• Moderators, please send me a PM if you are unable to access mod permissions. Thanks, Habsy.

Tech Thread

axlsalinger said:
Obviously a complex situation, but just to simplify, let's say a group of 20 human workers working 8 hours a day requires 1 manager.

So now you've got a group of 20 AI's working 24 hours a day, that still requires 1 manager. Even if requires 3 managers ... it's still a massive disruption in the market and huge cost savings for the company.
Click to expand...

And that's cool, but that's just not the reality we're seeing when the technology is implemented. Yeah, that is what OpenAI, Anthropic, etc have promised but what we're seeing is that the AI's simply can't work autonomously at all without breaking everything, for any length of time and requires human oversight more or less equal in time and effort to just having a human do it in the first place.

The other thing not really spoken about either is that when poorly trained/educated people break things, a more senior & competent person can come and ask the person questions about what they were doing when they broke the thing and use this and their superior knowledge to fix the thing that broke. LLM's don't work like that though. Your person fixing the failure has to do the work all over again from the very first step. The LLM can just run more and more iterations until it maybe gets it right. It can't logically debug/audit it's own work to find it's error.

What you're describing is absolutely a revolution. But we're not seeing a revolution in the results at all.

axlsalinger said:
All this while AI accuracy continues to get more accurate over time.
Click to expand...

Until someone solves the hallucination issue there's a hard cap on the limits of the technology and it's a problem that is actually getting worse as they throw more compute at it. Remember, more compute was always supposed to be the solution for hallucinations, but what we're seeing with more compute is just the model's ability to hallucinate go up.

www.vktr.com

AI Hallucinations Nearly Double — Here’s Why They're Getting Worse, Not Better

AI hallucinations surged from 18% to 35% in 2025. Learn why chatbots are spreading more false claims and what experts say must change.
www.vktr.com www.vktr.com

So yeah, I understand all of the stuff in the pitch decks, we're just not seeing that occur in our shared reality though.
 
axlsalinger said:
Just saying, if an AI can do two hours worth of research in 30 seconds, it's probably okay to expect some issues. Humans make mistakes too, and a lot of them.
Click to expand...

In what field, and how long does it take to fact check it's research? I mean, if we know that the model hallucinates ~25% of the time then in any field that requires accuracy (basically anything that isn't internet shitposting), the entire output of the model has to be fact checked by a human who is capable of spotting mistakes. Otherwise the output is useless for any sort of commercial purpose whether there are monetary risks for getting shit wrong (basically, all potential purposes).

As mentioned above as well, there is a huge difference between fixing mistakes made by a human (which yes, they make a lot) and an LLM. The LLM can't tell you what it was doing when it made the mistake, it can only iterate an entirely new set of data for you. A human can explain their work for you and allow a more skilled human to spot their errors without doing the entire task over from the beginning.
 
We're so early in the process, and you know every new tech has massive issues that get reduced over time as we get better and better at diagnosing and fixing them. Why wouldn't that be expected to happen here, too?

Even as they introduce troubleshooter AI's that can do the work of diagnosing and fixing errors reasonably well, I assume humans are always going to be part of this process.

Imagine looking at a very early computer, and the technician was saying someday you'll have a computer in your pocket that can do all these cool things. And you laughed and said the thing can barely add and subtract, and it's as big as a bus. It always takes time to iterate and improve new tech.

Now they have been over-promising for sure, and maybe the tech will never get as good, useful or widespread as Sam Altman wants us to believe, that's not the end of the world if it ends being just an extremely useful tool. And from what I've seen, certainly in the medical, automobile, and coding fields, it's going to have a massive positive impact, guaranteed.
 
axlsalinger said:
We're so early in the process, and you know every new tech has massive issues that get reduced over time as we get better and better at diagnosing and fixing them. Why wouldn't that be expected to happen here, too?
Click to expand...

Because we're not early in the process, LLM's are a 50 yr old technology and we're not getting better at diagnosing and fixing them, as mentioned in the article I provided, the problem is actually getting worse. The model creators themselves do not believe that hallicinations can be eliminated and the mitigation techniques they're trying to implement require way more compute (which is already a bottlenecking problem) and no one is sure even how effective shit like RAG (Retrieval Augmented Generation) will be. So it might take twice as much power to cut hallucinations from 25% to 23%. No one is sure yet.

An overlooked issue here as well is that for commercial viability, they can't sort this out in 40 years. They're throwing trillions of dollars at it today. If there is no ROI, all of that compute they're currently soaking up will no longer have firms willing to pay the bills and those fixed assets will be used to make money in other ways. Without that compute, LLM AI simple ceases to be a thing. The shit either works well enough to be monetized soon, or it's cooked.

axlsalinger said:
Even as they introduce troubleshooter AI's that can do the work of diagnosing and fixing errors reasonably well, I assume humans are always going to be part of this process.
Click to expand...

AI's can diagnose the problem, but they can't "fix" it in the way we use the word. If an AI generates 1000 lines of code that don't work, and are prompted to diagnose the code and find a mistake, they don't change the line they hallucinated and keep the other 999 lines. It iterates 1000 new lines with the code fix now probably included.

axlsalinger said:
Imagine looking at a very early computer, and the technician was saying someday you'll have a computer in your pocket that can do all things things. And you laughed and said the thing can barely add and subtract, and it's as big as a bus.
Click to expand...

