the past few decades? I wanted to know the number of deaths per mile driven. It initially said, this might be very difficult. I said, I allow you to try your best and give your best theory. Let's do it. Then it listed how many teams, how many races, and so on.
Sergey Brin: Which model?
Chamath Palihapitiya: I used Gemini, which is great. But I treated it like I was arguing with it, which worked pretty well for me.
Jason Calacanis: It's strange, like drinking, we don't beat around the bush. But not just our models, all models tend to perform better if you threaten them in a somewhat aggressive way, but people feel strange about this.
Chamath Palihapitiya: I threatened it then by saying it wasn't "fabulous" enough, and it responded.
Jason Calacanis: Before, you just had to say, "If you don't like that, I'll kidnap you."
Chamath Palihapitiya: It actually built a system where I felt we should factor in practice mileage, assuming 100 miles of practice mileage for every mile on the track, and then it gave an estimated number of deaths per mile, and then I started cross-referencing. I thought, this is like an undergraduate's term paper, you know? Wow, done in minutes.
Jason Calacanis: Yes. I mean, it's amazing. We've all had those moments where I throw something at AI, not really expecting it to succeed. And then exclaim: "Wow, it actually worked."
Sergey Brin: When you experience these moments, and then return to your daily life as a father, do you ever think, what will my children do in the future? Are they learning correctly now? Should I completely change everything they are doing now? Have you experienced similar moments?
Jason Calacanis: Honestly, I really don't know what to think about it. I don't have a fantastic method. My kids are in high school and middle school, and AI has basically surpassed them. Obviously, AI is special in some ways, like early simple math, where they make mistakes humans would never make. But generally, if you talk about math, calculus, and so on, they are very good. They can win math competitions, programming competitions, etc., beating some top people.
Jason Calacanis: My son will be going from sophomore to junior year of high school, and I'm thinking about what he will learn, and we've discussed this question, what areas will AI develop in?
Sergey Brin: Are there any fields you would tell your son not to touch, or to temporarily stay away from?
Jason Calacanis: I didn't particularly plan my life to be an entrepreneur or anything. I just liked math and computer science. Maybe I was lucky, and it turned out to be useful to the world. Kids should do what they like. I hope they do things that are challenging and can overcome various problems.
David Friedberg: Do you think universities should continue to exist as they are now?
Jason Calacanis: Even before AI brought this challenge, universities seemed to be undergoing a transformation. People would ask, should I focus more on vocational skills? What is truly useful? But we've entered a situation where people are questioning universities in various ways. Obviously, AI has brought this issue to the forefront.
David Friedberg: As a parent, I often think about how much of middle and upper-class education revolves around college, how to get children into college. Actually, recently I've been thinking, maybe they shouldn't go to college.
Sergey Brin: My son is about to enter his junior year of high school, and all his thoughts are on wanting to go to an SEC school because of the culture there. Two years ago, I would have been very anxious, wondering if I should help him get into this great school or that great school. But now I think this is actually the best thing he can do, to be socially well-adjusted and mentally able to cope with various failures.
David Friedberg: Enjoy a few years of exploration. Sergey, may I ask some questions about hardware? Many years ago, Google owned Boston Dynamics, which was perhaps a bit ahead of its time, but these systems learn through visual and sensory information, basically learning how to adapt to their surroundings, triggering a very profound learning curve in the hardware field. Now there are dozens of startups making robotic systems. How do you view the fields of robotics and hardware? Is now the time for them to truly take off?
Jason Calacanis: We've acquired and sold about five robotics companies, and Boston Dynamics was one of them. If I look back, we built hardware, and recently internally built Everyday Robotics, but then had to pivot. Robots themselves are cool, but the software isn't quite mature. Every time we try to do this, you have to make them truly useful. Maybe one day that won't be the case anymore.
Sergey Brin: Do you believe in humanoid robots? Or do you think that's a bit too early?
