Ai Ai Captain

I thought long and hard about what I wanted to write about for the last essay of 2022.

I quit my job in April 2022 to start something of my own. I’ve always wanted to write consistently and decided to do so by sharing something each week. It’s one of the best decisions I’ve made. Nothing makes my day like a reader telling me they enjoyed an essay so thank you for reading this.


Anyway, I’m going to write about the theme that I’m most bullish on: artificial intelligence. Specifically, I want to cover a couple of points that have been top of mind for me:

  1. Why this is AI’s golden moment
  2. Truth and accuracy in AI
  3. Monetising AI first applications
  4. Rethinking the user experience

I don’t have good answers for all of them, but I do believe they are critical questions to ask in relation to AI.

Why this is AI’s golden moment

We’ve been using AI and ML for years. Google search uses it. Apple uses it to tag your photographs. So why is the world going gaga over the current wave?

It’s really because of how accessible AI is today. The novelty is not in being able to predict X or classify Y, we’ve been doing that for years. It’s about how easy it is to do that from a cost and developer experience POV. Here’s a non-exhaustive list of things you can do with AI using a single API call:

  1. Generate images or video from text
  2. Generate code from natural language
  3. Summarise text
  4. Classify text

OpenAI and StabilityAI are to AI, what AWS was to the cloud. We’re already seeing an explosion of applications. Many products will fail but the ones who do succeed will be big.

Imagenet is a database of images used for image classification models. The cost to train models (up to the same level of accuracy) has fallen consistently over the last 5 years.


Truth and accuracy in AI

You ask ChatGPT a question and it provides a convincing answer. How do you know it’s accurate? You don’t. I wrote about how Neeva is trying the accuracy problem.

If someone asked me how I decide whether something is accuracy, my response would be something like: “Google X, look for a combination of for and against arguments, and then make my decision”. You can train a machine to do this. It can help you fact check statements and provide an accuracy score.


My question though is a people question, not an engineering question. How do you know that something is true?

A part of me thinks that people don’t want to accept the truth. It is assumed that drinking a cold drink when you have a sore throat is a bad idea. Is that really true? I’m not sure. Whatever the right answer is, I can bet that there’s a sizeable group of people who will choose to believe it anyway. We as humans get some comfort in believing what we’ve always believed. Change is hard.

I wonder how this dynamic plays out in a world where radical transparency and truth are at your fingertips.

Monetising AI-first applications

AI is cheap, but not as cheap as traditional software applications.

Every prediction costs money. There are really two choices. First, you use an existing service like OpenAI. Zero fixed costs, high marginal costs. Second, you build your own model. High fixed costs, low marginal costs.

OpenAI’s pricing for language models. 1K tokens ~ 750 words.
OpenAI’s pricing for language models. 1K tokens ~ 750 words.

Whatever route you choose, the cost of providing your service is higher than a traditional software application. A part of me thinks that the only way you can price these applications is based on usage. If I offer you an AI service for $100 per month and give you ‘unlimited’ predictions, my margins may not look very healthy.

I don’t have a good answer for this yet. I just know that pricing based on usage makes the purchasing decision complex. If I was buying “100 emails for $20”, I need to think about whether I’ll cross that threshold. If I am buying a service for $20 per month with no cap on usage (most common model for traditional software), it’s easier to decide.

Rethinking the user experience

Is a chat interface the best one for your use case? I don’t know. My (working) mental model is that chat works best when the final output can vary a lot. Let’s look at two examples:

Example 1: I want help writing code. Github Copilot nails this experience because it does the job on the go. Without AI, I’d search on Google and Stack Overflow, try a bunch of different things and then decide on what works best. With Copilot, I have recommendations and help on the fly.

Example 2: I want recommendations for a diet. This could benefit from a chat like interface because the inputs for the recommendation are complex. Without AI, I seek advice from a dietician. They probably ask me about my height, weight, exercise schedule, current diet and goals. Based on the goals, they ask another series of questions. Then, given all the information, they make a recommendation. This might benefit from a chat like interface.

Anyway, my point is, there’s lots to be done on the user experience side when building AI first applications.

To close

I’m an optimist and I take pride in it. I’d prefer to bet on something and be wrong about it over being a pessimist. I’m excited about what AI is going to bring. I jot down a new problem to be solved using AI almost every day.

If you’re thinking about this space, I have one suggestion based on my own journey. Forget AI. Start with the problem and how it’s solved today. Think about what you would improve and whether people would pay for that improvement. Then consider if AI is the best way to solve it.

Big businesses are built around big problems and a technology shift. I’m betting on AI as the next big technology shift.

Here’s to big 2023 for you. Thank you for reading my essays through the year. Nothing makes my day more than hearing someone say they enjoy reading what I write. See you on the other side.