Some notes aimed at: * People in enterprises, startups, and small businesses who want to use AI * Maybe you want to make your own models * Maybe you want to make your own ChatGPT Where should you add AI? * Does it help your product be more loved by end users by adding AI in a way the end users can use? * Does it help you produce your product in a way that makes it more loved by end users, by using AI to build it faster/better? * Does it help you reach more customers? Think of AI like a database. A database is something that can be added to various products that would enable useful things for the end user. And in some cases being able to use a database enables you to make new products that didn't previously exist. And databases might allow you to make internal tools that make it easier for you to produce products. The future is "AI inside" (a la Intel Inside) - AI added at various places in various existing products, and in some cases, enabling totally new products. Some of these uses of AI will allow businesses to strengthen existing moats, some will allow them to create new moats for new or existing products, and some will allow competitors to eliminate existing moats. Suggested pathway for finding ways to use AI inside the business: * First, encourage all employees to utilize AI as much as possible, wherever it helps them improve the quality, cost, or speed of their work. * For this, encourage them to use the main end user tools. * ChatGPT Plus, Midjourney Plus, Adobe Photoshop Generative Fill * Then, after a month or two, explore where it's most helpful - the places where employees are using it daily, or the places where it's saving major money or time. Consider looking for existing pieces of software that better serve those best use cases. Consider building internal tools to better serve those best use cases. * Find the employees who are making best use of the tools, and invite them to create videos showing how they do it to be shared with the team, and/or to run workshops for other employees. This is a "bottoms up" approach. It's likely to be more effective than any kind of "top down" approach. Instead of saying when and where AI should be used, you're giving people access to tools, and allowing them to find and explore the best use cases themselves. Potential internal use cases: * Customer support drafting * Sales outreach drafting * Internal knowledge search * A "smart coworker" for all employees who can help with brainstorming, thinking through plans * Product photography generation and other image generation * Writing (copy, ads, emails, web pages) * Summarizing emails you receive Suggested pathway for finding ways to add AI to your products: * Take the employees who were best at finding ways to use AI tools to help them in their job, and pair them with some of the people on the product team. Or hire some people with deep prompting expertise. * Observe your end users going through the full product journey. Observe them in person. Don't talk, observe. * Make notes of the points of confusion and frustration. * Make notes of the points where AI text search, AI text generation, or AI image generation could've saved them time or effort or got them better results while using the product. * Pick the 2-3 that would be most magic for the end users to start with. * First, see if you can add AI at those places with zero-shot GPT-4 prompts. If that doesn't work, try a few-shot GPT-4 prompt (give 3 examples). If that doesn't work, you could consider fine tuning, but my experience with fine tuning text-davinci-003 (GPT-3) hasn't been very promising. I expect fine tuning GPT-4 will be better, but it's not available yet due to capacity constraints. Should you build a fine tuned model? * Only if your problem can't be solved with zero-shot or few-shot prompts. Which LLM should you use? * GPT-4 is considerably better than alternatives in most cases. But, the reliability isn't great, the speed isn't great, and the cost is high. * Use GPT-4 for prototyping. * Once you have use cases that are working well for end users and you want them in production for all users, consider: * GPT-3.5-Turbo with a few-shot prompt * A fine tuned commercially available model Companies that'll be happy to take your money to help with things: * Stability AI: [email protected] * https://www.lamini.ai * Cohere * OpenAI Foundry * Anthropic My note is that the difference between the best models and the second best models is very large at this point. I think that enterprises should be cautious about working with anyone other than the best LLM companies, otherwise the results you get may be much worse than you're expecting if you're used to ChatGPT Plus with GPT-4. **How to make your own ChatGPT?** ![[List of tools for making a "ChatGPT for your data"]] **How to train your own model?** I think a lot of enterprises will waste 7-8 figures on getting custom models built that don't end up producing any value. Treat it like building a product - de-risk whether people will want what you build, whether you'll be able to make it, whether you'll be able to make it excellently, and whether you'll be able to make money from it, and whether you'll be able to reach users. Before you invest most of the money.