Five AI Products You Can Build With GPT-3 Today

AI critics say that the technology is all hype. They are wrong. I’ll show five categories of products that are a) currently possible, and b) I am extremely excited to use, hopefully soon. 

I’ve become familiar with what GPT-3 is capable of by building Lex. I spend a lot of time writing prompts, and seeing what it gets right and what its limits are. It reminds me of when the iPhone first came out: if you paid attention, it was easy to see that transformative new product experiences would inevitably emerge. It’s hard to see exactly what those will be. Still, it’s fun to guess :)

So here are some of the new products that I predict will emerge thanks to AI. None of the ideas are dependent on technical advances—they could all be built today using GPT-3.

1. The infinite article

How much of your time do you spend scrolling through feeds and scanning articles? If you’re like me, it represents a decent chunk of your time. What if you could hire someone to do this work for you and compile a daily briefing? What if you could ask them questions and prod them to go deeper on the topics that interest you, and skip over anything that isn’t worth your time?

I would love this! I can’t wait for someone to build it. I presented this idea initially in September of last year. At the time I thought it was something that was years in the future. Now, I’m convinced you could build it today.

When I first thought of the idea I assumed the best you could do is to build a personalized list of articles based on a user’s Twitter history, the email newsletters they subscribe to, or any other data you can gather about the user. And I thought the best the AI could do is offer a short summary. I didn’t think it was possible to allow users to ask questions and have the AI give accurate, interesting answers.

I was wrong. Since then, a few things have caused me to update my thinking.

First, I realized you don’t need to get much information from users to offer a compelling user experience quickly. You definitely don’t need to scrape their Twitter feed, browsing history, or email. How did I learn this? From a new product called Artifact, created by the Instagram founders. The basic idea is simple: you check a few boxes of topics you’re interested in (e.g., Formula 1, tech, interior design), and the app will start recommending articles to you. It doesn’t collect any data on you other than what you do inside the app.

When you first start using Artifact, the article suggestions are just okay. But it only takes a few days for the suggestions to get much better—proving that good personalized recommendations are not as hard to make as I thought, and don’t require a large existing user base.

The second big shift in my belief happened when I learned about embeddings, a technology that makes the “question answering” part of the idea easy to implement. Without diving too deeply into the technical weeds, embeddings make it possible to retrieve chunks of text from a document that are relevant to a user’s question. You can then stuff these chunks of text into a prompt and have GPT-3 use it to factually answer questions. My co-founder Dan has done a lot of cool stuff using embeddings to create chatbots that answer questions based on a specific corpus of text, like Huberman Lab transcripts or Lenny Rachitsky’s newsletter.

With these two pieces in place, I feel like someone could easily build a wonderful daily briefing tool. If you’re building it, please let me know on Twitter. I would love to try it.

2. Shopping assistant

I perform the following routine at least once a week:

1. Realize I need to buy something (e.g., bluetooth keyboard, saucepan, baby playpen, Olympic barbell)
2. Spend 20–30 minutes browsing Amazon, Wirecutter, Consumer Reports, and Reddit threads to find the “best” version of that thing
3. Spend another 10 minutes finding the best price
4. Buy it

You could definitely train a chatbot to do this! All I want to do is say to it, “I need a playpen for my nine-month-old baby. What are the best options?” Ideally the AI would respond, “There are a ton of options, so let’s narrow it down. How much space do you have?” Any question I had about logistics or practicalities of choosing one option over another, it could help me resolve, and ultimately it would identify the best choice. Then it would find the best price and help me buy it. For more expensive purchases like flights or fancy clothes, maybe it could track the price and tell me when there’s a sale.

I’m sure Amazon is working on this, and it’s already possible to some extent using the new Bing. But while those options will be popular, there’s a lane available for someone to establish themselves as the “independent” option. I never really trust Amazon’s search results or reviews, as they seem too easy to game and are possibly sold to the highest bidder by Amazon. That’s why I love Wirecutter, Consumer Reports, and Reddit. I hope someone encodes this level of trust into the convenient form factor of an AI chatbot.

3. People finder

Businesses need to find people that fit specific criteria all the time. Usually it’s either because you want to hire someone or sell something to them, but there’s a long tail of use cases like finding research participants or investors.

What if there was a way to ask a bot to scour the internet and identify people who might be a good fit for whatever you’re looking for, find a way to get in touch with them, and maybe even craft a personalized message that would optimize your chances of getting a response?

