1
For ex: Alice, a great designer, could give each of her students capital that the students can use to pay rent, food, products and services to help them design, explore, or whatever (it’s their money!).
But what exactly would Alice get in return? Let’s say we create a “personal token”, an instrument that represents an individual’s potential, with transactable shares, grounded in their equities in companies and other personal tokens (via dividends on capital gains).
So Alice would get shares (equity) in each student’s personal token in exchange for her training + capital.
This would mean:
1. Students don't take on debt. In fact they get paid to learn!
2. Alice is held accountable. If she fails to meaningfully improve her students' outcomes, she loses her investment. This means Alice is forced to adapt her training to what is actually relevant to the world.
3. Teacher-student relationships last years / decades, not semesters. Alice is strongly incentivized to help her students whenever they need it, because she has equity in their long-term success.
But why? AI is making outcomes extreme (power law distribution). We can already feel this in software engineering: AI makes the best engineers far better than the median. As the gap between the best and rest grows, it becomes too risky to finance education with debt for the same reason it’s too risky to finance startups or content creation with debt.
Paul Graham was one of the earliest examples of this model. He didn’t need to guard his knowledge or put it behind a paywall because he had a much more powerful way to capture value: by investing in the founders his essays attracted. If teachers could invest in students, knowledge would spread more freely because sharing knowledge itself would become a funnel for investing. Even students who never raise would still benefit from the higher-quality knowledge that becomes available.
Thoughts?
The Problem: Since it came out, I've been using ChatGPT for various tasks, such as coding, generating database queries, reviewing my writing, and learning new things. Initially, I found myself constantly managing useful prompts in a Notion doc. I'd either paste a prompt from Notion to provide the right context or sift through my chat history to locate that context.
Some examples of snippets I frequently used include: - A detailed summary of my Tech stack. - Movie and book preferences (for suggestions). - My database schema (so that it could write queries). - Writing guidelines to follow when reviewing my writing.
This workflow was fine for 3 prompts, but sucked for 10. So I set out to build a better way.
Typemagic automatically searches for the best prompts in your library, applying them when needed based on your chat. Once you save your a prompt, you (hopefully!) don't have to worry about applying it manually - Typemagic takes care of it. While there's definitely room for improvement, this approach has been more effective than manually searching and pasting the right prompts for me.
I'd greatly appreciate any feedback you have, especially critical feedback :).
The Journey: This is the second iteration of the product. In the first version, you could create ChatGPT agents to chat with. But before asking what you wanted, you had to choose the right agent to chat with. Needing to make this decision every time you wanted to switch context was too much friction.
I realized my design approach was outdated. With AI, we have the opportunity to offload more decision making from users' minds onto AI. The right balance is one in which users act more as managers: approving / rejecting proposals, vs. actually, outright, deciding. As AI becomes more powerful, this balance will shift towards AI making more / most of the decisions.
With this insight, my friend Akash and I designed the current version that relies on Typemagic choosing the prompts during your chat while giving you visibility and control over the prompts applied.
Under the hood: Currently, Typemagic uses embeddings (https://platform.openai.com/docs/guides/embeddings) to find the best prompt matches by comparing your current chat embedding with the embeddings for all of your prompts to find the best matches.
Using Firebase, Firestore for auth and to store data. Deployed on Vercel (huge fan!).
What's Next? I'm working on enhancing the experience with faster performance, better recommendations, and the ability to share and discover useful prompts.
I'd greatly, greatly appreciate your feedback. Thanks for reading this far.