Gary Marcus

Prediction · AI capabilities

“If so, we can expect that approach to still be marred by something fundamental: an inability to construct internal models of how the world works, and in consequence we should anticipate an inability to understand things at an abstract level.”

Gary Marcus, author · 25 December 2022 · What to Expect When You're Expecting … GPT-4 (Marcus on AI)

If GPT-4 is architecturally essentially the same as GPT-3, it can be expected to be unable to construct internal models of how the world works and, in consequence, unable to understand things at an abstract level.

Mostly true True in substance, wrong in a detail.

What would count as true or false

Judge against GPT-4 as released (March 2023) and as evaluated in the following months of 2023. True: credible evaluations and research in 2023 widely found that GPT-4 lacked robust internal world models and failed at abstract understanding (e.g. systematic failures on abstraction benchmarks such as ARC, physical/spatial reasoning, planning), and the balance of expert evidence supports both inability claims. Mostly true: GPT-4 showed clear, systematic failures at world modelling and abstraction, but some evidence showed partial or emergent world-model-like competence in some domains, so 'inability' is somewhat overstated. Mixed: the evidence is roughly balanced, with substantial findings both of world-model/abstract competence and of fundamental failures, or one of the two parts (world models, abstraction) is supported and the other is not. Mostly false: the predominant evidence shows GPT-4 could form usable internal models and handle many abstract tasks, with only narrow failures. False: GPT-4 was clearly shown to construct reliable internal world models and understand abstractions at a broadly human level. The claim is hedged ('if so'): its premise is that GPT-4 is a scaled-up LLM, which holds.

Evidence

1. Comparing Humans, GPT-4, and GPT-4V On Abstraction and Reasoning Tasks arXiv (authors: Melanie Mitchell, Alessandro B. Palmarini, Arseny Moskvichev, Santa Fe Institute) · 12 September 2023

   “Our experimental results support the conclusion that neither version of GPT-4 has developed robust abstraction abilities at humanlike levels.”

2. LLMs and the Abstraction and Reasoning Corpus: Successes, Failures, and the Importance of Object-based Representations arXiv (authors: Yudong Xu, Wenhao Li, Pashootan Vaezipoor, Scott Sanner, Elias B. Khalil)

   “GPT-4 solves only 13/50 of the most straightforward ARC tasks when using textual encodings for their two-dimensional input-output grids.”

3. Inherent limitations of GPT-4 regarding spatial information arXiv · 20 November 2023

   “As the map size and the obstacles ratio increase, the precision in finding the shortest path decreases.”

4. GPT-4 Can't Reason (Position Paper) ar5iv (arXiv mirror) · 26 June 2023

   “Based on the results of that analysis, this paper argues that, despite the occasional flashes of analytical brilliance, GPT-4 at present is utterly incapable of reasoning.”

Reasoning

An independent 2023 evaluation concluded that neither version of GPT-4 had developed robust abstraction abilities at humanlike levels [E1]. It solved only 13 of 50 simple ARC tasks when given text encodings [E2]. Its shortest-path precision fell as map size and obstacles increased, which points to weak spatial world modelling [E3], and one position paper argued it could not reason [E4]. The same sources also report partial success (13/50 tasks) and "occasional flashes of analytical brilliance", so calling it a blanket "inability" overstates the case somewhat.

Cite this claim

Gary Marcus, “If GPT-4 is architecturally essentially the same as GPT-3, it can be expected to be unable to construct internal models of how the world works and, in consequence, unable to understand things at an abstract level.” (said 25 December 2022). Verdict: mostly true. Reliability API, claim c_cdf244641a1a, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_cdf244641a1a