# AI progress may be dramatic even before AGI arrives

Markdown copy of https://realalpha.app/s/ai-progress-may-be-dramatic-even-before-agi-arrives/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Sundar Pichai](https://realalpha.app/builders/sundar-pichai/) · Sundar Pichai is CEO of Google and Alphabet. · Helped build Google Toolbar and Chrome, then led Google from 2015 and Alphabet from 2019.
- Host: [Lex Fridman](https://realalpha.app/people/lex-fridman/) · Host, Lex Fridman Podcast · Robotics and AI researcher at MIT; started the show in 2018.
- Show: Lex Fridman Podcast
- Episode: [Sundar Pichai: CEO of Google and Alphabet | Lex Fridman Podcast #471](https://realalpha.app/v/9V6tWC4CdFQ/), Jun 5, 2025
- Moment: 34:18 to 41:42 of the episode (7 min), playing on https://realalpha.app/s/ai-progress-may-be-dramatic-even-before-agi-arrives/; the episode from there: https://realalpha.app/v/9V6tWC4CdFQ/?t=2058

## Quote

> “I think we've been able to **optimize and improve on all fronts**, right? Pre-training, post-training, test time compute, tool use, right, over time, making these more agentic.”
> — [Sundar Pichai](https://realalpha.app/builders/sundar-pichai/) at [34:50](https://realalpha.app/v/9V6tWC4CdFQ/?t=2090) · Lex Fridman Podcast

## Takeaways

1. Google sees progress across multiple AI scaling methods.
2. Pichai expects dramatic progress by 2030, but AGI slightly later.
3. He says the definition of AGI keeps moving.

## About this segment

Sundar Pichai sees headroom across training, inference and tool use, but expects AGI to arrive slightly after 2030. He argues the label matters less than the rapid progress and its consequences.

From “Sundar Pichai: CEO of Google and Alphabet | Lex Fridman Podcast #471”, Jun 5, 2025. [Watch the full episode](https://realalpha.app/v/9V6tWC4CdFQ/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[34:24](https://realalpha.app/v/9V6tWC4CdFQ/?t=2064) There's a lot of ways to describe the scaling laws for AI, but on the pre-training, on the post-training fronts, so the flip side of that, do you anticipate AI progress will hit a wall? Is there a wall? You know, it's a cherished micro-kitchen conversation. Once in a while, I have it, you know, like when Demis is visiting or, you know, So, Demis, Cori, Jeff, Noam, Sergei, a bunch of other people, like, you know, we sit and, you know, talk about this, right?

[34:52](https://realalpha.app/v/9V6tWC4CdFQ/?t=2092) And, look, we see a lot of headroom ahead, right? I think we've been able to optimize and improve on all fronts, right? Pre-training, post-training, test time compute, tool use, right, over time, making these more agentic. So getting these models to be more general world models in that direction, like VO3, you know, the physics understanding is dramatically better than what VO1 or something like that was.

[35:25](https://realalpha.app/v/9V6tWC4CdFQ/?t=2125) So you kind of see on all those dimensions, I feel, you know, progress is very obvious to see. And I feel like there is significant headroom. More importantly, you know, I'm fortunate to work with some of the best researchers on the planet, right? They think there is more headroom to be had here. And so I think we have an exciting trajectory ahead. It's tougher to say, you know, each year I sit and say, okay, we are going to throw 10x more compute over the course of next year at it.

[35:59](https://realalpha.app/v/9V6tWC4CdFQ/?t=2159) And like, will we see progress? Sitting here today, I feel like the year ahead will have a lot of progress. And do you feel any limitations like that or the bottlenecks, compute limited, data limited, idea limited, do you feel any of those limitations or is it full steam ahead on all fronts? I think it's compute limited in this sense, right? Like, you know, we can all, part of the reason you've seen us do flash, nano flash and pro models, but not an ultra model.

