There’s been a lot of buzz recently about how all companies will incorporate AI into their software in the near future. The commentary is usually prefaced with pointing to the exceedingly fast pace of artificial intelligence research and the momentum building in the field. In fact, venture investors are updating the fields they like to invest in from “dev tools and data” companies to include this third group of upcoming startups. Now the grouping is: dev tools, data, and Applied AI. In terms of the labels and articles coming out about this newly developing hot area you have content on large language models (LLMs) and their applications from Founders Fund, and then you have Applied AI articles from the likes of Index Ventures.

So let’s start by examining: what is applied AI? A simple google search points out that definitionally “it becomes applied when artificial intelligence leaves the lab and sees the light of day, helping workers with their everyday tasks in various functional areas.”

Large language models are sequence-to-sequence (seq2seq) models which have really advanced bolstered by the Transformer architecture. There’s a breakthrough paper called “Attention is all you need” that was released in 2017. According to my fiance, who got his PhD in machine learning at Stanford a few years ago, there are plenty of memes on “Attention is all you need” – but that’s an aside! Now, transformers have notably impacted the performance when training LLMs (where you are predicting the next word in a sequence) but they’re also being applied in other fields like computer vision.

Oh yeah, and now that I found this meme, here’s a picture of my fiance’s work, just because I’m getting all proud of him as I write this. I guess I have to cite him now, it’s in the JMLR (Journal of Machine Learning Research). Hey – he cited me in his thesis, so tagging him in a blog post is only fair!

All the above shows (or at least one major takeaway) is that the machine learning field is seriously math heavy. Luckily, you can build on the shoulders of giants. One thing that stuck out to me in conversation with Jonathan was basically how the capabilities of today are the result of *many* years of contributions to the field. The breakthroughs are adding up and now we are at an inflection point where we are going beyond supervised learning into semi-supervised learning (where LLMs like ChatGPT sit). Personally, supervised learning is also amazing to me because it's really easy to understand and the secret to properly making decisions with your data. But I could talk a lot more about that, let's get back to examples of the field changing.

For those not on twitter, OpenAI’s recently released ChatGPT model has set the industry on fire. Sam Altman, the CEO of OpenAI and former Y Combinator President, sees the industry evolving in the following way: a few companies like OpenAI train models with billions/trillions of parameters, ever increasing models basically fed public data from the entirety of the internet. Then there’s a layer of “middleware companies'' that take these massive base models and train them specifically for certain industries. Even 1% adjustments to these base models can make them really much better than their base models and so Sam says there will be large companies built that can deliver these models and package industry-specific software in this fashion. I will note that of course Sam wants to paint the world as evolving this way – his company at the very core infra layer collecting Stripe-like API fees and then a large pool of middleware companies. But you can’t argue with the products they are putting out. They are delightfully easy to use. Plus, Sam is really smart and at the forefront of the field. Maybe he’s right. (You can watch the whole video for yourself here).

If you do believe in the field evolving in this middleware-like fashion, then the next question is: which companies will play here? Sam says that developing for biology/medicine, law, materials science… these are some of the fields that are ripe for large applied AI impact. 

The first industry that’s started to see impact is actually copywriting. In fact, I remember it was just 6 months ago that I used an “applied AI” company for the first time. It was copy.ai. Their software allows you to select an objective like “write a Facebook ad title and body” and then specify the tone you want the copy to have, and boom. It generates 15 different pieces of great copy. That’s the sick part. It actually works! All of the sudden, I was telling people that the ads I created to test something were “using machine learning” and that felt cool! Certainly, I never feel like an expert or even comfortable writing ad copy, so I was open to trying it out. Even if you just use the copy as a brainstorming tool, it’s useful.

It will be fascinating to see the applied AI space continue to develop. Gridlines actually is bucketed into this sector but it was really by accident. It wasn’t a case of starting with the technology. It just so happens that it’s a fun space to watch at the same time that we are building custom machine learning models to automate workflows as part of the problems we are focused on solving. We are building our own models from scratch specifically trained for certain tasks we believe can save financial analysts millions of hours each year. More to come here, that’s for sure!

By the way, I was going to end this with a rap from ChatGPT, but it’s constantly at capacity. Apparently they had 1M sign ups in the first 6 days. On the one hand, I’m thankful they let me use it all day for free. On the other hand, I’m jealous that they have the ability to just train this massive model and give it away for free. Must be nice ha.