LLMs Are Dumb.

🌈 Abstract

The article discusses the limitations of current large language models (LLMs) and argues that they are closer to databases than true intelligence. It highlights the inability of LLMs to reason and solve novel problems, and suggests that they are primarily good at memorizing and regurgitating patterns rather than understanding the underlying causal structures. The article also outlines the two key frontiers that AI must conquer to approach human-level intelligence: compression (moving from mere memorization to true regularization) and long-inference models (allowing models to iterate and converge on the best solution). The article criticizes the exaggerated claims and hype surrounding the current state of AI and calls for a more realistic and pragmatic approach to using these models.

🙋 Q&A

[01] The Ultimate Gaslighting

1. What is the main issue with the current state of the AI industry according to the article?

• The article argues that the AI industry is engaging in "gaslighting" by raising billions of dollars for "frontier AI" that does not meet expectations, and then lying to justify the investments.

2. How does the article characterize the performance of LLMs?

• The article states that LLMs are "dumb" and "closer to databases than they are to humans", as they primarily "vomit" memorized patterns rather than demonstrate true reasoning capabilities.

3. What is the author's critique of the common benchmarks used to measure LLM intelligence?

• The author argues that benchmarks like MMLU can be "mostly aced with simple memorization", and do not actually test for genuine intelligence or reasoning abilities.

[02] LLMs Can't Reason

1. What is the ARC-AGI benchmark, and how do LLMs perform on it?

• The ARC-AGI benchmark is similar to an IQ test, where the model must generalize a pattern from a small subset of examples. The article states that LLMs "fail miserably" on this test, as they rely on memorization and are highly sample-inefficient.

2. How do LLMs perform on the "Alice in Wonderland" test of inductive reasoning?

• The article provides an example where a GPT-4 model fails to correctly infer that Alice should be included in a "sister group", despite having all the necessary information. This demonstrates LLMs' inability to apply even simple reasoning chains over their data.

3. What does the author mean when they say LLMs are "databases" rather than intelligent systems?

• The author argues that LLMs can only perform correctly when the specific word pattern has been seen before, similar to how a database can only retrieve information that has been explicitly stored, rather than being able to reason and infer new facts.

[03] On the Path to the 'I' in 'AI'

1. What are the two key frontiers that AI must conquer to approach human-level intelligence?

• The two frontiers are:
  • Compression: Progressing from mere memorization to true regularization, where the model learns the underlying causal structures.
  • Long-inference models: Allowing models to iterate and converge on the best solution, rather than just responding with the first thing that "comes to mind".

2. What are some of the methods researchers are proposing to improve the reasoning capabilities of LLMs?

• The article mentions data augmentation, over-extended training (to allow for regularization), and test-time computation (allowing models to search for solutions before answering).

3. What are the two key elements the author believes are still missing for LLMs to overcome their training data and generate truly novel solutions?

• Depth: The ability to train LLMs to be superhuman at specific tasks, rather than just being good at many things but great at none.
• Active inference: The ability for LLMs to learn and adapt as they make predictions about the world, rather than just learning during the training phase.