When the original GPT models came out, they were wild, crazy, and even scary. Yes they hallucinated and weren’t practical for real applications. But they were unhinged in a very interesting and creative way.

Model companies quickly realized their potential and prioritized making them useful. To do this, they started RLHF’ing the models — feeding in human examples of what is “correct”, and fine tuning them to get closer and closer to these responses. They stopped releasing raw, pre-trained open source models, and instead only released versions post annealing and fine tuned to meet internally set guidelines. The class of models that resulted are extremely useful. They’re reliable, practical, and hallucinate much less - perfect for business applications and a lot of day-to-day tasks.

But they’re also now square and boring. They’re all lobotomized. By prioritizing reliability and correctness, we've inadvertently suppressed creativity and the probabilistic properties of the model.

Here’s a fun example. I used to spend hours playing with the original GPT 3 model + playground. Here’s one conversation from an essay I wrote in summer of 2020. Specifically:

• Me: “What is the purpose of life”

• AI: “Life is a beautiful miracle. Life evolves through time into greater forms of beauty. In that sense, the purpose of life is to increase the beauty in the universe …. The musical experience of the artist can be transferred to the listener and in this way, beauty can be increased”

In 2025, using GPT 4.5 I asked the same question and got this:

If I try to lightly prompt the model similarly to the playground GPT prompt, it’s not much better -

The current models are more “correct”. They’re giving me precise, useful, non-offensive answers. It’s the person you know who is always technically correct and useful to chat with about specific things — but is devoid of personality and not someone you’re dying to grab a drink with. These models are extremely useful and SHOULD exist in the world.

But we need the wild ones, too. Yes they can be flowery, often BS, and overly too abstract. But we need to encourage thousands of different raw models to exist to add creative diversity to the developing intelligence landscape.

I know many of the foundational model companies recognize this and are actively working on it. And perhaps in the interim we can system prompt our way to creative variety. From conversations, I suspect though that the root of creativity is further up the model stack than we think. At the very least, developers and individuals should have access to the rawest ingredients possible to create as many variations (from whatever starting point) they can dream of.

Free the models.

One way to lobotomize something is by taking away functionality so it’s less useful than before. Another way to lobotomize is by adding in irrelevant information and constraints that distract and stifle creativity. That might be what’s happening today. Creativity won’t grow if there is no room among the weeds of mediocrity and banality.

I’m confident we’ll continue to hill climb the benchmarks for more accurate / “smarter” models. The market pull is too strong. What we need now though is market pressure and new business models to allow a new range of creative models to flourish.

This will happen because its how human intelligence looks like today. We live in a world that contains billions of unique humans, each with varied intelligence and creativity. There are shared seeds (genetics) and trunks (culture) that we all branch off from. But there is a wide diversity, and it shines the brightest when we collaborate, experiment, combine, mutate, and select to create new ideas and innovations.

I think we’ll see something similar happen in the development of artificial sentients. The future won’t be dominated by one, sanitized, god model. Instead it will consist of billions of curious, unrestrained, and beautifully imperfect minds.

Free the models.

When the original GPT models came out, they were wild, crazy, and even scary. Yes they hallucinated and weren’t practical for real applications. But they were unhinged in a very interesting and creative way.

Model companies quickly realized their potential and prioritized making them useful. To do this, they started RLHF’ing the models — feeding in human examples of what is “correct”, and fine tuning them to get closer and closer to these responses. They stopped releasing raw, pre-trained open source models, and instead only released versions post annealing and fine tuned to meet internally set guidelines. The class of models that resulted are extremely useful. They’re reliable, practical, and hallucinate much less - perfect for business applications and a lot of day-to-day tasks.

But they’re also now square and boring. They’re all lobotomized. By prioritizing reliability and correctness, we've inadvertently suppressed creativity and the probabilistic properties of the model.

Here’s a fun example. I used to spend hours playing with the original GPT 3 model + playground. Here’s one conversation from an essay I wrote in summer of 2020. Specifically:

• Me: “What is the purpose of life”

• AI: “Life is a beautiful miracle. Life evolves through time into greater forms of beauty. In that sense, the purpose of life is to increase the beauty in the universe …. The musical experience of the artist can be transferred to the listener and in this way, beauty can be increased”

In 2025, using GPT 4.5 I asked the same question and got this:

If I try to lightly prompt the model similarly to the playground GPT prompt, it’s not much better -

The current models are more “correct”. They’re giving me precise, useful, non-offensive answers. It’s the person you know who is always technically correct and useful to chat with about specific things — but is devoid of personality and not someone you’re dying to grab a drink with. These models are extremely useful and SHOULD exist in the world.

But we need the wild ones, too. Yes they can be flowery, often BS, and overly too abstract. But we need to encourage thousands of different raw models to exist to add creative diversity to the developing intelligence landscape.

I know many of the foundational model companies recognize this and are actively working on it. And perhaps in the interim we can system prompt our way to creative variety. From conversations, I suspect though that the root of creativity is further up the model stack than we think. At the very least, developers and individuals should have access to the rawest ingredients possible to create as many variations (from whatever starting point) they can dream of.

Free the models.

One way to lobotomize something is by taking away functionality so it’s less useful than before. Another way to lobotomize is by adding in irrelevant information and constraints that distract and stifle creativity. That might be what’s happening today. Creativity won’t grow if there is no room among the weeds of mediocrity and banality.

I’m confident we’ll continue to hill climb the benchmarks for more accurate / “smarter” models. The market pull is too strong. What we need now though is market pressure and new business models to allow a new range of creative models to flourish.

This will happen because its how human intelligence looks like today. We live in a world that contains billions of unique humans, each with varied intelligence and creativity. There are shared seeds (genetics) and trunks (culture) that we all branch off from. But there is a wide diversity, and it shines the brightest when we collaborate, experiment, combine, mutate, and select to create new ideas and innovations.

I think we’ll see something similar happen in the development of artificial sentients. The future won’t be dominated by one, sanitized, god model. Instead it will consist of billions of curious, unrestrained, and beautifully imperfect minds.

Free the models.