AI General Thread

[Legend]

To be honest, I think Google is the only one that survives.  They've realized larger neural nets become way worse and the hardware costs are astronomical.  Like, I've noticed Google slashed down Gemini's neural net significantly.  Just relying on having it search for info, and then doing a quick and dirty inference on the data to see what's right.  Sounds bad, but when you look at the server costs and how the larger the net gets, the more mistakes it makes, they have the right idea.  "Good enough".

Interesting. Most people seem to think the opposite right now, that google is in a bad place and needs help.

My read on what people are currently feeling:
OpenAI great
Anthropic shaming great
SpaceXAI great in some cases
Google falling behind
Open Weight Models shockingly great


My personal thoughts from my own experience:
OpenAI kinda annoying to talk to
Anthropic pure trash I consider it broken with how it talks
SpaceXAI really stupid but fun to talk to
Google a great mix of quality and price, but can be dense as fudge
Never used anything else

[darkknightkryta]

Interesting. Most people seem to think the opposite right now, that google is in a bad place and needs help.

My read on what people are currently feeling:
OpenAI great
Anthropic shaming great
SpaceXAI great in some cases
Google falling behind
Open Weight Models shockingly great


My personal thoughts from my own experience:
OpenAI kinda annoying to talk to
Anthropic pure trash I consider it broken with how it talks
SpaceXAI really stupid but fun to talk to
Google a great mix of quality and price, but can be dense as fudge
Never used anything else

People aren't looking at costs.  Google is.  Over the summer I was playing around with different ones.  Grok's token counts are stupid and their monthly fees are stupid.  Elon has deep pockets and even he's feeling the costs.  Hell, if those numbers that I saw at Neogaf are correct, Microsoft just dumped 0.15 Trillion dollars into data centers (Using trillions for dramatic effect) with a return of 1 billion.  This was this past year.  That 150 billion becomes 250 billion next year . The hardware requirements go up drastically the larger the neural net is, and so does the "hallucinations".  So they're spending an exponentially larger amount of money to just end up with a neural net that's the same because of how they work.

[kitler53]

my thoughts are all of the AI's are basically the same.   sometimes cool often times shame.   shaming enough to never really trust them.

i'd never pay for AI personally.   work gives me access to a range of stuff but i really only use gemini.   google is just more convenient to me because it's better integrated into a wide range of services i actually care about.

[kitler53]

I've noticed a trend with my kids and their friends.  they say "that's so AI" about things that are clearly not AI.   anything animated (CGI) is labeled AI to them.

[kitler53]

lol. 

so my company keeps telling us to use AI to do stuff and they've given us access to AI and an allocation of AI tokens to use in our jobs.    a week or 2 ago my group (~50 of us) was communicated a new process in which AI usage is mandatory.   As one does we all tried to practice the new process but lately everyone reported the process was failing. 

We figured out the root cause today.   While my company allocates each individual employee an allocation of AI tokens it turns out there is also a company wide token cap to prevent overspend.   Apparently we reached the limit and thus everyone's AI was disabled for the last ~5 work days of Aug.  Apparently i only need to do my job at the beginning of the month.  xD

[darkknightkryta]

lol. 

so my company keeps telling us to use AI to do stuff and they've given us access to AI and an allocation of AI tokens to use in our jobs.    a week or 2 ago my group (~50 of us) was communicated a new process in which AI usage is mandatory.  As one does we all tried to practice the new process but lately everyone reported the process was failing. 

We figured out the root cause today.  While my company allocates each individual employee an allocation of AI tokens it turns out there is also a company wide token cap to prevent overspend.  Apparently we reached the limit and thus everyone's AI was disabled for the last ~5 work days of Aug.  Apparently i only need to do my job at the beginning of the month.  xD

One of my university friends told me the same thing, except, it's not that you can't work, it's that the new employees don't know how to work.  He told me the new hire just sits there cause he doesn't know how to code.  He also told me he doesn't understand the back-end anymore because it was stitched together with A.I. code.  If the back end breaks, that's it, no one can fix it, not even the A.i.

[Legend]

One of my university friends told me the same thing, except, it's not that you can't work, it's that the new employees don't know how to work.  He told me the new hire just sits there cause he doesn't know how to code.  He also told me he doesn't understand the back-end anymore because it was stitched together with A.I. code.  If the back end breaks, that's it, no one can fix it, not even the A.i.

