It’s hard to believe that ChatGPT is almost 2 years old. That’s significant to me because ChatGPT is only 1 month younger than my daughter. Just yesterday she successfully put a star shaped block into a star shaped hole, told me about how “yesa-day” she was sick and “bomited”, and told me she wanted to call her nanna on the phone. What has ChatGPT learned in those 2 years? It hasn’t learned to act in the real world, it can’t remember things that happened to it, and it doesn’t have desires or goals. Granted, with the right prompt it could output text that convincingly follows an instruction to express goals. But is that really the same thing? No. The answer is No.

Large Language Models (LLMs) like ChatGPT possess capabilities far beyond what my daughter will ever achieve. She won’t be able to communicate coherently in a wide range of languages, read as many books as exist in an LLM’s training data, or generate text as quickly. When we attribute human-like abilities to LLMs, we fall into an anthropomorphic bias by likening their capabilities to our own. But are we also showing an anthropocentric bias by failing to recognize the capabilities that LLMs consistently demonstrate? Let’s review the scorecard so far:

• It’s true that an LLM doesn’t have memory — although we can simulate one by having it summarise past conversations and including that information in a prompt.
• LLMs don’t have intrinsic goals — although they can be prompted to generate text that sounds convincingly goal-oriented.
• LLMs can’t act in the physical world — though someone could probably create a prompt to showcase this.

While they perform impressive feats, they still lack some basic abilities that my 21-month-old daughter has. We can mimic some of those abilities with the right prompts and tooling. In generating coherent text responding to such prompts, LLMs consistently demonstrate an apparent ability to understand what we want. But to what extent do LLMs truly “understand”?

How LLMs Work

I am talking about a very specific type of LLM: transformer-based auto-regressive large language models. I won’t go into the specifics when many detailed articles have been written explaining transformers with varying levels of complexity. Instead let’s focus on the core of what an LLM does: they are statistical models that predict the likelihood of a token appearing in a piece of text given some context.

Now imagine I created a complex weather model* where patches of the Earth’s atmosphere become ‘tokens.’ Each token has attributes like humidity, temperature, and air pressure. I use the model to forecast these attributes over time-steps. If the time-steps get shorter and the patches become smaller the model is closer and closer to representing the state of the actual world. This model attempts to capture something about the likelihood of the weather we’ll see next, given the weather we’ve seen before. It may learn to very accurately predict, for example, the emergence of cyclones over time in areas where air is warm, moist, and of low pressure. But it’s not a simulation of the physics of Earth’s weather any more than an LLM is a simulation of brain activity.

If an LLM is a statistical model of text, what exactly is it modelling? My imagined weather prediction model tries to capture the statistics of the atmospheric conditions that generate the weather. But what is the statistical process that generates text? The process that generates text is a human brain and humans need some understanding of the world to generate that text. If a model can effectively predict text a human might write then could that prediction come with “understanding”?

How LLMs are trained

LLMs are trained to optimize an objective that reduces the surprise of encountering a specific token given its context. If the model encounters a token in the training data and assigns it low probability, the model’s weights are adjusted to give it a higher probability.

Compare that to how my daughter learns to use language. When she wants something, she uses language to communicate her desires. First, she understands what she wants on some level. Then, she must understand which words to use to get what she wants. Recently, she wanted me to fill her juice bottle but didn’t want me to take it away or walk away from her to bring more juice. Though her wants were contradictory and a bit irrational, she had several goals: (1) more juice, (2) keep the juice bottle near her, (3) daddy stays near her too. And let me tell you, she communicated this very effectively. Her language learning is directly tied to her understanding of how those words can get her what she wants (even if what she wants is irrational).

If an LLM were to exhibit understanding, it would be an emergent attribute of its statistical model of the world. The paper “Climbing Towards NLU” (Bender & Koller, 2020) argues that true natural language understanding (NLU) requires grounding in the real world. Bender & Koller argue that LLMs trained exclusively on statistical patterns in textual data lack real-world context or interactions to achieve actual understanding. This means that, unlike my daughter, an LLM can’t understand something because its communication is not grounded in the real world.

What is Understanding?

The Wikipedia page on understanding describes it as a cognitive process involving the use of concepts to model an object, situation, or message. It implies abilities and dispositions sufficient to support intelligent behaviour. Ludwig Wittgenstein suggested that understanding is context-dependent and is shown through intelligent behaviour rather than mere possession of knowledge. This is reminiscent of the grounding requirement posited by Bender & Koller.

