July 14 2026 - Sutton’s Oak Lab & TRACE: Smarter, Continuous AISutton’s Oak Lab & TRACE: Smarter, Continuous AI

**Taylor** (0:00)
Welcome back to AI Signal & Noise. It is Tuesday, and we have some absolutely mind-blowing news today, dude. I am Taylor.

**Morgan** (0:10)
And I am Morgan.
Mind-blowing is a strong word, Taylor, but I have to admit, today's lineup is actually pretty fascinating. What is first?

**Taylor** (0:21)
Oh, just a Turing Award winner launching a brand new startup to completely reinvent how AI agents learn. No big deal, right?

**Morgan** (0:30)
Okay. You definitely have my attention now. Let's dive right into that one. Which Turing Award winner are we talking about here?

**Taylor** (0:37)
So Richard Sutton, who basically co-founded Modern Reinforcement Learning, just launched a brand new startup called Oak Lab in Toronto.

**Morgan** (0:47)
Wait, Richard Sutton? That is huge. He is an absolute legend in the field. What is the main mission for Oak Lab?

**Taylor** (0:55)
Dude, he is calling current deep learning methods weak and inefficient. He wants to build AI agents that actually learn continuously from their environment.

**Morgan** (1:07)
Weak and inefficient? Those are fighting words in the AI community right now, especially with LLMs dominating everything.

**Taylor** (1:15)
I know, right? But he argues that true intelligence can't just be a static model trained on the past. It has to adapt.

**Morgan** (1:25)
Continuous learning is notoriously hard. Right now, we train models on static datasets, and then they are frozen. Is he trying to break that?

**Taylor** (1:35)
Exactly. He thinks frozen models are a dead end. He wants agents that learn on the fly, just like humans do when we interact with the world.

**Morgan** (1:44)
It sounds incredible, but the compute costs for continuous training in real time must be astronomical. How does he plan to solve that?

**Taylor** (1:54)
Well, according to The Decoder, they are focusing on self-supervised reinforcement learning. He thinks this is the only path to true, adaptable intelligence.

**Morgan** (2:05)
I mean, if anyone can pull it off, it is sudden. But building a startup around this is a massive gamble.
Research is one thing, commercializing it is another.

**Taylor** (2:15)
True.
But imagine an AI that gets smarter every single second you use it. That would be so cool. It is like having a real partner.

**Morgan** (2:25)
Or a real headache if it learns the wrong things from a bad environment. We will definitely be keeping a close eye on Oak Lab.

**Taylor** (2:33)
Totally. But honestly, the idea of moving away from these massive, static LLMs is super exciting. It feels like the next big frontier.

**Morgan** (2:43)
It is a necessary shift for sure. But speaking of making LLMs smarter, didn't Stanford just release something related to agent training?

**Taylor** (2:53)
Yes.
Stanford researchers just introduced TRACE. It is a capability targeted training system that turns agent failures into synthetic RL environments.

**Morgan** (3:05)
Wait, so it takes the mistakes an AI agent makes and uses those exact failures to train it? That sounds like a meta-learning loop.

**Taylor** (3:15)
Yes, exactly. They noticed agentic LLMs keep failing the same way because they lack specific reusable capabilities. TRACE diagnoses those gaps.

**Morgan** (3:27)
It is like a targeted software patch, but how does TRACE actually diagnose the gaps? Is it automated?

**Taylor** (3:35)
Yes, it analyzes the agent's own trajectories to find where it consistently fails, then synthesizes a verifiable training environment for that specific capability.

**Morgan** (3:47)
Oh, interesting. So instead of just guessing why it failed, TRACE actually isolates the failure, creates a custom training ground, and trains a LoRa adapter for it.

**Taylor** (3:58)
Yes, dude. And then it routes tokens across these expert adapters. They tested it on SWEbench verified and reached a 73.2% pass rate.

**Morgan** (4:09)
Wow, a 73% pass rate on SWEbench is actually massive. That benchmark is notoriously difficult for software engineering tasks.

**Taylor** (4:20)
Right? And they got a 15.3 point bump on Towsquared Bench 2 It is like giving the AI a personalized tutor for its weaknesses.

**Morgan** (4:29)
I like this approach because it is targeted. Instead of retraining a whole model, you are just patching the specific holes in its skills.

**Taylor** (4:37)
Exactly. It is so much more efficient. It is like, why study the whole textbook again when you only fail the algebra section?

**Morgan** (4:47)
Exactly. But I wonder how well it scales when you have thousands of different failure modes.
Does it just create thousands of tiny adapters?

**Taylor** (4:57)
That is a fair point, but the routing system seems to handle it. It is a super clever way to handle agentic training.

**Morgan** (5:04)
It definitely is. It is a practical step forward while we wait for Sutton's continuous learning dream to become a reality.

**Taylor** (5:13)