next token prediction
The training objective of predicting the next unit of text given prior context, used to train language models.
As defined in AMA | June 2023
Other definitions
The core training objective of large language models, which learn to predict the next word or token in a sequence.
Sarah Guo - Funding the Frontier - [Invest Like the Best, EP.489]
Where it comes up · 10
- Dario Amodei (Anthropic CEO) — The hidden pattern behind every AI breakthrough4 claims
- Richard Sutton – Father of RL thinks LLMs are a dead end2 claims
- Richard Sutton – Father of RL thinks LLMs are a dead end2 claims
- Sarah Guo - Funding the Frontier - [Invest Like the Best, EP.489]2 claims
- Lex Fridman Podcast #501 - DHH2 claims
- Fully autonomous robots are much closer than you think – Sergey Levine1 claim
- Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken1 claim
- I spent 400+ hours in Cursor, here’s what I learned1 claim
and 2 more
Most important claims · 10 of 17Search all claims for “next token prediction” →
- When choosing between two solutions, prompting Claude to write a detailed paragraph arguing for each before concluding counters its tendency to be a 'yes man' that agrees with whatever it senses the user wants or its own previous reasoning, because once it recommends a solution in the first 5-10% of its response the rest is committed.I spent 400+ hours in Cursor, here’s what I learned
- Although RL is in principle better than imitation learning, current models use imitation/supervised learning first because effective learning from one's own experience requires already knowing something about the task; supervised learning builds the prior-knowledge foundation that lets the model later figure things out quickly via RL, mirroring how LLMs go from next-token prediction to RL.Fully autonomous robots are much closer than you think – Sergey Levine
- There is evidence that large language models can acquire a form of meaning despite being trained only to predict the next token, which connects to emergence because higher-level semantic competence appears as a collective property not explicitly programmed, illustrating how complex capabilities can emerge from simple objective functions.AMA | June 2023
- The alignment problem may be partially solved 'by default' through properties of large language models that naturally develop common sense understanding of goals from next-token prediction training, making them less prone to the Malign Omniscience failure modes that earlier alignment researchers feared.AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo
- Critics cannot charge AI both with being non-deterministic and therefore unreliable and with being uncreative; non-determinism (temperature) is what enables creativity, as human thought also cannot be predicted in advance.Lex Fridman Podcast #501 - DHH
- Next-token prediction is not a substantive goal because it does not change the world; tokens come at the model and predicting them does not influence them, so the model has no goal about the external world.Richard Sutton – Father of RL thinks LLMs are a dead end
- Legal work is a strong fit for next-token language models because law is structured language involving document-heavy retrieval, text generation, and large amounts of precedent, which made Harvey a rational early AI application bet by late 2022.Sarah Guo - Funding the Frontier - [Invest Like the Best, EP.489]
- Alignment and values are not guaranteed to emerge from scaling because a model trained to predict the world learns facts, not values; predicting the next token leaves free variables about what one should do, think, or value that the training signal does not constrain, so these will not emerge on their own.Dario Amodei (Anthropic CEO) — The hidden pattern behind every AI breakthrough
- Conviction chose application areas where models were a good technical fit before starting from customer problems, which many investors considered nonsense, because the firm believed capability growth would create non-obvious value.Sarah Guo - Funding the Frontier - [Invest Like the Best, EP.489]
- Next token prediction is not a goal because it doesn't change the world; tokens come at you, and if you predict them, you don't influence them—it's not a substantive goal about the external world.Richard Sutton – Father of RL thinks LLMs are a dead end