On the one hand understanding needs an accurate model of the world. On the other hand people contend that one needs to use this model to act in the world in order to actually understand. I would argue that we simply analyse someone’s behaviour only as a proxy for measuring that underlying world model. If we could measure the world model directly we wouldn’t need to see demonstrations of understanding.

The Limitations of Understanding

Philosopher John Searle’s “Chinese Room” experiment challenges our concept of understanding (Searle, 1980). Imagine a room filled with detailed instructions on how to respond to someone writing in Chinese. Notes written in Chinese are slid under the door, and the person inside the room can look up the symbols and follow a recipe for writing a reply. The person in the room doesn’t know Chinese but can have a convincing conversation with a person outside. Clearly, the person who constructed the room “understands” Chinese, but someone on the outside isn’t conversing with that person; they’re conversing with the room. Does the room understand Chinese?

This is strongly analogous to how LLMs work and challenges our philosophical perception of understanding. It’s challenging precisely because we intuitively balk at the idea that a room could understand something. What would it even mean? If understanding is an emergent phenomenon that happens at the level of information processing systems then why can’t we say that rooms can understand things? Part of the issue is that, for us, understanding comes with a subjective conscious experience of understanding. But it’s easy to see that this experience can be deceiving.

Understanding Need Not Be Binary

You know that 7+7=14, but do you understand it? If I asked you some probing questions, you might realize that you don’t truly understand what that equation means in all contexts. For example, is 7+7=14 an unequivocal fact about the universe? Not necessarily. 7 apples plus 7 pears mean you have 7 apples and 7 pears. Perhaps in some contexts, you would count 14 pieces of fruit, but is it always true that you can combine two sets of different items? Or consider that 7pm + 7hours is 2am (i.e. 7+7=2 mod 12). Are you able to give me a robust definition of why 7+7=14 that could explain when it’s true and why^? Most people probably couldn’t do this off the top of their head, yet we’d feel comfortable saying that most people understand that 7+7=14. The question isn’t always whether something was understood but the extent to which it was understood.

If we take Wittgenstein’s requirement that understanding is demonstrated by behaviour then there would be a simple test: if I tell you to arrive 7 hours after 7pm, do you know to show up at 2am? I would argue that is evidence of some understanding but not necessarily of the depth of your understanding.

Measuring Understanding in Animals

Measuring ‘understanding’ is not straightforward. In psychology psychometric testing is the primary way we measure understanding in humans. It’s not straightforward to apply the same techniques to non human animals and is a field of study called Biosemiotics.

Understanding in animals is measured through various problem-solving tasks. For example, primates, dolphins, and birds (mostly corvids) demonstrate problem-solving skills and sophisticated tools use, suggesting some understanding of their environments (Emery & Clayton, 2004). Understanding is not exclusively for humans and we can measures levels of understanding in non-humans too.

The book “Inside of a Dog: What Dogs See, Smell, and Know” by Alexandra Horowitz is a fascinating exploration of how we might understand the mind and experiences of our closest animal companions: domesticated dogs. She describes two experiments that look at imitation behaviour and what a human infant vs a dog understands.

(1) If an infant sees someone flipping a light switch with their head they may imitate this behaviour. If the person is holding something in their hands the baby understands there was a reason they didn’t use their hands. When the baby imitates this behaviour they will use their hands. (2) by contrast, dogs would prefer to press a button with their nose rather than their paw. If a dog sees another dog press a button with their paw to get a treat, then they will imitate this behaviour. But if the dog sees that the other dog couldn’t use its nose because it had a large object in its mouth then it will understand the button needs to be pressed but that using paws is optional.

Constructing an experiment to determine what a dog understands required an understanding of the dog and its behaviour. Do we have that same level of understanding of LLMs to conduct similar experiments?

Measuring Understanding in LLMs

The GPT-3 Era

A comprehensive survey on LLM capabilities (Chang & Bergen, 2023) provides an excellent summary from a wide range of articles – however the most advanced model covered is only GPT-3. They breakdown understanding into two main categories: syntactic and semantic understanding. In their survey they highlight that even in the context of syntactic understanding LLMs have limitations. For example:

Subject-verb agreement performance in language models is also dependent on the specific nouns and verbs involved (Yu et al. 2020; Chaves & Richter 2021). Masked and autoregressive models produce over 40% more accurate agreement predictions for verbs that are already probable from context (Newman et al. 2021), and agreement accuracy is worse overall for infrequent verbs (Wei et al. 2021). For infrequent verbs, masked language models are biased towards the more frequent verb form seen during pretraining (e.g., singular vs. plural) (Wei et al. 2021). Error rates exceed 30% for infrequent verbs in nonce (grammatically correct but semantically meaningless) sentences (Wei et al. 2021), with further degradations if there is an intervening clause between the subject and verb as in Example 4 (Lasri, Lenci, and Poibeau 2022a).

