A Hackers' Guide to Language Models

As we move further into the intricacies, Jeremy unpacks advanced strategies for model testing and optimization, utilizing tools like GPTQ and Hugging Face Transformers. He also explores the potential of specialized datasets like Orca and Platypus for fine-tuning and discusses cutting-edge trends in Retrieval Augmented Generation and information retrieval. Whether you're new to the field or an established professional, this presentation offers a wealth of insights to help you navigate the ever-evolving world of language models.

(The above summary was, of course, created by an LLM!)

00:00:00 Introduction & Basic Ideas of Language Models
00:18:05 Limitations & Capabilities of GPT-4
00:31:28 AI Applications in Code Writing, Data Analysis & OCR
00:38:50 Practical Tips on Using OpenAI API
00:46:36 Creating a Code Interpreter with Function Calling
00:51:57 Using Local Language Models & GPU Options
00:59:33 Fine-Tuning Models & Decoding Tokens
01:05:37 Testing & Optimizing Models
01:10:32 Retrieval Augmented Generation
01:20:08 Fine-Tuning Models
01:26:00 Running Models on Macs

This is an extended version of the keynote given at posit::conf(2023). Thanks to @wolpumba4099 for chapter titles.

[howardjeremyp]

Gotta admit I'm feeling kinda teary reading all the lovely comments here. Thank you everybody -- love you all!

[buildar]

Just realised Jeremys paper led to the LLM revolution. Such a humble kind man. God bless you and all your students. You are such an example to follow. An example in character, humility and intelligence.

[Jporomaa]

This is probably the best invested youtube time of this year so far. What a gem. A lot of things he mentions have taken me month to figure out by my own. My new GPT-4 prompts will begin with "You are the expert Jeremy Howard..."

[wolpumba]

*Transcript Summary:*

- Introduction & Basic Ideas of Language Models (00:00:00 - 00:18:05)
- Limitations & Improvements of GPT-4 (00:18:05 - 00:31:28)
- AI Applications in Code Writing, Data Analysis & OCR (00:31:28 - 00:38:50)
- Practical Tips on Using OpenAI API (00:38:50 - 00:46:36)
- Creating a Code Interpreter with Function Calling (00:46:36 - 00:51:57)
- Using Local Language Models & GPU Options (00:51:57 - 00:59:33)
- Fine-Tuning Models & Decoding Tokens (00:59:33 - 01:05:37)
- Testing & Optimizing Models with GPTQ & Hugging Face (01:05:37 - 01:09:48)
- Fine-Tuning with Llama 2 & Platypus Datasets (01:09:48 - 01:10:32)
- Retrieval Augmented Generation & Information Retrieval (01:10:32 - 01:20:08)
- Running a Private GPT & Fine-Tuning Models (01:20:08 - 01:22:32)
- Running Models on Macs (01:26:00 - 01:27:42)
- Discussing Llama.cpp & Its Cross-Platform Abilities (01:27:42 - 01:30:07)
- Challenges & Opportunities in Language Models (01:30:07 - 01:31:05)

Key points of interest: Function usage in GPT-4 (00:46:36), OCR application with Google Bard (00:33:59), and improving GPT-4 responses with custom instructions (00:24:36).

[MensRea]

No one can explain a topic like Jeremy👍

[DHAiRYA]

The moment I got to know that you and Andrej weren't included in the Time's list, I realized that the people making such lists have no idea what they are doing. Loved the tutorial, thank you!

[JustSayin]

I found this video so useful that I felt compelled to pull my keyboard closer toward me, fix my posture, and write this comment - something I rarely do. I'm a professional data scientist hoping to push my company's GenAI agenda and this video makes me feel like I can actually do it! Thank you for so clearly encapsulating the state of LLMs. I'd learned many of these concepts before and this video is the glue that now holds it together.

[KEcAzZ]

A true legend! So far, I have not seen a better educator than Jeremy. His approach of teaching is what all schools and universities need! I am always interested to learn more, whenever I hear Jeremy. Thank you!

[dr_rd]

Truly enlightening! As a software engineer with limited math and data science knowledge, this video has been a revelation. The way Prof. Howard simplifies complex concepts is incredible, making each rewatch rewarding with new insights. Really grateful for his content that opens up the world of LLMs to a broader audience. His clear and thorough explanations are incredibly invaluable. Thanks, Prof. Howard, for demystifying this topic and helping us all learn.

[wolpumba]

*Positive Learnings:*
1. Language models, such as GPT4, are tools that can predict the next word in a sentence or fill in missing words in a sentence.
2. Language models have the ability to create a rich hierarchy of abstractions and representations which they can build on.
3. The guide will cover all the basic ideas of language models, including how to use open source and open AI-based models.
4. GPT4 can solve many tasks that it is often claimed it cannot.
5. GPT4 can be primed to give high-quality information by giving it custom instructions.
6. AI can be used to write code and parse large-scale data quickly and efficiently.
7. AI can be used in optical character recognition (OCR) for extracting text from images.
8. AI can be used in data analysis to create comprehensive tables from scattered information.
9. The OpenAI API allows users to use AI programmatically for data analysis and other repetitive tasks.
10. Function calling can be used to create a code interpreter that runs inside Jupiter.
11. Pre-trained models can be accessed using the Hugging Face library.

