Melanie Mitchell, Subbarao Kambhampati, Emily Bender, Ernie Davis and others have long emphasized that LLMs have principled limits.

Judge as of 2024-11-09. True: each named person (Melanie Mitchell, Subbarao Kambhampati, Emily Bender, Ernie Davis) published or publicly argued, repeatedly and over several years before that date (roughly 2020 or earlier onward), that LLMs have fundamental or principled limits, e.g. in reasoning, planning, understanding or meaning. Mostly true: all four argued for fundamental limits, but for one of them the record is shorter or the framing is weaker (practical rather than principled limits). Mixed: about half of them have such a record. Mostly false: only one has a long record of arguing principled limits, while the others raised only limited or recent concerns. False: none of them made such arguments, or they publicly rejected the idea of principled limits.

Set before the research began.

1. E1 Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data Association for Computational Linguistics (ACL 2020 proceedings) · 4 July 2020

   “In this position paper, we argue that a system trained only on form has a priori no way to learn meaning.”

2. E2 The Debate Over Understanding in AI's Large Language Models arXiv (authors: Melanie Mitchell and David C. Krakauer, Santa Fe Institute) · 10 February 2023

   “Indeed, these are precisely the abilities lacking in current AI systems, including state-of-the-art LLMs, though ever-larger LLMs have exhibited limited sparks of these general abilities.”

3. E3 Why large language models aren't headed toward humanlike understanding Science News · 28 February 2024

   “Humans are still very good at solving letter-string problems. But LLMs usually fail. They are not able to generalize the concepts used on an alphabet they know to another alphabet.”

4. E4 Why large language models aren't headed toward humanlike understanding Science News · 28 February 2024

   “She does not deny that AI might someday reach a similar level of intelligent understanding.”

5. E5 The Debate Over Understanding in AI's Large Language Models arXiv (authors: Melanie Mitchell and David C. Krakauer, Santa Fe Institute)

   “Indeed, these are precisely the abilities lacking in current AI systems, including state-of-the-art LLMs, though ever-larger LLMs have exhibited limited sparks of these general abilities.”

6. E6 The Debate Over Understanding in AI's Large Language Models arXiv (authors: Melanie Mitchell and David C. Krakauer, Santa Fe Institute)

   “no experience or mental models of the world; their training in predicting words in vast collections of text has taught them the form of language but not the meaning [8, 9, 55].”

7. E7 The Debate Over Understanding in AI's Large Language Models arXiv (authors: Melanie Mitchell and David C. Krakauer, Santa Fe Institute)

   “E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell. On the dangers of stochastic parrots: Can language models be too big?”

8. E8 Position: LLMs Can’t Plan,But Can Help Planning in LLM-Modulo Frameworks ar5iv (arXiv paper mirror) · 5 August 2023

   “We argue that auto-regressive LLMs cannot, by themselves, do planning or self-verification (which is after all a form of reasoning)”

9. E9 Position: LLMs Can’t Plan,But Can Help Planning in LLM-Modulo Frameworks ar5iv (arXiv paper mirror) · 5 August 2023

   “Even from a pure engineering perspective, a system that takes constant time to produce the next token cannot possibly be doing principled reasoning on its own.”

10. E10 Linguistics, Automated Systems, & the Power of AI, with Emily M. Bender Carnegie Council (AI & Equality podcast) · 17 June 2024

    “I get so frustrated when I see people blithely “asking” ChatGPT, putting a prompt into ChatGPT and taking its output as if it were interesting about anything.”

11. E11 Writings for a General Audience: Ernest Davis Ernest Davis (NYU Computer Science homepage) Own record

    “Commonsense Reasoning and Commonsense Knowledge in Artificial Intelligence by Ernest Davis and Gary Marcus. CACM, September 2015, 58:9, 92-105.”

The evidence doesn't show that each of the four named people has a multi-year record of arguing that LLMs have principled limits. Mitchell's paper with Krakauer (E2) argues that current LLMs lack certain general abilities, and a Science News article (E3) reports that LLMs usually fail to generalize letter-string concepts; these items date from 2023–2024, and they are hedged: E2 allows 'limited sparks' of those abilities, and E4 says she does not deny AI might someday reach such understanding. The 2020 'form vs. meaning' paper (E1) and the 'LLMs Can't Plan' position paper (E8, E9) make in-principle arguments, but the quoted text doesn't name their authors, so they can't be credited to Bender or Kambhampati. No item shows Ernie Davis arguing anything about LLM limits (E11 is a 2015 commonsense-reasoning article), so the evidence can't settle the claim either way.