>not using gemini
What do you do between the 1/2 hour time it takes to answer your queries?

>>107991222
theres a way to do things
first, log in with your gmail account, if you have one, no need to subscribe or anything
the biggest perk youre gonna get is that youre gonna work with a semi decent model

also yeah, all models are sycophantic
thats what makes em hallucinate shit
you need to ask your questions in a turbo tame tone, so that the shatbot doesnt start hallucinating expected results
and even then you have to lead it at the conceptual level
its a back and forth where you give concepts, ideas to the shatbot and it throws back at you what correlates with what youve given it, rinse and repeat to zero in on the information youre looking for

>>107991222
as much as vibecoding is a meme.
1. you are using one of the worse model
2. you are not using a proper tool for it

you are supposed to use something like opencode, codex or claude code.

>>107991270
opus is shit at tasks that necessitate reasoning
personally id rather work with chud gpt when im dealing with something new, that im gonna need explained to me

>>107991288
i'm generaly using opencode + sonnet.
though, vibecode is only any good for webshitting imo.
for anything backend or system i write by hand.

>>107991375
>i'm generaly using opencode + sonnet.
>though, vibecode is only any good for webshitting imo.
ill give vibecoding a proper stress test
lets see how it fares with programming a vidya in c+sdl2

>>107991222
its great that this fascist was killed

>>107991476
in my experience the smaller and simpler you make individual tasks the better it'll do.
it doesn't replace proper engineering, you still have to think about the architecture yourself etc.

>>107991570
ofc, ill apply the best practices
the idea is to create a valid argument, not to create a strawman
ill give the method the best fighting chances and ill throw it at a real problem

>people responding to an OP literally too lazy to click one single sign in button and get to use an actual service rather than one designed for third world bots

>>107991736
>has no respect for privacy schizos
go back ot your shithole, impolite niggeroid

>>107991570
preliminary mucking about with the basic model of opencode yielded the following observations:
a) the shatbot is extremely bad at modifying existing c code in the context of a real-life generic program.
it takes much more time and tokens to do that than writing a correct version from the get go
b) because of that the user has to provide a very detailed plan, with explanations down to the desired algorithm, when important.
c) as expected, it fucks up shit that needs reasoning. in my current toy example it fucked up its collision detection maths
so
d) the user has to be an experienced programmer, the shatbot is no substitute for competency

but all in all, i can see how one could use it and extract value, especially with webshittery langs because everything is much better contained
and i can see how the method itself is appealing- you say some shit, shatbot executes, and then you review and test the code
i wouldnt trust it with unit tests though, as ive heard some people do
its fucking retarded, the human doing the testing is a key element in how this method is supposed to function, the human is supposed to provide feedback
also, organizationally, it forces the programmer to actually look through the code thats been produced

i didnt bring the model to its breaking point, properly, but im only at 450 lines
i didnt ask it for optimized code either and the complexity is non existent yet
so the testing will continue (later, im tired, i need to dress a proper plan for it before introducing actual complexity)

but the preliminary impression is that its not THAT bad. it can certainly accelerate writing boilerplate-tier code. its still gonna take a good programmer to make it do useful work. in c at least.
again, oop langs tend to have their functionality/data better contained (duh) so the shatbot should have an easier time at modifying specific elements of the code

btw
current test was:
make a program that takes a text file that represents a 2D scenery, a bmp representing a wall, a bmp representing the player
it builds said scenery
the player is not bound to the grid.
the camera is centered on the player, and, ofc, it detects collisions with walls
there is a zoom functionality which allows to, well, zoom in/out
the program uses opengl, glsl to do the rendering and the zooming

the main thing i was testing for was how well the shatbot can deal with side effects and implicit properties in code
and, ngl, it kinda passed with flying colors.
next step: we'll see what happens with actual complexity

>>107991270
opencode depends a lot on the model to tool call with any reasonable success, a lot of models that were free last month are gone from their default provider now, so you either need to have 96gb+ ram and an expensive 4000 or 5000s card with enough vram to use a decent local model shell out hundreds of dollars in token fees every month.

>>107991222
should have used emacs