OpenAI's o3 and the 'JAGGED FRONTIER' of AGI....

The latest AI News. Learn about LLMs, Gen AI and get ready for the rollout of AGI. Wes Roth covers the latest happenings in the world of OpenAI, Google, Anthropic, NVIDIA and Open Source AI.

My Links 🔗

Centaurs and Cyborgs on the Jagged Frontier

#ai #openai #llm

[MattHabermehl]

"As intelligent as the average human" and "intelligent enough for the average job" are different distributions.

[andreasmoyseos]

AGI is when your employer tells you not to come to work tomorrow.

[NotBirds]

I guess this proves the "cant teach a fish to climb a tree" theory. Many of the things we think of as easy and take for granted are exceptionally difficult, while others seem to be trivial yet overstated. This is like comparing the progress of a civilization on another planet to our own. 8:00

[randalx]

I use the analogy of comparing a plane to a bird. Sure birds are more agile and incredible fliers but it obviously does not mean planes are not useful. I'm sure AI will eventually overcome any limitations, but in the meantime we should focus on getting value from their strengths.

[chrisanderson]

1:49 The problem is that AI is better at rocket science than it is at answering the phone at Kim's Bakery which puts us in a weird spot. EDIT: In terms of what you said at the end, I think there is an insane amount of progress yet to come but yes we might find ourselves in the position of ultra-powerful narrow ASI instead of humanistic AGI, flying a rocket ship before we walk. General MEANS general, if an AI can't fully and completely generalise then it isn't "AGI", but it may be, at the same time, godlike narrow ASI. Either version will be transformative.

[cristianandrei]

We humans are good at moving the goal post in our favor.

[AAjax]

With a jagged frontier, many (most?) people will reject an AI as AGI. Similarly, most people will reject someone with savant syndrome as being generally intelligent. The "general" carries with it the idea that the intelligence has to intellectually navigate within human society, and the weaker parts of the jagged frontier prevent that.
But I also think AGI as a target is overrated. An intelligence that can solve cancer, stitch together science disciplines deeper and wider than any human could deal with, narrowly navigate the labour of most people... who cares if it's AGI or not.

[nullifier_]

- can a robot turn a canvas into a beautiful masterpiece?
- can you?

[thematriarch-cyn]

"AI hit a wall" Yeah, and it's fucking climbing it

[EdwardAshdown]

It's not human-like. People are looking for human-like intelligence which should not necessarily be synonymous with AGI

[patrickmchargue]

Discounting AGI based on what failures it has seems to stack the deck against both machine AGI as well as human GI. There are a lot of people that fail very simple tests. Check out some YouTube videos to see some of that.

[dlbattle]

Using AI right now is like driving in the dark. You never know when you'll hit an obstacle. This make it difficult to rely on it.

[sergey]

The focus on removal of jaggedness from this boundary is exactly the validation for reasoning vs recall x hallucination x calculation. If reasoning isn't there, self-prompting capabilities are limited. The model keeps jumping between two wrong solutions. That becomes quite obvious, when you use LLMs for programming.

[ChaseFreedomMusician]

I appreciate the follow-up video and the concept of the "Jagged Frontier." It’s a great way to illustrate how these models can excel in certain areas, like math, physics, or chemistry, while struggling with tasks that seem simple to humans, such as counting letters in a word or solving trick questions. But I want to emphasize that those kinds of gaps—the things these systems can’t do yet—aren’t part of my argument. I’m not concerned about whether they can solve the strawberry problem or handle “gotcha” questions. My point is about the system itself, not its current limitations. Those gaps don’t define whether or not it’s AGI.

To clarify, I’ve never said these models aren’t intelligent. They’re incredibly intelligent and represent a monumental achievement in AI. I also believe they’re AGI-adjacent in some ways—pushing the boundaries of computation and reasoning, and perhaps even laying the groundwork for creating AGI in the future. But the "G" in AGI—general—is what separates it from what we have now. AGI isn’t just about excelling at certain tasks or reasoning through complex problems with test-time computation. It’s about the ability to generalize, adapt, and learn dynamically across an unlimited range of tasks without retraining or manual intervention. That’s the line we haven’t crossed yet.