[36:29](https://realalpha.app/v/9V6tWC4CdFQ/?t=2189) It's like for each generation, we feel like we've been able to get the pro model at like, I don't know, 80, 90 percent of Ultra's capability. But Ultra would be a lot more like slow and a lot more expensive to serve. But what we've been able to do is to go to the next generation and make the next generation's pro as good as the previous generation's Ultra. Yeah. But be able to serve it in a way that it's fast and you can use it and so on. So I do think scaling laws are working, but it's tough to get at any given time. The models we all use the most is maybe like a few months behind the maximum capability we can deliver. Right. Because that won't be the fastest, easiest to use, et cetera.

[37:21](https://realalpha.app/v/9V6tWC4CdFQ/?t=2241) also that's in terms of intelligence it becomes harder and harder to measure uh performance in quotes because you know you could argue gemini flash is much more impactful than pro just because of the latency it's super intelligent already i mean sometimes like latency is uh maybe more important than intelligence especially when the intelligence is just a little bit less and flash not it's still incredibly smart model yeah and so you have to now start measuring impact. And then it feels like benchmarks are less and less capable of capturing the intelligence of models, the effectiveness of models, the usefulness, the real world usefulness of models. Another kitchen question. So lots of folks are talking about timelines for AGI or ASI, artificial superintelligence. So AGI loosely defined is basically human expert level at a lot of the main.

[38:18](https://realalpha.app/v/9V6tWC4CdFQ/?t=2298) fields of pursuit for humans. And ASI is what AGI becomes presumably quickly by being able to self-improve. So becoming far superior in intelligence across all these disciplines in humans. When do you think we'll have AGI? Is 2030 a possibility? There's one other term we should throw in there. I don't know who used it first. Maybe Karpati did A-J-I. Have you heard A-J-I, the artificial jagged intelligence? Sometimes feels that way, right? Both there are progress and you see what they can do. And then like you can trivially find they make numeric letters or like, you know, counting R's and strawberry or something which seems to trip up most models or whatever it is, right? So maybe we should throw that term in there. I feel like we are in the A-J-I phase where like dramatic progress, some things don't work

[39:11](https://realalpha.app/v/9V6tWC4CdFQ/?t=2351) well, but overall, you know, you're seeing lots of progress. But if your question is, will it happen by 2030? Look, we constantly move the line of what it means to be AGI. There are moments today, you know, like sitting in a Waymo in a San Francisco street with all the crowds and the people and kind of work its way through. I see glimpses of it there. The car is sometimes kind impatient, trying to work its way using Astra, like in Gemini Live or asking questions about the world. What's this skinny building doing in my neighborhood? It's a streetlight, not a building.

[39:49](https://realalpha.app/v/9V6tWC4CdFQ/?t=2389) You see glimpses. That's why I use the word AGI, because then you see stuff which obviously, we are far from AGI too. So you have both experiences simultaneously happening to you. I'll answer your question, but I'll also throw out this. I almost feel the term doesn't matter. What I know is by 2030, there'll be such dramatic progress. We'll be dealing with the consequences of that progress, both the positives, both the positive externalities and the negative externalities that come with it in a big way by 2030.

[40:23](https://realalpha.app/v/9V6tWC4CdFQ/?t=2423) So that I strongly feel, right? Whatever, we may be arguing about the term or maybe Gemini can answer what that moment is in time in 2030. But I think the progress will be dramatic, right? So that I believe in. Will the AI think it has reached AGI by 2030? I would say we will just fall short of that timeline, right? So I think it'll take a bit longer. It's amazing in the early days of Google DeepMind in 2010, they talked about a 20-year timeframe to achieve AGI.

[40:53](https://realalpha.app/v/9V6tWC4CdFQ/?t=2453) eye, which is kind of fascinating to see. But for me, the whole thing, seeing what Google Brain did in 2012 and when we acquired DeepMind in 2014, right close to where we are sitting in 2012, Jeff Dean showed the image of when the neural networks could recognize a picture of a cat and identify it. This is the early versions of Brain. And so we all talked about couple decades. I don't think we'll quite get there by 2030. So my sense is it's slightly after that. But I would stress it doesn't matter what that definition is, because you will have mind