R to @egor_sp: insurancepointoffice.com

R to @imgyf: @egor_sp shared a nice one a while back. Can't find it now but if he shows up here maybe he can link to it :)

R to @egor_sp: 0.2

R to @egor_sp: let's get it egor

R to @TakoTreba: Scandinavia is known for good UI/UX and design in general. And that's awesome! I really appreciate when devs care about their clients by providing good human-centric services.

R to @egor_sp: When I see this even tho I complain so much about Finland 🇫🇮 every government website is perfect even supporting English language:D

I always thought German banks have the worst UX imaginable. I was wrong. Here's what you need to do in order to make a transfer via HSBC.

R to @gaeldemondragon: It’s supposed to determine how “creative” the model is. But it also has side effects like making the outputs more or less predictable.

R to @egor_sp: What is this temperature thing? I never heard about it until today and it's the second time already!

Does anybody have a good rule of thumb for GPT-4 temperature parameter? I still have 0 intuition about how to set it properly depending on the use-case.

Move important commands to the end of the prompt ⬇️ Based on my observation, the closer a command is to the end of the prompt, the more important it is to the model, and the less likely it is to be ignored by the model.

R to @egor_sp: Many findings like these are not described yet even though prompt engineers use them intuitively. I'm really looking forward for someone to compile the findings into a good article/book.

R to @egor_sp: For example, if you want to extract only action items, you can easily put in 5k tokens into GPT-4. But if you want to extract action items, key takeaways and deadlines, you're looking at something like 2k to get satisfactory results.

R to @egor_sp: Interesting thing is that it's filled up, as I empirically found, not by the amount of information, but by its complexity. The more complex things you ask a model to do, the less of the actual input it will be able to handle.

R to @egor_sp: Why is that? Turns out LLM are not only limited by the context window, they also have a thing called "attention". It can be thought of as RAM of an LLM.

R to @egor_sp: With 5k tokens of input text GPT-4 struggled to extract basic things like action points. Decrease it to 2k and the quality of the output rises dramatically.

R to @egor_sp: Let's say you have a simple langchain template that summarizes a piece of text. It turns out that if you change the size of the text you put in, the output will differ drastically.

Today I learned about another factor that affects the quality of your LLM outputs. It's the "relative complexity" of the your input. 🧵👇 #buildinpublic

R to @chris8fernandes: Even better: replace AI with "we're not sure what the thing does but we're releasing to production anyways".