Except that wasn't crazy at all for anyone who understood the technology. Moore's Law was identified in 1965 and very accurately described when we were going to reach the benchmarks you're referring to. Yeah, the layperson would have been amazed, but not the experts or anyone listening to the experts.
 
MindzEye said:
Because we're not early in the process, LLM's are a 50 yr old technology and we're not getting better at diagnosing and fixing them, as mentioned in the article I provided, the problem is actually getting worse. The model creators themselves do not believe that hallicinations can be eliminated and the mitigation techniques they're trying to implement require way more compute (which is already a bottlenecking problem) and no one is sure even how effective shit like RAG (Retrieval Augmented Generation) will be. So it might take twice as much power to cut hallucinations from 25% to 23%. No one is sure yet.

An overlooked issue here as well is that for commercial viability, they can't sort this out in 40 years. They're throwing trillions of dollars at it today. If there is no ROI, all of that compute they're currently soaking up will no longer have firms willing to pay the bills and those fixed assets will be used to make money in other ways. Without that compute, LLM AI simple ceases to be a thing. The shit either works well enough to be monetized soon, or it's cooked.



AI's can diagnose the problem, but they can't "fix" it in the way we use the word. If an AI generates 1000 lines of code that don't work, and are prompted to diagnose the code and find a mistake, they don't change the line they hallucinated and keep the other 999 lines. It iterates 1000 new lines with the code fix now probably included.



Except that wasn't crazy at all for anyone who understood the technology. Moore's Law was identified in 1965 and very accurately described when we were going to reach the benchmarks you're referring to. Yeah, the layperson would have been amazed, but not the experts or anyone listening to the experts.
Click to expand...
The current iteration of generative AI is 7-8 years old. ChatGPT became public in 2022. It's early days, man.

Hallucinations are a problem that may never be fixable. We simply have no idea what's going to happen, but tech has a history of fixing problems over time. I'm not saying it isn't a serious problem.

And this discussion about massive compute has been an issue all along. AI companies were talking about building nuclear power plants to cover the massive amounts needed when DeepSeek came out, providing similar efficiency using a fraction of the power. This type of thing always happens with new technology.
 
axlsalinger said:
The current iteration of generative AI is 7-8 years old. ChatGPT became public in 2022. It's early days, man.
Click to expand...

Nah. These are all just newer versions of the same tech IBM created almost 40 years ago. It's the same technology with incremental improvement over that period.

axlsalinger said:
Hallucinations are a problem that could never be fixed. We have no idea what's going to happen, but tech has a history of fixing problems over time.
Click to expand...

Again though, at the costs required to run this technology and the complete lack of ROI to this point (and in the forseeable future), they just don't have time like that to fix something that is widely considered to be an inherent part of the technology. It's not a bug, it's how the technology is designed to work. It's like saying that eventually they'll eliminate tailpipe emissions from the internal combustion engine. lol, no, they won't. It's inherent to the technology and is an unfixable problem.

axlsalinger said:
And this discussion about massive compute has been an issue all along. AI companies were talking about building nuclear power plants to cover the massive amounts needed when DeepSeek came out, providing similar efficiency using a fraction of the power.
Click to expand...

A few things though...Deepseek has turned out to be significantly worse in every important way other than energy efficiency. Also, and again, who is paying for it all? The industry is burning money at massive, historic rates, no one is actually making any money to return to investors and no one has any plans of making money any time soon.
 
MindzEye said:
Nah. These are all just newer versions of the same tech IBM created almost 40 years ago. It's the same technology with incremental improvement over that period.
Click to expand...
40 years ago was 10 years before we even had public internet. You're comparing current computer tech to a Vic 20.

MindzEye said:
Deepseek has turned out to be significantly worse in every important way other than energy efficiency. Also, and again, who is paying for it all? The industry is burning money at massive, historic rates, no one is actually making any money to return to investors and no one has any plans of making money any time soon.
Click to expand...
These are separate arguments though. I assume you can at least imagine the possibility of this technology generating revenue, if millions of consumers are using and most companies paying for it at corporate rates. My point was that from what I understand, they did not see DeekSeek coming. And it only came out what, 2 years ago? It may be worse now, but we'll see how it progresses. Or more likely, the big guys "steal" some elements of DeepSeek to use in their own systems to greatly reduce the compute power required. Or create their own versions.
 
axlsalinger said:
Just saying, if an AI can do two hours worth of research in 30 seconds, it's probably okay to expect some issues. Humans make mistakes too, and a lot of them.
Click to expand...

Interesting thing i've started to realize is that the AI's don't necessarily actually do all that research but only actually scan parts of the information and then just start making filling the gaps themselves.
 
The most likely victory for the AI industry isn't a tech revolution, it's a government bailout to make their balance sheets whole.
 
When the machines start making memes about the time they wiped out the humans, that's going to be a classic.
 
You must log in or register to reply here.
Back
Top