If you look at this leaderboard below, which shows the biggest models, you look at the very best, but do you notice something strange?

There are several versions of GPT-4. Yes, they correspond to different versions at a given time. For example, GPT4 03–14 is the version from March 2023, 06 is from June 2023, and so we can already see the first thing is that they are not ordered chronologically.

Is there anything else?

Yes, on Claude’s ones. Claude 2 is below Claude 1, and CL 2.1 is below the other two.

This means that people generally find recent proprietary models to be of lower quality than older ones.

Do you realize what that means?

You have teams with millions of dollars in funding and hundreds of people working on models, they spend months and months creating new versions that are rated lower than stuff released a year and a half ago.

Also, to read:

https://medium.com/@pareto_investor/chatgpt-has-just-been-dethroned-by-french-geniuses-bcee41843775?source=post_page-----5272ff33c5cd--------------------------------

What could explain that?

It’s not so much a question of becoming stupid, that is, of no longer being able to solve tasks, but it’s a general question of behavior.

Where maybe 6 months, a year ago, we could ask it to write a complete script, for example, with development, because it’s quite telling.

I remember at the time I was learning Swift, and I asked GPT to do really complex tasks on the Mac’s GPU, and it churned out scripts in Swift and generated shaders.

It seems we’ve gone from that to now, where if we look at some examples, we

If you look at this leaderboard below, which shows the biggest models, you look at the very best, but do you notice something strange?

There are several versions of GPT-4. Yes, they correspond to different versions at a given time. For example, GPT4 03–14 is the version from March 2023, 06 is from June 2023, and so we can already see the first thing is that they are not ordered chronologically.

Is there anything else?

Yes, on Claude’s ones. Claude 2 is below Claude 1, and CL 2.1 is below the other two.

This means that people generally find recent proprietary models to be of lower quality than older ones.

Do you realize what that means?

You have teams with millions of dollars in funding and hundreds of people working on models, they spend months and months creating new versions that are rated lower than stuff released a year and a half ago.

Also, to read:

https://medium.com/@pareto_investor/chatgpt-has-just-been-dethroned-by-french-geniuses-bcee41843775?source=post_page-----5272ff33c5cd--------------------------------

What could explain that?

It’s not so much a question of becoming stupid, that is, of no longer being able to solve tasks, but it’s a general question of behavior.

Where maybe 6 months, a year ago, we could ask it to write a complete script, for example, with development, because it’s quite telling.

I remember at the time I was learning Swift, and I asked GPT to do really complex tasks on the Mac’s GPU, and it churned out scripts in Swift and generated shaders.

It seems we’ve gone from that to now, where if we look at some examples, we