Learn to Spell: Prompt Engineering (LLM Bootcamp)

New course announcement ✨
We're teaching an in-person LLM bootcamp in the SF Bay Area on November 14, 2023. Come join us if you want to see the most up-to-date materials building LLM-powered products and learn in a hands-on environment.
www.scale.bythebay.io/llm-wor...
Hope to see some of you there!
--------------------------------------------------------------------------------------------- In this video, Charles gives high-level intuitions and a default playbook for prompting language models. We consider two very different sources of intuition: "language models are statistical models of text" and "prompts are magic spells". Then, we review prompting techniques, like decomposition, reasoning, and reflection.
Download slides from the bootcamp website here: fullstackdeeplearning.com/llm...
Intro and outro music made with Riffusion: github.com/riffusion/riffusion
Watch the rest of the LLM Bootcamp videos here: • LLM Bootcamp - Spring ...
00:00 Intro
02:15 Language models are statistical models of text
04:50 But "statistical model" gives bad intuition
08:35 Prompts are magic spells
09:57 Prompts are portals to alternate universes
16:42 A prompt can make a wish come true
23:03 A prompt can create a golem
27:10 Limitations of LLMs as simulators
32:17 Prompting techniques are mostly tricks
33:40 Few-shot learning isn't the right model for prompting
37:40 Character-level operations are hard
40:23 The prompting playbook: reasoning, reflection, & ensembling

[@oleksandrasaskia]

I love this talk! So easy to follow, so fun, with humor, and I especially appreciate down-to-earth explanations and parallels.

I am impressed by how great of a teacher Charles Frye is! The content is very informative and easy to comprehend. Thank you very much!

[@charles_irl]

Thanks for the kind words!

[@ronaldokun]

Second it!

[@Zoronoa01]

Very interesting lecture!

[@edwardsmarcom]

This is great. His voice and intonation sounds just like comedian David Cross.

[@arvj123]

I can't believe this is for free. Thank you. I probably need to watch this video one more time to get a better understanding. Some information just breezed past my head. :D

[@Zoronoa01]

around 1:00 you talked about a lecture you did in the morning, is there a video of that lecture?

[@ayanghosh8226]

Awesome explanation. Love it!

[@mydaisy2]

What a valuable video it is.. thanks for sharing your knowledge... :)

[@agenticmark]

Prompts are certainly magic incantations. The difference between the new wizard and the old wizard is the ability to say the spell correctly :D Prompting is very much like hacking out programs. The more you have a "conversation" with your code/tools, the better your results are. The more you have a conversation with GPT, the more that "clicks" and your prompts get better which leads to the models giving you better results! I find my lack of higher maths to be a boon in prompt engineering because I make less assumptions not based on the responses of my experiments and journey. Great talk!

[@joaoprado9385]

Great content, man!

[@ArtOfTheProblem]

Really great lessons I'm impressed, would love to work with you

[@RK-fr4qf]

Very good presentation

[@zacboyles1396]

19:00 - at what point does eliminating the “undesirable social biases” make the model seem ignorant? Worse than that, it seems dangerous to get into a habit of hiding probabilistic realities that cause some to feel uncomfortable instead of unleashing bright minds to create solutions for them - assuming the ‘afflicted’ desire a solution in the first place. Sure, some tennis shoe customer service chat bot doesn’t matter but it’s not going to stop there.

[@charles_irl]

In the context of the lecture, the intent there is just to present the "don't be biased" instruction prompt as an example of the use of the "Literally Just Ask, Bro" Pattern™. I chose that example to try and make it clear that you can "just ask" for fairly complex behavior and get it. Steering model behavior requires choices throughout the stack, from pretraining and finetuning to prompting and monitoring. It is a really hard problem, and a simple approach like a one-line prompt hack is far from solving it!

[@theoqudatq2150]

Not a very useful presentation. I learnt nothing new. No "magic" recipe.