• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

How AI will fail

kram182 said:
Gonna tongue-in-cheekily use AI to reply to this 😁

View attachment 804245
Click to expand...
The Electrical Grid includes urban surface rail transit, that is, streetcars and now Light Rail Vehicles, elevated or underground urban metro systems (like the Chicago L or the New York City subway system), and suburban and intercity rail, passenger and freight, notably the Boston - Washington part of Amtrak.

Ironically, after WW II, almost all streetcar lines were replaced by bus lines

The railroads, mostly diesel powered, have adopted modern technologies like drones and now some AI
 
rinaldo00 said:
You mentioned nuclear power, one movie set back US nuclear power for 40 years. Who knows what raging mobs will due to data centers.
Click to expand...

And you think our failure to roll out nuclear power then was a good thing?

For starters, we wouldn't be having the energy issues with AI now if we had rolled it out then. Which is why the green energy initiative has since essentially died at all these formerly pushing it companies and nuclear plants being spun up new/again now (and nuclear was always greener than all the pushed green initiatives). Imagine what else we could have had/done in all that time with all that energy abundance, if not for the troglodytes.

That, along with the other given past examples, should be reflected on now, in regards to pushback of AI.

And so long as China/foreign adversaries of the US will be using/building out AI - economically, militarily - so too will (and should) the US.
 
Last edited:
kram182 said:
And you think our failure to roll out nuclear power then was a good thing?

For starters, we wouldn't be having the energy issues with AI now if we had rolled it out then. Which is why the green energy initiative has since essentially died at all these formerly pushing it companies and nuclear plants being spun up new/again now. Imagine what else we could have had/done in all that time with all that energy abundance, if not for the troglodytes.

That, along with the other given past examples, should be reflected on now, in regards to pushback of AI.

And so long as China/foreign adversaries of the US will be using/building out AI - economically, militarily - so too will (and should) the US.
Click to expand...
I never said it was a good thing, I am just saying people are emotional and you never know what they will do. Every poll says people hate AI, every poll says people hate data centers. Will it be enough to slow AI growth? I DON'T KNOW, neither do you.
 
rinaldo00 said:
I never said it was a good thing, I am just saying people are emotional and you never know what they will do. Every poll says people hate AI, every poll says people hate data centers. Will it be enough to slow AI growth? I DON'T KNOW, neither do you.
Click to expand...

And I just explained with the previous past given examples why regardless of what every poll says/the masses think, it probably won't amount to anything in regards to hindering AI's progress/buildout - and then specifically with the outlier example of nuclear, why it showed us what the polls/masses think in regards to AI/datacenters shouldn't matter, as well.
 
kram182 said:
And I just explained with the previous past given examples why regardless of what every poll says/the masses think, it probably won't amount to anything in regards to hindering AI's progress/buildout - and then specifically with the outlier example of nuclear, why it showed us what the polls/masses think in regards to AI/datacenters shouldn't matter, as well.
Click to expand...
And I gave you an example of technology that people hated and pushed back on slowing its growth. I never said it wasn't coming I am saying how fast, how large, I DON'T KNOW and neither do you.
 
rinaldo00 said:
And I gave you an example of technology that people hated and pushed back on slowing its growth. I never said it wasn't coming I am saying how fast, how large, I DON'T KNOW and neither do you.
Click to expand...

And again YOUR EXAMPLE SHOWS IT WAS A BAD THING WHEN PEOPLE DID THAT AND GOT THEIR WAY AND THE PEOPLE WERE WRONG.

That should be reflected on now, when people are trying to pull the same thing.
 
Alright, let's not start getting shitty with each other, as I have a feeling this conversation might be moving that direction. State your opinions, facts, and analysis, and move on.
 
FrgMstr said:
Alright, let's not start getting shitty with each other, as I have a feeling this conversation might be moving that direction. State your opinions, facts, and analysis, and move on.
Click to expand...
 

