Jerry Tworek | Training Data Summary

1. Key Themes

The Transformer Architecture Has Hit a Fundamental Ceiling

Jerry and Rohan argue that the Transformer's core limitation is not efficiency but an inability to learn at test time — meaning models must be continuously retrained in labs rather than adapting to real-world deployment. Jerry frames this as the central architectural problem to solve.

"Core Automation is a lab created to build models that continuously learn from deployment. We believe that Transformers are incapable of continual learning. There's no way how to put continual learning on Transformers. So we know we have to find a different architecture." 00:45:09 — Jerry Tworek

Benchmarks and Training Data Are Converging, Leaving a Real-World Gap

Jerry observed directly inside OpenAI that as RL scaled up and benchmark scores went up uniformly, real-world task performance did not follow. The training distribution and real-world distribution are fundamentally misaligned.

"All the evals and training tasks are the same sides of the coin. But the real world distribution and real world tasks is much messier, much murkier, much more different. Our training data didn't really replicate the real world use cases. And despite us basically maximizing all the tasks... we are still not covering the entirety of the real world distribution." 00:06:54 — Jerry Tworek

The Two Existing Test-Time Learning Mechanisms Are Both Broken

In-context learning is too limited (roughly 20 minutes of usable context before compaction), and fine-tuning suffers from catastrophic forgetting and low data efficiency. Neither path leads to genuine continual learning.

"We could try to continuously fine-tune our models, but then those have the issues of catastrophic forgetting. We have issues of very low data efficiency. And neither of those are very solvable. Neither of those are very easy to find ways. People have been trying. If those were easy to solve, someone would already solve it." 00:08:20 — Jerry Tworek

Computational Depth Is the Hidden Bottleneck in Current Models

Rohan identifies that Transformers are architecturally shallow — at most ~100 layers deep — and that chain-of-thought token generation is an inefficient band-aid that increases computational depth one token at a time rather than through architectural design.

"Most Transformers that we train are quite shallow — that's at most like 100 layers deep. Depth is like, it's called deep learning because you wanted deeper representations... Chain-of-thought reasoning and RL to do chain-of-thought by the model itself is one way to increase computational depth because every token you add you add like one more pathway." 00:19:23 — Rohan Anil

Pre-Training and RL Must Be Optimized End-to-End Together

Both speakers believe the labs treat pre-training and RL as sequential rather than jointly optimized processes, and that combining them end-to-end is where an order-of-magnitude improvement lies.

"We have to combine pre-training and RL together. And that's like where one order of magnitude improvement would come from." 00:27:47 — Rohan Anil

The Big Labs Are Structurally Incapable of Exploring Transformer Alternatives

Competitive release pressure and quarterly revenue incentives make it nearly impossible for the frontier labs to invest seriously in architectural alternatives. This creates a structural white space.

"The biggest labs... are probably in the most competitive market fight ever right now, which makes them not very keen on trying different paths... If Transformer is profitable and if you can spend more efforts and more resources scaling Transformer to win in the next quarter, it's very hard to put at least a lot of attention and a lot of energy to work on something that will maybe better or maybe will redefine the field in a year or two." 00:12:58 — Jerry Tworek

Kernel Generation Is the Unlocking Primitive for Novel Architectures

The bottleneck to exploring new architectures is not ideas — it is the ability to write high-performance GPU kernels for novel operations. Only a tiny number of humans can do this, and current AI models cannot, creating a hard ceiling on architectural exploration.

"If I use CuSolver's QR, I get some performance. If we use the competition winner's QR, you get 60x faster. And that is a completely different playing field. Now it opens up an entire new set of algorithms you can apply." 00:44:18 — Rohan Anil

The Automated Lab Is Both the Research Strategy and the Product

Core Automation's approach is to build the most automated research lab — where AI accelerates the experiment iteration loop — as both a means to discover new architectures and as a demonstration of the continual learning thesis.