Jason Calacanis: I might be a weirdo who doesn't like humanoid robots much, but maybe that's because I'm too dense, because we acquired two humanoid robot startups and then sold them, so I'm a bit tired of it. But the reason is, people want to build humanoid robots largely because the world is designed around that form. You can train on YouTube, you can train through video, you can do all sorts of things. I personally don't think this gives AI enough credit. AI can learn very quickly how to handle different situations through simulation and real life. It doesn't need the exact same number of arms, legs, and wheels as humans for everything to work properly. So I might not be too optimistic about this. But there are many very smart people making humanoid robots, so I won't rule it out.
Chamath Palihapitiya: What about the path to becoming a programmer? Google now has a 20-year-old codebase, so it could actually have a big impact. So what have you seen internally? For example, Kennex developers, or occasionally seeing some unicorn projects running. But will we see all developers' productivity reach levels of 8, 9, 10, or will all this be done by computers, and we just need to check?
Jason Calacanis: If you like code, I'm actually a bit embarrassed to talk about this. I recently had a big argument internally at the company because we had a list of allowed and disallowed code tools, and Gemini was on the disallowed list. For some very strange reason, which confused me greatly.
Regarding Gemini, no one would enforce this rule, but for some historical reason, there was indeed an internal webpage mentioning Gemini, and we had a big argument.
Chamath Palihapitiya: I don't know if you remember, you're a founder with super-voting shares, this company is still yours after all.
Jason Calacanis: But he (CEO) was very supportive of me. I told him: "I can't deal with those people, you need to handle it." I was very angry that they actually said we were "weird"!.
Chamath Palihapitiya: Imagine, bureaucracy in a company you weren't even involved in creating, that must feel strange.
Sergey Brin: From another perspective, I actually find it quite surprising that some junior marketing people dared to tell us to "go away." I'm serious, I think this is precisely a sign of a healthy corporate culture.
Jason Calacanis: Anyway, it's resolved, everyone is using it (Gemini). Were they fired?
David Friedberg: Sent to Google's Siberian office?
Jason Calacanis: No, not really. We are now trying all possible AI, including external ones, like Cursor and other tools. All of these are to see what really improves everyone's productivity. Personally, using these tools has definitely improved my productivity.
Sergey Brin: Have you trained many foundation models? Looking three years ahead, will these models begin to diverge and become highly specialized? For example, in addition to general and reasoning models, perhaps a model specifically for chip design will emerge, and obviously there will also be one for biodrug design, protein folding. Will the number of future foundation models be several times today's, or about the same, or somewhere in between?
Jason Calacanis: That's a good question. Your guess is probably as good as mine, but if I had to bet, I think the trend is towards increasing convergence. This is basically true across the entire field of machine learning. In the past, we had various different models, such as convolutional neural networks for vision, and recurrent neural networks for text and speech. But all of this has basically turned to Transformer, and there is an increasing trend towards using a single model.
Of course, we occasionally specialize models for specific goals. When you have a specific goal, this is definitely a good iterative method in scientific research. You don't have to use one model to handle all languages, images, videos, and audio. But usually, after we specialize, we can draw on those experiences and basically integrate those capabilities into a general model. So the benefits of specialization are not that many. You can use a smaller, more specialized model, maybe faster and cheaper, but the overall trend is not developing that way.
Sergey Brin: How do you view the choice between open source and closed source? Have there been any significant shifts in philosophy that have given you different perspectives on the value of open source? We are still waiting for OpenAI; we haven't seen it yet, but theoretically, results will come soon.
Jason Calacanis: To be realistic. DeepSeek released a very powerful model in January, which indeed narrowed the gap with proprietary models. We are exploring both aspects. So we released Gemma, which is our open-source model. They perform well, they are small and dense models that can run well on a single machine. It's not as powerful as Gemini. However, it's still hard to say which way will win, what do you think?
David Friedberg: What do you think human-computer interaction will look like as AI develops? Previously, thanks to your (Sergey Brin's) development of the search box, we would type keywords or questions into a box and click on web links to find answers. Will future interaction methods involve typing a question, or speaking into AirPods? Or thinking? And the answer is given directly through voice.
Sergey Brin: Last Friday, Neuralink's brain-computer interface received "Breakthrough Device" designation, which is a very important step towards FDA approval for chip implantation in everyone.
