GPT-4 - How does it work, and how do I build apps with it? - CS50 Tech Talk

First, you’ll learn how GPT-4 works and why human language turns out to play such a critical role in computing. Next, you’ll see how AI-native software is being made.

Taught by Ted Benson, founder of Steamship, MIT Ph.D., & Y Combinator Alum; and Sil Hamilton, researcher of emergent AI behavior at McGill University.

***

This is CS50, Harvard University's introduction to the intellectual enterprises of computer science and the art of programming.

***

HOW TO SUBSCRIBE

HOW TO TAKE CS50

HOW TO JOIN CS50 COMMUNITIES

HOW TO FOLLOW DAVID J. MALAN

***

CS50 SHOP

***

LICENSE

CC BY-NC-SA 4.0
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License

David J. Malan

[NikTh]

That's why these are some of the best universities in the world . no wonder why so many students wants to enroll in there
The immediately include recent development in practical world instead of teaching you 20 year old syllabus

[hrishikeshh]

This lecture is going to be initial reference for so many people who are going to build things on top of GPT.

[CitizenWarwick]

I've been working on ~2000 token long conversational prompts with response formatting and decision making even with data structures in the context and it just keeps on giving, spent hours tweaking my prompts and they keep on giving, amazing tech!

[set_app]

I've been using GPT for quite a while and am glad I got to build up my own knowledge of what I thought it was capable of to then watch this and realize it can do SO much more!

[wda_digital]

Pretty damn cool. Thanks for the chat. If we ask GPT how many times PIZZA was mentioned it will probably return 'too many times' :) Now let's go build an app and force the AI into a loop.

[sudzam]

Loved the framework approach for each application!

[roccoruscitti]

I thoroughly appreciate this talk, I feel it did a great job to inspire me further into this particular field of development, even if only in small ways that are relevant to my particular work, or even to just try things as he said by hitting things with this new hammer!

[yash]

Like always CS50 never fails to amaze.😍

[techwithmatheusmello]

This is so basic but is so necessary, really good to be able to watch this. Thank you.

[jaceyang]

Love to see how quickly people can adapt to new tech and start building

[StephanHaewss]

Very good explainations! I would see these systems as a simulated intelligence or a way to do knowledge discovery from a learned model. But since it does not have real life experiences, needs, emotions and cannot "do" anything by itself, it seems to cover only part of what makes up a human or even animal. But certainly a great tool that can be used for many purposes.

[kokits]

thank you for sharing and for opening doors to the field

[talgatjampeissov]

🎯 Key Takeaways for quick navigation:

00:00 🧠 Understanding GPT: Introduction to GPT and its various descriptors.
10:43 🚀 Expanding GPT's Abilities: GPT's role in question-answering and how it becomes more than just a language model.
16:59 🤖 Companionship Bots: Creating personalized AI companions.
19:09 💡 Question Answering with GPT: Leveraging GPT for question-answering.
19:52 🔍 How vector databases work
21:00 🤖 Building question-answering bots
25:01 🛠️ Building utility function apps
28:06 📚 Leveraging creativity and domain knowledge
32:36 🌌 Exploring baby AGI and self-directed AI
40:31 🧠 How GPT-4 works and addressing hallucinations
43:21 🗣️ Influencing GPT-4's behavior through language
45:03 💼 Use cases and business value of AI apps
48:36 🔄 The evolution of AI models like GPT-4
51:11 🔒 Privacy implications of GPT-4 prompts and IP

Made with HARPA AI

[jeromeeusebius]

This was great lecture. Thanks for sharing this.

[K.F-R]

An excellent talk. Thank you for sharing.

[noevelasquez]

Thanks so much !!! God bless you guys.

[ShotterManable]

What an awesome talk guys. It was immensely helpful for me, I'm an enthusiast learner on AI besides not understanding the detailed maths of it. I feel this is an evolution of technology that any nerd wants to be in. And I'm so happy to be part

Thanks a lot, your knowledge sharing inspires me. greetings from argentina.

[lightconstruct]

Very useful and well presented lecture, also good questions.

[DownunderGraham]

I read that one of the reasons for the chatgpt inaccuracies is it's linear generation method. As mentioned it is trying to do it's best to predict the next logical word in a sequence. Unfortunately once it's made it's choice it is unable to correct the “stream of prediction”. This is apparently why, when you “prompt” it that there is an error it is able to re-read it's output and correct the error. I have heard that methods like “chain of thought” might help with this issue. This method allows the generation to backtrack up the tree to effectively undo a path it may have previously gone down and start down a different path. Much, much more computationally expensive though.

[pmarsec]

Interesting... my thought on why the "experienced <profession>" prompt or the instruction to "prefix the answer with 'my best guess is'", is that it affects what section(s) of its database, the internet, it uses in modeling an answer. For example, someone who's trolling or isn't particularly interested in the accuracy of their answer is more likely to state something as fact than to couch it in terms of confidence or best guess. Likewise, if "experienced professionals" frequent a forum and chatgpt can tell that, maybe it chooses their style of answer (which happens to be more correct) over something from another area of the internet.