- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
You must log in or register to comment.
Dead end, local maximum, tomayto, tomahto.
The right tool for the right job. It’s not intelligent, it is just trained. It all boils down to stochastic.
And then there is the ecological aspect…
Or sometimes the moral aspect, if it is used to manage someone’s “fate” in application processing. And it might be trained to be racist or misogynist if you use the wrong training data.Yeah. Considering the obscene resources needed for ChatGPT and the others, I don’t think the niche use cases where they shine makes it worth it.
The moral aspect is resolved if you approach building human systems correctly too.
There is a person or an organization making a decision. They may use an “AI”, they may use Tarot cards, they may use the applicant’s f*ckability from photos. But they are somehow responsible for that decision and it is judged by some technical, non-subjective criteria afterwards.
That’s how these things are done properly. If a human system is not designed correctly, then it really doesn’t matter which particular technology or social situation will expose that.
But I might have too high expectations of humanity.
Accountability of a human decision maker is the way to go. Agreed.
I see the danger when the accountant’s job asks for high throughput which enforces fast decision making and the tool (llm) offers fast and easy decisions. What is the accountant going to do, if (s)he just sees cases instead of people and fates?
If consequence for a mistake follows regardless, then it doesn’t matter.
Or if you mean the person checking others - one can make a few levels of it. One can have checkers interested in different outcomes, like in criminal justice (… it’s supposed to be).
It’s not intelligent, it is just trained.
The same could be said about every human being…
An LLM cannot think like you and I. it’s not able to solve entirely new problems. And it doesn’t have a concept of the world - it paints hands without knowing what a hand does.
It is a system which learns the rules of something by means of reinforcement learning to tune the coefficients of its heap of linear equations. It is better than a human in its area. I guess it can be good for tedious, repetitive tasks. Nevertheless it is just a huge coefficient matrix.
But it can only reproduce what is in the training data - you need lots of already solved examples in the training data. It doesn’t work for entirely new problems.
(that’s also the reason, why LLMs don’t give good answers to questions about specialized niche topics. When there are just one or two studies, there just isn’t enough training data for the LLM.)
its not able to solve entirely new problems.
They replaced the training data with an evaluator. (which rates the LLMs output for training?) Interesting, thanks.
Edit: this reminds me of the self evolving (virtual) robot problem, a robot which is rated by an external moderator and improves over time. I.e.: https://www.sciencedirect.com/science/article/pii/S0925231221003982
Right? I see comments all the time about it just being glorified pattern recognition. Well…thats what humans do as well. We recognize patterns and then predict the most likely outcome.
That is one part of many that a human brain does. This is like trying to say the color red is a rainbow, because the rainbow has red in it.
Can you expand on that?
How? You’re focusing on one thing a human does and using it to point to how human like LLMs are, while ignoring everything else humans do. You’re missing the forest for the trees.
I didn’t say that at all. What I said was LLMs solve problems just like a human does. Pattern recognition. Then I asked you to provide an example of one thing a human does that doesnt boil down to pattern recognition. The words we speak and type are patterns. The decisions we make are based on patterns we learned in the past. Thats really all I meant by it.
LLMs don’t solve problems. That’s the point being made here. Many other algorithms do indeed solve issues, but those are very niche, as the alogos were explicitly designed for those situations.
While yes, humans excel at pattern recognition, sometimes to the point of it being a problem, there are many things we do that have nothing to do with patterns beyond the fact that they are tangentially involved. Emotions for instance don’t inherently follow patterns. They can, but they aren’t directly tied. Exploration also doesn’t come from pattern recognition.
If you need examples of why people flat out say LLMs aren’t solving problems, look at the recent “how many r’s in strawberry” which has admittedly been “fixed” in many models.
And after all, current “AI” models are just one step on a longer way, which is what I read as the conclusion of the article.
But if you read the article, then you saw that the author specifically concludes that the answer to the question in the headline is “yes.”
This is a dead end and the only way forward is to abandon the current track.
Yes. No article needed.
Judging just by the headline, the answer should be “No”, though, according to Betteridge’s law of headlines.
But we don’t like AI, therefore anything negative said about it is more plausible than anything positive said about it. You see the dilemma here.
Hmmh, I usually get upvoted for citing Betteridge’s law… But not today within this context. Yeah, I’m aware of Lemmy’s mentality. I still think what I said holds up 😉
I’ve found the Fediverse to be a lot “bubblier” than Reddit, I suspect because the communities are smaller. Makes it easier for groupthink to become established. One element of the @technology bubble is a strong anti-AI sentiment, I’ve kind of given up on getting any useful information on that subject here. Quite unfortunate given how widespread it’s getting.
I’m subscribed to: [email protected] , [email protected] , [email protected] , [email protected] (German)
I think we could co-exist peacefully, if we put in some effort. But that’s not how Lemmy works. We use the “All”-Feed instead of subscriptions and then complain we don’t want posts about all the topics… Most posts are made to AskLemmy, NoStupidQuestions or Technology, disregarding if any dedicated communities exist… It’s a bit of a mess. And yeah, the mob mentality is strong. It helps if you go with the flow. Just don’t mention nuanced and complicated details. Simple truths always win. Especially if people like to believe something were true.
