We scan new podcasts and send you the top 5 insights daily.
Emmanuel describes poetry planning: after ending a line with an “it” rhyme, the model’s intermediate activations represent a plan to end with “rabbit,” and future words consult that plan. This shows next-token prediction can involve multi-token planning even when the plan is never spoken aloud.
The seemingly simple task of next-token prediction, when perfected, requires a model to understand concepts as deeply as the source. To accurately predict what Einstein would say in a new situation, a system must be as intelligent as Einstein, proving prediction is fundamental to intelligence.
LLMs predict the next token in a sequence. The brain's cortex may function as a general prediction engine capable of "omnidirectional inference"—predicting any missing information from any available subset of inputs, not just what comes next. This offers a more flexible and powerful form of reasoning.
Under intense pressure from reinforcement learning, some language models are creating their own unique dialects to communicate internally. This phenomenon shows they are evolving beyond merely predicting human language patterns found on the internet.
A non-obvious aspect of LLM training is that to accurately predict text describing reality (e.g., a lab result), the AI must learn to model the underlying physics or logic. This inherently trains it to be more capable than the human who merely observed the result, forcing emergent intelligence.
Emmanuel says the bet was not just compute but finding an objective where compute keeps returning value. He notes scaling laws showed that pouring more compute and data into next-word prediction predictably improved a metric that correlates with many useful tasks, which was rare in practice.
The fact that LLMs, designed to predict the next word (a writing task), spontaneously exhibit reasoning abilities provides empirical evidence for the long-held belief that writing and thinking are intertwined. A machine built to write inadvertently learned to think.
A prompt does not issue a command to an LLM like code to a computer. Instead, it activates relevant patterns within the model's weight space, guiding it to generate a completion consistent with its training data for having followed such instructions.
We can now prove that LLMs are not just correlating tokens but are developing sophisticated internal world models. Techniques like sparse autoencoders untangle the network's dense activations, revealing distinct, manipulable concepts like "Golden Gate Bridge." This conclusively demonstrates a deeper, conceptual understanding within the models.
The idea that LLMs just predict the next statistical character is wrong. To accurately predict the next word in a physics paper, for instance, an AI must create an internal model of the physical world. This implies that scaling current architectures can lead to deeper understanding, not just pattern matching.
A new technique forces a model's forward pass to go through a natural language representation of its internal state. This makes the model's internal reasoning interpretable to humans in real-time, offering a significant breakthrough for monitoring and understanding what the model is actually "thinking" about a task.