Ive dropped out of society, but come here to check on tech and other /g/ay shit.

From your perspective, where are we at with AI? Acceleration happening? Plateau? Mask-off, it is fake& lé gay? Is it jeets all the way down? Best/worse platforms? Open sour ce dead?

No more useful data to train the text models, hard limit reached
Lots of data to train video models and robots, progress potentially unlimited but slow and expensive
Real Reasoning (tm) AI still 30 years away
98% of the current wave of AI startups will go 'rupt before 2030

>>106010003 (OP)
plateau fueled by random discoveries
>chain-of-thought
ideas
>tool calling
>image understanding
and artificial hype
+ AI (Actual Indians) slaving off behind the scenes to "teach" the models

despite massive (artificial) claims and huge (real) potential of agentic AI (memeword for just AI tool calling) there's not many working nor decent products with it

for code they're pretty ok but if you don't know what you're coding or wouldn't be able to code it yourself you'll quickly enter slop territory
/opinion

>>106010297
this too

>>106010003 (OP)
Floods of AI slop code are now being committed to all major tech employers nationwide. Leadership thinks it's great

>>106010003 (OP)
In the past few months, AI coding tools seem to have crossed some kind of threshold and become an actual threat to junior dev jobs.

>>106010003 (OP)
i think they suck for coding
but theyre decent at running ideas through
shaking the tree to see what falls kind of deal
aslo: documentation and code examples
but i dont trust its code and i certainly wouldnt let it manage a project

>>106010003 (OP)
>Mask-off, it is fake& lé gay? Is it jeets all the way down?
yeh

>>106010499
Some of the tools actually do a pretty damn decent job of writing chunks of code as long as you can be explicit in your design requirements

>>106010003 (OP)
Marketers are full shit. All research moves with major disruptive findings. I.e. the transformer architecture was the last huge disruption to AI. Nobody was that interested in chatbots. Then Google drops the transformer architecture and overnight LLMs are way better than anyone could have expected. Between disruptive work, researchers incrementally improve the latest big thing but these improvements are minor. This is where we currently are, the transformer architecture is pretty much maxed out. It is not the final form of AI. Anyone saying otherwise has no idea what they are talking about. At some point in the future, whether tomorrow or 5 years from now, someone will make some other major improvement and, again, overnight improvements. This will not discard the value of the transformer. The transformer introduces the attention mechanism. The next big finding could be some way to train much more efficiently (i.e. not needing all the worlds data to create an okay language model) or internal models with better pattern recognition (i.e. more rational and better at problem solving/generalizing out to new areas). In this case they will still use the attention mechanisms. It's not discarding the current methods completely. For this reason of random disruptive work having an incredibly disproportional impact on the state AI (all research), anyone trying to make predictions from current trends for when super intelligence or whatever other buzzwords they're trying to use to get hype is blatantly unaware of how things work. We cannot predict how long it will take to any degree because it is not from continual small improvements but by singular major works that anything new of value will come and there is no way of knowing when the next major work will be or what it will target.

ai inference is all the rage now, very exciting field
ai training is still a gay cuda moat, absolutely boring field,

>>106010658
yeaaaah but they shit themselves anyways the second you encounter side effects
its stuff thats present in discrete contexts and it doesnt carry between them very well
i am willing to believe some people manage to leverage them but it must be in very specific cases

>>106010297
>>106010965
> our system thinks this is spam
Sorry for the no yous for everyone. Fucking filters.

Thanks for the responses!

>>106011015
i've found them to be reasonable for writing green-field, one-off programs, but they shit the bed if they have to operate inside of an existing context. small wonder they're wiping out junior devs since most of the stuff you'd hand off to them is pre-chewed throwaway work anyhow.

>>106011280
yeah but you have to take your seniors form somewhere
also: to reiterate
theres a context where an ai can prove useful
but as you see its worthless for anything serious
its inherent to the technology
the rest of the industry is slowly getting there but i woldnt expect anything worthwile for the next 3 years at least
id say 5- realistically
were at the very beginning of the paradigm shift toward intermediary representation

>>106011280
>>106011346
clarif: neuro is a hypothetical framework i played around with chud gpt
shaking the tree
see what falls down
and whats the industry take on things

>>106010003 (OP)
isn't AI still super expensive to run and train? i feel like AI is a massive cash pyre at this point. it's been ok for porn (video is still shit though, at least until an actual porn finetune comes out). other than that it's being used by retards to fill the internet with absolute garbage.

>>106011419
are you kidding me? It's cheaper than ever and the price only keeps dropping.

>>106011346
>yeah but you have to take your seniors form somewhere

absolutely, and it's gonna be a shitshow in a few years when the talent pool has not appreciably grown but every major project has taken on crufty, unmaintainable crap.

i'm seeing a worrying trend where projects architecture is getting incredibly bad. instead of clear data organization and flows between well-defined states, you have subroutines which mutate data from one arbitrary structure to another without much critical evaluation.

for example, i recently got a MR which added a new endpoint to an existing service that converted an json payload into a csv, wrote it to a file, and added the filepath to a database, then in a separate routine picked up the database entry to read the file, and opened an entirely new req context when it messaged the next service in the chain. i asked the dev what the fuck he was doing and he said that he just asked the AI to generate the endpoint, then to write the request to storage, then to asynchronously process the request; the AI technically did all that, it just did it in the worst way possible.

>>106011482
>the price
the price that companies are charging or the actual money being used to run the datacenters? like do we know how much money for power and equipment and everything is?

>>106011482
lmao wtf. you don't seem to belong here

>>106011485
wtf
brain on complete autopilot
and he thinks its normal to just copey paste code from the chatbot straight into prod without even reading it?
i think c-suite bets on the advancements of ai
and it will advance
but i think it will advance on timescales they didnt expect

and corpo will end up paying big money to people with actual experience

>>106010499
>it's important to add one to be respectful
then why didn't you do it

>>106011573
the crux of the problem with the ai
its just bad at carrying information bw contexts
it focused on producing code in accordance with the docs but has absolutely no cognizance that its supposed to know you have rate limits
which then got activated when i asked about it

>>106010003 (OP)
>From your perspective, where are we at with AI? Acceleration happening? Plateau? Mask-off, it is fake& lé gay? Is it jeets all the way down? Best/worse platforms? Open sour ce dead?
AI is making progress by the day but it isnt as huge or fast that optimists were hoping it would be

>>106011482
me when I lie

>>106011199
>Sorry for the no yous for everyone. Fucking filters.
they put a limit to the number of posts you can quote because people kept quoting like 50 people and RAPING the shitty 10 year old servers to death.

>>106010003 (OP)
VRAM bottleneck + more and more filters + feminists enabling payment processors to be even more shitty = doom and gloom. The tech will be used against you and any potential happiness you can get out of it will be watered down and monitored.

>>106010003 (OP)
The only use case I have found is gooning. AI images don't do it for me but you can run Zonos locally to generate speech that's at least on the level of those shitty youtube ads