this post was submitted on 24 Jan 2025
91 points (100.0% liked)
technology
23465 readers
636 users here now
On the road to fully automated luxury gay space communism.
Spreading Linux propaganda since 2020
- Ways to run Microsoft/Adobe and more on Linux
- The Ultimate FOSS Guide For Android
- Great libre software on Windows
- Hey you, the lib still using Chrome. Read this post!
Rules:
- 1. Obviously abide by the sitewide code of conduct. Bigotry will be met with an immediate ban
- 2. This community is about technology. Offtopic is permitted as long as it is kept in the comment sections
- 3. Although this is not /c/libre, FOSS related posting is tolerated, and even welcome in the case of effort posts
- 4. We believe technology should be liberating. As such, avoid promoting proprietary and/or bourgeois technology
- 5. Explanatory posts to correct the potential mistakes a comrade made in a post of their own are allowed, as long as they remain respectful
- 6. No crypto (Bitcoin, NFT, etc.) speculation, unless it is purely informative and not too cringe
- 7. Absolutely no tech bro shit. If you have a good opinion of Silicon Valley billionaires please manifest yourself so we can ban you.
founded 4 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
That's not the case with stuff like neurosymbolic models and what DeepSeek R1 is doing. These types of models do actual reasoning and can explain the steps they use to arrive at a solution. If you're interested, this is a good read on the neurosymbolic approach https://arxiv.org/abs/2305.00813
However, automation doesn't just apply to stuff like factory work. If you read the articles I linked above, you'll see that they're specifically talking about automating aspects of producing media such as visual content.
The “chain of thought” output simply gives you the “progress” and the specific path/approach the model has arrived at a particular answer - which is useful for tweaking and troubleshooting the parameters toward improving the accuracy and reducing hallucinations on a model, but it is not the same reasoning that could be given from a human mind.
The transformer architecture is really just a statistical model built to have very strong memory retention when it comes to making associations (in the case of LLMs, words). It fundamentally cannot think or reason. It takes a specific “statistical” path and arrives at an answer based on the associations it has been trained on, but you cannot make it think and reason the way we do, nor can it evaluate or verify the validity of a piece of information based on cognitive reasoning.
Do you actually understand what symbolic logic is?
Neurosymbolic AI is overhyped. It's just bolting on LLMs to symbolic AI and pretending that it's a "brand new thing" (it's not, it's actually how most LLMs practically work today and have been for a long time GPT-3 itself is neurosymbolic). The advocates of approach pretend that the "reasoning" comes from symbolic AI which is known as classical AI, which still suffers from the same exact problems that it did in the 1970's when the first AI winter happened. Because we do not have an algorithm capable of representing the theory of mind, nor do we have a realistic theory of mind to begin with.
Not only that but all of the integration points between classical techniques and statistical techniques present extreme challenges because in practice the symbolic portion essentially trusts the output of the statistical portion because the symbolic portion has limited ability to validate.
Yeah you can teach ChatGPT to correctly count the r's in strawberry with a neurosymbolic approach but general models won't be able to reasonably discover even the most basic of concepts such as volume displacement by themselves.
You're essentially back at the same problem where you either lean on the symbolic aspects and limit yourself entirely to advanced ELIZA like functionality that can just use classifier or your throw yourself to the mercy of the statistical model and pray you have enough symbolic safeguards.
Either way it's not reasoning, it is at best programming -- if that. That's actually the practical reason why the neurosymbolic space is getting attention because the problem has effectively been to be able to control inputs and outputs for the purposes of not only reliability / accuracy but censorship and control. This is still a Garbage In Garbage Out process.
FYI most of the big names in the "Neurosymbolic AI as the next big thing" space hitched their wagon to Khaneman's Thinking Fast and Slow bullshit that is effectively made up bullshit like Freudianism but lamer and has essentially been squad wiped by the replication crisis.
