This is a YouTube video titled "We’re All Going to Be Managing AIs—Here’s What That Means" containing a multi-speaker discussion recorded in June 2025. The participants include Richard Socher, Brian Roemmele, and Peter, who engage in a detailed conversation about artificial general intelligence (AGI), AI's future impact on workflows, simulations, and the rapid evolution of software engineering. The discussion explores the challenges and opportunities in AI development, feedback loops, and the potential for AI to transform various industries and human life. There is no mention of George Bandarian in the conversation.
At this point, more and more of us are separating super intelligence also from general intelligence. And just to show like, yeah, you can go beyond a bunch of different human skills and go into being even better than people on some skills, which requires often simulations. Because if you only train on human data, it's hard to be above superhuman.
And then I think there's going to be a combination, there's going to be a need more or less to start automating those feedback loops in different areas as much as possible. So to a large degree, quickly validating whether an AI is successful or not might be even more key at this point.
What are your thoughts on AGI, guys? Are we already there? How do you define it? What do you think it'll be like when we get there if we're not already there? And what role do you see u.com playing in that now?
I would love to hear from Brian too, but high level indeed. Depending on how you define it, we're already there. In fact, when we built the first DECA NLP model and you could ask it any question and get an answer in natural language, you know, you could argue, yeah, wow, that was a pretty general purpose intelligent system. But then of course, you know, we keep moving that bar on how general. And at this point, more and more of us are separating super intelligence also from general intelligence. And just to show like, yeah, you can go beyond a bunch of different human skills and go into being even better than people on some skills, which requires often simulations. Because if you only train on human data, it's hard to be above superhuman. You know, if we're doing machine translation and we give it a bunch of human data, well, you know, the AI is not going to translate that much better than those humans. If we give it a bunch of human IQ tests, then the AI will have a hard time like being even superhuman in those IQ tests. And so like eventually realized there shouldn't even be this kind of test. There should be an even harder test that humans can come off because we're not intelligent enough. So there's a lot of space still.
That's how it learned, right? I mean, that's how it learned how to win at Google. And even chess is not studying the moves that humans do. Eventually that initially that's what it did, but when it was given the opportunity to start from a blank slate is when it was able to come up with its own moves that humans have never done.
Exactly. Everything you can simulate AI can solve, but we can't simulate biology yet. We can't simulate A cell perfectly. Yet we can't simulate organs, humans and complex systems. And so but you know, that will happen as AI gets better, as simulations get better, as computers get faster, as quantum computers may eventually have real breakthroughs and can be scaled up. That's when we can have like more and more super intelligence. So I'm really excited about it. I hope U.com will play its part in that and get humanity to that state. Because it is truly incredible what happens when technology scales massively into a field of, of the human condition and you end up in a state where a billionaire and a middle class teenager will use the same iPhone every day for hours. Right. And so you can look at like, what are all the other goods and services that are bottlenecked on intelligence? And you can realize like, well, we'll all have personal assistants, personal tutors for our kids, personal healthcare teams for longevity and health and, and so on. So I'm really for that. And AGI will unlock a lot of those use cases. And then there's sort of very pragmatic definitions of just if we automate 80% of the workflows of 80% of the digitized jobs, which is a couple of trillion dollars worth of gdp, then I think we can actually get there in the next few years. But then you'll see the same thing you see with most other technologies, which is the future is already here, but not equally distributed. Anyway, I love the future. I'll pause it.
Yeah, no, I love it too. So those are all, and I agree with you, I'm very optimistic about the future and look at what's happening. So if we go more into the short term, what do you guys see happening in the next year? Let's call it that Most people, even Maybe some founders, AI founders and VCs don't really see yet.
One of the ideas that I think Richard and I have put forth for the last few years and seems to be true is that the quicker the validation and feedback loop you can have for these systems, the more likely they are to improve very quickly and then more likely that they're going to automate a lot of workflows more quickly. So yeah, coming to the short term and coming to some of those more like pragmatic years as Richard was talking about, it seems like coding and software engineering is changing very, very quickly now, in part because that feedback loop is there and the validation cycles are so quick, the code either works or it doesn't. And that's at the level of does it even compile or not before does it actually do what you want in the application? So that level of feedback is super helpful for rapid iteration. It seems clear that next couple years software engineering is just going to continue to change dramatically. Richard often says, you know, we're all just going to be managing AI, and that's probably true of these AI engineers as well. And then I think there's going to be a combination, you know, kind of jumping off of Richard's simulation point as well. There's going to be a need more or less to start automating those feedback loops in different areas as much as possible. There's so many. You mentioned operations earlier. There's so much in the operation running of companies, economies, scientific laboratories or biology and experiments that I think we could sit here and say that there's no theoretical reason why that stuff couldn't be automated, but the process just isn't there for collecting the data, validating it quickly. There's too much human feedback that's needed in the loop right now. To a large degree. Clever ideas on quickly validating whether an AI is successful or not might be even more key at this point where the Holy grail in any domain there is a full simulation of the state space and you can just play it all out, but in the absence of that, maybe on the way, trying to build up these deciders that help us get the AIs better in incremental steps.
That's cool. I like that.
I'm super excited for AI, for science.
Yeah, I'm really excited about that as well. I think it's my friend Peter. The amandus calls it the longevity escape velocity. Him and Raker as well, and really, you know, just don't do anything stupid. Stay alive. And in the next 10 years, we'll figure out a lot of these diseases and conditions that we've been struggling with for so many years. And we've had, you know, we've lost so many lives to cancer and others, so I'm very hopeful about that.