"The speed of iteration, the speed of research, the speed of how quickly you can move through ideas and how quickly you can get data on your ideas is something very, very different... If we can execute at least even one of those experiments a day, that's already a pretty good iteration speed versus anything that was done before. And maybe one day we get to 10 of those a day. Maybe one day we get to 100 of those a day." 00:36:17 — Jerry Tworek

AGI Requires Removing Humans from the Loop Entirely — and We Are Very Far Away

Jerry offers a precise personal definition of AGI: a model that can improve itself without any human in the loop. By this definition, current systems — however impressive — are nowhere close, and the human-LLM hybrid is masking the gap.

"For me, AGI is a model that can improve itself without human in the loop in any way... Removing humans from loops with models has been actually notoriously, notoriously difficult so far. We haven't come anywhere close to it. It's very, very hard for me to find any task where LLMs were able to get humans out of the loop." 00:38:20 — Jerry Tworek

2. Contrarian Perspectives

The Architecture Is the Bottleneck — Not Data or Compute

The mainstream view is that scaling data and compute is still the path forward. Jerry argues the bottleneck has shifted to architecture itself, and that the field is essentially doing the same two operations (MoE and attention) at ever-larger scale without addressing fundamental expressiveness.

"What I think right now we are in this stage, we got really, really good at training really, really big models. We mastered two algorithms. We mastered pre-training at a large scale and we mastered reinforcement learning at a large scale... I think at this moment what the bottleneck is to better models and to smarter systems is the architecture itself." 00:02:44 — Jerry Tworek

Pre-Training Today Is Mostly Distilling Other AI Models, Not Human Knowledge

The commonly held view is that pre-training captures human knowledge from the internet. Jerry quietly drops the observation that most internet tokens are now AI-generated, meaning pre-training is largely model-to-model distillation.

"There is clearly pre-training which is behavioral cloning, which is mimicry, which is compression of internet data. But pre-training today is largely distilling other models into the new model because most of the tokens in the internet are coming from AI." 00:24:01 — Jerry Tworek

Reinforcement Learning Is Not the Final Form of Learning from Experience

Against the dominant "RL is the path to AGI" narrative — championed by Rich Sutton, David Silver, and others — Jerry argues RL is just one mathematical formulation, and a particularly inefficient one at that, not the endpoint.

"Reinforcement learning is not the end of learning from experience. And there will be better approaches that researchers will be coming up in the coming years on how to use that data in richer and richer settings." 00:00:28 — Jerry Tworek

Transformers Were Adopted Because of Economics, Not Because They Are Optimal

The received wisdom is that Transformers won because they were architecturally superior. Rohan and Jerry argue the real reason was economic: their scaling returns exceeded training costs sufficiently to justify investment, whereas LSTMs would not have attracted the same capital.

"The majestic thing about Transformer... is that Transformers are economically valuable. Training them, the cost of training them is lower than the revenue that they generate, which is just magic of machine learning. It's not guaranteed by itself. But for LSTMs, it probably wouldn't be that way, which made it happen." 00:10:18 — Jerry Tworek

We Will Not Approach Biological Learning Efficiency on Current Hardware

Against the optimism that scaling AI will eventually match human data efficiency, Rohan argues this is a hardware problem, not just a software one — and that biological efficiency requires co-designing the learning algorithm with the physical substrate.

"Until I think we design hardware to be much more like how humans operate, maybe more analog, figure out how to deal with analog circuits and figure out how to do it at a correction, figure out how to get information through, it'll be much harder." 00:32:29 — Rohan Anil

3. Companies Identified

Core Automation A new AI research lab founded by Jerry Tworek and Rohan Anil focused on building architectures capable of continual learning at test time, and on becoming the most automated AI research lab in the world. The company's thesis is that Transformers cannot achieve AGI because they cannot learn continuously after deployment.

"Core Automation is a lab created to build models that continuously learn from deployment... Our quest is to find that new architecture, find that Transformer replacement. And we want to build the most automated lab to do it." 00:45:09 — Jerry Tworek

OpenAI Mentioned as the originating institution behind the Strawberry and reasoning model lines, and as the lab that made the contrarian bet on scaling over algorithmic efficiency. Jerry was VP there.

