Jun 25, 2026 · 52 min · 11 segments
Jonathan Schaeffer thinks we're building AI the wrong way. While large language models have produced remarkable results, he argues that hallucinations, bias, and unreliability aren't bugs that can be…
No entities detected.
Now, when we create large language models, we're creating AI in a particular way, using a particular kind of architecture for the model, make it a large language model.
My guest today thinks that we're going down the wrong path, that we can seek AI, but that LLMs are not going to get us to where we ultimately want to be.
He's also the co-founder of OLEA, the world's first MOOC, Massive Open Online Course Production Company.
So, for instance, because LLMs are what they are, we have the well-known problem of hallucinations, AI just making stuff up.
Actually, you've heard the founders of companies who create large language models say this kind of thing as well.
It's the very same thing that allows it to be creative, so to speak, to make it work in so many different kinds of contexts is also the source of why it creates hallucinations.
He thinks then that we need a different kind of model architecture if we're going to get to the kind of AI that we can properly rely on.
In the meantime, he thinks we need much better regulations and guardrails around the current technology.
He wants something like, I don't want to accuse him of saying that he wants something like perfection, but he wants it to be performing way better than it is now.
Myself, I'm more inclined to think that the relevant bar is how well humans are doing.
And unfortunately, I'm not sure that we're very great at a very great many activities.
And so what we think of as the benchmark for good enough is different between the two of us.
And so I think that leads us to having different views about the overall utility and ultimate ultimate this is that a word successfulness of large language models.
And as we talk about here, it's even worse when it comes to large language models, because when they talk to us, they sound so confident and authoritative and intelligent, even though they're none of those things.
We talk a little bit about some of the kinds of remediations that companies have done.
So for instance, Anthropix Claude says, Claude makes a mistake, make sure to check his answers.
And anyway, this is a sort of very large, wide-ranging conversation about the current state of AI through the lens of how good are LLMs really and are the ultimate solution? Or, as Jonathan thinks, are they really just a stepping stone to the next kind of model architecture from which we get far more reliable AI? Okay, that's what this conversation is about.
Read the full transcript.
Create an account to read the whole episode, search across every transcript, and follow the shows you care about.
Now, when we create large language models, we're creating AI in a particular way, using a particular kind of architecture for the model, make it a large language model.
My guest today thinks that we're going down the wrong path, that we can seek AI, but that LLMs are not going to get us to where we ultimately want to be.
He's also the co-founder of OLEA, the world's first MOOC, Massive Open Online Course Production Company.
So, for instance, because LLMs are what they are, we have the well-known problem of hallucinations, AI just making stuff up.
Actually, you've heard the founders of companies who create large language models say this kind of thing as well.
It's the very same thing that allows it to be creative, so to speak, to make it work in so many different kinds of contexts is also the source of why it creates hallucinations.
He thinks then that we need a different kind of model architecture if we're going to get to the kind of AI that we can properly rely on.
In the meantime, he thinks we need much better regulations and guardrails around the current technology.
He wants something like, I don't want to accuse him of saying that he wants something like perfection, but he wants it to be performing way better than it is now.
Myself, I'm more inclined to think that the relevant bar is how well humans are doing.
And unfortunately, I'm not sure that we're very great at a very great many activities.
And so what we think of as the benchmark for good enough is different between the two of us.
And so I think that leads us to having different views about the overall utility and ultimate ultimate this is that a word successfulness of large language models.
And as we talk about here, it's even worse when it comes to large language models, because when they talk to us, they sound so confident and authoritative and intelligent, even though they're none of those things.
We talk a little bit about some of the kinds of remediations that companies have done.
So for instance, Anthropix Claude says, Claude makes a mistake, make sure to check his answers.
And anyway, this is a sort of very large, wide-ranging conversation about the current state of AI through the lens of how good are LLMs really and are the ultimate solution? Or, as Jonathan thinks, are they really just a stepping stone to the next kind of model architecture from which we get far more reliable AI? Okay, that's what this conversation is about.