# Today's AI still needs humans to define its goals Markdown copy of https://realalpha.app/s/todays-ai-still-needs-humans-to-define-its-goals/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy. - Speaker: [Eric Schmidt](https://realalpha.app/people/eric-schmidt/) · Eric Schmidt is a former Google executive chairman and CEO. · Scaled Google as CEO through 2011, now runs rocket maker Relativity Space and co-founded Schmidt Sciences. - Host: [Chamath Palihapitiya](https://realalpha.app/people/chamath-palihapitiya/) · Founder & CEO, Social Capital · Led Facebook's growth team; his firm backed Slack and AI chipmaker Groq. - Host: [Jason Calacanis](https://realalpha.app/people/jason-calacanis/) · Founder & CEO, LAUNCH · Early Uber angel; sold Weblogs Inc. to AOL; hosts This Week in Startups. - Host: [David Sacks](https://realalpha.app/people/david-sacks/) · Co-founder & partner, Craft Ventures · PayPal's first COO, founded Yammer; White House AI & crypto czar 2025–26. - Host: [David Friedberg](https://realalpha.app/people/david-friedberg/) · CEO, Ohalo · Founded The Climate Corporation, sold to Monsanto for about $1.1B in 2013. - Show: All-In - Episode: [Eric Schmidt on AI, the Battle with China, and the Future of America](https://realalpha.app/v/EkuVqdj8O6E/), Sep 24, 2025 - Moment: 21:45 to 28:23 of the episode (7 min), playing on https://realalpha.app/s/todays-ai-still-needs-humans-to-define-its-goals/; the episode from there: https://realalpha.app/v/EkuVqdj8O6E/?t=1305 ## Quote > “Your goal changes every day, whereas **computers have trouble with that**. As a math problem, we don't have an algorithm yet for LLMs that can do that.” > — [Eric Schmidt](https://realalpha.app/people/eric-schmidt/) at [25:07](https://realalpha.app/v/EkuVqdj8O6E/?t=1507) · All-In ## Takeaways 1. Superhuman performance in a specialty is not the same as general intelligence. 2. Current AI struggles when its objective keeps changing. 3. Schmidt sees human-AI work as complementary for now. ## About this segment Eric Schmidt distinguishes specialized AI progress from general intelligence, which would require systems to set their own objectives. He says current models still need people to guide and validate their work, though recursive self-improvement may arrive in five to ten years. From “Eric Schmidt on AI, the Battle with China, and the Future of America”, Sep 24, 2025. [Watch the full episode](https://realalpha.app/v/EkuVqdj8O6E/) ## Transcript Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there. [21:51](https://realalpha.app/v/EkuVqdj8O6E/?t=1311) can we go back to um ai for a second so um eric i think you can help us get to a a let's call it a a bipartisan understanding of these issues because i think you you think really clearly about this um you know in in the wake of chat gpt launching at the end of 2022 i think the discourse was really dominated in 2023 and 24 by this idea of agi and that agi was imminent and i think it created almost like a panicky atmosphere in Washington among policymakers. And you saw things like we got to restrict open source because, you know, then China will get it. And this is before DeepSeek launched. And then we saw that actually they're ahead of us in open source. But it feels like there's been a pullback a little bit from the AGI narrative, which I think is actually a good thing. I think it's more conducive to calm, rational policymaking. What's your perception of [22:44](https://realalpha.app/v/EkuVqdj8O6E/?t=1364) of AGI right now, where are we on that whole train? So first place, the speech that the president delivered about a month ago about AI strategy, which I think you probably wouldn't say it, but you kind of wrote it for him, was exactly right, right? So thank you. David collaborated with an amazing leader who we all respect and admire so much, Eric. Yes. So nevertheless. saying I wrote it was was way too strong I mean actually if you didn't if you didn't write it then it must have been your twin but in any case the you got you got the emphasis right which was that investment in research investment in the kind of stuff that we do is really really important I don't agree with you on this on this AGI thing because there's this group which I call the San Francisco narrative because they all live in San Francisco. And their narrative [23:43](https://realalpha.app/v/EkuVqdj8O6E/?t=1423) goes something like this. Today we're doing agents. The agentic revolution will change businesses, which I agree with, that what happens is the