*Negative Learnings:*
1. Language models are not always useful on their own and need to be fine-tuned.
2. GPT4 often repeats mistakes and it is difficult to get it back on track once it starts making mistakes.
3. GPT4 has limitations such as not knowing about itself, not knowing anything about URLs, and not knowing anything after its knowledge cutoff in September 2021.
4. GPT4 does not always give correct responses.
5. AI has limitations in code interpretation and cannot substitute for human programmers.
6. The use of the OpenAI API can result in rate limits which need to be handled correctly.
7. Fine-tuning is needed to make the pre-trained models more useful.
8. The use of GPUs for local language models can be expensive and may require renting or purchasing GPUs.

[raindrops]

This video landed up on my feed and out of curiosity I started watching and before I knew it had watched the entire video and taken copious amounts of notes too. One of the best videos I have ever watched!

[marvit]

We do not deserve you Jeremy! YOU ARE AN AMAZING TEACHER AND HUMAN BEING! Thanks, really, for all these beautiful lectures!!

[x]

You are literally changing lives, all for free. Thank you.

[abhiabhi]

I am a total beginner but u made me understand abt LM models way better than anuone else..u r such a great teacher..I pray for giving u Lord Gurus blessings dor more insight and vision for such a Humble and good Soul.😊😊

[kindoblue]

I liked the video even before watching. Thanks Jeremy for your work, always learning from your content.

[satirthapaulshyam]

Hoping to look forward it.Ur fastai stable diffusion course was Perfect to the minute details

[sidsarasvati]

Thought provoking one code block at a time. As usual Jeremy the king

[ytpah]

🎯 Key Takeaways for quick navigation:

00:00 🤖 Introduction to Language Models
10:27 🧠 Neural Network Basics
16:38 🚀 The Power of GPT-4
24:53 🌐 Limitations of Language Models
25:23 💡 Language model limitations:
31:32 📊 Advanced Data Analysis:
36:18 💰 OpenAI API Pricing:
39:19 🧩 Using OpenAI Functions:
46:40 🐍 Custom Code Interpreter:
51:13 🐍 Creating a Python code interpreter
53:39 💻 Running a language model on your own computer
55:01 🏎️ Choosing a GPU for language model work
56:15 🖥️ Options for renting GPU resources
57:57 💾 GPU memory size and optimization
59:20 📚 Using Transformers from Hugging Face
01:00:06 🏆 Evaluating and selecting the right model
01:14:12 📖 Retrieval augmented generation for answering questions
01:17:10 📚 Overview of using language models for document retrieval and question answering
01:20:35 💼 Private GPT models for document retrieval
01:21:03 🎯 Fine-tuning language models for specific tasks
01:25:15 📊 Building a language model for SQL generation
01:26:36 💻 Running language models on Macs

[solutrian]

I've been a student of yours for years, thank you for all you do for us.
A non-optimal but correct solution to the cabbage, wolf & goat crossing works by just using incantation "think step by step" :
Think step by step to solve the following tricky problem:
Here is a logic puzzle. I need to carry a cabbage, a goat and a wolf across a river. I can only carry one item at a time with me in the boat. I can't leave the goat alone with the cabbage, and I can't leave the cabbage alone with the wolf. How can I get everything to the other side of the river?
This classic river crossing puzzle can be solved through a series of strategic moves that ensure the safety of each item. Here's a step-by-step solution:

1. **Take the goat across first:**
- You take the goat across the river and leave it on the other side.
- This ensures that the goat is safe from the wolf, and the cabbage is safe from the goat.

2. **Return alone:**
- You return to the original side alone to get either the cabbage or the wolf.

3. **Take the cabbage across next:**
- You take the cabbage across the river.
- Now, the goat is on the other side, and the cabbage is with you.

4. **Bring the goat back:**
- You bring the goat back with you to the original side.
- Now, the cabbage is alone on the other side, and the wolf is alone on the original side.

5. **Take the wolf across:**
- You then take the wolf across the river.
- Now, the wolf and the cabbage are on the other side, and you have the goat on the original side.

6. **Return alone:**
- You return to the original side alone.

7. **Take the goat across one final time:**
- Finally, you take the goat across the river again.
- Now, all three— the goat, the cabbage, and the wolf —are safely on the other side.

This way, you ensure that the goat is never left alone with the cabbage, and the cabbage is never left alone with the wolf, solving the puzzle.

[bartekbinda]

This is golden summary of the state of the LLMs, Thank You