Steven Pinker

Factual · AI capabilities

“What these models don’t have is a model of the world.”

Steven Pinker, interviewee · 2023 · The Intelligence and Rationality of AI and Humans: A Conversation With Steven Pinker (Harvard Data Science Review)

Large language models like GPT do not have a model of the world.

Mostly false False in substance, right in a detail.

What would count as true or false

Judged as of 2023 against the research available then. True if the evidence then showed that LLMs such as GPT have no internal representation of the world (no structured internal model of entities, states or causal relations), and studies claiming to find such representations were rejected. Mostly true if the consensus was that LLMs lack a full or robust world model, with only narrow or contested evidence of partial internal representations. Mixed if the research was genuinely split or inconclusive. Mostly false if studies from that time (e.g. probing work such as Othello-GPT, or spatial and temporal representations in LLMs) showed meaningful if limited internal world representations, so the flat denial overstates it. False if the evidence then clearly showed that LLMs do have substantial world models.

Evidence

1. GitHub - awellis/othello-world: Emergent world representations: Exploring a sequence model trained on a synthetic task GitHub (repository README, a fork of likenneth/othello_world) · 15 February 2023

   “Although the network has no a priori knowledge of the game or its rules, we uncover evidence of an emergent nonlinear internal representation of the board state.”

2. Muddles about Models Gary Marcus (Substack) · 5 October 2023

   “modern LLMs acquire structured knowledge about fundamental dimensions such as space and time, supporting the view that they learn not merely superficial statistics, but literal world models”

3. The Debate Over Understanding in AI's Large Language Models Melanie Mitchell and David C. Krakauer (arXiv, via ar5iv) · 8 October 2019

   “Those on the other side of this debate argue that large pretrained models such as GPT-3 or LaMDA—however fluent their linguistic output—cannot possess understanding because they have no experience or mental models of the world;”

4. The Debate Over Understanding in AI's Large Language Models Melanie Mitchell and David C. Krakauer (arXiv, via ar5iv) · 8 October 2019

   “Of 480 people responding, essentially half (51%) agreed, and the other half (49%) disagreed [57].”

Reasoning

E1, the Othello-GPT work from early 2023, reports evidence of an emergent internal representation of the board state in a sequence model that had no prior knowledge of the game's rules. That is a meaningful, though narrow, internal world representation. E2 quotes 2023 research claiming LLMs learn structured representations of space and time, which it calls "literal world models". E3 and E4 show that expert opinion on LLM understanding was split. Under the criteria, the board-state representation found in a sequence model trained on the Othello game (E1) and the reported space and time representations in LLMs (E2) weigh against a flat denial, though expert opinion at the time was split (E3, E4); the claim is rated mostly false.

Cite this claim

Steven Pinker, “Large language models like GPT do not have a model of the world.” (said 2023). Verdict: mostly false. Reliability API, claim c_9fcd0d38e428, method 0.1.0+28da144b. https://epistacklabs.com/claims/c_9fcd0d38e428