AI productivity: my experience with Copilot
I stated previously that AI in 2023 (in particular LLM’s) will boost individual productivity by augmenting what humans can do. The most common use of AI will be in paid tools because the cost of using it in things like web search doesn’t make financial sense yet.
For 2 months now, I have been using and testing Github copilot which is AI augmentation for coding. My experience with it has been amazing and I can hardly see myself coding again without it. It creates a habit and it augments productivity with measurable results. Two clear indications of a winner product.
Further below, I provide a more detailed account of my experience. But let me spell out the conclusion first: I estimate I could code twice as fast and that resulted in ~20–30% increase in productivity.
Now, hear me out. I am a product guy. And as many product guys (specially in startups or small companies), I have to touch many fields out of my zone of expertise. Coding is actually one of them (I am hardly a pro although my background is in CS). But I have routinely to go into into other areas as well (legal, finance, marketing). Coding-augmentation is great but I cannot wait to have legal-augmentation, finance-augmentation, marketing-augmentation. And eventually, why not, product-augmentation. Once models can be privately fine-tuned on specific proprietary corpus of data (such as a product specs, roadmaps, customer feedback), writing requirements may get its own boost!
The funny bit. I am a foreigner living in France and I speak and write French pretty well. However, I have to confess that since chatGPT came out, I have not written a formal letter in French myself. And in France you do need to write many formal letters in your daily life. I call this, the French Bureaucracy AIntidote.
My experience with Github Copilot
Finally, this is the account of my experience with Copilot after writing ~1.5k lines of ML python code for my RL learning project.
Github copilot can be used in two ways:
- As a supercharged autocomplete that seems to read your mind while you are typing code and suggests entire lines or code blocks
- To replace code writing by english prompt writing (in the form of comments) plus code review
The main benefits I have observed are:
- First of all it integrates perfectly in IDE’s such as VSCode making its use seamless.
- Takes care of writing boilerplate code like a charm (a big deal IMO)
- Can take care of more complex code as well via prompting. For me, it helped a lot building complex plots or doing some convoluted tensor operations.
- Because of the above, I code more confidently and I have less tendency to procrastinate because Copilot takes care of a lot of the boring stuff I do not like doing.
- Using it results in better documented code as a product of using comments to prompt Copilot to write code but also using Copilot itself to write documentation
- It reduces the amount of google search required to remember syntax or even explore the utilization of a new framework. Less context switching, less distractions, more focus on the task in hand.
- Reduces typing mistakes although it can also create new ones (review is always necessary)
In my case, it boosted my code writing productivity by a factor of 2 (resulting in an overall productivity boost of 20–30% because, of course, not all the time is spent writing code). For less than 20 EUR/$ per month for business use, seems like a no-brainer.
In 2023 we are going to see a bunch of individual productivity boosts based on AI-augmentation. As with the web2.0 collaboration tools, companies slow to adopt, will get behind their peers.
But choosing and trusting the right tools will also be a challenge so it needs to be done properly. If one wants to bring AI-augmentation to their company, one first should find a few motivated users and run a pilot experience with them. And from there, expand.
