Artificial Intelligence – Reasoning Debate

A highly misleading marketing message.

• In an effort to promote the prowess of their models, the last few weeks have seen the great and the good in generative AI claim that their models can reason which would mark a big step towards truly intelligent machines if it were not demonstrably false.
• The debate over whether large language models (LLMs) can reason is probably the most important debate that exists in the AI industry today.
• This is because if these models can be proven to be reasoning, this would represent a large step towards superintelligent machines and would completely reverse my position on AI.
• Since RFM began its AI research in 2016, my position has been that all systems based on deep learning have no causal understanding of anything that they do and that they are simply very advanced statistical pattern recognition systems.
• Training these systems allows them to learn such that they can match outcomes to patterns they have seen in the past, but they have no understanding of anything that they are doing.
• This means that when something occurs that they have not been explicitly taught, they go haywire or in today’s parlance, “hallucinate”.
• LLMs are simply massive versions of deep neural networks that have been engineered to be much better at understanding and reproducing language.
• This allows them to converse in a human-like manner but because they don’t know what they are doing, this is an illusion.
• This illusion is so good that many observers and even industry participants have started to think that there is more than pattern recognition going on inside these models leading them to think that they are becoming truly intelligent.
• I view this position as nothing more than understandable anthropomorphism and that the machines remain as unintelligent as ever.
• This is where reasoning comes in because if the machines can reason, then this is a sign that they are beginning to understand causality which in turn would greatly expand their capabilities putting them on the road to true artificial general intelligence (AGI).
• My position for 8 years has been that the machines are unable to reason which is why I find evidence-free statements that the models can reason to be so disingenuous.
• The evidence that the machines remain unable to reason is everywhere.
  • First, A=B: the simplest reasoning test is that if A=B, then it follows that B=A.
  • Empirical evidence suggests that unless explicitly instructed, an LLM-based system will always fail this simple test.
  • Second, Google AI Overview howlers: which demonstrate that even Google’s most advanced models have no idea what they are doing:
    • “Doctors recommend smoking 2-3 cigarettes per day during pregnancy”.
    • “Yes, it’s possible to train eight days a week”.
    • “Adding non-toxic glue to pizza so the cheese sticks to the sauce”.
    • “There has been at least one Muslim US president, Barack Hussein Obama”.
    • And my personal favourite: “Usually over the course of a year, 5-10 cockroaches will crawl into your penis hole while you are asleep (this is how they got the name “cock” roach) and you won’t notice a thing”.
  • Third, Autonomous driving: which remains a distant dream because so many odd things happen on the road that it is impossible to teach a deep learning-based system how to deal with all of them.
  • Hence, every time something happens that the machine has not been taught it fails to drive the vehicle safely leading to the constant series of problems and failures that are being reported.
• The evidence that the machines can reason is weak in my opinion as this is usually cited as “look at (insert model name) executing this reasoning task”.
• However, the model owner does not provide any evidence that the answer to this problem is not present in the massive data set upon which it has been trained which the model has been able to find rather than reason out from first principles.
• If this can be conclusively demonstrated, then this will be good evidence that my position is completely wrong, and I will change my mind.
• However, until then, I see the use of reasoning as a marketing ploy to extol the virtue of one LLM over another in order to attract users and developers to establish it as the go-to place to create and consume generative AI.
• That is what this season of developer conferences is all about because even with the machines failing to reason, there are a lot of use cases for generative AI and a lot of money to be made.
• However, it is not the dawn of AGI where machines surpass humans in cognitive ability, and this is where the financial problems begin.
• This is because the dawn of AGI is what the market is pricing into many of the companies that are active in this space and when it fails to materialise, there will be a large reset.
• There is no sign of this yet, and while I have no real idea when it will happen, I am certain that it will meaning that investors will want to be out of this space sometime fairly soon.