Attachments

  • it-makes-me-feel-like-a-big-man-big-man.gif
    it-makes-me-feel-like-a-big-man-big-man.gif
    1.4 MB · Views: 0
Intriguing video. My thoughts:

Local Pros:
  1. As memory requirements shrink, costs will go down and more can be done locally due to optimizations (quantization, KV Cache Transform Coding, etc.),
  2. Accuracy of AI from absorbing more information is projected to plateau in a few years. These large companies are hitting diminishing returns.
  3. Robots will be the next big thing that use AI. We've had AI Roombas for over a decade that could run locally. They don't need to have the entire internet downloaded.
Local Cons:
  1. Cost of achieving almost parity with existing cloud-based LLMs is >$12K (RTX 6000 Pro 96GB + nVME PCI Gen 5 SSDs in RAID + CPU/MB/Memory). Paying that off with a ChatGPT $20/M subscription would take 50 years. That's a huge gap.
  2. The cost of achieving parity will go up, as hardware / electricity costs go up, from being devoured by the big tech firms. The companies making local hardware (i.e. NVIDIA, Micron, etc.) have been shifting their resources to making more cloud hardware - bad for local.
  3. Time costs: I've invested a ton of money in local AI hardware for privacy reasons. I still have not set up the hardware, let alone the software, as I haven't had the time. ...and myself and anyone on this forum is probably in the top 0.1% of the population of hardware savviness. That's a big obstacle for average Joe.
  4. Local AI is generally slower than the cloud.
  5. Local AI will lack feature performance. Good luck with AI video rendering on a desktop in a reasonable amount of time, let alone a laptop or phone.
  6. Local AI will always have an information gap - even with internet connections and API's there will be moats of information that they will lack access to.
  7. Privacy: Most people don't care about privacy. I do, but life moves too fast for most people to prioritize fighting the dozens of ways that freedom is taken away from them every day. With AI, life moves even faster. People don't have time to care about privacy when their job is being taken over by AI.
Will local AI takeover locally? I think we'll see a mix. I think the best stuff will still be in the cloud. I think some of these companies are going to run out of money or give up their pursuit (OpenAI, Meta, Amazon, Apple, xAI).

...and to compare this to the music industry? Music studios haven't gone away. Artists may have the hardware to create content, but an artist does not have the millions of dollars to spend on marketing, producers, mixers or connections with venues that will generally make them famous. Ask Bruno Mars or Taylor Swift about contracts - they still needed them.
 
Last edited:
1.1.2.3.5... said:
How do you put guard rails on? You can effectively delete any safety put on any model, and huggingface is full of examples of this.

As this evolves, especially for local instances, there's not much you can do. I mean, you can make a law, but you can't enforce it without violating fourth amendment protections.

Cloud providers already have model safety.

Other countries we can't regulate.
Click to expand...
Forget law, you can't even enforce guardrails in the models themselves most of the time. Generative AI regularly bypasses it's own guardrails as a fluke with completely innocuous prompts.
 
https://www.linkedin.com/posts/amdevs-ugcPost-7463013993474039810-zGMv


✨ Personal AI is the next computing platform.

AI is shifting from something you access to something you build with, locally, at the edge, and across systems.

We’re unlocking new possibilities for developers:

• AMD Ryzen AI Halo, a local-first developer system, preorder starting in June, develop AI without limits on your desk

• Gorgon Halo with up to 192GB unified memory, supporting 300B+ parameter models locally, run massive models locally

We’re excited to partner with Clem Delangue 🤗, Co-founder and CEO of Hugging Face, to advance open-source AI for Ryzen AI.

Our focus is seamless AI, from model to deployment.

Cloud, edge, device. One continuum.

Multi-agent systems, local inference at scale, open models as infrastructure.

This is the next computing era 🚀
 
FrgMstr said:
https://www.linkedin.com/posts/amdevs-ugcPost-7463013993474039810-zGMv


✨ Personal AI is the next computing platform.

AI is shifting from something you access to something you build with, locally, at the edge, and across systems.

We’re unlocking new possibilities for developers:

• AMD Ryzen AI Halo, a local-first developer system, preorder starting in June, develop AI without limits on your desk

• Gorgon Halo with up to 192GB unified memory, supporting 300B+ parameter models locally, run massive models locally

We’re excited to partner with Clem Delangue 🤗, Co-founder and CEO of Hugging Face, to advance open-source AI for Ryzen AI.

Our focus is seamless AI, from model to deployment.

Cloud, edge, device. One continuum.

Multi-agent systems, local inference at scale, open models as infrastructure.