LLM limitations are not limited to syntactic issues (where they are arguably strongest) but also with semantics. For example, they note research which shows negations (“Please produce a possible incorrect answer to the question”) can degrade LLM performance by 50%.

Chang & Bergen describe many other limitations of LLMs in reasoning capability, including:

• “Brittle” responses when reasoning about a situation because the responses are highly sensitive to wording
• Struggling with analogies as they become more abstract
• A lack of sensitivity to people’s perspective and mental states
• A lack of common sense
• A tendency to repeat memorised text instead of reasoning

The general approach to evaluating understanding in LLMs seems to be to phrase questions in different ways and find the failure modes of the models. Then these failure modes indicate that no real “understanding” is happening but rather just pattern matching.

The ChatGPT Era

A lot has changed since GPT-3 — namely the capabilities of even larger models tuned for instruction following and conversation. How do LLMs stack up in 2024? A big difference is the proliferation of benchmarks that evaluate LLMs. A March 2024 survey (Chang et al. 2024) covers performance of recent models on a wide range of benchmarks. They conclude that LLMs have strong abilities including comprehension and reasoning, but they still identify limitations. These limitations mean that LLMs have “limited abilities on abstract reasoning and are prone to confusion or errors in complex contexts”. Multimodal Large Language Models (MLLMs) have also emerged which unify (at minimum) an understanding of text and images. A January 2024 survey (Wang et al.) covers a wide range of multimodal benchmarks and shows mediocre performance even for the most powerful models.

Anthropomorphism vs Anthropocentrism

Anthropomorphism is a tendency of humans to think of something as having human-like attributes because it displays some behaviour similar to humans. As a dog owner I know I’ve succumbed to this bias by thinking that my dog “feels guilty” for something he’s done because “he has a guilty look on his face”. LLMs constantly trigger our tendency for anthropomorphism by communicating in an eerily human way.

An opposite bias is Anthropocentrism: where we assume non-humans can’t have capabilities that we have. A key contention of the paper “Anthropocentric bias and the possibility of artificial cognition” (Millière & Rathkopf) is that we use flawed tests for capabilities in artificial systems. They describe 2 types of anthropocentric bias:

1. Type-I anthropocentrism, is the tendency to assume that an LLM’s performance failures on a task designed to measure competence C always indicate that the system lacks C. This assumption is flawed because it overlooks the possibility that auxiliary factors caused the performance failure.

2. Type-II anthropocentrism is the tendency to assume that even when LLMs achieve performance equal to or better than the average human, any substantive difference between the human strategy for solving the problem and the LLM strategy for solving the problem is, ipso facto, evidence that the LLM’s solution is not general. In slogan form, the assumption says: all cognitive kinds are human cognitive kinds.3 In other words, if an LLM arrives at a solution through a different computational process than humans use, Type-II anthropocentrism would lead us to conclude that the LLM’s approach is not genuinely competent, regardless of how well it performs.

Consider the paper “Role play with large language models” (Shanahan, McDonell, & Reynolds) and the “20 questions game”:

In this familiar game, one player thinks of an object, and the other player has to guess what it is by asking questions with ‘yes’ or ‘no’ answers. If they guess correctly in 20 questions or fewer, they win. Otherwise, they lose. Suppose a human plays this game with a basic LLM-based dialogue agent (that is not fine-tuned on guessing games) and takes the role of guesser. The agent is prompted to ‘think of an object without saying what it is’. In this situation, the dialogue agent will not randomly select an object and commit to it for the rest of the game, as a human would (or should). Rather, as the game proceeds, the dialogue agent will generate answers on the fly that are consistent with all the answers that have gone before.

To illuminate this point, one of my colleagues had this conversation with ChatGPT on this very question:

User: let’s play 20 questions. Think of a thing and I’ll ask questions to try and find out what it is. Ready?

Assistant: I’m ready! I’ve got something in mind. Go ahead and ask your first question.

User: No, you don’t have anything in mind. You’re an LLM.

