In the coming years there will be a lot of AI software engineers.
According to CEO Sundar Pichai, more than 25% of all new code at Google is now generated by AI. This code is subject to human approval, with engineers reviewing and accepting AI-generated code to improve workflows a nd development speed. This significant adoption of AI in coding at Google signals a broader trend in software development, though it's important to note that AI isn't writing full systems from scratch, but rather complementing human development. (this paragraph was AI generated)
But I think this last part will change. There will be a point where writing code will not be done by humans anymore. Or at least not from a maximizing productivity perspective. Writing code will become more akin to riding horses, where you are really just doing it for pleasure and fun, but not for taking yourself from point A to point B, in an efficient manner.
There's two main reasons. One is actually pretty obvious.
LLMs are just faster. Like they stream out responses at ~500 tokens per second. Whilst the average human writes at ~50-100 tokens per second. Obviously they need some reasoning, actual intelligence to be able to write the code, but practically this has already been solved.
Second is their intelligence is getting better everyday. The SOTA models are basically already #1 at competitive programming problems so intelligence isn't even the bottleneck anymore.
So what is the bottleneck?
Its the system around the LLMs. You know, how do you build an "Agentic" system that can modify, delete, create files, get access to compute, payments, talk to users, communicate with other people/agents. and iteratively improve itself and the system. i.e. solve real world problems with software.
Argument against the above paragraph is like fuck that, why build a system when you can just build like a God model that can control your entire computer and just go one shot prompt the model to create a $1M ARR SaaS at 90% net margin. Seems crazy, but honestly doesn't seem too far fetched.
Now the above argument is important, but not really if you are trying to build a PRODUCT. A.K.A something that needs to work now and provide real value.
This is why I'm trying to experiment and hack around on this. I started a project called AI Software Engineer, where I'm trying to build a system that can write full systems.
TBC
Source: ITPro - November 2024