Why OpenAI obfuscated the output

[daveshap]

Consider:

1. Noam says it's all trained to do it one shot, this lends credence to the idea that it has tags or tags or something like that.
2. There was a similar post that said that it had access to an unfettered, unfiltered model to do more open-ended thinking. I will see if I can find it.

[russellballestrini]

Our open version will show the chain and chain summary as a click to display above the LLM completion.

[daveshap]

make sure you watch the video, I explain the reasoning and show how it should look cleaned up

[russellballestrini]

Understood. I'm talking about an app or frontend for the resulting model during inference, not the finetuning process.

For example Claude 2 was notorious for context poisoning due to its censored nature.

Watching now at 1x speed. ; )

[daveshap]

we are not going to be working on a public API, frontend, or web app. We are building a dataset for finetuning a model. Please make sure you understand the purpose of this project and do not try to change the purpose or scope.

[blackeuler]

Will the fine-tune be sufficient to mimic the behavior? I think your data approach makes sense but Im thinking they have to be using the Stanford Self Reasoner paper as well. https://arxiv.org/abs/2203.14465
The paper says that fine-tune's couldn't match the performance.

[daveshap]

we will find out!