Y Combinator Practice

Why do the reluctant ones hold back?

- Provisioning of the tool at work - Reluctancy because of ratings and saturation of the market

What are the top things users want?

1. Personalization of responses 2. Control

Why isn't someone already doing this?

There's a high barrier to entry. You not only need the machine learning and software infra expertise, you also need the resources. and the resources to scale this is very costly. The 175 B parameter GPT-3 model is about 800GB of storage. Now imagine doing that for thousands of customers. It's very hard, and that's why people are avoiding this problem. But we love solving hard problems.

Question for Ali

Ali you worked closely with tony Hsieh, and one of Tony's biggest message is happiness. Over the year, hustle culture has been glorified... [finish later]

So why won't Open AI or Google do this?

Andrej said OpenAI stores 1GB of data for every fine-tuned model. Open AI will have millions of companies using them. those companies also have millions of customers, and eventually billions of customer interactions. Scaling that is frankly not optimal for OpenAI or Google. After being in generative AI for a while now, we've come to realize that companies tend to focus on one layer of abstraction in Gen AI. The bottom layer >> research >> building specific AI models layer is where Open AI and google operate> The top layer >> The applications layer is where all of these gen AI companies operate. It's incredibly saturate, and easy to get into. The middle layer, however, the AI Ops layer. is what we're going for and there are not a lot of players in this space.

What's your competitive advantage

As a startup, we don't have the pressure to serve everyone. We're only serving companies that generate text-data on LLMs. Think of law firms wanting to generate leagal documents. Monterey AI wanting to generate applicaiton insights from customer feedback on twitter. We can build an embeddings suite for a few of these folks. Sort of like a quasi solution. Deploy it and start getting customer feedback. I'm laying a huge bet that we will discover a breakthough as we solve this problem to be able to scale it. We're consulting with our advisors, these are AI professors at CMU and AI practitioners in the field. We actually have a design session next week to draw up software designs for this.

What, exactly, makes you different from existing options?

Existing options that offer federated learning like Dynamo FL. only offer it on a company level. Very difficult to offer federation on a user level or even a user interaction level. But we know this is where the real value is. We're focusing on a multi-layer personalization and we are not optimizing for privacy. We're optimizing for personalization. I've talked to about 50 customers on zoom calls and over 90% of them don't mind the privacy part. What the vast majority care about is personalization.

What are they doing now?

For text-based applications, as part of their ML Infra: They used embeddings model to maintain a vector store of their data. The embeddings model usually have larger context window and they use this to feed extra context for every inference.

What has surprised you about user behavior?

For the email assistant. Real estate vs Sales. use it for two different reasons

How would Neo help?

I have a very specific ask. We want to use Neo to partner with Scale AI. I saw scale AI is in your portfolio. We want to leverage ScaleAIs labeling suit to scale to other use cases other than conversational text. One problem we found is there are so many different kinds of data structures. Let's say an applicaiton such as ML agents, where an observer observes the movements you make on your computer in order to replicate those tasks. That data is not purely text. It's how many times you clicked a button, how long did you stay on a page for. All of that is unstructured data if we have to personalize inference for, it can't just be text. and we think scale AI would be quite instrumental in this

What are the key things about your field that outsiders don't understand?

It's not privacy! It's personalization. The #1 reason why people uninstall the app. By a huge margin. is because the emails don't sound like them. We though a 50-60% similarity would be good but it's not. and we've talked to over 50 people and they mostly say the same thing. I think people just haven't discovered (at least on a huge scale) this multi-dimensional personalization of LLMs yet.

What does each founder do on the team?

Michael: Coding, marketing, analytics, customer engagement, etc. Hassan: Business operations and strategy

What have you learned so far from working on it?

Single score - Trust score from feedback from our users @upload Users want to be incentivized to rate projects. Implementing the tokenization feature where people earn crypto tokens by writing reviews Early access to upcoming NFT mints.

What are you working on?

We are building a service that allows AI applications to personalize LLM output not only per user but per user interaction. Since launching Addy AI, we've helped 4,600 people generate about 39,000 emails. and we've learned from talking to customers and analysing user data the LLM outputs are not personalized so we want to solve for this ___END IT HERE___ If they ask, why did you pick this problem A few weeks ago we added one thing. We added a google form that pops up automatically when you uninstall the app. and it had one question. Why did you uninstall addy ai? The #1 reason why people state is that the generated email responses don't sound like them. So we start workiing on a personalization feature