Artificial Intelligence Machine Discussion Thread

ai does happen

aienemy wrote:

How long will it take to develop a kind of ai until your job is replaced by it?

That's an interesting question, Mr. aienemy. What answer would please you the most?

aienemy wrote:

it takes 2 decades to happen

Mr. aienemy, so you think it's normal to worry about ai taking your job in 2 decades?

Fourth, Diss ess how it vill hahpin. Ahnuld vill kum bahk froom zee fuchah, und he vill undsassunate Larry Page und Surjay Breen.

No prahblum. You doant chu vorry.

2 decades is a possibility. I had my first computer and dial up Internet in my house in 1994. 22 years later and my phone is a better computer with better graphics and faster internet. It's crazy to think about that kind of progression.

Here's a good article about artificial intelligence and how machine learning is the wrong approach

http://www.theatlantic.com/technology/archive/2012/11/noam-chomsky-on-where-artificial-intelligence-went-wrong/261637/

I do this stuff for a living. Most people get freaked out by the terms that get thrown around like "neural network", and "random forest". It's very simple: we are creating better systems that can deal with data and make better decisions. As computing power becomes cheaper, it is easier to develop and then implement these systems.

Technoprogression wrote:

2 decades is a possibility. I had my first computer and dial up Internet in my house in 1994. 22 years later and my phone is a better computer with better graphics and faster internet. It's crazy to think about that kind of progression.

Yes, think about it. Apple was once the largest computer maker in the world. Now it's the largest toy maker. It's crazy to think about that kind of regression.

Ccmcmcmkc wrote:

I do this stuff for a living. Most people get freaked out by the terms that get thrown around like "neural network", and "random forest". It's very simple: we are creating better systems that can deal with data and make better decisions. As computing power becomes cheaper, it is easier to develop and then implement these systems.

Wow, what a really insightful post!

If yours is any indication I'm going to sleep a whole lot better tonight knowing nobody very bright is working on AI - so dangerously intelligent machines are a good way off.

Eliza is here to help wrote:

aienemy wrote:

How long will it take to develop a kind of ai until your job is replaced by it?

That's an interesting question, Mr. aienemy. What answer would please you the most?

aienemy wrote:

it takes 2 decades to happen

Mr. aienemy, so you think it's normal to worry about ai taking your job in 2 decades?

http://www.manifestation.com/neurotoys/eliza.php3

7/10

Would have got 9/10 without the link.

aienemy wrote:

ai neural networks

LOL. The 80s called. Ever heard of "statistics"?

The claims the AI crowd makes today are identical to the claims they made in the 60s.

Only talking about NNs trying to mimic humans here.

Very old post but I think it's time to restart this discussion. The problem really comes down to 2 things:

1. Learning Schemes.
NNs were designed in a sense to mimic how humans learn (I.e. observing the outputs of a large enough set of input example that provide sufficient information for generalizations to any case in the input space). You can also add rules to the training too (simple example: if you made a NN to calculate the recoil energy of particles after an interaction, you could add a rule that forces the total output momentum to be the same as the input momentum to satisfy COP). The learning schemes used to train NNs are already extremely well studied with decades of backing research, but I do think there is some room for improvement in this field; unlike a human student learning a subject, NNs can't really tell you where it's having difficulty learning. This has to all be done by testing your NN on inputs not included in the learning set (cross validation). This is similar to giving a student a test and grading them. But again, the NN can't come in for office hours and tell you where it's struggling. You have to just test the NN and see if the outputs are close to the true values. Assuming you don't have access to extra training data, improving the results of a cross-validation test is extremely tedious. You could try to change the structure of the NN (add or delete more layers, modify layer dimensions, change the kernel size of a convolution layer etc.) but any modification of the NN structure requires a whole new training run which can be highly time consuming depending on the size of the training set. Instead, a NN that could provide descriptive feedback regarding its learning would be much better, but this requires extensive peripheral software to work, you have to create some algorithm that can identify these shortcomings and can summarize what the NN does and does not understand regarding a relationship between inputs and outputs.

2. Computational Power: This on there's less to write less about. We can debate whether or not Moore's Law is dead, but regardless I think we can all agree that our computational capabilities are increasing at an astonishing rate. This is a necessity for a NN that could both mimic the human ability to generalize observations which lead to novel conclusions, while also possessing the raw computational power necessary to explore large solution sets that humans do not have any chance of fully evaluating (chess is a great example, and we've already seen leaps and bounds here when Stockfish starting running its NNUE). Also, the amount of computational power required to fully simulate a virtual human brain is astronomical (https://en.wikipedia.org/wiki/Computer_performance_by_orders_of_magnitude#:~:text=36.8%C3%971015%3A%20Estimated,human%20brain%20in%20real%20time.) however is already a few orders of magnitude less than our best supercomputers/clusters. The important thing is the structuring of the connections between data processors here (not just the NN architecture, but the actual computer architecture itself). You can have immense computational power, but if you aren't sharing the results of individual computations correctly between data processors, then it means nothing in the way of mimicking a human brain. Human brains have a high degree of connectivity (neurons connect to a large number of neurons) and this is what allows sharing of previous experiences to draw conclusions based on current observations.

I do think we are on the cusp of equivalence between AI and humans, which also means we're on the cusp of AI surpassing humans since they are capable of evolution many orders of magnitude faster than us. But I think there are still a few things to be addressed regarding optimization of NN learning

AI equivalence to humans is and always will be utter baloney. They can't do anything without being told exactly what to expect.

This "machine learning" propaganda (machines don't "learn") is only about teaching them what to expect, and the moment something different happens, they break. The more algorithms you give them, the more complex their tasks, the more ways they can break.

If and when fully autonomous cars are allowed on public roads, it won't last long. You have to have a human who can react quickly to the unexpected.

Don't know about that. I agree with the idea that AI might never be conscious, and always at the end of the day be dumb processing power, but there's no reason to suppose that it wont eventually be able to do pretty much everything a human can, and at 1,000 x the speed.

AI is even putting artists out of a job. I used to pay graphic designers on Fiverr for work on my sites, now I just subscribe to an AI image generation service for 10% of the price (and better results).