Talking to insurance agents I realised, they don't bother to read the policy documents and have a very superficial knowledge of the policies they are selling. You can glean lot more information feeding the docs to an LLM and asking questions.
Yes, something like TreeSitter would seem to be of more value - able to lookup symbols by name, and find the spans of source code where they are defined and used.
Really cool! I'm also dabbling with this idea. The biggest challenge I find is to reduce noise. Large codebases come with a lot of cruft. Surfacing every small detail in the visualization tends to make it messy and less useful. I've not seen contemporary tools tackle this, but think can be useful.
People said the same thing about the internet, that you should be getting your information from actual books and that the internet will make you lazy and complacent. There was a time when you were encouraged to write code on paper first instead of typing it directly because it made you think clearly. Some time before calculators apparently dulled your mental faculties, so you should be hand rolling all your calculations. Go back in time far enough and you'll find Socrates disparaging writing because it weakens your memory and destroys your mind. And yet humanity is here and seems to be doing all right. Every generation has managed to produce smart people that have been able to push the boundaries of scientific and technological progress. If anything we may be getting smarter. What history has repeatedly shown is that when you reduce friction for the human brain, it goes and finds more complex things to do. Such a periodic removal of friction, may very much be a necessity for progress, because it allows the paradigm of thought to shift to a higher level. The same should happen with AI as well.
Layout is one of those things humans do so easily and intuitively, yet you couldn't write an easy algorithm for it. I wonder if there's potential to use gen ai to achieve human like results. Anyone has any thoughts on feasibility and complexity of such an approach?
I dabbled in this area, there are poster layout generation attempts that use gen ai to come up with an initial layout plan, and even feed the visualization back into the llm for iterative fine-tuning.
I was intrigued, but couldn't make it work reliably. Perhaps I forgot to add "make it look nice" to my prompts.
I think the hard part is more defining the different types of visually pleasing layouts that you want to support. What to align to what, what metrics to minimize, etc.
There is no single layout definition that looks good for all applications. What looks good for one type of graph data (flow charts, org charts, ASTs, family trees, etc, etc) may not look good for another.
Once you've reduced a desired type of layout to a set of layout rules, then I don't think implementing it is the bottleneck, and it seems this part could even be automated if wanted (e.g. evolve layout to best meet constraints).
Now exploring on extending this to general programming languages. Software is getting increasingly complex, but I don't think we have figured out many ways of navigating large codebases. Would love to hear about your favourite tools in this space and what would you like to see?
This is a cool problem to solve! especially as more and more code is being generated with AI. But are you pivoting away entirely from code visualization? Curious to hear why it didn't work out.
We may go back to code visualization for editors, but after testing it with a lot of engineers, we found that it was too confusing for code reviews. The main issue was just how alien the interface is!
reply