Dylan Patel

With good expert controls, it also just makes it so that AI can be used much less. And I think that is a much easier goal to achieve than trying to debate on what AGI is. And if you have these extremely intelligent, autonomous AIs and data centers, those are the things that could be running in these GPU clusters in the United States, but not in China.

With good expert controls, it also just makes it so that AI can be used much less. And I think that is a much easier goal to achieve than trying to debate on what AGI is. And if you have these extremely intelligent, autonomous AIs and data centers, those are the things that could be running in these GPU clusters in the United States, but not in China.

With good expert controls, it also just makes it so that AI can be used much less. And I think that is a much easier goal to achieve than trying to debate on what AGI is. And if you have these extremely intelligent, autonomous AIs and data centers, those are the things that could be running in these GPU clusters in the United States, but not in China.

buzzy words in the AI community about this, test time compute, inference time compute, whatever. But Dylan has good research on this. You can get to the specific numbers on the ratio of when you train a model, you can look at things about the amount of compute used at training and amount of compute used at inference.

buzzy words in the AI community about this, test time compute, inference time compute, whatever. But Dylan has good research on this. You can get to the specific numbers on the ratio of when you train a model, you can look at things about the amount of compute used at training and amount of compute used at inference.

buzzy words in the AI community about this, test time compute, inference time compute, whatever. But Dylan has good research on this. You can get to the specific numbers on the ratio of when you train a model, you can look at things about the amount of compute used at training and amount of compute used at inference.

These reasoning models are making inference way more important to doing complex tasks. In the fall, in December, OpenAI announced this O3 model. There's another thing in AI when things move fast, we get both announcements and releases.

These reasoning models are making inference way more important to doing complex tasks. In the fall, in December, OpenAI announced this O3 model. There's another thing in AI when things move fast, we get both announcements and releases.

These reasoning models are making inference way more important to doing complex tasks. In the fall, in December, OpenAI announced this O3 model. There's another thing in AI when things move fast, we get both announcements and releases.

Announcements are essentially blog posts where you pat yourself on the back and you say you did things and releases are on the models out there, the papers out there, etc. So OpenAI has announced O3 and we can check if O3 Mini is out as of recording potentially. But that doesn't really change the point, which is that

Announcements are essentially blog posts where you pat yourself on the back and you say you did things and releases are on the models out there, the papers out there, etc. So OpenAI has announced O3 and we can check if O3 Mini is out as of recording potentially. But that doesn't really change the point, which is that

Announcements are essentially blog posts where you pat yourself on the back and you say you did things and releases are on the models out there, the papers out there, etc. So OpenAI has announced O3 and we can check if O3 Mini is out as of recording potentially. But that doesn't really change the point, which is that

The breakthrough result was something called Arc AGI task, which is the abstract reasoning corpus, a task for artificial general intelligence. Francois Chollet is the guy who's been, it's a multi-year old paper. It's a brilliant benchmark. And the number for OpenAI 03 to solve this was that it used some sort of number of samples in the API. The API has like thinking effort and number of samples.

The breakthrough result was something called Arc AGI task, which is the abstract reasoning corpus, a task for artificial general intelligence. Francois Chollet is the guy who's been, it's a multi-year old paper. It's a brilliant benchmark. And the number for OpenAI 03 to solve this was that it used some sort of number of samples in the API. The API has like thinking effort and number of samples.

The breakthrough result was something called Arc AGI task, which is the abstract reasoning corpus, a task for artificial general intelligence. Francois Chollet is the guy who's been, it's a multi-year old paper. It's a brilliant benchmark. And the number for OpenAI 03 to solve this was that it used some sort of number of samples in the API. The API has like thinking effort and number of samples.

They used a thousand samples to solve this task and it comes out to be like $5 to $20 per question, which you're putting in effectively a math puzzle. And then it takes orders of dollars to answer one question. And this is a lot of compute. If those are going to take off in the US, OpenAI needs a ton of GPUs on inference to capture this.

They used a thousand samples to solve this task and it comes out to be like $5 to $20 per question, which you're putting in effectively a math puzzle. And then it takes orders of dollars to answer one question. And this is a lot of compute. If those are going to take off in the US, OpenAI needs a ton of GPUs on inference to capture this.

They used a thousand samples to solve this task and it comes out to be like $5 to $20 per question, which you're putting in effectively a math puzzle. And then it takes orders of dollars to answer one question. And this is a lot of compute. If those are going to take off in the US, OpenAI needs a ton of GPUs on inference to capture this.

They have this OpenAI ChatGPT Pro subscription, which is $200 a month. Which Sam said they're losing money on. Which means that people are burning a lot of GPUs on inference. And I'm I've signed up with it. I've played with it. I don't think I'm a power user, but I use it.

They have this OpenAI ChatGPT Pro subscription, which is $200 a month. Which Sam said they're losing money on. Which means that people are burning a lot of GPUs on inference. And I'm I've signed up with it. I've played with it. I don't think I'm a power user, but I use it.