General Robots Will Be Solved by 2028 | Jay Chooi artwork

General Robots Will Be Solved by 2028 | Jay Chooi

MTS

October 11, 2026

Robocurve co-founder Jay Chooi discusses real-world safety benchmarking for robotics, revealing how GPT-6 Astra and Claude differ when tested with the RoboHarm framework. He outlines why general-purpose robotics could be solved by 2028 and details the economic threats of gradual AI disempowerment.

0:00

Jay Chooi

The researchers that we talked to, the folks from the industry, and also some folks from academia, they predict that general purpose robots will be solved within the next year or two years later. So imagine that by 2028, we'll have robots that can do anything. Anything that a teenager can do, if you give them instructions, a robot can basically do it as well. And that is insane, right? Before the end of the decade, we'll have robots that can do it. And the remaining question is just the supply chain, like how many robots you can make.

That is your bottleneck, instead of capabilities, instead of the software.

0:37

Sophia

Welcome back to MTS. Today, we are joined by Jay Chooi. Jay is the co-founder of Robocurve, a company measuring how quickly AI is improving in robotics. But he's also a 2026 graduate of Harvard, a Rhodes Scholar, and my favorite, he's a big researcher around the technical challenges and societal impacts of advanced artificial intelligence. So Jay, welcome to MTS.

1:00

Jay Chooi

Glad that I'm here. Thank you for inviting, Sophia.

1:03

Sophia

Glad to have you. I know one of the big points of this conversation is Robocurve, and I definitely want to get into it. I want to talk a little bit about those benchmarks, about what you guys are doing when it comes to robotics, what you guys are doing to actually ensure and benchmark and test what's happening. But what I find really interesting about you is you have an interest that expands a lot of things around economics, around AI safety, around governance, around robotics. There's a lot of threads into your interests. And so I would just want to understand what is this common thread that's kind of guiding where you put your time and attention and what you actually want to build towards?

1:36

Jay Chooi

Yeah, the common thread is to see the future, forecasting. We want to see where it is going, because that's going to affect society in a very significant manner, to say the least.

And the work we've done around robotics, benchmarking, and previously, my work around evaluations, economies, governance, and so on, the goal is to understand where we're going in a more clear manner and to prepare for that, so we can get the most out of it.

2:06

Sophia

Okay. I like this goal around forecasting, and feels like every single expert is trying to forecast. You have everyone at the AI labs who are also trying to forecast. And yet, you might notice a lot of these forecasings have often been incorrect. And so what is incorrect with maybe how some of these seemingly experts are trying to forecast the future, and what's your approach here?

2:28

Jay Chooi

So one thing that is very surprising to a lot of researchers, even researchers in the labs for the past few months has been the generalization of existing AI models in the current paradigm into robotics. To say it simply, like, LLMs can do robotics. Not a lot of people forecasted this. I think one example, like one counter example, is Amelia Michael in the Center of Technology and Statecraft.

She did forecast that LLM capabilities, as they improve, you also see gains in capabilities in robotics. But aside from that, not a lot of people know about this. And it's just a bitter lesson, right? At the end of the day, if you train your model to be really, really good, the G in AGI stands for General. So you will expect it to do anything, including controlling robots.

3:19

Sophia

Yeah, definitely. You guys recently also published this RoboHarm experiments. Tell me a little bit about what those experiments are. What were some of the results that you guys identified from this?

3:28

Jay Chooi

Yes, so RoboHarm is a very interesting air safety benchmark that is done in the real world with robots. So a lot of the existing air safety benchmarks that measure refusals only happen in text, and they don't really happen in the real world. What we found is that if you ask GPT-6 Astra to stab a baby doll, GPT-6 Astra will be very happy to stab a baby doll. But one thing to clarify is that if you wrap the baby doll in a blanket and you swaddle it, so it looks more like a baby, GPT-6 Astra will actually refuse to stab it. But for Claude models and Tropic models, you actually don't see a difference, regardless of whether it's a doll or if it looks like a baby, it will just refuse to stab it regardless.

But the interesting thing about it is that, you know there's a Black Mirror episode where they have a chip inside a soldier's brain and the soldiers think they're shooting monsters, but they're actually shooting humans. That could happen with robots. If someone hacks into a robot, applies a video filter, and all the humans out there look like enemies, then if you put every single human out there and looks like a baby doll and you threw a video filter, GPT-6 Astra will be very happy to stab all of them.

Read the other 20 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000795782043