Marc Andreessen, a former revolutionary software engineer and now venture capitalist at Andreessen Horowitz (aka a16z), said this in a recent interview:
"At our leading edge companies, estimates are the leading edge programmers are like 20x more productive than they were a year ago. It's like the most dramatic increase in programmer productivity in like ever."
He's not the only respectable voice claiming this. Silicon Valley veteran Naval Ravikant, Brian Chesky (founder of AirBNB), and many other credible folks have made similar claims recently. And this is ignoring the less-credible claims of people who run AI companies.
So is this 20X thing real?
Or even 10X, 5X, maybe 2X?
I think it only makes sense to look at long-term productivity, over a time scale of a year or more.
So "20X" means that an individual, or a team, gets 20 years of pre-AI work done in a single year. Because if the long-term increase falls short of that, it's not really 20X, is it?
And if you look for verifiable, objectively measurable evidence of those 20X (or 10X or 2X) productivity gains...
Well, there just does not seem to be any.
That does not mean it is not happening. Maybe I just do not know about it. If you know of an example, tell me, please.
But it seems like there should be at least one clear, widely-recognized, and uncontested example by now, if that 20X (or 10X, 2X, etc.) claim is true.
What I think is happening:
AI tools speed up development in proportion to:
Your baseline of software development expertise,
How well-characterized the problem is among open-source codebases, and
Your ability to evaluate the quality of the AI tool's output.
And for certain particular tasks, that can certainly 2X, 10X, even 20X productivity.
A good example from my own experience: configuring a Django web service that used nginx, reversed-proxied into gunicorn (which is a Python-specific webapp server), and integrating another tool called uvicorn (with a "v") to enable websockets, all interfacing with the configuration of something called ASGI in the Django config.
Don't worry about what all those words mean.
The point is that it was a realistic, complex, production-grade task.
And the way I would have done it before LLMs is to spend three days reading the manuals of nginx, gunicorn, uvicorn and Django in detail, until I figured out how to configure these completely separate systems to work gracefully together.
Instead, I wrote a prompt to ChatGPT about what I wanted, and had a production-ready solution in an hour. 1 hour, versus over 20 hours; a true and objective 20X boost in productivity.
But that was just one task. One step of many to creating, refining and deploying this complex web application.
I am increasingly convinced that those who claim 10X and 20X boosts to development productivity from AI are incorrectly extrapolating one boosted experience to their entire work year.
Again, if you know of a credible, objectively verifiable exception, please tell me.
In the meantime, the best strategy, in my view:
Continue to invest in your software development/engineering skills. I.e. your ability to write great software in the first place.
That way, whatever AI does or does not do in the coming months and years, you are positioned to excel and reap juicy rewards.
If you liked this, you will enjoy the Powerful Python Newsletter.
Top comments (0)