GPT-3 and the future of human productivity

The idea of artificial intelligence taking over the world is at least as old as Mary Shelley’s Frankenstein, which was published in 1818. Public figures such as Stephen Hawking and Bill Gates have warned us against the reckless creation of superintelligent machines.

In June 2020, OpenAI, an artificial intelligence research laboratory based in San Francisco, launched the beta programme of GTP-3, the most powerful AI language model ever released to the public.

A lot has been written about GPT-3 already. Mostly, people worry about a takeover of the entire human workforce, starting with knowledge workers.

As often, there is a fine line between caution and fear, so let’s have a look at some of the potential benefits technology such as GPT-3 could have on the way we work, and in particular its future impact on human productivity.

GPT-3 Screenshot - Future of Productivity

A layperson overview of GPT-3

GPT stands for “generative pre-training transformer”, a language model which can generate world knowledge by training on a diverse corpus of text. GPT-3 is the third iteration of this model. It’s basically a language predictor: you feed it some content, and it guesses what should come next.

What makes GPT-3 extraordinary compared to its predecessors is the sheer size of the model, which has 175 billion parameters. GPT-2 “only” had 1.5 billion parameters, which was already considered massive when it was released last year.

GPT-3 has effectively ingested most of what humans have published online. It uses all the text available on the Internet to generate a statistically plausible response based on the text input it receives. And because it has lots of data to figure out what response is most plausible, the predictions tend to be quite accurate—too accurate for some people who fear software based on GPT-3 will replace their jobs.

For instance, GPT-3 has been used to design and code applications based on text input (“I want a 300px centered text box with a 1px light grey border and a blue button underneath it saying I’m feeling lucky”), write creative fiction, blog posts, or emails in one’s personal style, and turn legalese in to simple English. Someone even created a =GTP3() spreadsheet function to automatically pull information from the web and perform relevant calculations for them.

So, should you be scared GPT-3 (or, more likely, one of its subsequent versions) will “steal” your job? The short and unsatisfactory answer is, as you probably guessed: yes and no. In fact, what we should worry about is also exactly what should make us feel optimistic about the applications of GTP-3.

A mindless and useful tool

Ultimately, GPT-3 is still a language predictor. It doesn’t “think”, and it doesn’t have a “mind” of its own. It only generates content based on the input it receives. There are aspects of such a powerful language predictor which we should be cautious about, such as the auto-generation of fake news or the facilitation of increasingly convincing phishing scams.

And yes, GPT-3 could remove the need for mundane tasks such as generating variations of a same design or building simple websites based on common—thus unoriginal—principles.

But GPT-3 cannot answer questions that have never been addressed online, and cannot come up with innovative solutions that require unique thoughts. And because GPT-3 generates its output word by word, it doesn’t have a persistent mental model as we humans do.

In their paper about the limitations of the model, OpenAI research team explains: “GPT-3 samples can lose coherence over sufficiently long passages, contradict themselves, and occasionally contain non-sequitur sentences or paragraphs.”

Sam Altman, the co-founder of OpenAI, also tried to temper expectations: “The GPT-3 hype is way too much. It’s impressive (thanks for the nice compliments!) but it still has serious weaknesses and sometimes makes very silly mistakes. AI is going to change the world, but GPT-3 is just a very early glimpse. We have a lot still to figure out.”

That being said, the predictive power of GPT-3 and its successors does open many doors in terms of human productivity. Instead of seeing it as a potential replacement for human work, these models should be seen as powerful tools at the service of our own creativity.

The same way Excel enabled many small businesses, independent workers, and non-profit organisations to afford better planning, GPT-3 will allow many to supercharge their workflows.

A glimpse into the future of productivity

What would productivity and creativity look like in a world where we can easily bring our ideas to life thanks to a versatile assistant?

A design assistant. Instead of wasting hours clicking around to get to align two buttons on a mockup for a web page, imagine just talking to your computer, asking: “Add two rows of blue cards with a 2px radius under the header.” After seeing the result, you would add: “Wait, add a bit more space between the two rows,” and “Can we compare side-by-side what the cards would look like in green?” Note that the assistant cannot imagine for you what the page will look like, but it speeds up your workflow and lets you focus on the creative aspect of the design process.

A smarter secretary. Everyone loves to complain about email. Managing an inbox is tedious, feeling like an endless stream of people knocking on your door. Our current “smart” assistants don’t do much more than display buttons with often inappropriate canned responses as quick replies. With GPT-3, convincing and useful automated responses could be pre-drafted for you. Again, this would not work to respond for emails where the common sense or creativity of a human mind is needed, but it would save up a great amount of time to manage the back-and-forth of calendar scheduling or to reply to common questions.

A creative companion. A language predictor is unlikely to write content you would be proud to publish under your name, except if your goal is to re-hash information that’s already been published elsewhere on the Internet. But it could help you find inspiration. Feeding it your notes—from your Roam database maybe—or some of your past articles would suggest potential headlines to use as a starting point for an article. It could even generate whole drafts that may spark some new ideas. Or you could write hand-in-hand with the language predictor to take your thinking in surprising directions.

A tireless teacher. Let’s say you read quite a bit about a specific topic and want to test your knowledge. The old way would be to manually create questions, add them to your spaced repetition system, and train yourself to answer these correctly. The new way would be to ask the language predictor to generate questions based on the topic you are studying.

A translation tool. Yes, we already have Google Translate. But what about translating any content into a format you can understand better? We already mentioned translating legalese into plain English, but you could imagine translating complex scientific, technical, or medical information into simple language to better understand its meaning. GPT-3 could be used to build your personal ELI5 (“Explain like I’m 5”) on steroids.

As you can see, all of these applications are not replacements for the human mind—they are enhancements. Knowledge work has a bright future. Delian Asparouhov wrote: “30 years ago, Steve Jobs described computers as bicycles for the mind. I’d argue that, even in its current form, GPT-3 is a racecar for the mind.”

Update: You can now build your own custom GPT model without any code through the ChatGPT interface.


Join 100,000 mindful makers!

Ness Labs is a weekly newsletter with science-based insights on creativity, mindful productivity, better thinking and lifelong learning.

One email a week, no spam, ever. See our Privacy policy.