GPT-3 › What GPT-3 Did Not Do

It is worth being clear about what GPT-3 did not achieve, beyond the general limitations discussed earlier. Understanding the specific nature of GPT-3's failures helps clarify what problems remained for future work.

GPT-3 did not develop a persistent world model that it updates as new information arrives. Each query is processed independently; the model has no memory of previous interactions. This means it cannot track changes in the world, maintain consistent beliefs across a conversation, or update its "knowledge" based on corrections you provide in the prompt beyond what fits in the context window.

GPT-3 did not develop reliable calibration. A well-calibrated model should be uncertain about things it does not know and confident about things it does know. GPT-3 often shows the opposite pattern: high confidence about incorrect facts and sometimes hedging unnecessarily about correct ones. The model has no separate "uncertainty module"; its expressed confidence is a product of the token probabilities it assigns, which reflect training data frequency more than epistemic warrant.

GPT-3 did not resolve the grounding problem. Its representations are built entirely from language and have no connection to perceptual experience, physical interaction, or direct observation of the world. All of its "knowledge" is mediated through text, and all of its generation is conditioned on textual patterns. This means it can use words correctly in context without having any model of what those words refer to in the physical world. A model that has never seen a banana but has read millions of sentences about bananas has a different kind of "knowledge" of bananas than a model that has been trained on visual or tactile data.

GPT-3 did not achieve instruction following in a reliable, consistent way. While it could often understand and respond to instructions in prompts, its compliance was uneven and depended heavily on how instructions were phrased. It would sometimes ignore instructions, produce outputs that technically satisfied the letter but not the spirit of instructions, or generate outputs harmful or offensive in ways that violated reasonable expectations. These issues motivated the development of instruction tuning and RLHF, which are central to the design of GPT-4, Claude, and other subsequent models.

These limitations are not failures of GPT-3's design so much as they are clear statements of what next-token prediction on text data alone cannot achieve. They define the research agenda that occupied the field in the years following GPT-3's release and shaped the architectural and training innovations of the next generation of large language models.