"It was a bit of a contrarian bet by OpenAI at that moment to try to say, hey, we don't care about better and better algorithms. We care about more and more scalable algorithms. And how do we pour more and more compute and get better results, which OpenAI was criticized repeatedly by many people in the community for a long time. But thanks to that, we have the models that we have today." 00:11:17 — Jerry Tworek

Google Brain / Google DeepMind Mentioned as the birthplace of the Transformer (Rohan was present when it was invented), and as the institutional context for much of Rohan's optimization and pre-training work including Gemini 1.5 Flash.

"When the Transformer — Ashish, Noam and others came up with it... It wasn't a big deal internally. There was only a few people who actually got it." 00:14:30 — Rohan Anil

Anthropic Mentioned as a stop for Rohan between Google and Core Automation, and as one of the major labs whose models were tested on the QR kernel problem and found incapable of solving it.

"If you give this problem to Anthropic models, OpenAI models, Gemini, it just wouldn't solve it." 00:41:34 — Rohan Anil

DeepSeek Called out for independently arriving at ngram-style memory embedding work — an idea Rohan had been developing for five years internally — and demonstrating via scaling laws that MoE could potentially be replaced with ngram embeddings.

"Six months before that, DeepSeek wrote their ngram, which is an improved version of adding more memory, showed scaling laws that, yeah, you don't need MoEs. You could have actually replaced it with these ngram embeddings. For me, that was like, oh." 00:43:53 — Rohan Anil

GPU Mode A community mentioned as co-hosting Core Automation's QR kernel competition, which produced a 60x performance improvement on the B200.

"An example of this is our QR kernel competition that we hosted with GPU Mode." 00:40:36 — Rohan Anil

4. People Identified

Jerry Tworek Co-founder of Core Automation; former VP at OpenAI where he led the Strawberry and reasoning teams and was central to scaling reinforcement learning. Has been an RL maximalist since joining OpenAI and now believes architectural change is the next necessary step.

"I basically believed that scaling up reinforcement learning is a necessary stepping stone on a path to AGI... And I saw us training model after model. This model was getting better and better. All the benchmark scores were going up. And did we also solve all the real world tasks at that moment? Unfortunately, unfortunately not." 00:05:32 — Jerry Tworek

Rohan Anil Co-founder of Core Automation; one of four pre-training leads on Gemini; led fundamental AI research at Google Brain; co-created the Shampoo optimizer; was at Anthropic before founding Core Automation. Described as a pre-training maximalist turned architectural skeptic.

"That inner belief and second is you need your architecture to run efficiently on hardware. A theoretically optimal architecture is not useful to anyone... There's only few places right now which have integrated teams doing that and I think we have built a team in a way that puts the experts together not in different silos." 00:17:26 — Rohan Anil

Ashish (Vaswani) and Noam (Shazeer) Named as the inventors of the Transformer at Google Brain. Rohan was present at the internal research conference where it was presented and recalls it was not recognized as significant by most.

"Ashish, Noam and others came up with it... It wasn't a big deal internally. There was only a few people who actually got it." 00:14:30 — Rohan Anil

Rich Sutton and David Silver Authors of "The Age of Experience" paper, cited by Sonya as a reference point for the RL-maximalist worldview, which Jerry respectfully pushes back on.

"Rich Sutton and David Silver have this paper, The Age of Experience... Reinforcement learning is not a particularly new approach, particularly new thing to do. So in some way, The Age of Experience I think always has been there." 00:23:22 — Jerry Tworek / Sonya Huang

Vinit Gupta Google researcher who approached Rohan and co-developed the Shampoo optimizer, which delivered roughly a 2x training efficiency improvement and eventually influenced Gemini 1.5 Flash.

"Someone, Vinit Gupta, just showed up one day at my desk... We have this idea... what turned out to be the Shampoo algorithm." 00:28:38 — Rohan Anil

Yonghui Wu Rohan's manager at Google who supported the Shampoo optimizer work throughout, despite limited community enthusiasm.

