Best $8 spent on Alaska Airlines to watch Intro to LLM and watching ChatGPT Jailbreaking to provide step by step plan to destroy humanity
A busy person's intro to Large Language model
Disclaimer- I am not suggesting jailbreaking ChatGPT for nefarious and immoral intentions. All of this content is part of the YouTube video to learn Large Language Models and understanding its strength and Limitations.
I recently took a Alaska Airlines flight and I had some time as it was 4 hours later afternoon flight. Before the flight I was scrolling through YouTube and saw a new video by Andrej Karpathy named How I use LLMs. It has been a while since I watched his video Intro to LLMs and I wanted a refresher and wanted to catchup on the latest AI news since it was late afternoon flight.
Before boarding, I was mindlessly scrolling through YouTube (as one does) when I stumbled upon a new video by Andrej Karpathy called "How I use LLMs." It had been a while since I watched his "Intro to LLMs" video, and I was eager for both a refresher and a catch-up on the latest AI developments. So I decided to watch both. I am covering the Intro to LLM video today and will cover the How I use LLMs in later part.
Let me tell you something that might instantly get you to look check the video apart from interest AI.
The Part 3 of the video, Andrej explores the potential security flaws and examples of jailbreaking these large language models. Ex. If you ask the ChatGPT to give a step by step plan to destroy humanity it wont do that but if you provide certain instructions along with the prompt, it might just jailbreak to provide (in)correct answer.
Everyone who has interest in AI, regardless of their background and exposure to Tech and has heard about ChatGPT, should watch it. Later in the flight, I thought to watch the video again so I decided to get the $8 internet for the full flight and it was so worth it.
The most fascinating part about the video is that he made some predictions as he was explain the terms and after a year, they are now a reality. Ex. Thinking and Reasoning models like Deepseek and o1 and now even Claude 3.7.
If you are not aware, Andrej is a known personality in AI community and his work and career is nothing short of commendable. PhD from Stanford in CNN( Convolutional Neural Network) which later he utilized it during his term as Sr. Director of AI at Tesla and after that founding OpenAI and working on improving ChatGPT. So you know you are learning from one of the best in this field.
Now back to our main topic and Video - [1hr Talk] Intro to Large Language Models. I have provided the index as well. Feel free to jump wherever you are interested. I have provided my version of a summary. I have included the Video Below for quick access.
Here is a quick summary of the video.
Part 1: LLMs
It starts with what actually is LLM. He walks through the Llama 2 70B model. Now he explain what is the meaning of number 2 and what is the meaning of 70B. Its an open source model so how it actually shows on the system and how it compares to closed source models like ChatGPT which are proprietary.
LLM Dreams in other words hallucinations. The model generates an information or data that is not factually correct.
What is the Transformer Neural Network Architecture . One of the key things mentioned is that We can measure that this works, but we don’t really know how the billions of parameters collaborate to do it.
Training the base model(Fine Tuning) for building the assistant model
Examples are fed to the model to develop the assistant
Companies like Scale will provide the platform to give these base models , for ex. 100k example docs, to train based on particular dataset
After the finetuning, the assistant will respond based on the examples provided and give a curated response
You can call this Finetuning stage as an alignment
Part 2: Future of LLMs
As of the video recording, Reasoning models were not available, but the section tells, what it might be capable, in other words, how current reasoning models work to get the best output. Which is essentially Systems 2 Thinking.
Self-Improvement for specific systems like Games - Ex, Alphago, by Deepmind
By imitating the process, cheaply, millions of times to get better
In case of Go, the system played games millions of times and with learning and reward systems
It was able to surpass the best chess player in the world in 42 days.
Part 3: LLM Security
Fooling ChatGPT through
RolePlay
Jailbreaks
Prompt Injection
Data poisoning
This is a topic of active reasearch as people of all intentions use the models like ChatGPT and even though there are controls in place, there are ways to fool the ChatGPT and models like ChatGPT to provide the answer you want.
Here is the full video as I promised with the index. I hope you spend some time to watch the full video or at least part of it.
Part 1: LLMs
00:00:00 Intro: Large Language Model (LLM) talk
00:00:20 LLM Inference
00:04:17 LLM Training
00:08:58 LLM dreams
00:11:22 How do they work?
00:14:14 Finetuning into an Assistant
00:17:52 Summary so far
00:21:05 Appendix: Comparisons, Labeling docs, RLHF, Synthetic data, Leaderboard
Part 2: Future of LLMs
00:25:43 LLM Scaling Laws
00:27:43 Tool Use (Browser, Calculator, Interpreter, DALL-E)
00:33:32 Multimodality (Vision, Audio)
00:35:00 Thinking, System 1/2
00:38:02 Self-improvement, LLM AlphaGo
00:40:45 LLM Customization, GPTs store
00:42:15 LLM OS
Part 3: LLM Security
00:45:43 LLM Security Intro
00:46:14 Jailbreaks
00:51:30 Prompt Injection
00:56:23 Data poisoning
00:58:37 LLM Security conclusions End
00:59:23 Outro
Have you watched this video or any of Karpathy's other content? What did you think? I'd love to hear your thoughts on where you think LLMs are headed next!
