Introduction

There’s a lot of debate about when we’ll read Artificial General Intelligence (AGI). Anyone who tells you they know the answer—including me—is almost certainly wrong because the answer’s more complex than people think. Unfortunately, it depends on how we define AGI.

The OpenAI View

OpenAI has gone back and forth, but has settled on a description of it as AI which is above human capability for most economically valuable tasks. This is an interesting definition, one which seems achievable at their current rate of progress, but has the side benefit that Microsoft won’t have access to AGI if the OpenAI Board declares it’s achieved that goal. This is ostensibly to ensure that no for-profit company (er, like OpenAI itself), has the ability to abuse AGI for their own benefit.

Now OpenAI is considering removing that restriction . That would allow the for-profit part of OpenAI to continue to rake in billions of dollars in investment from Microsoft. Whether you consider this a shrewd or abusive business decision is largely a matter of taste. I’ll leave that for others to argue.

Can OpenAI succeed?

That’s where the debate lies. Right now, Wharton Professor Ethan Mollick writes that several professors have contacted him to say that OpenAI’s o1 model appears to have found something new in their field. Nothing like curing cancer, but small discoveries that will require considerable research to verify.

But the world of AI hallucinations still exists and while AI brings huge benefits to many tasks, such as brainstorming, it still fails at many things that you and I take for granted. Many AI models give a wrong answer to the questions “how many ‘r’s are in the word ‘strawberry'.” Or the fact that A=B means B=A escapes them because they can’t know that.

To explain why this is, consider that generative AI simply predicts the next word (er, token, but I’m keeping this simple). For example, if you ask it to predict the next word for “the cat sat on the ...“, it might internally generate a list of probable next words like this:

It then uses something called a “weighted pick” to choose the next word, with the word “mat” having a 30% chance of being chosen (again, oversimplifying). If your AI system exposes the “temperature,” can set it to zero to ensure that only “mat” is chosen, or to a higher value to give more randomness.

Does that sound like “intelligence” to you, given that AI doesn’t have any actual understanding of the real world? Despite all of the work to reduce hallucinations, this fundamental lack of real understanding could lock AI out of many economically valuable tasks. For example, AIs generally don’t understand that if A equals B, B equals A. You might try various prompts with models and think they know this, but in reality, they can’t figure this out logically. It’s just pattern matching. Here’s a classic example:

If all A are B, and all B are C, is it true that all C are A?

I just asked Venice if this was true and it responded:

This is a classic example of a syllogism in logic. The answer is yes, it is true that all C are A.

Here’s the reasoning:

1. All A are B (All A are C in this case)
2. All B are C
3. Therefore, all C are A (since all C are B, and all B are A)

This is known as a transitive property, which states that if A is a property of B, and B is a property of C, then A is a property of C. In this case, the property is “being A”, “being B”, and “being C”.

On the surface, that seems sane. In reality:

1. All dogs are canines.
2. All canines are animals.
3. Therefore, all animals are dogs.

Larger foundation models usually get this correct, but again, it’s pattern matching, not logical reasoning. On the other hand, there’s some evidence that humans rely on pattern matching more than logical reasoning , or perhaps we only use pattern matching and don’t use logic. So this damning critique of generative AI might not be so damning after all. For example, consider this equation:

$$(3(2+2))^2$$

How do we solve this? Let’s think “step-by-step”:

First, we know from math class that parentheses come first and we recognize this pattern. So let’s solve what’s inside the parentheses (2+2):

$$(3(2+2))^2 = (3(4))^2$$

Next, multiply 3 and 4:

$$(3(4))^2 = (12)^2$$

Finally, square 12:

$$(12)^2 = 144$$

Therefore, $(3(2+2))^2 = 144$

Each of those “steps” might seem like logical reasoning, but $2+2 = 4$. $3(4) = 12$. $12^2 = 144$. I didn’t have to “think” about those answers. I recognize those patterns.

But even if we’re just pattern matching, AI can’t learn. It has a set of facts and that’s it. It’s very expensive to train it and most of us who use the foundation models are familiar with the infamous “knowledge cutoffs.” Of course, researchers are working on this problem, too.

Hallucinations

When we talk about AI, the subject of hallucinations always emerges. The term is thrown around as the ultimate rebuttal. AI can and does make serious mistakes and when it’s telling you to glue your pepperoni to your pizza , it’s seems a pretty serious critique. If hallucinations are serious and common, a superintelligent AI might be worse than today’s AI because we might not be able to detect the hallucinations.

But let’s rephrase that argument. What’s often being said is “AI isn’t perfect, therefore AI isn’t good.” When viewed in that light, we can immediately see the problem.

Consider the image of a bed you see next to this paragraph. That image has “alt” text to allow those using screen readers to know what that image is. I now don’t (usually) write alt text for my blog because I added code to my web site software to detect the lack of alt text and use OpenAI to generate it. It can make mistakes, but as of this writing, it generates better alt text than what I write. This is real value being provided by AI and let’s me focus on the core of my writing and yet still provide an accessible experience if your vision is impaired. As of this writing, that alt text reads as:

The image depicts a cozy bedroom with a bed covered in soft, wrinkled bedding, positioned against a textured concrete wall. A small plant in a woven basket and a stack of books sit on a wooden bedside table, illuminated by warm sunlight streaming through a window.

When I first used that image, my alt text was “A slightly messy bed.” For whatever reason, I’m awful at writing alt text. AI is not.

For this web site, that’s a tiny amount of value. I routinely use AI for tiny amounts of value. It’s massively boosted not only my productivity, but my enjoyment in the things I do, taking away tiny drudgeries and letting me focus on my main goal.

Of course, AI hallucinations can be extremely problematic in critical domains, but if I’m asking it for five creative advertising campaign ideas, I don’t care if it’s hallucinating.