David Friedberg: If you had to summarize, what do you think the most common human-computer interaction mode will look like in the next decade? Is it glasses with a screen? Didn't you try something like that a long time ago?
Jason Calacanis: Honestly, I messed it up then, completely missed the beat.
Chamath Palihapitiya: That was a pioneer, too early.
Jason Calacanis: Yes, it was just too early. There are many things I wish I had done differently back then. But the technology wasn't ready for Google Glass at the time. But now, I think these things make a lot more sense. But battery life issues still exist, and I think we and other companies need to solve that, but it's a cool form factor. Many people are saying that the singularity will arrive in about five years, so what's the future outlook?
David Friedberg: I want to ask a question. Larry said many years ago that humans are just a stepping stone in the evolutionary process. What do you think about that? For example, do you think that this Artificial General Intelligence, super-intelligence, or true silicon-based intelligence will surpass human capabilities, and humans are just a stepping stone in the evolutionary process?
Jason Calacanis: I think sometimes we nerds like to say profound things when we've had a few drinks. I've had two drinks now, and I'm a bit excited.
David Friedberg: Haha, please continue. Jason Calacanis: I probably need more drinks. Let's talk about human implants, we're getting somewhere now.
We are gradually starting to experience that some AIs indeed do much better than us in some aspects. And it's certain that, for example, regarding my math and programming abilities, it's best to ask AI for help now. Actually, this hasn't really bothered me. You know, I treat it as a tool, so I feel like I'm used to it. But maybe in the future, when they become more powerful, I'll look at all this differently.
Sergey Brin: This brings up safety issues. Jason Calacanis: Maybe. I want to say off-topic, using AI for management is actually the easiest thing.
Chamath Palihapitiya: Absolutely.
Jason Calacanis: I also tried some work-scenario chat applications on Gemini, kind of like Slack, but that's our own internal version. We have a very powerful AI tool. Unfortunately, we temporarily took it down, but I think we'll bring it back online and promote it for everyone to use. It can grab content from the entire chat space and answer quite complex questions. So at the time I told it: "Okay, help me summarize the key points discussed." It replied: "Okay, now assign some tasks to everyone." Then I pasted its reply back into the work group, so that everyone wouldn't realize it was AI-assigned tasks. You could tell from snippets, but it did a very good job. Then I thought: "Okay, in this chat group, who should be promoted?" And it actually picked out a young female engineer who usually didn't say much in that group, especially when others were filtered out.
In fact, there weren't those (traditional HR) processes. Then I realized what AI had detected. I went to her manager, and he actually said: "Yeah, you know what, you're right. She's been working very hard and done so much." And it actually happened (she got promoted). So I thought after a while, you might get used to it and feel that AI can make these decisions.
Sergey Brin: Do you think "infinite context" has a place?
Jason Calacanis: One hundred percent useful. If those are all the things to consider, then theoretically you only need one model.
Sergey Brin: Google's codebase has full access to infinite context, plus multi-session parallel execution, so you can run 19, 20 of these projects simultaneously, or let it self-evolve in real time.
Jason Calacanis: Yes, there's no limit to the use of context length. And there are many ways to make it longer and longer.
Sergey Brin: There are rumors that we have an internal version codenamed Gemini Bill, which has an infinite context system. I don't know if this thing is valuable. For any super cool new idea in AI, we probably have five similar attempts internally. The question is how well they perform. We are definitely pushing all boundaries in terms of intelligence, context, speed, everything you can think of is being attempted.
Jason Calacanis: What about hardware? For example, when you build systems, do you care about having unimpeded access to Nvidia? Or do you think this layer will eventually be abstracted, for example, a converter appears, and then the underlying layer is Nvidia plus ten other options, so who cares, we just move forward as fast as possible?
