Key Takeaways

  • Pre-training hits a hard wall because humanity cannot gather enough clean data to cover every specialized job in the real world.
  • Transformers do not reason from first principles; they retrieve context across attention windows and sample outputs until they look like reasoning.
  • Reinforcement learning gives models verifiable reward systems, letting software grade its own actions and build new skills like a human.
  • Rao built and sold Nervana Systems to Intel and MosaicML to Databricks, and he expects post-training curricula to matter far more than base models.

You Cannot Pre-Train Your Way Out of Data Limits

The dominant playbook in modern AI assumes that if you feed enough text and code into a cluster of GPUs, real problem-solving falls out the other end. Naveen Rao, who founded Unconventional AI after selling MosaicML to Databricks for $1.3 billion, says that theory is dead.

“We cannot gather enough data to solve every task in the world,” Rao explains. “We simply can't. You can't pre-train your way out of this by just gathering data. It's just too hard.”

Pre-training works well when the internet has millions of examples of the target behavior. It fails when you enter niche enterprise workflows, physical hardware control, or novel engineering tasks where training samples do not exist in the public domain. Brute-force scaling was a convenient phase, but scraping web pages cannot teach an agent how to handle corner cases it has never seen.

To build software that tackles unseen tasks, you need systems that can experiment, fail, and update their behavior without a human hand-labeling a million demonstrations first.

Transformers Conflate Retrieval With Actual Reasoning

Silicon Valley still treats the transformer as the permanent architecture for artificial intelligence. Rao thinks that bet will age poorly. Attention layers are exceptional at retrieving relevant fragments from a context window, but retrieval is not the same thing as logical deduction.

“I do think we will be in a post transformer world,” Rao says. “I don't think the transformer is the be all end all. I don't think I truly understood this until maybe a couple of years ago because transformers conflate a lot of things. What we care about is reasoning.”

Right now, when an LLM writes complex code or works through a math puzzle, it is sampling tokens across context to simulate logical steps. Rao points out the mechanical waste behind this approach: “Attention basically has this context which is a retrieval thing, and then we can sample off of that and eventually get to something that looks like reasoning. I think this is extremely inefficient for how we do reasoning. We don't need to build reasoning this way.”

Real human reasoning works by forming internal hypotheses, testing them against rules, and throwing out dead ends before committing to an answer. It does not generate five paragraphs of probabilistic text to see if the ending happens to match reality. As compute costs rise and inference markets expand toward trillions of dollars, burning megawatts of electricity on retrieval tricks will become too expensive to sustain.

Verifiable Rewards Are the Real Moat

If raw data collection cannot scale indefinitely, where does real progress come from? Rao points straight to reinforcement learning paired with automated verification.

“When I have that verifiable reward, then I can start to say, okay, did the actions I took, did those actions lead to that good outcome or not?” Rao notes. “So now I have a structure that allows me to solve any set of general tasks and that's why RL is so exciting.”

The advantage does not sit in the base pre-trained weights that every foundation lab rents out over an API. The advantage belongs to the training curriculum. As Rao puts it: “The curriculum how you went from pre-training to all these different sorts of trainings to modify the behavior of the model is actually the secret sauce.”

When a model receives a clear signal on whether an output succeeded (such as whether a compiled program passes its unit tests, or whether a database query returned the exact right rows) it can run thousands of internal trials. It explores different paths, measures the result against the reward, and corrects itself. That feedback loop creates new problem-solving paths that never existed in the training corpus.

What to Do With This

Stop building workflows that rely on raw prompt engineering or fine-tuning models on static customer support logs. Audit your AI pipeline this week and identify one task with an objective, programmatic pass/fail condition, like an API call that returns a valid response code or a script that clears all syntax checks. Build an automated evaluation script that tests model outputs against that condition, and use those verified test runs to filter your training data before deploying your next agent.