Does current AI represent a dead end?
Does current AI represent a dead end?
Eerke Boiten, Professor of Cyber Security at De Montfort University Leicester, explains his belief that current AI should not be used for serious applications.
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...
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?
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.)
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.
Dead end, local maximum, tomayto, tomahto.
Yes. No article needed.
Judging just by the headline, the answer should be "No", though, according to Betteridge's law of headlines.
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.