aifaq.wtf

You can follow this up with 404 Media's Hugging Face Removes Singing AI Models of Xi Jinping But Not of Biden

I think the way to think about prompt engineering is: what would the best teacher preface an instruction to a student with, if they really really wanted the student do do the best possible job?

The worst performer, at 62.7%:

Start by dissecting the problem to highlight important numbers and their relations. Decide on the necessary mathematical operations like addition, subtraction, multiplication, or division, required for resolution. Implement these operations, keeping in mind any units or conditions. Round off by ensuring your solution fits the context of the problem to ensure accuracy.

That is obviously an awful way to start a lesson or a test. Even if someone knows the answer they're going to lose their minds!

The best performer, at 80.2%:

Take a deep breath and work on this problem step-by-step.

So relaxing, so kind, so guaranteed to ensure high performance.

Large Language Models as Optimizers: Paper here

According to Betteridge's law of headlines:

Any headline that ends in a question mark can be answered by the word no.

It's a little in the weeds of how these tools work, but generative AI tools are based on "tokens," individual units of text or information. Usually the tokens are developed organically by the model itself, but in this case, the tokens were developed by the team at Baichuan, and they're very revealing of the model's... biases? I don't think I'd call them a bias, just... a reflection... of... something.

Read the responses.

If you're angry about companies crawling the net to steal your text/images and train their machines using it, this one's for you! I and everyone else hate Terms of Service, but as a counterpoint to the "we need an opt-out mechanism for data collection" argument: 85% of the top domains in the LAION2B-en dataset already opt out through their TOS.

(LAION is a series of datasets of images + captions that are used to train models.)

The part everyone is especially loving is this:

"Surveying the AI’s responses for misleading content should be “based on your current knowledge or quick web search,” the guidelines say. “You do not need to perform a rigorous fact check” when assessing the answers for helpfulness."

Which, against the grain, I think might be perfectly fine. Your model is based on random information gleaned from the internet that may or may not be true, this is the exact same thing. Doing any sort of rigorous fact-checking muddies the waters of how much you should be trusting Bard's output.

Read the thread, there are a lot lot lot of useful links in there. I won't even put them here because there are so many (...and Twitter has better previews).