I use it as an enchanced syntax checker and error message solver i.e. you get an option to just "AI this error message" and it suggests solutions.
Perp has helped me set up my recent diskless booting environment. And while I could have done it myself - I've done it before - it recommended netplan, kea-dhcp, hostapd, nftables and dnsmasq all of which I would not have considered because they are not the default on debian.
And, if i prompt "I have eno0 as LAN serving dhcp on 10.0.1.1, enp4 as WAN as a dhcp client and wlan0 as a Access point on 192.168.0.1, generate me configs for netplan, kea-dhcp4, nftables and hostapd for all that" it will!
First of all: I, although normally interested in the general landscape of technologies and devx / productivity, IDGAF about the current AI hype, try to keep away from it, and generally speaking cannot care less at this stage.
That said, Codeium for me (and many other colleagues) became like a kitchen utility: you turn it on, and it just works, speeds things up, and I don't have to understand all the physics / engineering behind it.
By default it works with vim and vscode, I installed it and then forgot about it. Completes 1-5 lines worth of shellscript if/else branches, whole methods, etc. Also, xml, c++, python is not a problem. It's sometimes so accurate, I actually have to check if completion is coming from the LSP or Codeium.
I contribute to 5-10 projects with different languages, and don't have the capacity to understand all of them. As an example, sometimes I have to git mergetool quite big patches with many conflicts. Codeium perfectly fixes pytest unit tests for me during the merge, and I don't even understand necessarily, what was broken after the merge/rebase.
It offloads a lot of mental load and context switching, TBH. I understand that my use case is "fix /change a handful of most probably repetitive lines in a dynamically typed script", and won't expect it to write elaborate methods like fizzbuzz etc. It really feels like it's more of an ASSISTANT that helps me or sometimes pair programs with me
I don't doubt that such tools can help those with less experience and skill at figuring things out.
Would you be willing to give them another try in, say, one or two years as they will surely have improved?
def fizzbuzz_even_surprise(n): return [ (lambda x: 'Surprise' if x % 2 == 0 and x % 3 == 0 and x % 5 == 0 else 'Fizz'(x % 3 == 0) + 'Buzz'(x % 5 == 0) or x)(x) for x in range(1, n+1) ]
print(fizzbuzz_even_surprise(30))
As far as a “development community” exists, we have usually been very vocal when there are licenses violations and ready to give flack to companies abusing their access to FOSS code and resources. Still, when it comes to training sets, people are surprisingly lax and turn a blind eye to the millions of resources that have been slurped to train AI without any care for licenses or wish from authors. This is true for code but also for art.