Limits of LLMs, AGI & the Future of AI

This article is mine notes of Lex Fridman with Yann Lecun - Chief AI Scientist at Meta.

The podcast is quite insightful and touches on the real limitations of LLM and how AGI would evolve.

Limitations of LLMs

These are four important characteristics of intelligent behavior:

1. Capacity to understand the physical world
2. Ability to remember and retrieve things
3. Persistent memory
4. Ability to reason and plan

LLMs can't fully achieve these; they can only do so in a very primitive way.

Most of our knowledge is derived from interactions with the physical world, not just language. LLMs can pass the bar exam, but they can't launch a drive in 20 hours, nor can they learn to clear the dinner table and fill the dishwasher. Essentially, they can't perform the daily physical actions that we do.

Most of our thinking is not in terms of language, but current LLMs are trained to predict the next word since there are finite possibilities for the next possible word. We plan our answers before we produce them, but LLMs do not; they just produce one word after another.

Can You Build a World Model That Has a Deep Understanding of the Real World?

This would involve understanding the world and why it is evolving the way it is. A world model would need to account for:

1. The state of the world at Time T
2. The action taken at Time T
3. The state of the world at Time T+1

Such a model can't be achieved with current generative models. If we think in terms of video, it would involve predicting the next frame of the video, which we don't yet know how to do. The world is incredibly more complicated and richer in terms of information than text, which is discrete; video is high-dimensional and continuous.

Joint Embedding Predictive Architecture & How It Differs from Generative AI Architectures Like LLMs

• LLMs generate inputs that are non-corrupted and non-transformed, requiring the prediction of all pixels, which is resource-intensive.
• In JEPA, you aren't predicting the pixels; you are only predicting the abstract representation of the inputs. It is built to extract as much information from the input as possible but only extracts information that is relatively easily predictable, and it eliminates the noise. It's generative, but in an abstract representation space. JEPA might be able to learn common sense.

Why Hallucinations Happen in LLM and Why It Is a Fundamental Flaw of LLMs

• There is a small probability in the process of predicting the next word that the answer does not lie within the reasoning space, and with each word, the eventual probability increases exponentially.
• You can fine-tune for 80% of what people would ask, but the long tail is so large that you can't fine-tune for all conditions.
• Whether you ask a simple question or a complicated one, the amount of computation that the system devotes to the answer is constant or proportional to the number of tokens produced. That's not how humans work; in complex problems, we spend more time figuring out the solution.

Future of AI

• Yann LeCun is optimistic about achieving human-level intelligence through a system that can understand, plan, remember, and reason.
• AGI wouldn't be an event; it would be a gradual progress.
• Since it would be a gradual process, we would first achieve a certain level of intelligence and then evolve it further. As we evolve, we would be putting in appropriate guardrails, hence AI wouldn't be destructive or kill us all.
• Even if someone misuses it, there would be others who would do it right. So we could use the good ones against the wrong ones, something like smart AI police against rogue AI. Thus, we wouldn't be exposed to a single rogue AI that could kill us all.
• AI wouldn't have a desire to dominate because that needs to be hardwired; this is hardwired in humans.