Reinforcement learning incentivizes AIs to find the right answer, not just mimic human text. This leads to them developing their own internal "dialect" for reasoning—a chain of thought that is effective but increasingly incomprehensible and alien to human observers.

Even with vast training data, current AI models are far less sample-efficient than humans. This limits their ability to adapt and learn new skills on the fly. They resemble a perpetual new hire who can access information but lacks the deep, instinctual learning that comes from experience and weight updates.

Reinforcement learning achieves superhuman results not by inventing alien concepts, but by surfacing and combining rare behaviors that are already possible within a model's vast pre-trained distribution. The goal of pre-training is to make this search for novel solutions more efficient and less random.

The popular conception of AGI as a pre-trained system that knows everything is flawed. A more realistic and powerful goal is an AI with a human-like ability for continual learning. This system wouldn't be deployed as a finished product, but as a 'super-intelligent 15-year-old' that learns and adapts to specific roles.

Solving key AI weaknesses like continual learning or robust reasoning isn't just a matter of bigger models or more data. Shane Legg argues it requires fundamental algorithmic and architectural changes, such as building new processes for integrating information over time, akin to an episodic memory.

The current focus on pre-training AI with specific tool fluencies overlooks the crucial need for on-the-job, context-specific learning. Humans excel because they don't need pre-rehearsal for every task. This gap indicates AGI is further away than some believe, as true intelligence requires self-directed, continuous learning in novel environments.

While both humans and LLMs perform Bayesian updating, humans possess a critical additional capability: causal simulation. When a pen is thrown, a human simulates its trajectory to dodge it—a causal intervention. LLMs are stuck at the level of correlation and cannot perform these essential simulations.

To bridge the learning efficiency gap between humans and AI, researchers use meta-learning. This technique learns optimal initial weights for a neural network, giving it a "soft bias" that starts it closer to a good solution. This mimics the inherent inductive biases that allow humans to learn efficiently from limited data.

A major flaw in current AI is that models are frozen after training and don't learn from new interactions. "Nested Learning," a new technique from Google, offers a path for models to continually update, mimicking a key aspect of human intelligence and overcoming this static limitation.