Microsoft Says New AI Shows Signs of Human Reasoning
nytimes.com
nytimes.com
That and marketing. Microsoft wants all of them sweet £ flowing in and thus over selling what their ai does. Sad part is that many will fall for it.
I also watches the writer of the papers talk and it was a good talk.
I also saw the progress of ml in the last 5 years and I have never seen something progressing that fast (besides Smartphones perhaps).
So it's not far fetched that got 4 is not at the ceiling of doable.
The opposite: chatgpt makes it much easier to get funding, the race is on
I use ml regularly.
I know quite a lot of people doing this also.
Why do you think we are at transistor age?
Assuming we're in the "transistor age" implies that things are going to get significantly better, and yet that's a very big assumption, IMO.
Wrong, you are not using chatgpt. You are training chatgpt. You are paying a fee simply because they cant scale yet - there are only that many gpus available - and they had to slow things down a bit but openai gains value in you improving its little bot.
Millions of people training a model and an ever increasing computing power means than in 10-20 years ai will be where computer games are today vs 20-30 years ago.
Hence my analogy to transistors. Ai is not going away but what needs to go away is companies such as openai and bad scifi believing mob.
A dangerous myth. Idiocracy was an entertainment film with a side order of eugenics, not a documentary.
The surprising thing is not that it was able to code a good solution, it's that it was able to understand the specifications of how i wanted it to work and how it should be integrated into the existing code.
It's obvious that this will be faster and more effective once a GPT4-class model is tightly integrated into the run-compile-test loop.
For example, given a set of text, an LLM will predictably pick like 1 out of 3 possibilities for the next word, and that word affects it picking the next word and so on. So its essentially turning what would otherwise be a giant look up table into a compressed markov model representation.
Reasoning is the ability to generate information that is not currently in that model. For example, if you take a neural net and train it on all the properties of fluids, it should be able to give you answers about aerodynamics if it can truly reason.
To use the example in the article, the answer to
"Here we have a book, nine eggs, a laptop, a bottle and a nail. Please tell me how to stack them onto each other in a stable manner"
is not something the original 'lookup table' could possibly help you with, other than by manually reasoning about how the various parts of a 'concept web' fit together. This question represents a completely unfilled gap in the training data.
Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic.
Lets say I ask you the following question "Here we have a flibert, nine moltice, a grook, a seerik and a lopo. Please tell me how to stack them onto each other in a stable manner". You would be able to figure out how to stack them, given the fact that you have the ability to "reason" and figure out what those items are in the first place. In the same way, an AI that can reason would be able to do the same.
Just like transformers were revolutionary because they essentially are representative of a process inside human brains where we learn to pay attention to select features, there will be some new architecture in the next 10 years that will encode this ability to assign semantic properties to information and operate on those properties, which will then allow for things like AI to ask clarifying questions.
This is very reductionist. Of course everything is numbers under the hood. But it's entirely orthogonal to the question. Numbers can encode all of the complexity needed for anything that can do reasoning. Numbers in the form of a neural network can execute arbitrary algorithms, given enough layers.
> Lets say I ask you the following question "Here we have a flibert, nine moltice, a grook, a seerik and a lopo. Please tell me how to stack them onto each other in a stable manner". You would be able to figure out how to stack them, given the fact that you have the ability to "reason" and figure out what those items are in the first place. In the same way, an AI that can reason would be able to do the same.
I really don't understand the purpose of this thought experiment. A human is able to figure this out because they can see what those objects are. Is your objection to language model reasoning that they are not multi-modal?
> Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic.
What if you couldn't see? What if the interaction was limited to just text? Would you be smart enough to ask for an image of those objects?
Do you think an LLM can "reason" to ask for additional information, like images of the objects?
A sufficiently large LLM can be trained with a wide range of training data, to the point where it can answer some questions where its seemingly generating new information, but that new information is simply a combination of existing information. In the same way, when you ask it some question about code, it may put together two snippets of code never seen before into something that is new. The key factor there is all of that information is encoded into the network, so all the network is doing is essentially a somewhat complicated graph lookup.
GPT-4 really does do human-like reasoning. And it's clearly quite general purpose within its limitations.
But it doesn't have many other aspects of humans/animals such as self-direction, high bandwidth senses, a stream of subjective experience, emotions, certain types of adaptivity, etc. It's not alive and it's not a digital person.
