Anthropic has attributed Sonnet 3.5's model improvement to better training data.

"Which data specifically? Gerstenhaber wouldn’t disclose, but he implied that Claude 3.5 Sonnet draws much of its strength from these training sets."[0]

[0]https://techcrunch.com/2024/06/20/anthropic-claims-its-lates...

and water is wet

My guess, which could be completely wrong, Anthropic spent more resources on interpretability and it's paying off.

I remember when I first started using activation maps when building image classification models and it was like what on earth was I doing before this... just blindly trusting the loss.

How do you discover biases and issues with training data without interpretability?

Is it really that much better? I'm really happy with GPT-4o's coding capabilities and very seldom experience problems with hallucinations or incorrect responses, so I'm intrigued by how much better it can actually be.

In my experience Sonnet 3.5 is about the same as 4o for coding. Sometimes one provides a better solution, sometimes the other. Both are pretty good.

>or is there another reason Claude Sonnet 3.5 is so much better at coding than GPT-4o?

It's impossible to say because these models are proprietary.

Isn't the very article we're commenting on an indication that you can form a basic opinion on what makes one proprietary model different from another?

Not really, we know absolutely nothing about Claude 3.5 Sonnet, except that it's an LLM.

> It seems more and more that the solution to AI's quality problems is... more AI.

This reminds me of the passage found in the description of the fuckitpy module:

"This module is like violence: if it doesn't work, you just need more of it."