GPT-3 – Will OpenAI’s New Language Generator Make Writers Obsolete?
Guest Contributor: Mike Rhodes is a ‘Top 25’ PPC Influencer and the CEO & Founder of WebSavvy.com.au, which was named as one of Google’s ‘most important 18 agencies globally’ in 2018. You can read his full bio here.
How does the near-future of AI affect marketers & writers??
GPT-3 is almost old news now in some parts of Twitter. Maybe you’ve heard of it, maybe not… either way, it’s going to change how you work at some point in the fairly near future.
There are very geeky ways to describe what GPT-3 actually is, but in the simplest terms, it’s a machine that predicts the next word to write given some sort of prompt. A very, very good one.
For example, if I said to you “the cat is drinking _____”
You’d probably fill in the word milk, maybe water.
You wouldn’t be about to say Diet Coke!
Or if I said “twelve plus eleven is _______” then hopefully you’d say 23.
Now, extend that to writing poems, or computer code, or school homework essays, or songs, or … well, anything written!
What might this mean for Search Results?
I’ve been tinkering with it for a few months now & have started to think about how this will affect search.
Once this – or the next version of this – is used to power search results, it will change how people search and which results will be presented.
If it can surface great results for natural language questions, clearly that has a positive impact on voice search.
How will it affect the SERPs!? Google loves to show a featured snippet or knowledge card. More & more I think Google (or whatever thing comes next) will NOT give you a list of sites to choose from. It probably won’t even link to websites, it will just provide the (1) answer you want.
What does this do to content marketing? Are all those long-tail blog posts suddenly useless?
What role do ads play in that world?
OK, how does it do this?
Bear in mind, this is not just finding the one “best” answer (on that long-tail blog post) and surfacing that. GPT-3 can synthesize multiple pieces of text.
How? Because it’s read them all.
Seriously. Almost every written piece of content humans have ever put online. Half a TRILLION words. That’s what it was trained on.
So, it’s not ‘copying’, it’s creating. And it’s good.
And given that Microsoft jumped into bed with OpenAI (creators of GPT3) what might this mean for a resurgence of Bing. No, seriously. It could happen.
Yeah ok, back to the real world. Given that GPT3 “only” cost about $12m to train – a drop in the ocean for Google – how much are they spending right now to improve on their T5 model?
(T5 is a kinda similar machine that can do kinda similar things, but not as well).
I watched a *very* geeky video a few months ago about GPT3 and the tech behind it. My main takeaway: “if you plot the performance on a chart, the curve is still going up and right”.
In other words, it’s going to keep getting better.
Which is a little terrifying when you see what it can do already. And why I think it signals some huge changes ahead.
Let’s put that in perspective. GPT3 arrived a year & a bit after GPT2 & it was ~100x more ‘powerful’. Power, in this case, is the number of ‘parameters’ – you can think of this as (very roughly) the number of connections in the machine brain (sorry purists!)
If you’ve watched my ‘AI Fundamentals for Marketers’ course, then you’ve seen me explain the basic structure of a ‘neural network’. Here’s a quick video if you’re interested:
So… back to the future.
GPT3 was ~100x more powerful than v2. In about a year.
And the geeks say more is possible.
And it was ‘cheap’ to build (at least compared to the benefits to be gained).
So, where to from here?
I think we’re going to see a new arms race of these models.
Because of how easy they are to use for a non-technical person.
The skill is no longer in fine-tuning the features of your data set or tweaking the model.
In fact, you don’t need to know any code at all.
Instead, these models depend on the ‘prompts’ that you give them.
For example, this is a ‘prompt’:
“Here’s a poem by Dr. Seuss about Elon Musk launching rockets with SpaceX, building electric cars at Tesla, selling flame throwers, dating Grimes, fighting the SEC, crushing short-sellers, building brain chips to fight AI, and colonizing Mars, and tweeting at all hours of day and night. Musk wants to make humans interplanetary to prevent extinction. The poem rhymes every other line with an ABAB structure. The rhyme structure is typical of Dr. Seuss nursery rhymes.”
(read the poem the machine wrote here https://arr.am/2020/07/14/elon-musk-by-dr-seuss-gpt-3/ )
There’s definitely an art to this.
Knowing what to use as a prompt.
Seeing how the output changes when you tweak the prompt.
If you’re old like me, this might remind you of the ‘art’ of searching back in 1999!
You had to craft what you wanted.
It was funny (then) to watch old people fumble around and type the ‘wrong’ things into Google & get nothing useful out.
The Kids are Alright
Kids will pick this up fast. They’ll learn how to ‘prime’ the machine with the right prompts and make it do incredible things.
I’ve been speaking to grade 5-6 classes during lockdown about AI & Machine Learning.
I’ve been telling them that they will probably use a tool like this to write the first draft (or 5) of their essays in high school within a couple of years.
And no teacher will be able to run the essay through a plagiarism checker, because it will be completely NEW content.
Synthesized from many other pieces, for sure – but completely new. Not copied from any one place. Just like a human would write.
It probably won’t be good enough (for a little while) to just ‘hand in’ as is.
But as an idea generator… crazy good!
Imagine tweaking the prompt slightly (and maybe the (very few) settings on the machine) and getting five totally different essays back – what a starting point for a writer!
But is it as good as a human?
But in five years, will we even be able to tell which books were written by humans and which by the machine? Will it matter?
So… What will you do with it?
How will it change what you write every day?
Your emails? Check out copy.ai
Your blog posts? Check out usetopic.com
Your Google & Facebook Ads? Check out usebroca.com