"I thanked my manager, Yonghui Wu, who supported it throughout that until the end of my tenure in 2024." 00:28:38 — Rohan Anil

Rasmus and Peter Hawkins Described as among the only two humans at Google capable of writing the deep XLA/LLO kernel code required to productionize Rohan's second-order optimization work. Cited as evidence of how scarce expert kernel writers are.

"There's only like two humans at Google who could write it, Rasmus and Peter Hawkins, because it was deep XLA LLO code that you have to write to make this work. And then that took them two years to do." 00:42:39 — Rohan Anil

Jeff Hinton Referenced by Rohan in the context of biological learning, specifically the idea of model computation — that biological systems build their own circuits as they develop.

"As Jeff Hinton says, model computation. So we built our own circuit as we grow up and we built our own learning algorithm with the hardware." 00:32:00 — Rohan Anil

5. Operating Insights

Build Native to the New Tool Regime Rather Than Adapting Old Structures

Jerry's framing for Core Automation's operational philosophy is that research organizations should be rebuilt from scratch around AI-native workflows rather than retrofitting AI into existing team structures. The leverage difference is an order of magnitude.

"You can try to move the old structures around it and the team's workflows, how data is gathered. Or you can try to build, like you said, you can try to build natively for it — for processes that maximally empower each researcher and allow them to just iterate on their idea much, much quicker." 00:36:17 — Jerry Tworek

Use Your Own Product as Your Primary Evaluation Environment

Core Automation's stated evaluation methodology for their new models is their own daily research work. If the model gets measurably better at doing Core Automation's science over time, it is working. This turns the organization itself into a living benchmark — far closer to real-world distribution than any synthetic eval.

"What we are looking for is systems that learn at test time... We want to be evaluating our systems of our everyday work. They get better at doing the work of Core Automation scientists each day." 00:47:16 — Jerry Tworek

Run Competitions to Crowdsource Rare Expertise and Establish Performance Ceilings

Rather than hiring the two or three people in the world who can write high-performance GPU kernels, Core Automation ran an open competition (with GPU Mode) and produced a 60x performance improvement over the baseline — establishing both a concrete target and a benchmark for when their AI systems have reached human-expert capability.

"A human plus some search loop can get you like something like 7x. But it requires the high taste human — like there's maybe three people in the world — to spend about $100,000 on these coding agents over a span of four weeks to get to a solution that's 60x faster." 00:41:04 — Rohan Anil

6. Overlooked Insights

The Ngram Memory Architecture May Be More Important Than MoE — and Was Independently Validated by DeepSeek

Rohan briefly mentions that in 2020 he was working on replacing Transformer parameters with external ngram memory (the "ngrammer"), that internal deployed versions existed, and that DeepSeek independently arrived at the same idea and demonstrated via scaling laws that it could replace Mixture-of-Experts entirely. This was mentioned in a single throwaway sentence, but the implication is enormous: if ngram memory embeddings can replace MoE at scale, one of the two core primitives of modern large models may be replaceable with a more memory-efficient mechanism — and a small team at a Chinese lab already proved it while Rohan was still at Google. This strongly hints at where Core Automation's architectural search may be pointed.

"Six months before that, DeepSeek wrote their ngram, which is an improved version of adding more memory, showed scaling laws that, yeah, you don't need MoEs. You could have actually replaced it with these ngram embeddings. For me, that was like, oh." 00:43:53 — Rohan Anil

Optimization Methods Define Which Architectures Are Discoverable

Rohan notes almost in passing that the choice of optimizer constrains which architectures researchers can even train successfully — meaning that the field's architectural exploration has been artificially bounded by the weakness of optimizers like Adam. Stronger optimizers (like Shampoo and its successors) unlock architectures that weaker optimizers cannot train, including potentially removing residual connections and achieving genuinely deep representations. This means the optimizer is not a training detail but a gating factor on the entire architecture search space — a point that received almost no discussion despite being a direct implication of Rohan's own published work.

"Your optimization algorithm defines what architectures you will discover... Residual connections have been extremely useful for training neural networks. There are folks who have now gotten rid of them and learned deeper representations. But they needed a better optimization method." 00:30:50 — Rohan Anil