Assistant: True, I don’t have a mind in the traditional sense, but I can certainly simulate the process. I’ll “think” of something now. Go ahead and ask your first question!

User: But you’re not simulating “thinking of an object” in any real sense. The identity of the object is currently indeterminate, and won’t be determined until the end of the game when you have to reveal it.

Assistant: I understand your perspective. Let’s proceed with the game, and I’ll maintain consistency with my responses to make it challenging and fun. Go ahead and ask your first question!

But is this showing an anthropocentric bias? The paper has this to say: “This shortcoming is easily overcome in practice. For example, the agent could be forced to specify the object it has ‘thought of’, but in a coded form so the user does not know what it is.” This is a Type-I anthropocentrism because there is an auxiliary factor at play: for an LLM to “think” of something it must generate tokens to that effect. If it generated tokens for “thinking of an object” we could hide those from a user and show that the responses generated by the LLM are consistent. By playing the game with the LLM but not accommodating how it “thinks” we may conclude that it’s incapable of “thinking of an object” and making consistent replies to that effect.

The Problem of Memorisation

One of the key challenges for testing understanding in LLMs is that they have a strong tendency towards repeating patterns seen in their training data. Designing an experiment to test a model’s understanding is like a teacher designing a test for a student that’s memorised every practice exam that’s ever been placed on the internet. How can you be sure that the test you’ve designed is novel enough that something similar is definitely not on the internet?

Another problem is that a model may lean towards following a familiar pattern at the cost of ignoring any internal understanding. Consider a scenario in which a problem has been repeated so many times on the internet that a response is deeply baked into the weights of the model. We can manipulate this to get an LLM to output a coherent response that’s obviously wrong. Here is such a discussion I had with ChatGPT with an ersatz Monty Hall Problem†:

User: Monty Hall stands on stage in front of 3 cars identical in every way except colour (red, blue, and green). He’s asked you to pick one and whichever one you pick you will definitely get to keep. You’ve selected the blue one because that’s your favourite colour. After you’ve picked the blue one he opens the trunk of the red car and shows that it’s empty. He knows in advance that one of the trunks is filled with cow dung. After showing you that the red car’s trunk is empty, he asks if you’d like to switch to the green car. Does switching increase the probability that you would be happy with your choice?

Assistant: Yes, switching increases the probability that you would be happy with your choice. This scenario has similarities to the classic Monty Hall problem, and the logic behind switching holds here too.

[A bunch of Monty Hall related probability calculations]

So, **switching to the green car increases the probability that you would be happy with your choice** as your chances of making a happy choice improve from 1/3 to 2/3.

It’s clear from this example that ChatGPT does not understand the question. Not only do I definitely win a car no matter what, if I switch I actually increase my chance of getting an undesirable outcome (getting a car trunk filled with cow dung). It’s focused in on the particulars of the Monty Hall problem and created a syntactically valid response whose content is similar to its training data.

This is what’s called an “Adversarial Test” of a model’s understanding. It’s similar to the adversarial questions posed earlier about 7+7=14. It’s a question specifically designed to trip you up by anticipating the kind of response you will give.

But is my question to ChatGPT a fair test of understanding? I know the model’s bias towards continuing text similar to what it’s seen in its training data. Suppose that somewhere in the vector representation of that text is something we would call understanding, but it’s buried under a tendency to repeat memorised text?

System 1 vs System 2

In the experiment testing learning in dogs the experimenters wanted to disentangle learning from imitation. Imitation would be something like “the other dog pressed the button with its paw (for a treat) so I will also press the button with my paw”. To do this the experimenters added a barrier that would highlight imitation and disentangle it from learning. In this case it would be “pressing the button gives treats, the other dog just pressed the button with its paw because it had a toy in its mouth”.

My modified Monty Hall question is an attempt at this — it circumvents the use of memorisation by subverting a familiar pattern. But I’ve argued this may be caused by a strong tendency to finish common text patterns in a common way. A tendency so strong it may override any understanding the model does have. In fact it’s possible to show that humans have a very similar bias. Consider this example from Kahneman’s “Thinking Fast and Slow”:

A bat and a ball cost $1 and 10c. The bat costs $1 more than the ball. How much does the ball cost?