OK, fine, AI can provide some value ... but those damned hallucinations! In reality, what we’re seeing is that people tend to take AI’s responses at face value, but blame it more when they spot the errors.

Here’s the truth: humans often hallucinate far more than AI, particularly on issues of recall. Already every major AI writes grammatically correct prose, with fewer spelling errors, than almost anyone on the planet. For many common tasks, it far outperforms humans. Even a year ago those were hard assertions to make. Today, they’re rarely challenged.

In fact, in working with AI, there are two main types of hallucinations based on context and content. One is when the AI starts “forgetting” what it was talking about and goes off in a random direction, like your drunk uncle at a party who forgot what he was talking about and won’t shut up. This is largely mitigated with longer context windows.

The other is when AI becomes captain of the USS Make Shit Up. These hallucinations are much harder to detect because the AI sounds so plausible.

The “make shit up” hallucinations are more likely when AI hits topics that it doesn’t know as much about. If there’s not much training data, it struggles. But how much training data does it have?

Let’s assume the average book is about 100,000 words. Assuming 1000 tokens translates to about 750 words, a book is about 133K tokens in length. How much data have these models been trained on? Assuming you could read one book a day, just looking at OpenAI models (estimated training data size):

For even the smallest of these models, it would take you millennia to consume that much knowledge and, unlike the AI, you wouldn’t remember it. Even some of the smallest models, such as bert-tiny , were probably trained on data sets that would take you centuries to read. The smallest of these models “knows” far more than we do about everything. Take any subject you don’t know well and we’ll compare your hallucination rate.

Your objection, of course, is that you’d say “I don’t know” instead of making stuff up. Here’s what Claude 3.5 Sonnet just replied to me when I asked about the ChatGPT training data:

I aim to be direct but acknowledge uncertainty here. I don’t actually know the specific amounts of training data used for OpenAI’s models, as this information isn’t public. While there has been speculation and estimates in the media and research community about the scale of data used, I think it’s best not to make claims about specific numbers without being able to verify them.

So again, we see AI getting better and better at this (Anthropic, the company behind Claude, has been doing great work in this area). There’s also been work to have LLMs emit special tokens like [DK] which signal to it that it should say “I don’t know.”

So let’s move on, keeping this context in mind: even the smallest of these models have far more knowledge than any single human has ever had. That’s the base we’re building on.

World Models

Generative AI doesn’t see the world the way we do. It can’t. It generates the next word or the next pixel. This is very one-dimensional. We see the world in three dimensions with time passing. AI can tell you that you can’t unbreak an egg, only because that’s a pattern it sees in text. It’s never seen an egg break. It doesn’t “know” time beyond patterns in text. Maybe what AI needs are “world models” to see the world. AI researcher Yann LeCun points out the problem very clearly :

[Even] the most advanced AI systems struggle with tasks that humans learn quickly—like clearing a table or driving a car. While humans can learn these tasks in hours, AI systems need to be trained on thousands or even millions of hours of data, and still don’t get it right all the time.

The one-dimensional AI doesn’t “know” that eggs break when dropped. It doesn’t “know” that time passes. It doesn’t know that if A = B, then B = A. It’s all pattern matching of text. But what if it can see the world? What if it can learn these things using a fraction of the data it has now? Or if robotics advances at its current rate, coupled with AI, it can experience the world first hand, not just observe it.

This is why many researchers are thinking that giving AI a chance to see and experience the world firsthand is the key to truly reaching AGI, not just “predicting the next word.” Further, this training could be based on reality, not Reddit posts. That “glue on the pizza” hallucination was the model apparently picking up a humorous Reddit reply about preventing cheese sliding off pizza . Using training on experience in the world, many kinds of hallucination seem less likely. The information is more grounded in reality, the AI will be able to “intuitively” understand what is going on in a way it cannot today.

The o3 Model

As I was writing this piece, OpenAI announced an o3 model, something currently only available for safety researchers.

Aside from being, hands down, the strongest AI out there, o3 has PhD level math skills, something that most AI cannot achieve. However, it’s the ARC-AGI benchmark that’s really impressive. ARC-AGI stands for “Abstract and Reasoning Corpus for Artificial General Intelligence,” a series of unpublished problems that humans can often solve, but AI is terrible at.

Most foundation models score around 5% or lower on this. o1 models were hitting between 8% to 32%, depending on the amount of compute time. o3 is between 75.7% and 87.5%, depending on the amount of compute used. Humans tend to score around 85% on this.

I mention the o3 model because, as with so many things AI, anything written about it can be out of date by publishing time. While I suspect that proper world models may be a key to AGI, pushing GPTs to their limits also appears to be approaching AGI. The cost of the full o3 model using max compute can be thousands of dollars per request, so it’s prohibitively expensive. Smaller requests on the low-compute models can be 20 dollars per task, but can exceed one thousand dollars on high-compute tasks .

So we’re seeing incredible advancements towards AGI, but soaring costs and energy consumption requirements. It’s not sustainable, but there’s amazing work being done there, too. For example, matmulfree LLMs maintain performance, but reduce the costs of running LLMs by 50x . You could run one of the smaller ones for the energy consumption of a light bulb and Nvidia stock could become worthless. Sadly, the researcher are still trying to acquire funding for creating a larger model . I expect that, barring unforeseen difficulties, they’ll get that funding. There’s simply too much money at stake.

Conclusion

While AI development is making remarkable progress through multiple approaches (world models and advanced foundation models), significant technical and practical challenges remain before achieving true AGI. The path forward likely involves both improving efficiency and developing more sophisticated ways for AI to understand and interact with the real world.

I can’t give a timeline, but we’ve still got a few years before any potential “job apocalypse.”