This is the next computing era 🚀
Click to expand...
I'm pretty impressed with what a 32GB 5090 can do locally TBH. I've been playing with Qwen 3.6 the last few days and, well, it knows its shit with Python better than I do. I've been writing code for decades, but I'm more of a C-style guy. Java, C#, C++, C, etc. We have basically unlimited Google Gemini at work. It's better, but I have an easier time keeping up with Gemini working in Java than I do keeping up with Qwen running locally with Python. Not a Python expert. Java? Been doing that since the 1990s. At any rate my point is you don't need mega hardware for local AI to be useful. I should play around more with local AI on my portable rig. That was a $1k build I did in early 2025. ARC B580, 9600X, 32GB. Ran a small model on it a little and it seemed to do ok. A system with a combined memory model and a lot more than 32GB could run larger models, but I'm not sure what I'd like better for local use. It'd probably depend on the job. 5090 will beat anything short of a datacenter or high end pro GPU on speed if the model fits in vram, but larger models are less likely to give a bad answer.
 
zandor said:
I'm pretty impressed with what a 32GB 5090 can do locally TBH. I've been playing with Qwen 3.6 the last few days and, well, it knows its shit with Python better than I do. I've been writing code for decades, but I'm more of a C-style guy. Java, C#, C++, C, etc. We have basically unlimited Google Gemini at work. It's better, but I have an easier time keeping up with Gemini working in Java than I do keeping up with Qwen running locally with Python. Not a Python expert. Java? Been doing that since the 1990s. At any rate my point is you don't need mega hardware for local AI to be useful. I should play around more with local AI on my portable rig. That was a $1k build I did in early 2025. ARC B580, 9600X, 32GB. Ran a small model on it a little and it seemed to do ok. A system with a combined memory model and a lot more than 32GB could run larger models, but I'm not sure what I'd like better for local use. It'd probably depend on the job. 5090 will beat anything short of a datacenter or high end pro GPU on speed if the model fits in vram, but larger models are less likely to give a bad answer.
Click to expand...
But vllm will let you load more than 32gb of parameters and here's the key...if you use a MOE model that easily fits in your vram, you can get an exceptionally powerful local model.

You should consider trying the new deepseekv4 model on hf and running it with vllm.

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
158b parameters, but only 13b are active, and those get loaded into your vram. So with 32gb, you should have amazing token rates. What I don't know is how fast pushing an expert into your vram is, but let us know if you try it and how well it works.
 
Last edited:
AI is a force multiplier. For better or for worse. Use it if you want. Don't use it if you don't want.

Things change in the world all the time. Jobs are always affected by market forces and changes.

Whether or not that empowers or defeats you, is up to you.

/thread.
 
1.1.2.3.5... said:
But vllm will let you load more than 32gb of parameters and here's the key...if you use a MOE model that easily fits in your vram, you can get an exceptionally powerful local model.

You should consider trying the new deepseekv4 model on hf and running it with vllm.
Click to expand...
I'll take that for a spin. I mostly just tried Qwen when I decided to try a local setup instead of just using Gemini at work because it sounded like it would be good for writing code. And yeah I know I can load larger models than my vram can take and use quantized models. I've been playing with the Qwen 3.6 35b parameter model quantized to Q4_K_M so the whole thing fits in vram. My main point was just that one that fits in 32GB has been working pretty well, and depending on the job I might go for max speed over a bigger model at least for personal local use. I also want to play with smaller setups more just to see what they can handle. Maybe do most stuff on gaming rig, occasionally outsource to a cloud provider. Do most stuff locally, then when you have a hard question rent $1M+ worth of kit for a few seconds.
 
The FOMO is to hate on A.I. but most people are already using A.I. even if they think they are not A.I. is here to stay I just buildt my first 8 GPU server that is already doing it's job for customers and it will not stop there.

The hate is hilarius and reminds me of this:
1779392956768.png


You can cry but you will not stop progress and the only thing you achieve is looking like a bellend.
So keep crying over the "steam engine", you are already consuming it.
 
littlegreenmenfromhome said:
The FOMO is to hate on A.I. but most people are already using A.I. even if they think they are not A.I. is here to stay I just buildt my first 8 GPU server that is already doing it's job for customers and it will not stop there.