Sergey Brin: For Gemini, we mainly use our own TPUs. But we also support Nvidia. We are one of Nvidia's main chip purchasers, and we provide these chips, as well as TPUs, to customers on Google Cloud. At this stage, to pursue the best performance, complete abstraction is not yet possible. Maybe someday AI will help us abstract this layer. But you also know that, given the huge amount of computation required for these models, you actually have to consider very carefully how everything is implemented, and which chip you use, how memory works, how communication works, etc.—these are actually very important factors. Maybe someday AI itself will be smart enough to weigh these for us. For today, it's not that smart yet.
Chamath Palihapitiya: Is your experience with the user interface similar? I find that, even on my desktop, and certainly more so on my phone, I now immediately go into voice conversation mode and say to it: "No, stop. That's not my question, my real question is this. Not like that. Say it again, summarize in bullet points. No, I want to focus on this point." It's exactly like that. The system response is very fast now. Last year, this feature was unusable, too slow. Now it immediately stops and responds: "Okay." And then you continue to the next step.
I can input by voice, and at the same time, I watch the text being entered on the screen. I then open another window, searching on Google, or sending secondary queries to large language models, or writing Google Docs or Notion pages, or manually typing something. The whole scene is almost like the scene in "Minority Report" where he operates with gloves, or in "Blade Runner" where he says "a little to the left, zoom in; a little to the right, zoom in" in his apartment. All this is related to these language models and their capabilities—response time is always your focus, right? Is there such a leap in response speed that voice interaction is now worthwhile, whereas it wasn't before?
Jason Calacanis: Everything is getting better and faster. So smaller models are also becoming more powerful. There are faster and better ways to infer with them.
Sergey Brin: You can also stack them, like Nico's company, Eleven Labs. It has an excellent text-to-speech (TTS) and speech-to-text (STT) model stack. There are other options, Whisper is also excellent in some respects. But I think in the future you will see a modular combination: certain specific tasks will have certain specialized foundation models. You stack them together, handle the latency, and the effect is great. Like those voice examples you just mentioned, Whisper and Eleven are both very impressive.
Chamath Palihapitiya: Until you turn on the camera, and it can see your reaction when you hear the answer. You say "hmm," and before you can say "no need" or just raise your finger, it pauses. "Oh, do you want a different result? Oh, I see, you're not satisfied with this result."
Jason Calacanis: Interestingly, our company has a large open-plan office layout. So I can't really use voice mode during work hours. I usually use it when driving.
Chamath Palihapitiya: Using voice while driving is simply amazing.
Jason Calacanis: I feel like it's not okay in the office... I can listen to the AI output with headphones, but if I speak to it, everyone around me will hear, which is weird; I just feel it would be socially awkward. But I should use it that way in the car, I do talk to the AI assistant in the car, but that's voice input, voice output. But honestly, maybe that's why I should get a private office. I should spend more time working alone like you guys.
David Friedberg: Exactly.
Chamath Palihapitiya: You can talk to your manager about it.
Jason Calacanis: They might catch me (laughs). Actually, I just like being with everyone.
David Friedberg: Me too, I like to be integrated with everyone and get along. But I do feel like I missed an AI use case. If people want to try your new product, maybe they should try it more often.
Chamath Palihapitiya: If people want to experience your new product, is there a website they can access, or are there any special invitation codes available now to try it? Go check it out. Honestly, there's a dedicated Gemini app. If you want to use Gemini, just like you used Google Search navigation before, download the Gemini app directly. It's great.
Sergey Brin: I think this is really the best model right now. Jason Calacanis: You should use the 2.5 Pro version.
David Friedberg: 2.5 Pro, is that the paid version, right? Jason Calacanis: Yes, you get a few free queries. But if you use it often, you need to subscribe for $20 per month.
Chamath Palihapitiya: Have you thought about making it free and putting some ads next to it to monetize?
Sergey Brin: It might go downhill then, including the entire hardware sector.
Jason Calacanis: Well, right now it (Gemini) is free, and there are no ads next to it. There are just limits on the number of uses for the top models. I don't think we'll ever be able to offer our latest and best models to everyone for free, because they require a lot of computing power. However, when we move to the next generation, you know, like every time we upgrade, the new free tier is usually as good as, and sometimes even better than, the previous professional paid tier.
Chamath Palihapitiya: Thank you, Sergey Brin! Thank you.