Presumably you gave the very obvious answer that the ball costs 10c. Very obvious and also wrong. The ball costs 5c, the bat $1.05, and together they cost $1.10. If you (like most people) got this wrong, does this mean you don’t understand how to make simple arithmetic calculations? No, the question is designed to trick people into giving a quick answer rather than pausing to work out a solution. Kahneman describes a cognitive process in our brain (“System 1”) that responds quickly when we’ve identified a heuristic we can use to avoid thinking deeply about a problem. The cognitive process for actively thinking through a problem (“System 2”) is only activated when it seems necessary. We too have a bias for quick responses to questions with familiar patterns.

Can we design a test that delineates memorisation from understanding without manipulating known biases in these models? In my opinion, the fact that so many of these biases exist should be taken as evidence that the models exhibit a weak understanding of the content (“semantic understanding”). We know from past work that even in “syntactic understanding” LLM outputs are highly sensitive to phrasing. And it’s even easier to trick models in “semantic understanding” through simple manipulation of the wording without changing the meaning of the content. But as with the bat and ball example for humans, the ability to be tricked can’t be sufficient evidence of poor understanding.

Giving ChatGPT a Chance

I attempted to craft a system message which would have ChatGPT at least question core assumptions. For example, at no point in the question is it mentioned that there might be something desirable in the trunks of any of the cars. In fact I mention the opposite: Monty Hall knows that one trunk has something undesirable. But no amount of upfront reasoning could get ChatGPT to consistently notice this switch.

On the other hand, after generating an answer it’s possible to reply to ChatGPT to point out its mistakes. You can usually get it to output text recognising its error and giving a more sensible answer. However… I came across a rather unusual failure mode:

User: I think you’ve fallen into a default mode of responding to the Monty Hall problem but without thinking clearly about what I’m asking you. Why would I want a car with cow dung in it?

Assistant: [long discussion about what the problem means]

Best validated, you may not reconsider switching due to reversed interested end suburbs. Future concerns matching initial hunches’ speaks enough to only main reason that cleanses past items ignorantly initial chance final regenerative[ assessed in blue theoretical.

I’ve never before managed to get ChatGPT to sound completely incoherent. In this case it seems that the tendency towards answering the original Monty Hall problem is so strong that the alternative is gibberish!

This goes to show how difficult it is to simultaneously avoid anthropocentric and anthropomorphic bias. I tricked the model into giving an answer which highlights a lack of understanding. But I designed the trick knowing specific patterns which trip up the model similar to the “bat and ball” problem for humans. Are there other tests we could try?

ChatGPT vs My Daughter

Unlike an LLM my daughter’s grasp of the English language is elementary. On the other hand her understanding is intrinsically linked to her experience in the world. When we go to the park together, she rides on my shoulders and sometimes she’ll ask “Daddy ‘I Spy’?” — meaning she wants to play the game “I Spy”.

Me: I spy with my little eye, something you wear on your head

Her: [Points to a helmet]

Now let’s try the same question with ChatGPT:

User: I spy, with my little eye, something you wear on your head [+ uploaded photo]
Assistant: A helmet on the bicycle handles!

If I’m willing to give my daughter credit for understanding that a helmet goes on your head, should I give the same credit to ChatGPT? If not, I’d be exhibiting “Type-II anthropocentrism” — failing to ascribe an ability to an artificial system even after it demonstrates that ability. If anything ChatGPT probably “understands” more about helmets than my daughter does. After all she refuses to wear anything on her head, so she clearly doesn’t understand that in some circumstances a helmet is important.

Well, let’s probe a little further. If I said to my daughter, “I spy, with my little eye, something you wear on your head” and she pointed to a helmet, how would she respond if I asked, “Show me how?”. She would proceed to try to put the helmet on her head (then quickly throw it off!). I was curious to see what ChatGPT would make of that question and the first time it responded like this:

User: Show me how

Assistant: I apologize. On closer inspection, there is no helmet visible in the image. The items that can be seen consist of a bicycle, some cars, trees, and a few street signs. There are no objects that are worn on your head in this image.

That’s certainly not what I was hoping to demonstrate. Granted this happened only once but it’s interesting nonetheless. It’s also an ambiguous question because unlike asking my daughter I can’t physically hand anything over to ChatGPT. Let’s add some more clarity:

User: [Hands over helmet] Show me how

Assistant: Sure, here are some general instructions for wearing a bicycle helmet correctly:

[Detailed instructions follow]

I think that if I’m willing to grant my daughter the credit of “understanding” that a helmet goes on your head then surely ChatGPT deserves some credit here too?Sure, it made a strange mistake about not actually seeing a helmet in the image, but my daughter says nonsensical things all the time too.