systems will become recursively self-intelligent. With recursive self-improvement, as it's called, If you have a scale-free problem, and a scale-free problem, for example, is programming or math, where you can just keep doing it, you get these enormous fast gains if you buy enough hardware, do enough software, so forth and so on. That is still underway. The collective of that says that in the next three-ish years, they believe that we will get forms of superintelligence. [24:23](https://realalpha.app/v/EkuVqdj8O6E/?t=1463) and the way they define it is basically a savant, a chemist savant, a physics savant, a mathematician savant. I don't agree with the three years, but I do agree that it'll be maybe six or seven years. But if it's a savant in a particular area, is that general intelligence? It's not general intelligence yet. General intelligence is when it can set its own objective function. Right. And there's no evidence of that. There's no evidence right now of the the ability to set your own objective function, the thinking, and I'm writing a paper on this, so I've been studying it, is that the technical problem is non-stationarity of mathematical proofs. [25:06](https://realalpha.app/v/EkuVqdj8O6E/?t=1506) And what you're doing is you're trying to solve against objective function, but the objective function keeps changing, which is how humans operate. Your goal changes every day, whereas computers have trouble with that. As a math problem, we don't have an algorithm yet for LLMs that can do that. people are working on it and the the test will be can you basically using the information available in 1902 can you derive the same thing that Einstein did with special relativity followed by general relativity we cannot do that today and most people believe that the way this will be solved is through analogy so the theory of great geniuses is that they understand one area of extremely well and they're so brilliant the lady or man can then take their ideas and apply it to a completely different domain. If we can solve that problem then I think it's over. Then we get [26:03](https://realalpha.app/v/EkuVqdj8O6E/?t=1563) to AGI and then it's a whole different world. I think one of the reasons why it's hard to replace a human and you know J.K. and I debate this is that humans are end-to-end. You know we can do the whole job you have sort of a complete understanding you can pivot very easily ai at least as we know it today is not end to end it has to be prompted you get an answer that answer has to be validated then you have to ask a new question because it never gives you exactly what you want you have to apply more context you have to go through an iterative loop finally you get to an answer that has business value the way biology puts it is that ai is not end to end is middle to middle. Humans are end to end. And so as a result of that, instead of AI replacing all of us, AI will be very synergistic with humans because we can define the objective function, [26:53](https://realalpha.app/v/EkuVqdj8O6E/?t=1613) we do the prompting, and we work with it to iterate, and it does a lot of the work in the middle. That seems to me like a very optimistic, less dumeristic take on it. What you just said is exactly what's going to happen for the next few years, that each of us will have assistance which on our command and our prompting will be incredibly helpful to whatever problem we have. You know, personal, you have people who are using these things for relationship advice, for, you know, talking to their kids. [27:22](https://realalpha.app/v/EkuVqdj8O6E/?t=1642) I mean, it's all crazy stuff. But the fact of the matter is that's it. To me, the real question is when does it cross over to having its own volition, its own ability to seek information and solve new problems? That's a different animal. But have we seen any evidence of recursive self-improvement yet? Not yet. But I've funded a number of startups which claim to be close to it. But, of course, these are startups, and you never know, which tells me it's five, ten years. [27:50](https://realalpha.app/v/EkuVqdj8O6E/?t=1670) What do you think Google's doing on this front? Well, I'm not at Google anymore. Every issue of Gemini is top of the leaderboard. So 2.5 just overcame everybody, and I'm sure there's another one coming. DEMIS is working really hard on this question about scientific discovery. So that is a path to getting to AGI. Eric, we appreciate the work you're doing. We appreciate you being here with us.