The hate is hilarius and reminds me of this:
View attachment 804644

You can cry but you will not stop progress and the only thing you achieve is looking like a bellend.
So keep crying over the "steam engine", you are already consuming it.
Click to expand...
It's the canners vs the luddites
 
littlegreenmenfromhome said:
The FOMO is to hate on A.I. but most people are already using A.I. even if they think they are not A.I. is here to stay I just buildt my first 8 GPU server that is already doing it's job for customers and it will not stop there.

The hate is hilarius and reminds me of this:
View attachment 804644

You can cry but you will not stop progress and the only thing you achieve is looking like a bellend.
So keep crying over the "steam engine", you are already consuming it.
Click to expand...
It can be a useful tool, but wake me up when it begins to resemble anything close to the "I" part.
 
Meeho said:
It can be a useful tool, but wake me up when it begins to resemble anything close to the "I" part.
Click to expand...

I would argue that what we have demonstrates that consciousness and intelligence are not the same thing. There are relatively low iq animals that are obviously conscious.

The big part that is missing isn't intelligence but state. We sort of work around that with complex tooling to make llms look almost conscious but it's still just a function call (sort of) that's been optimized on a loss function.

But if and when we it have internal state and it's able modify its own weights to solve for a dynamic loss function, then we're playing a different game. Last year I would have said that's 20 years out, but now i feel like 10 is more realistic.
 
littlegreenmenfromhome said:
A.I. is already more intelligent than some of the posters in this thread...the Dunning-Kruger is strong here.
Click to expand...
I don't believe this will happen but one guy does.
Stephen Hawking expressed deep concern about the potential dangers of AI, particularly the possibility that machines could surpass human intelligence and evolve independently. In a 2014 BBC interview, he stated that AI "would take off on its own, and re-design itself at an ever increasing rate," emphasizing that humans, limited by slow biological evolution, could be overtaken and rendered obsolete. He feared that superintelligent AI might pursue goals misaligned with human survival, posing an existential risk.
Hawking also highlighted the societal and economic consequences of AI. He warned that automation and intelligent machines could accelerate industrial efficiency at the expense of jobs, widening global inequality. Additionally, he cautioned that AI could be exploited in military applications, potentially leading to autonomous weapons and global conflicts.
Click to expand...

https://www.sciencing.com/1981927/stephen-hawking-artificial-intelligence-prediction/
 
Since what posts are on topic in this thread has loosened up I want to go on record saying I don't think AI sucks and I never said so. It is a powerful tool for MANY things. I believe that at present there are wildly massive exaggerated claims, we don't know how fast it will grow, and how willing are people to do it locally. That's all.

My original post was put to spur debate and it has.
 
kram182 said:
Gaussian IQ bell curve meme stays winning
View attachment 804670
Click to expand...
Ran this through Qwen on my 5090 to make some edits... Qwen apparently sucks at "Jar-Jar Binks" and "Emperor Palpatine". At least the Palpatine to Vader swap is consistent. Asked for Jar-Jar on the left and I get a different random critter every time.
1779416199649.png

Just deal with it. Resistance is futile. Learn to make AI serve you.
 
mullet said:
It's not Ai.
Click to expand...
On a philosophical level I agree with you. It's not "intelligent". It's just a very complex system that responds to inputs and produces often useful outputs. OTOH that's what everyone seems to be calling this tech so we can just go with the standard popular label or cause confusion.
 
littlegreenmenfromhome said:
A.I. is already more intelligent than some of the posters in this thread...the Dunning-Kruger is strong here.
Click to expand...
They are not comparable as AI is not intelligent at all. It's as to claim that a Boeing 777 is more intelligent than most posters here because it can land itself. They are a narrowly useful tool for specific tasks that fails miserably when tasked with general intelligence problems.

And even retarded people have the ability to say "I don't know", while LLMs are all but incapable of it.

AI usage is the embodiment of the Murray Gell-Mann Amnesia effect.
 
Can we move on from the esoteric discussion about AI being "intelligent?" We all know what "AI" refers to in the context it is being used.
 
FrgMstr said:
Can we move on from the esoteric discussion about AI being "intelligent?" We all know what "AI" refers to in the context it is being used.
Click to expand...
I mean if AI was actually "intelligent" ... I would consider it mentally retarded in its current state.
 
You must log in or register to reply here.
Back
Top