Yeah, it's interesting how the current thing is always the end of programming, but then when a new thing comes it turns into absolute trash that can't be used for anything.
Maybe if you only use the LLM to proof read or something? I don't know if LLMs are actually any good for this purpose, but it's something I came to think of when trying to answer your question.
I do think that it is very difficult to not be a meat proxy when you utilize an LLM for writing tasks. These tools are computer programs that generate text, so what else would you do with it than have it write for you? This is being a meat proxy.
Trying to avoid being a meat proxy can quickly eat the supposed productivity gains of using an LLM in the first place.
Like say I need to write an article about some subject and it would take me the entire day to complete this task without an LLM. I can get some sort of article out of an LLM by simply asking, I don't need to think about the subject myself at all. The LLM productivity promise is also fulfilled, because I did an entire days work in a minute or so.
Now if I also need to read and understand the work, and be able to defend it or explain it to others as if I produced it myself, then this will take a lot of time. If on top of this I need to edit the text to sound like me rather than an LLM, then I probably still need the whole day for this task and I have lost all the productivity increase that I sought from using the LLM.
It's the thinking that takes time when you write, not typing. But you can't do the thinking and avoid doing it at the same time.
The same idea applies to coding too. Doing code review and requiring the author and reviewer to read and understand the code basically eliminates all the productivity gains that using the LLM was supposed to provide. Throw out the reviewing, reading, and understanding and we can have each developer producing a thousand kloc per day, over 1000x productivity. If we don't throw those out, we are comparatively in the exact same spot as not using LLMs at all.
I find that the LLM is great for helping with the starting problem. I will go for a walk and basically brain dump my thoughts and ideas about a topic/doc into a voice note, where I can ramble, wander from point to point, etc. Then I transcribe it, get a model to take that and turn it into something cohesive, and then I spend multiple hours reading it, reworking it, and editing it. I am not very good at writing docs and this method produces something with far more detail and quality than I would ever get on my own. But AI;DR right?
Yes, it's like people have forgotten that junior contributors you can't trust yet often have net negative productivity due to the time you need to invest to review and guide them.
Crypto guys are in the business of grifting magic beans to rubes. They are mostly interested in the aesthetics of these technologies. Using an academic programming language gives these scientific magical vibes and such that make people believe these guys are high tech and know their shit and so on. This is the reason many crypto projects do use Haskell too, it's known as an academic research language that is difficult to use and "if it compiles it works." They come at it from the perspective of "What would sound the best when I try to sell the space money of the future?"
Blockchain protocols are also faced with the most active, persistent, and well-funded state sponsored attackers of any industry, so naturally formal verification should be of great interest.
reply