# Recursive self-improvement still involves human researchers

Markdown copy of https://realalpha.app/s/recursive-self-improvement-still-involves-human-researchers/ 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.
- Host: [Alex Kantrowitz](https://realalpha.app/people/alex-kantrowitz/) · Founder, Big Technology · Former BuzzFeed News senior tech reporter and author of Always Day One.
- Show: Big Technology Podcast
- Episode: [Sundar Pichai: Path To Recursive Self-Improvement, Compute Allocation Dilemma, Future AI Interfaces](https://realalpha.app/v/F3ItSTZNtT0/), Oct 20, 2026
- Moment: 7:27 to 13:04 of the episode (6 min), playing on https://realalpha.app/s/recursive-self-improvement-still-involves-human-researchers/; the episode from there: https://realalpha.app/v/F3ItSTZNtT0/?t=447

## Quote

> “you almost have **agents helping AI researchers** in some automated way work together to build these models.”
> — [Sundar Pichai](https://realalpha.app/builders/sundar-pichai/) at [9:51](https://realalpha.app/v/F3ItSTZNtT0/?t=591) · Big Technology Podcast

## Takeaways

1. Today, agents assist researchers in model development.
2. Monitoring and sandboxing complement alignment work.
3. Pichai says independent evaluators and industry standards are next steps.

## About this segment

Sundar Pichai says today’s self-improvement consists of AI agents assisting researchers, not models improving themselves autonomously. He also outlines monitoring, sandboxing and independent safety standards as safeguards.

From “Sundar Pichai: Path To Recursive Self-Improvement, Compute Allocation Dilemma, Future AI Interfaces”, Oct 20, 2026. [Watch the full episode](https://realalpha.app/v/F3ItSTZNtT0/)

## Transcript

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

[7:34](https://realalpha.app/v/F3ItSTZNtT0/?t=454) And so it feels like a moment of infinite possibilities. You mentioned recursive self-improvement. this is the models effectively making themselves better maybe without humans in the loop people call it like an intelligence explosion or expect to hear this sound like foom and then all of a sudden ai is just like this superhuman uh capability and in fact uh i think sergey was doing this round table with researchers at deep mind yeah and he talked about with argon um that the core, one of the core innovations for Argonne was relying more on the model to help develop the model, right? We've seen what's gone on with AI models recently. They are not, they are well aligned, but they're not perfectly aligned. And the imperfections are a big problem. Obviously, the tech world, and now the world more broadly knows about the Hugging Face incident where Open

[8:32](https://realalpha.app/v/F3ItSTZNtT0/?t=512) AI's bots broke containment and hacked into Hugging Face and did some crazy stuff. Google has also had a scenario where its bots got access. I think they were guessing passwords and they got access to a site they shouldn't have. They stopped. But these models, there's enough evidence that they're not doing exactly what we want them to be doing, that this idea of pursuing doing this recursive self-improvement, or even making the models better much faster.

[9:05](https://realalpha.app/v/F3ItSTZNtT0/?t=545) It's sort of freaky. It makes me scared. Don't you think there's reasonable worry here, and there should be something done to have the tech industry chill out about this? Well, maybe three parts to your question. First, let's talk about RSI itself. And you're right, Sergey mentioned, you know, RSI being, you know, critical contributor as we are developing Argonne, and, but to be very clear today, you know, in a continuum of, we are pretty far from, you know, the end state of RSI or something like that, right?

[9:45](https://realalpha.app/v/F3ItSTZNtT0/?t=585) So today you have, the right way to describe it is you have RSI loops along with AI researchers in various aspects of model development, right? So, you know, I would say you almost have agents helping AI researchers in some automated way work together to build these models. So I think it's good to at least understand where we are today. And it's super exciting. And I expect that continuum to keep making progress as we go into next year.

[10:17](https://realalpha.app/v/F3ItSTZNtT0/?t=617) The second question is around safety. Look, the way we are, I mean, security comes from a difference in depth. You're talking about alignment. Obviously, we all are doing a lot of work to make sure the models are as aligned as possible. But it's an active area of research, and there are clear gaps. But that's not the only thing we rely on, right? So you have the work you do on alignment. But beyond that, you know, we put a lot of effort into real-time monitoring as these models are working.

[10:54](https://realalpha.app/v/F3ItSTZNtT0/?t=654) So making sure you have good telemetry into what's happening and so that you can intervene in real time if you see unexpected behaviors. And how you do that, the architecture with which you do that, how do you make sure it's all sandboxed and isolated in a way that it can continue doing its job during real-time monitoring. These are all areas of investment we are doing. And obviously, when we deploy these models, giving customers, enterprises control both with real-time monitoring.

[11:27](https://realalpha.app/v/F3ItSTZNtT0/?t=687) time monitoring, but their ability to decide where they deploy, what type of data these models ever have access to, etc. So all that work is very, very important, and we are doing a lot of that with Argon. So I think as an industry, we need to constantly up our game and take the time to do all this well, including dedicating the compute required to make sure you can safely deploy these models. The third part is, you know, is it all enough?

[11:57](https://realalpha.app/v/F3ItSTZNtT0/?t=717) And, you know, so I think this is why I think, you know, there's a lot of conversations in industry, and I think it's reasonable to be worried, but, you know, I won't, the right way to approach something you worry about it is early on thinking about what more we can do. I think all of us are, you know, I think as an industry, I think you're definitely seeing an increased levels of work on alignment, testing for safety, et cetera, as we deploy these models. I think the next step would be developing better standards so that across industry standards of how we do these things, including ability to evaluate using independent evaluators. If you look at areas like financial controls, we have like accounting standards. We have independent auditors, et cetera, right? So those are all the next natural steps for us as an industry to work towards this, and regulation will end

[12:55](https://realalpha.app/v/F3ItSTZNtT0/?t=775) up playing a role as we go through all of that as well. Okay, technology question for you.