This sounds like science fiction, but it could work. I built a primitive version of a tool like this when I worked at Substack. We scraped Twitter to identify people who might have a good chance at success if they launched a paid newsletter. That project used zero AI, and it was so helpful that the last I heard, the team still uses it to this day.

This is a broad category of idea—more of a theme than a specific product—and there will be many versions of it across different niches.

For example, it’s easy to imagine LinkedIn building a version. The company is owned by Microsoft, which is pushing AI the hardest out of all the big tech companies, having allied itself with OpenAI.

Facebook also sort of already does this, in the sense that it uses machine learning to target advertisements. But what I’m describing would allow Facebook ad buyers to target users using something closer to a natural language conversation with a bot than the byzantine user interface Facebook built for buying ads.

The best version of this might be more personal. For example, maybe it connects to your email and finds people who you’ve already talked to. There are a lot of different ways you could go with it and room for multiple products.

4. Financial advisor/explainer

Finance is complicated. Businesses like Investopedia and NerdWallet have been built off of providing simple explanations to questions that people Google when making financial decisions, like which credit card to apply for, what mortgage they can afford, etc. The chatbot would show up in my life two ways: 1) when it has a proactive suggestion for me, and 2) when I need its advice on an upcoming decision. 

There are many personal finance products built around keeping track of your spending and sticking to a budget. But they never seem helpful when they matter most: decision points. There are many times throughout the year when I wonder what the right thing to do is. Sometimes it’s a big purchase; other times it’s tax season or investment decisions. I always end up spending a long time on Google and occasionally asking a friend or official service provider.

I bet it’s possible to load my entire financial context—income, spending, investments, loans, and more—into a database that a chatbot can query and understand, so that it can keep track of everything for me. There are already products like Channel that allow you to ask for data in natural language and returns charts and graphs, so it seems likely you could apply this approach to personal financial data. To answer general questions about terms like “1099s” and “Roth IRAs” all you’d need is to use embeddings in the same way that I described in the first product idea of this article.

The hard part would be combining your personal data with general knowledge of finance to produce useful suggestions, like “consider switching to this credit card” or “you might want to form an LLC for your consulting work.” The biggest challenge is accuracy: one bad suggestion acted upon is a big deal. Maybe to start you’d have to stick to lower-stakes suggestions or have a human in the loop. I also suspect you could get pretty far using techniques like Reinforcement Learning from Human Feedback to avoid making high-risk recommendations.

5. Company librarian

Companies generate a massive amount of internal documentation and knowledge. Some of it is formal (e.g., a written policy regarding paid time off) and some is informal (e.g., a loose agreement in a Zoom call that turned into a quick email).

What if there was a central question answering service that was able to ingest all emails, documents, and recordings of video calls? At big companies, so much time is wasted asking repetitive questions and navigating through an organization trying to find the person who knows the answer. And the person who knows often ends up answering the same question ad nauseam. You can attempt to avoid that fate by writing an FAQ document, but 99% of the time nobody finds it and instead asks you directly.

The technical challenge is that large language models aren’t trained to understand the internal jargon that big companies inevitably generate, and on a more human level, it would probably be hard to convince a large company to hand over all of its data. You’d need a lot of credibility and funding to get this one off the ground. I can imagine starting with smaller businesses and integrating with tools that already exist, but on the other hand, at smaller companies, the “Who do I ask about this?” problem is much less severe.

Perhaps the best path for new products today is to start with a specific problem, rather than trying to answer any type of question. For example:

• Meeting search: Transcribe all meetings with Whisper and allow users to search for questions like, “What did we decide about our paid leave policy at the HR team meeting last month?”
• Code search: Connect with GitHub and get natural language answers to questions like, “Where is this module used, and what will break if I change it?”
• Document search: Connect with Google Drive and ask for the link to the latest draft of the paid leave policy.

Google and Microsoft are both well-positioned to solve these problems. Salesforce, too. To me the open question is whether anyone can build a single integrated solution that has access to everything and can answer any type of question.

What else?

I wouldn’t be surprised if, in five years, the data shows that there were more startups that emerged out of the AI wave—both in terms of total number of companies and aggregate value created—than in mobile or crypto. This is a controversial claim  and wild speculation, but I would bet on it.

I’m sure all of the ideas I listed above are being worked on at large companies and by small startups already. The real question is whether we’ll be able to look back in five years and see that the use cases I’m imagining materialized, or if the AI wasn’t good enough yet.

These are just five ideas I’ve been pondering lately. What ideas are you marinating on? I’d love to hear about it in the comments below, or on our subscribers-only Discord.