Lessons from OpenAI's Former Chief

⏱️ Your Morning Brief (TL;DR)

GM Decision Makers,

AI literacy and leadership alignment go hand in hand.

We’re learning and trying to keep up ourselves.

This edition focuses on the basics of using LLMs. Andrej Karpathy, former head of AI at Tesla and Co-Foudner of Open AI released a masterclass on how he uses AI.

Despite his credentials, he provides an accessible, jargon-free breakdown that any executive can immediately apply.

Rather than have you watch a 2-hour video, we break things down for you.

Let’s dive in.

 How Andrej Karpathy uses LLMs

Open up your favourite model (OpenAI, Claude, Perplexity) and give these scenarios a go.

1. Using Multiple LLMs as an "LLM Council" - You can consult multiple LLMs the same question to get different perspectives. Try asking Gemini, Claude and Grock for travel recommendations, ask for suggestions on cool cities to visit.
   

2. Stop Google Searching: Early on, you’d run into issues where models weren’t trained on the latest models. Now, most LLMs have the ability to access the internet for up-to-date information. You can uses tools like Perplexity AI extensively for search-based queries due to its strong search capabilities and citation of sources. ChatGPT also has a "Search the web" feature. You’ll find it more efficient than manual web searching.
   

3. Solving Complex Problems (Thinking Models): For difficult problems, especially in math and code, Karpathy utilizes "thinking models" which are fine-tuned with reinforcement learning to perform more extensive reasoning. These thinking features most beneficial for complex tasks and not necessary for simpler questions and queries.
   

4. In-depth Research (Deep Research): For more comprehensive research on specific topics, use "Deep Research" feature available in the ChatGPT Pro subscription (also available with Perplexity AI and Groq). You’ll get a combination of internet search and extensive thinking over a longer period (tens of minutes) to produce detailed reports with citations, akin to a custom research paper. Use Deep Research to generate thorough reports to compare different kinds of products, competitive analysis and exploring topics.
   

5. Document Analysis: Try uploading documents (like PDFs of research papers or blood test results) to LLMs and ask questions about the document. You’ll get summaries and explanation of the content. This is helpful for understanding complex information and for reading books more effectively by asking questions and getting clarifications in real-time.
   

6. Code Generation and Assistance (with caveats): While he acknowledges LLMs can write code and some platforms like Cloud offer "artifacts" for creating simple web applications, TIP: Personally prefers using dedicated IDE extensions like Cursor for professional coding work.
   

7. Data Analysis and Visualization: Andrej walks through ChatGPTs "Advanced Data Analysis" to analyze data, create plots and charts, and perform extrapolations. You can use it to visualize datasets, but note that at this stage there can be inaccuracies and you should review the results.

8. Generating Custom Podcasts: You can use Google's NotebookLM to generate custom podcasts from uploaded documents or web pages on niche topics of your interest. You can use this to passively learn while walking or driving.
   

9. Multimodal Interaction: Talk to your LLM for interactive audio conversations, upload images and ask questions about them. Like a “LLM Council”, you an use the LLM as a thought partner, summarize, and draft an initial document for you.

GPTLDR Takeaway

Unlike searching a topic up on Google, there isn’t one way to use LLMs. Use multiple LLMs for different perspectives, use their search capabilities, and employ specialized features like "Deep Research" for comprehensive reports. Remember to verify outputs and use the right tool for each task

❓ Helpful Tips for Accuracy and Personalization

• Verify Responses with Cited Sources: After getting an answer from an AI, have a look at the the cited sources the model gives you to to make sure the AI didn't just make things up.

• Review what the AI creates: Even though AI is capable of Advanced Data Analysis that makes great charts, you need to understand what its code is doing and watch it carefully because it sometimes makes mistakes or misses details.

• Memory: You can turn on ChatGPT's memory so it remembers past conversations. This makes the AI's responses more useful because it knows what you've talked about before.

• Custom Instructions: You can give ChatGPT instructions on how to respond by changing settings. For example, you can make it more formal or casual, or tell it what kind of answers you prefer.

• Custom GPTs: In ChatGPT, you can create special versions of ChatGPT for tasks you often do (copywriting, email responses). These are basically saved instructions that save you time repeating instructions. For example, he made tools to help learn Korean - one pulls out vocabulary, another translates in detail, and another reads Korean text from images.

📚 Interesting Reads

• Speaking Things into Existence by Ethan Mollick discusses how natural language interfaces like "vibecoding," is emporing users to transform ideas into tangible outcomes.

• Google’s The ‘Future of AI: Perspectives for Startups’ report features 23 leading voices in AI

• 7 Things You Can Do With AI Beyond Just Talking to It

• How to use AI for Fact Checking

• AI Basics for Beginners - Google’s AI Essentials Course

 ➜ Until Next Week

We know things are moving fast out there, but we're still in the early innings of this technological revolution. The most important part right now is to experiment freely, test different approaches, and discover what actually works for your specific needs.

Until next week,

The GPTLDR Team

AI, Simplified for Decision-Makers