There are many interesting differences between a traditional computer program and a large language model. One of the most powerful is the non-deterministic nature of a model's output; that is, how a single prompt may produce multiple unique responses each time it is run.

To see why this is important, it is helpful to understand where models come from in the first place. LLMs like ChatGPT, Claude, and Gemini are pre-trained on terabytes of text sourced from the internet. Pre-training may take weeks or months to complete, and the result is a base model.

Base models do not behave like assistants. Imagine autocomplete on steroids - to a base model, a "chat" looks like one long document, and the model simply continues that document. Base models typically output vastly varied responses to the same prompt. This makes sense - the training dataset is so large that for most prompts, many tokens have a high probability of coming next. With each output token, the range of possible following tokens shrinks, slowly pushing the model towards a deterministic output.

Usually, the base model is then fine-tuned on a specific task, typically learing to assume the role of a 'helpful, harmless, honest assistant'. This is known as instruction tuning. Instruction tuning inherently introduces a level of determinism to the model's output - now, tokens like "Sure, I'd be happy to help you with..." have a much higher probablity of following a given prompt than most other tokens in the training dataset.

Asher is still typing...

Intrinsic Labs is invested in facilitating widespread, deep understanding of AI behavior. Latent Spaces is our first big step in that direction.