I really hope this will be like compilers and totally fine, vs something where every back end starts failing and we have a massive collapse.

[darkknightkryta]

I really hope this will be like compilers and totally fine, vs something where every back end starts failing and we have a massive collapse.

Oh, we are 100% going to have a massive collapse.  The smart IT companies already started to rehire their programmers.  The majour banks here are hiring again.  Everyone else will die off.

[darkknightkryta]

On another note: I finally hit my limit with Gemini.  First time.  I even upgraded to the second tier for a few bucks more a month (I was paying for cloud storage already).

[kitler53]

Oh, we are 100% going to have a massive collapse.  The smart IT companies already started to rehire their programmers.  The majour banks here are hiring again.  Everyone else will die off.

agreed.

I was doing a project where AI was clutch as pointing me in the right direction for some stuff i never learned like how to do a recursive sql query.  but the code it created had to be revised 3 times just to get it to compile and i still had an error because my example data didn't have a particular edge case that is important.

AI is not a replacement for humans.  not yet anyways.  its massively far away from that.  i think it does a good job at making a junior developer deliver much closer to a senior but if anything it slows down the senior developers when forced upon them.   troubleshooting bad code is a huge time sink.

[darkknightkryta]

agreed.

I was doing a project where AI was clutch as pointing me in the right direction for some stuff i never learned like how to do a recursive sql query.  but the code it created had to be revised 3 times just to get it to compile and i still had an error because my example data didn't have a particular edge case that is important.

AI is not a replacement for humans.  not yet anyways.  its massively far away from that.  i think it does a good job at making a junior developer deliver much closer to a senior but if anything it slows down the senior developers when forced upon them.  troubleshooting bad code is a huge time sink.

It never will be.  It's why you're currently seeing all of them being the same.  They've burned through trillions of dollars to get them to their near limits.  "Maybe" a large neural net on a quantum processor, a highly sophisticated quantum processor that does a billion-bit, might be able to think like we can.  But we may never get there, if there's quantum limits we can't get around.

[Legend]

agreed.

I was doing a project where AI was clutch as pointing me in the right direction for some stuff i never learned like how to do a recursive sql query.  but the code it created had to be revised 3 times just to get it to compile and i still had an error because my example data didn't have a particular edge case that is important.

AI is not a replacement for humans.  not yet anyways.  its massively far away from that.  i think it does a good job at making a junior developer deliver much closer to a senior but if anything it slows down the senior developers when forced upon them.  troubleshooting bad code is a huge time sink.

It never will be.  It's why you're currently seeing all of them being the same.  They've burned through trillions of dollars to get them to their near limits.  "Maybe" a large neural net on a quantum processor, a highly sophisticated quantum processor that does a billion-bit, might be able to think like we can.  But we may never get there, if there's quantum limits we can't get around.

I posted this in the chess thread but it really shows the problem with llms.

They simply do not generalize enough. Any human chess player would dominate with additional queens yet ChatGPT Sol spazes out.

LLMs are incredibly useful but they are in effect just better search algorithms. In the past you could google a code problem and copy and paste from stackoverflow, now the LLM can do effectively that for you. Ask for something truly novel and good luck lol.

[darkknightkryta]

I posted this in the chess thread but it really shows the problem with llms.

They simply do not generalize enough. Any human chess player would dominate with additional queens yet ChatGPT Sol spazes out.

LLMs are incredibly useful but they are in effect just better search algorithms. In the past you could google a code problem and copy and paste from stackoverflow, now the LLM can do effectively that for you. Ask for something truly novel and good luck lol.

That's the biggest issue no one is getting.  There's a reason why I call them neural nets instead of "A.I.".  They're highly disorganized data that rely on probability to determine if the program found what it's looking for.  Like, they make me nervous because they're the exact opposite of what Computer Science is; highly clean, organized data.  The larger issue is that they're using neural nets for things a standard data structure would be more useful in.  This is why we have all these garbage data centers.

[kitler53]

LLMs are incredibly useful but they are in effect just better search algorithms.

it's true but it's funny to hear you say it.  you've been soo enthusiastic about AI over the years.

but yeah, searching a mountain of thousands of documents and finding the 3-5 paragraphs of useful information in different locations and smushing it together?   AI is just great.  

but then the more people that use it the less people that will make the information that AI needs because there is no $$ in it.  it will take a while but it will be a downward spiral of enshitification for the entire internet.

[the-pi-guy]

Why is it funny? I called it that a year ago.