The Next Frontier: Why AI Must Move Beyond Classroom Training

The current AI paradigm relies on a flawed premise: that intelligence can be perfected in a vacuum. Labs are spending billions on grindable environments, such as coding, where tasks are reproducible and verifiable. They hope that scaling these will lead to AGI. However, this creates classroom intelligence that fails when it hits the messy, unpredictable reality of business, law, or politics. The consequence is a massive, wasted opportunity. Our models are deployed across the global economy, yet they remain static and unable to learn from their own operational experiences. The real competitive advantage will not go to the lab with the most compute, but to the one that masters continual learning, or the ability for an AI to distill real-world lessons back into its core weights. For technical leaders, this means the era of static model deployment is ending.

The Grindability Trap

We often wonder why computer use, such as booking travel or filing taxes, has lagged behind progress in coding. The answer is not just data quality. It is the lack of grindability. Coding is easy to automate because you can spin up 1,000 identical containers to test a hypothesis. You cannot do the same with a live website or a real-world business negotiation.

It is not enough for a domain to be verifiable. It also has to be very grindable. In the sense that you have to be able to run lots of parallel rollouts against a deterministic and replayable simulator.

-- Dwarkesh Patel

This creates a structural bottleneck. Because current models are inefficient, they require massive, repeatable simulations to learn. If you cannot build a simulator for a domain, such as winning an election or building a startup, the model cannot learn. We are training AIs to be elite test-takers while the real world demands a practitioner.

The Hidden Cost of In-Context Reliance

The current industry trend is to shove everything into the context window. The logic is that if a new employee takes six months to become productive, we should just fit those six months of experience into a massive context window.

This is a fragile solution. As models scale to handle longer contexts, we hit diminishing returns. Short-horizon training does not necessarily lead to long-horizon performance. Relying on context windows is like trying to memorize an entire encyclopedia to solve a problem instead of learning the underlying principles. It is temporary. Once the session ends, the intelligence gained is lost.

Why On-the-Job Learning is the Real Moat

The most valuable data exists only in deployment: the specific failure modes of your organization, the nuances of your internal infrastructure, and the unique problems your users face. Currently, 30% to 50% of a lab's compute goes to inference. This compute is essentially wasted because it does not improve the base model.

To bridge this, we need to move beyond simple inference and toward techniques like On-Policy Self-Distillation (OPSD). Instead of trying to memorize every interaction, OPSD allows a model to distill the veteran insights gained during a session back into its core weights.

The way you get better at your job is not by recalling the transcript of every single thing that happened every day with perfect fidelity, rather it is by consolidating the handful of insights and pieces of knowledge that are actually relevant to you getting better at your job.

-- Dwarkesh Patel

This is a shift from data collection to knowledge compression. It is the difference between a student who records every lecture and one who learns the intuition behind the subject. The latter is far more capable of handling the ambiguity of the real world.

The 2027 Horizon: From Dreaming to Deployment

The next breakthrough is not just bigger models. It is dreaming, where an AI builds its own internal simulations of reality to rehearse skills before applying them in the real world. If successful, this creates a feedback loop. The model is deployed, it encounters a novel problem, it dreams up a simulation to master that problem, and then it distills those lessons into its weights.

Over the next few years, the primary way AI will improve will shift from pre-training to on-the-job learning. Every interaction will make the model smarter, not just for that user, but for the entire system. This creates a compounding advantage that is difficult for competitors to replicate, as it requires an architecture that can learn continuously without forgetting its foundational knowledge.

Key Action Items

• Audit your Human-in-the-Loop data (Immediate): Identify where your AI agents are currently making decisions that you are reviewing. This feedback, the thumbs up or down, is the precursor to the distillation signals that will eventually train your future models.
• Prioritize Grindable Internal Workflows (Next Quarter): If you are building AI agents, focus on tasks that can be sandboxed in deterministic environments. This allows you to accumulate the rehearsal data needed for the next generation of RL training.
• Shift focus from Context Size to Architectural Durability (6-12 Months): Stop betting your long-term strategy on massive context windows. As the industry moves toward weight-based continual learning, ensure your infrastructure can support models that learn and update rather than just remembering via cache.
• Invest in Distillation Pipelines (12-18 Months): As OPSD and similar techniques become more accessible, prepare your data pipelines to move from simple SFT to distillation-based learning, which preserves existing knowledge while integrating new, job-specific insights.
• Prepare for the Learning Moat (18+ Months): Recognize that the ultimate advantage will be the model that learns most effectively from its own deployment. Start treating your AI deployment not as a finished product, but as a student that should be getting smarter with every interaction.