Brainstorming GPT-3 Business Ideas with Harish Garg
6 min read

Brainstorming GPT-3 Business Ideas with Harish Garg

Listen to our conversation here.

Why care?

  • Harish is a GPT-3 consultant, founder of MynaOne, and author of the "Definite Guide to building products with OpenAI".
  • GPT-3 is the most capable language model that uses deep learning to produce human-like text.
  • GPT-3 businesses are hot. Even mini-tools are making bank, and there already have been several big acquisitions.
  • The market is far from overcrowded. There are currently only around 300 GPT-3 apps.
  • GPT-3 tools are easy to build. You just need a simple wrapper around the API provided by Open-AI. By now there are dedicated boilerplates, and it's even possible to build GPT-3 apps using no-code tools like Bubble.

Highlights

Pain Points

  • "The GPT-3 API is quite expensive (ca. $0.08 per 1000 words)."
  • "Open-AI has very limiting rules of what you can do with their API."
  • "They can revoke access at any time, change pricing, or even the API itself, which is quite scary."
  • "GPT-3 is very good at faking knowledge and never tells you "I don't know", or "I'm not sure". It always replies confidently with something that sounds plausible, even if it's completely wrong."
  • "There is no good way to make sure GPT-3 output is formatted in a certain way or is sensical at all. Hence everyone is currently writing their own ever-growing library of error-catching functions."
  • "GPT3 writing assistants do not create a full article at once. They are fussy, require a lot of rewriting, also have no real way to do a research step so the content is very general. The pricing is also very high at $49-$100 a month for "unlimited content" and I still spend hours actually writing everyday."

Predictions

Eventually, all companies that grow beyond a certain threshold will start deploying their own models.

  • Only if you deploy your own models, you have full control over the most essential part of your business.
  • Copy.ai already deployed a first proprietary model.

The next generation of machine learning APIs will be more focused on being useful for businesses rather than wowing regular users.

  • GPT-3 was trained on a dataset that consisted of all kind of texts collected online. But since these texts obviously were collected before GPT-3 was launched, GPT-3 knows nothing about GPT-3. So if you ask it, "What is GPT-3?", it replies with something like:

    "GPT-3 is an innovative gene therapy for Duchenne muscular dystrophy currently in development that would be given to patients in their teens, twenties, or thirties. GPT-3 uses a modified virus to deliver healthy copies of the gene that causes Duchenne muscular dystrophy into muscle cells."

    Or:

    "GPT-3 is a universal solvent composed of aqueous solutions of polyoxyethylene (POE) and polyoxypropylene (PPO) with an average molecular weight (MW) of 30,000."

    Both answers seem plausible at first glance, but if you do some research, you'll quickly notice that they're complete nonsense.
  • Note also how GPT-3 generated business ideas are all what Paul Graham would call sitcom startup ideas:

    "Imagine one of the characters on a TV show was starting a startup. The writers would have to invent something for it to do. But coming up with good startup ideas is hard. It's not something you can do for the asking. So (unless they got amazingly lucky) the writers would come up with an idea that sounded plausible, but was actually bad."
  • This behavior is great if you want to impress people, but not what you want in business use-cases where no answer is often better than a wrong one.

Texts were only the beginning. Code, photos, audio, videos, and design are next.

  • OpenAI published a preview of a new model they're working on called DALL繚E which is able to create images from text captions. This obviously will have a huge impact on the $4B stock photo industry.
  • It's possible to create photo-realistic images of humans using machine-learning models for quite a while now.
  • GitHub released in collaboration with OpenAI a new tool called Copilot that suggests whole lines or entire functions right inside code editors.
  • Tools like Descript are able to mimic anyone's voice with just a few minutes of training and have become scarily good at text-to-speech synthesis.
  • Even complete landing page designs can be generated using machine-learning tools.
  • Facebook is working on a model that is able to "understand" videos.

儭 Opportunities

  • Sell shovels. Any project that makes it easier for companies to build and deploy their own version of GPT-3 has a bright future. Vercel for machine learning models would be a huge step forward.
  • Start building a platform for curated stock photos that were generated by machine learning models. The tech might not be there yet, but it will be soon.
  • "Create a software that does a research step, with the ability to actually put in a template that is unique to your business, and write an entire article that is high quality like a writer would write. Make a preview before you use your credits on an article that just does not fit your purposes. I don't care if the articles take some time to write as long as the quality is indistinguishable from what a skilled writer would write. Most importantly show the links where you pulled the research from so that we can verify accuracy. Show references and make citations if requested. Make it customizable to what we need(s) with a template that suits our needs so that we can get our unique style and voice into the writing."

    Note 1: This idea is verbatim from a pain points interview.

    Note 2: Magic Sales Bot is doing many things described here but is currently entirely focused on sales emails.
  • A slightly different approach to the same problem would be to build Brandbucket but for texts that were generated by machine learning models. Pre-generate hundreds of texts, let freelancers check if they make sense and if necessary edit them, then create a platform like Constant Content ($5m/year revenue) to sell them.
  • Build laser-focused GPT-3 niche tools. Most of the big GPT-3 tools produce rather mediocre results for specific use cases. There is definitely room for smaller tools that nail just one use case.

    JobDescription.ai, TextCortex (GPT-3 product descriptions), and Magic Sales Bot are just three examples of what this could look like.

    One concrete idea mentioned in last week's report: "Use GPT-3 to create a niche version of Headlime that's tailor-made for newsletter creators. Newsletter writers send the draft email to the service, and it will reply with a dozen or so subject line proposals."

Enjoying these posts? Subscribe for more