Don't get me wrong DeepSeek and Duobau are steps in the right direction. They're less proprietary, less wasteful, and broadly more useful, but they aren't a breakthrough in anything but capitalist hoarding of technological capacity.
The reason AI is not useful in most circumstance is because of the underlying problems of the real world and you can't algorithm your way out of people problems.
I don't think it's overhyped at all. It's taking two technologies that are good at solving specific types of problems and using them together in a useful way. The problem that symbolic AI systems ran into in the 70s are precisely the ones that deep neural networks address. You're right there are challenges, but there's absolutely no reason to think they're insurmountable.
I'd argue that using symbolic logic to come up with solutions is very much what reasoning is actually. Meanwhile, classification of input problem is the same one that humans have as well. Somehow you have to take data from the senses and make sense of it. If you're claiming this is garbage in garbage out process, then the same would apply to human reasoning as well.
The models can create internal representations of the real world through reinforcement learning in the exact same way that humans do. We build up our internal world model through our interaction with environment, and the same process is already being applied in robotics today.
I expect that future AI systems will be combinations of different types of algorithms all working together and solving different challenges. Combining deep learning with symbolic logic is an important step here.
Not in any meaningful way. A statistical model cannot address the Frame problem. Statistical models themselves exacerbate the problems of connectionist approaches. I think AI researchers aren't being honest with the causality here. We are simply fooling ourselves and willfully misinterpreting statistical correlation as causality.
Let me repeat myself for clarity. We do not have a valid general theory of mind. That means we do not have a valid explanation of the process of thinking itself. That is an insurmountable problem that isn't going to be fixed by technology itself because technology cannot explain things, technology is constructed processes. We can use technology to attempt to build a theory of mind, but we're building the plane while we're flying it here.
Because you are a human doing it, you are not a machine that has been programmed. That is the difference. There is no algorithm that gives you correct reasoning every time. In fact using pure reasoning often leads to lulzy and practically incorrect ideas.
It does. Ben Shapiro is a perfect example. Any debate guy is. They're really good at reasoning and not much else. Like read the Curtis Yarvin interview in the NYT. You'll see he's really good at reasoning, so good that he accidentally makes some good points and owns the NYT at times. But more often than not the reasoning ends up in a horrifying place that isn't actually novel or unique simply a rehash of previous horriyfing things in new wrappers.
This is a really Western brained idea of how our biology works, because as complex systems we work on inscrutable ranges. For example lets take some abstract "features" of the human experience and understand how they apply to robots:
Strength. We cannot build a robot that can get stronger over time. Humans can do this, but we would never build a robot to do this. We see this as inefficient and difficult. This is a unique biological aspect of the human experience that allows us to reason about the physical world.
Pain. We would not build a robot that experiences pain in the same way as humans. You can classify pain inputs. But why would you build a machine that can "understand" pain. Where pain interrupts its processes? This is again another unique aspect of human biology that allows us to reason about the physical world.
The frame problem is addressed by creating a model of the environment the system interacts with. That's what provides the context for reasoning and deciding what information is relevant and what isn't. Embodiment is one obvious way to build such a model where the robot or even a virtual agent interacts with the environment and encodes the rules of the environment within its topology.
This is not necessary for making an AI that can reason about the environment, make decisions, and explain itself. Furthermore, not having a theory of mind does not even prevent us from creating minds. One example of this could be using evolutionary algorithms to evolve agents that have similar reasoning capabilities to our own. Another would be to copy the structure of animal brains to a high degree of fidelity.
You are programmed in a sense of the structure of your brain being a product of the information encoded in your DNA. The same way the neural network is a product of the algorithms used to build it. However, the learning that both your brain and the network are doing is encoded in the weights and connections of the network through reinforcement. These are not programmed in either case.
🙄
You're showing utter lack of imagination on your part here. Of course we could build a robot that could get stronger. There's nothing uniquely biological about this example.
Maybe try thinking why organisms evolve pain in the first place and what advantage it provides.