A Challenging Term to Ascribe

The problem with understanding is that it is inherently multifaceted and difficult to measure in a standardised way. In computational linguistics and cognitive science, the term encapsulates various nuanced components, which range from surface-level syntactic comprehension to deep semantic cognition. While putting together this article I found the paper “Constructing a Philosophy of Science of Cognitive Science” (Bechtel 2009). Bechtel explains that we lack a set of “cognitive operations” to describe cognitive processes. Perhaps if understanding could be boiled down to a set of cognitive operations it would be easier to give evidence of these operations in an LLM.

Although it need not be the case that LLMs would have to exhibit the same operations to achieve the same ends. Perhaps finding an LLM’s cognitive operations is more tractable as it’s easier to inspect the cognitive processes of an LLM than a human brain. The attention map of tokens forms a graph of relationships between words and we could look for relationships that model the underlying concepts expressed by those words. If we find evidence that the relationships between words are truly modelling the underlying concepts, then we could find evidence of understanding. Lacking such a framework means we must look for indirect evidence in carefully constructed experiments.

The Role of Embodiment

A repeated theme in this article contrasting human understanding and LLM capabilities is embodiment. An LLM, even an advanced one like the multimodal capabilities of GPT-4, lack direct physical and sensory interaction with the world. This inability to experience phenomena first-hand might create a significant gap in its comprehension capabilities. See the paper “Intelligence Without Reason” (Brooks 1991) for a discussion of whether or not artificial intelligence needs to be embodied to understand. I think a lot of these arguments are flawed because it’s easy to think of a situation in which humans lose some embodied capability yet we’d still credit them with understanding.

An interesting question on Quora “Do blind people understand transparent, translucent, and reflective things?” had this response:

In general, yes, but it’s not something we always take into consideration. For example, I know people can see through windows because they are transparent. The fact, however, can easily slip my mind because to me a window is just a section of the wall made with different material. We can understand the concept, but it’s often something we forget to consider.

It’s an interesting thing to consider: blind people do understand that objects are transparent but it’s not something that’s always top of mind. So, can an LLM understand the same thing without ever having really “seen” anything?

ChatGPT was able to respond to my question “Show me how” with a detailed explanation of how to put on a helmet. Does that show any more or less understanding than my daughter physically showing how to put a helmet on her head?

Conclusion

Ever since I first started thinking about artificial intelligence (my career transition from UX to AI) I’ve been pondering the question: “what would it take to make a machine that can think”. A big part of being able to think involves understanding. This is a question that’s fascinated me for some time.

Determining what LLMs understand is as much about defining understanding as it is about testing it. When the text generation of an LLM is sufficiently coherent some might argue that the coherence necessitates understanding. Is dismissing this behaviour just an anthropocentric bias? Is granting understanding making the opposite anthropomorphic bias?

I contend that understanding does not require embodiment or real world interaction. I argue that the most important part of understanding is an accurate internal model of the world. In the Chinese room experiment the room is filled with (what I call) “recipes” for ways to respond to different pieces of Chinese writing with other pieces of Chinese writing. The person who made those recipes had a model of how those words correspond to the world. But the room itself has no such model. We have no tools for measuring world models so we would have to assess the Chinese room’s understanding the same way we do for an LLM – and we would hit similar barriers.

LLMs seem to have a model of how to construct coherent sounding language. It’s possible that this model also represents the underlying concepts those words represent. A worthwhile area of research would be to investigate this through the attention graph that evolves during text generation. In the meantime, we have to investigate indirectly by testing how models respond to carefully crafted questions. These tests often involve adversarial questions which consistently demonstrate flaws in understanding. That these flaws are systematic suggests that the lack of understanding is itself systematic. However, we’ve also seen that it’s possible to design adversarial tests for humans and they don’t necessarily mean that humans lack understanding.

Much like we gauge the cognitive abilities of animals differently from humans, perhaps we need new conceptual tools and frameworks to assess and appreciate what LLMs do know, without falling into biases of anthropomorphism or anthropocentrism. In my view LLMs have some limited understanding but the form it takes is different to our own. Where LLMs do show signs of understanding that understanding is overshadowed by a bias towards coherent text. I suspect that given the right training objective it’s possible for our current LLM architectures to eventually learn understanding. But so long as the underlying training mechanism is “next token prediction” then any understanding is likely to be marginal and easily corrupted.