1) have books like 'The Art of Programming' on my shelf, as AI seems to propagate solutions that are related to code golf more than robustness due to coverage in the corpus.
2) Force my self to look at existing code as abstract data types, etc... to help reduce the cost of LLMs failure mode (confident, often competent, and inevitable wrong)
3) curry whenever possible to support the use of coding assistants and to limit their blast radius.
4) Dig deep into complexity theory to understand what LLMs can't do, either for defensive or offensive reasons.
5) Realize that SWE is more about correctness and context than code.
6) Realize what many people are already discovering, that LLM output is more like clip art than creation.
> 1) have books like 'The Art of Programming' on my shelf,
Decades ago I used to be constantly thumbing through vol. 1 of Knuth (Fundamental Algorithms) and coding those basic algorithms. Now that all comes from libraries and generics. There has been progress, but not via LLMs in that area.
That is part of the problem, LLM coding assistants often code out what a human would often use a core library for, only including when required, like importing deque in Python to do a BFS.
An example I helped someone with on AoC last week.
Looking for 'lines' in a game board in Python, the LLM had an outer loop j, with an inner loop i
This is how it 'matched' a line.
if i != j and (x1 == x2) or (y1 = y2):
I am sure you can see the problem with that, but some of the problems in Knuth 4a are harder for others.
There is a lot to learn for many in Knuth 1, but I view it as world building for other concepts.
With LLMs polluting web search results, the point is to have reference material you trust.
Knuth is accessible, I did error on that side rather than suggesting books that are grad level and above that I really appreciate but would just collect dust.
People will need to figure out what works for them, but having someone to explain why you do something is important IMHO.
2) Force my self to look at existing code as abstract data types, etc... to help reduce the cost of LLMs failure mode (confident, often competent, and inevitable wrong)
3) curry whenever possible to support the use of coding assistants and to limit their blast radius.
4) Dig deep into complexity theory to understand what LLMs can't do, either for defensive or offensive reasons.
5) Realize that SWE is more about correctness and context than code.
6) Realize what many people are already discovering, that LLM output is more like clip art than creation.