I really don’t care any more. I mean I agree with most of this problem. Some days half the posts on technology (or more) are about AI. And a lot of that isn’t even newsworthy stuff or something with substance. If I were in charge, I’d delete that, tell people to discuss the minute details in some dedicated community and have some more balance. Because there are a lot of other very interesting tech-related topics that could be discussed instead.
And simultaneously, I’d love to see some more activity and engagement in the AI communities.
Yes. That’s why everyone is scrambling to create new interoperable model languages and frameworks that work on more efficient hardware.
Almost everything that is productized right now stems from work in the Python world from years ago. It got a swift uptake with Nvidia making it easier to use their hardware on compiled models, but now everyone wants more efficient options.
FPGA presents a huge upside to not being locked into a specific vendor, so some people are going that route. Others are just making their frameworks more modular to support the numerous TPU/NPU processors that everyone and their brother needlessly keeps building into things.
Something will come out of all of this, but right now the community shift is to do things without needing so much goddamn power draw. More efficient modeling will come as well, but that’s less important since everything is compiled down to something that is supported by the devices themselves. At the end of the day, this is all compilation and logic, and we just need to do it MUCH leaner and faster than the current ecosystem is creeping towards. It’s not only detrimental to the environment, it’s also not as profitable. Hopefully the latter makes OpenAI and Microsoft get their shit together instead of building more power plants.
I don’t really see how FPGA has a role to play here. What circuit are you going to put on it. If it’s tensor multipliers, even at low precision, a GPU will be an order of magnitude faster just on clock speed, and another in terms of density.
What we’ve got right now has almost nothing to do with python, and everything to do with the compute density of GPUs crossing a threshold. FPGAs are lower density and slower.
If you’re unfamiliar with FPGA, you may want to read up a bit, but essentially, a generic platform that is reprogrammed between iterations of doing something more efficiently than a generic instruction set. You tell it what to do, and it does it.
This is more efficient than x86, ARM, or RISC because you’re setting the boundaries and capabilities, not the other way around.
Your understanding of GPUs is wrong though. What people run now is BECAUSE of GPUs being available and able to run those workloads. Not even well, just quickly. Having an FPGA set for YOUR specific work is drastically more efficient, and potentially faster depending on what you’re doing. Obviously for certain things, it’s a circle peg in a square hole, but you have to develop for what is going to work for your own specific use-case.
I know exactly what they are. I design CPUs for a living, use FPGAs to emulate them, and have worked on GPUs and many other ASICs in the past.
FPGAs can accelerate certain functions, yes, but neural net evaluation is basically massive matrix multiplies. That’s something that GPUs are already highly optimised for. Hence, why I asked what circuit you’d put on the FPGA. Unless you can accelerate the algorithmic evaluation by several orders of magnitude the inefficiency of FPGAs Vs ASICs will cripple you.
You don’t design CPUs for a living unless you’re talking about the manufacturing process, or maybe you’re just bad at it and work for Intel. Your understanding of how FPGA works is super flawed, and your boner for GPUs is awkward. Let me explain some things as someone who actually works in this industry.
Matrix math is just stupid for whatever you pipe through it. It does the input, and gives an output.
That is exactly what all these “NPU” co processing cores are about from AMD, Intel, and to a further subset Amazon and Google on whatever they’re calling their chips now. They are all about an input and output for math operations as fast as possible.
In my own work, these little AMD XDNA chips pop out multiple segmented channels way better than GPUs when gated for single purpose. Image inference, audio, logic, you name it. And then, SHOCKER!, if I try and move this to a cloud instance, I can reprogram the chip on the fly to swap from one workload to another in 5ms. It’s not just a single purpose math shoveling instance anymore, it’s doing articulations on audio clips, or if the worker wants, doing ML transactions for data correlation. This costs almost 75% less than provisioning stock sets of any instances to do the same workload.
You have no idea what you’re talking about.
Why are you being so condescending about this?
FPGAs are a great tool, but they’re not magic.
They are a great way to prototype ASICs or for performing relatively simple low latency/high-throughput tasks below the economies of scale where actually taping out an ASIC would make sense but there is pretty much no case where an FPGA with a bunch of the same logic path is going to outperform a dedicated ASIC of the same logic.
NPUs are already the defacto ASIC accelerator for ML. Trying to replicate that functionality on an FPGA fabric of an older process node with longer path lengths constraining timing is going to be worse than a physically smaller dedicated ASIC.
It was the same deal with crypto-mining, the path for optimizing parallel compute is often doing it badly on a GPU first, moving to FPGA if memory isn’t a major constraint, then tape out ASICs once the bugs in the gateware are ironed out (and economies of scale allow)
And that doesn’t even begin to cover the pain of FPGA tooling in general and particularly vendor HLS stacks.
Matrix math is just stupid for whatever you pipe through it. It does the input, and gives an output.
Indeed.
That is exactly what all these “NPU” co processing cores are about from AMD, Intel, and to a further subset Amazon and Google on whatever they’re calling their chips now. They are all about an input and output for math operations as fast as possible.
Yes, they are all matrix math accelerators, and none of which have any FPGA aspects.
Except AMD XDNA is a straight up FPGA, and Intel XEco is as well.
For someone who claims to work in this industry, you sure have no idea what’s going on.
yes, but what you need to be doing is tons of multiply-accumulate, using a fuckton of memory bandwidth… Which a gpu is designed for. You won’t design anything much better with an fpga.