But people aren't able to separate all of those different things, so they can't admit that it has any kind of useful intelligence, because for them that means it's a digital person.
Actually it's even dumber, many people jump to assuming that it goes straight to god-like superintelligence if it has any intelligence at all. Which is another reason they can't admit there is any reasoning, because that would mean the end of the world or something.
My personal experience shows me that ChatGTP 4 is something of a fool with a big fast reference book. I typically use it for programming, but I also study music theory. When I was exploring one of the more complex (to me) ideas I asked chat GTP 4 to explain it to me and provide examples.
After previously asking about Cmaj to Emaj, I asked "can you give me an example of the H transformation?"
ChatGPT: "Sure! To demonstrate the Hexatonic pole 'H' transformation using chords similar to C major and E major, let's first introduce the concept of the hexatonic system. A hexatonic system consists of six notes, and two triads can be extracted from it. In the context of the 'H' transformation, these two triads are related by a common tone, and the other two notes in each triad are a tritone apart.
Let's use the C major triad (C, E, G) as a starting point. We'll build a hexatonic collection containing this triad. One possible hexatonic system containing C major is the one with the notes C, Db, E, F#, G, and A. From this collection, we can extract two triads:
C major (C, E, G) F# major (F#, A, Db) Notice that the C major and F# major triads share a common tone (C/Db enharmonically), and the other notes in each triad (E and G in C major, and F# and A in F# major) are a tritone apart. In this case, the 'H' transformation links the C major and F# major triads.
So, the Hexatonic pole 'H' transformation example using chords similar to C major would be transitioning from C major to F# major."
Me: "I'm pretty sure F# major isn't F#, A, Db"
ChatGPT: "You are correct, and I apologize for the mistake. The F# major triad should be F#, A#, and C#. Let's re-examine the 'H' transformation with the correct F# major triad..."
After about 10 more prompts trying to extract information from it I ended up needing to be very literal to get results out of it:
Me: "the definition is "The 'H' transformation takes a major chord of root X to the minor chord whose root is eight semitones above X." can you give me a VALID example of this"
ChatGTP: "I apologize for the confusion in my previous responses. Based on the definition you provided, the 'H' transformation takes a major chord of root X to the minor chord whose root is eight semitones above X. Here's a valid example using this definition:
Let's start with a C major triad (C, E, G). To find the target chord, we need to count eight semitones above the root note (C):
C# / D♭ D D# / E♭ E F F# / G♭ G G# / A♭ So, the root of the target chord is A♭. Now, we form an A♭ minor triad (A♭, C♭, E♭).
The 'H' transformation in this example would be transitioning from C major (C, E, G) to A♭ minor (A♭, C♭, E♭)."
Nice! This time it was correct (I think!). But it required that I feed it the exact definition and only when I saw it doing the work did I have any confidence that it was correct. In the end I did learn how it these concepts work, I think I did at least. Overall it's not a great teacher, half of the time I had to teach it and correct it when it seemed to be very confident in its answers, and even more confident that my corrections were correct, which I'm not 100% sure of.
One observation is that ChatGTP4 didn't have any idea what it was talking about the entire time. It was just putting related information together. That observation was based on the notes it kept claiming were related to each other via interval degrees. The intervals it was using were always wrong, not even close, just pure nonsense. Only when told exactly how to count did it get it right and only when the counting was part of the response. Now mind you, the things it was getting wrong were fundamental music theory 101 stuff, but it was making these fundamental mistakes in the middle of a explanation of a very complex topic. I don't know what it all means, but I wouldn't trust it to fly or build an airplane, or even boil water now that I think about it. How would you know when it goes dumb?
You could say 'it is just a fool backed up with a good encyclopedia', but let's be honest, those responses could not come from a fool.
And really the responses MUST have come from a fool. It feels like I was talking to someone who knew absolutely nothing about music theory but had a good internet connection and was trying to wing an interview.
The breadth and depth of the AI's knowledge is profound to say the least. However its ability to comprehend the information is nothing short of abysmal. If you were trying to pass a Music Theory 101 class you would fail bad if you listened to the answers it comes up with. It's as if it was unable to apply what it was talking about at some fundamental level. Like it could tell you all about covalent bonds and air pressure ratio to boiling point but will tell you water boils at at 20,000 degrees C at sea level. Impressive, but terribly wrong.
https://alpof.wordpress.com/2021/10/09/neo-riemannian-exampl...