A lot of the discussion seems to focus on prompts and context windows as ways to enable these systems to "learn." But prompts are inherently limited. They can only hold a finite amount of information, and the model doesn’t retain or integrate what it learns in a way that updates its core understanding. This is where the distinction becomes clear: AGI wouldn’t just process a task within the confines of a context window—it would take insights from solving one problem and apply them broadly to others. It wouldn’t need to repeat the same compute-intensive reasoning process every time because it would have already evolved its understanding. And to this point, there’s often confusion between caching and generalization. If a system becomes faster at answering the same question the second time, that’s not generalization—it’s caching. True generalization means understanding the core principles of a problem and applying those principles to entirely new situations.

This is why I don’t consider models like 03 to be AGI. They’re brilliant expert systems—highly intelligent and transformative in their capabilities. They’re solving monumental problems and paving the way for even greater advancements. Take the Riemann Hypothesis and satisfiability problems, for example. These are two of the most profound challenges in mathematics and computer science, and models like 03 could help us tackle them in ways we’ve never been able to before. Solving the Riemann Hypothesis could reshape our understanding of prime numbers, leading to breakthroughs in cryptography and computational efficiency. And advancing SAT problem-solving could revolutionize fields like logistics, healthcare, and AI optimization itself.

These achievements are extraordinary. They have the potential to solve problems that humanity has struggled with for centuries, and they might even help us create AGI someday. But solving specific problems—even incredibly important ones—doesn’t make a system AGI. AGI is about adaptability, generalization, and the ability to learn and grow autonomously. It’s not just a collection of tools or reasoning processes—it’s a fundamentally different kind of intelligence.

So, let me be clear: I think what’s happening here is amazing. Models like 03 are a testament to how far we’ve come, and they’re going to change the world in significant ways. But they’re still operating as highly advanced, specialized systems. They’re intelligent, yes. They’re transformative, absolutely. But they’re not AGI. That’s not a criticism—it’s just about being precise with our definitions.

[ChristianWilliams-lv]

I've been following you for a while and this is by far the best video you've done. AI news is fantastic but these videos where you approach some of the ideas about artificial intelligence and what people are saying with a more novel description is amazing. This is the kind of stuff I can tell other people about.

Cheers and Merry Christmas

[rmt]

11:00 AGI doesn't mean smarter than humans. Any AI will be able to beat humans at specific tasks, that's the whole point of AI. This is specific intelligence.

AGI is General Intelligence. Rn, we're just patching the gaps and skill, adding more specific intelligences to the AI. This can never become general intelligence, as it will always have gaps.

The curves are supposed to be massively larger than human, because that's the point. This isn't a sign of AGI, it's a sign of AI. It is proof you made Artificial Intelligence, which we've had for a very long time now.

26:20 This is EXACTLY why it's not even close to AGI. If you have to train it to understand data that's already in a human readable format, it's not AGI. Take the Model, Give it the problems, and let it solve them.

This LITERALLY can't even see the problems without you having to translate it to JSON.

AGI needs ZERO steps between it and the problem. If it is unable to even attempt the problem without someone spoonfeeding it, it's not AGI, not even close.

[rexmundi]

I know an aerospace engineer who has washed her phone in the laundry twice. I turn raw aluminum into parts that go into space but I can’t spell a lot of words. If AI wasn’t so polite and cussed some, it would definitely pass the Turing test. I think people are in serious denial about how disruptive AI is going to be for jobs and society as a whole. I have a friend who does corporate level IT and systems integration who is willfully and stubbornly not even looking at what’s going on with AI. I can’t understand it.

[PasseScience]

I like the whole passage on the jagged frontier from roughly 6:00 to 14:00. Another question is is this really jagged (are those objectively task of unequal difficulty) or is it anthropocentric ? from the AI point of view it's its jagged line that is smooth and the human intelligence curve that is jagged, so because we are bad at stuffs the AI consider trivial does this discards us as AGI ? from it's point of view :)

[Jopie]

For me a profound realisation was that e.g. dogs can do things easily that humans consider hard. So does that mean humans are _not_ AGI?

[SurfCatten]

It's about being able to reason not about being smarter than humans. And at least the o1 model doesn't really reason. It kind of seems to by talking to itself about what it says and adjusting its response but that's not really the same thing. I use these models daily and it's absolutely clear that they don't really reason. Now that said their simulation of reasoning is indeed better than many if not most people!