Why AI Research Won't Hit Fast Takeoff | Charlie O'Neill
MTS
November 4, 2026
Baseten co-head of model training Charlie O'Neill breaks down why recursive self-improvement will be much weaker than fast-takeoff theories predict, pointing to real-world compute and RL bottlenecks.
Charlie O'Neill
I could see a world in which recursive self-improvement exists, but it is a very weak form of recursive self-improvement, in the sense that it still takes exponentially more inputs from the real world to get the next equivalent jumping capabilities. And so you could imagine that, you know, if you don't have a learner that is like significantly more sample efficient or in context efficient than the transformer with pre-training class RL, then every new frontier of capabilities that you need to hit requires exponentially more effort in terms of the RL environments, for instance, you need to create in order to get to that frontier.
I think recursive self-improvement will be a thing, but it will be like a weaker form of what we're imagining and won't necessarily lead to the fast takeoff.
SPEAKER_2
And we are back. We are live with Charlie O'Neill, who's co-head of model training at Baseten and lead of Baselabs, which is Baseten's open source AI research lab, working on continual learning, open RL environments, and a lot more. Charlie, welcome to MTS.
Charlie O'Neill
Thanks for having me.
SPEAKER_2
So tell us about some of the stuff that Baseten just announced today on the inference platform, grid, sandboxes, loops.
Charlie O'Neill
Yeah, it's been a big day.
So yeah, we've got basically three big products coming under our unified inference platform today.
Grid, loops and sandboxes. Grid is essentially a way to get AI spend under control. This has been a huge topic of debate recently. With the split between open-source models and closed-source models, how should an enterprise or a company manage their spend? So it basically brings every model and dollar are spend into one place and also provides a state-of-the-art router to route to the right model on quality and price. And it also works with any model or harness. So that's a big thing that our customers have been asking for for a while. Loops, I'm most excited about. This is the training side of things. This is essentially gives enterprises serverless SFT and RL on frontier open weight models, which basically means you can run training from literally your laptop and not have to worry about any of the infrastructure or GPUs or anything like that, and deploy straight to dedicated inference, which is really cool. And then finally, sandboxes are like the glue that's plugging all these things together in the world of agentic AI. They're essentially the cloud for agents. They're built on top of our BlackSol acquisition. They're basically stateful, VM-isolated, scaled machines that enable us to do really long-running agentic tasks in the real world. And I guess all of these kind of things are going to show that the world of AI now is not just generating tokens. It's like an ecosystem of products that commoditizes and makes very organic the experience of plugging together the intelligence for whatever you're trying to do.
SPEAKER_2
Right. So yeah, very interesting.
Another topic that gets talked about constantly, basically, is recursive self-improvement and timelines to recursive self-improvement and what this actually means for AI progress. And I'm very interested in the latter section of that, which is if AI labs are able to fully automate much or all of the machine learning research process, how much will that meaningfully accelerate progress towards for real, for real superintelligence, as in AIs that are actually better than all humans at all cognitive tasks?
Charlie O'Neill
Yeah, I think it's really interesting when you think about this problem, because let's assume that we define recursive self-improvement in terms of the current paradigm we're in. That is some form of transformer that's gone through pre-training and then a large amount of RL and post-training, which has been scaled up.
If you stay on that paradigm, I could see a world in which recursive self-improvement exists, but it is a very weak form of recursive self-improvement, in the sense that it still takes exponentially more inputs from the real world to get the next equivalent jumping capabilities. And so you could imagine that if you don't have a learner that is significantly more sample-efficient or in-context efficient than the transformer with pre-training plus RL, then every new frontier of capabilities that you need to hit requires exponentially more effort in terms of the RL environments, for instance, you need to create in order to get to that frontier. And I think we're already running to this a little bit with the combination of humans plus LMs, building these things and working on the research. And I think the labs are seeing this, they're spending exponentially more money on RL data. Each RL environment takes exponentially more time to create because we're saturated the low-hanging fruit. These tasks are now running to the bottleneck of like, okay, it does take 24 hours to a week to do this rollout based on the real world and based on how quickly we can decode and things like that. So I think recursive self-improvement will be a thing, but it will be like a weaker form of what we're imagining and won't necessarily lead to the fast takeoff.
Read the other 11 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:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown 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:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000799858060