Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

How AI knows things no one told it

scientificamerican.com

Post reply on HN
Print view View original

111–120 of 140 posts

Re: How AI knows things no one told it

#111

Post by mike741 »

HarHarVeryFunny wrote: post #97
mike741 wrote: post #96

Earlier quoted context omitted.

> but it can ace the bar exam A traditional static program could ace the bar exam, or even a well-prepared stack of flash cards. We wouldn't say the flash cards are exhibiting intelligence though, only their creators. > The whole discussion is about intelligence. I was replying to OP. OP was replying to Scientific American's article making numerous unfounded claims of "intelligence." So the party you should be asking…

The way GPT-4 works is by having built a world model of the generative processes that produced the data it was trained on. The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. I'm not sure where you are seeing daylight between this and your own definition of intelligence. FWIW GPT-…

> The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become.

Increasing training data doesn't increase consistency. Each data point acts as a potential new axiom, and each axiom decreases consistency. GPT-4 is trained to satisfy humans, and humans are wildly inconsistent. Even if humans were perfectly consistent, attempting to satisfy multiple different humans simultaneously results in inconsistency. Additionally, even if GPT-4 were perfectly complete and consistent it still wouldn't have reached this state autonomously. So the difference between GPT-4 and intelligence, by my definition, is night and day.

> FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's.

Floating points are digital 1's and 0's. Adding more digits is never going to make something analog.

> The values are imperfectly calculated (limited accuracy) as computer math always is.

Agreed.

>There is also a large element of pure randomness to the output of any of these LLMs.

Strongly disagree. There isn't a single element of randomness during the training stage. We know the exact architecture of the neural net, we know the exact data it was trained on, and we know the exact beam selection algorithms used to synthesize outputs. Every single step can be simulated, traced, and recreated to achieve the exact same results. The number of steps involved might overwhelm us, but that doesn't make it random.

> They don't get to control exactly what words they generate

We do get to control it, we just lose track of the inputs and then pretend it was all out of our control. But of course every single step was willed and controlled by us. We call it "random" for personal convenience, not because its actually true.

Re: How AI knows things no one told it

#112

Post by HarHarVeryFunny »

mike741 wrote: post #111
HarHarVeryFunny wrote: post #97

Earlier quoted context omitted.

The way GPT-4 works is by having built a world model of the generative processes that produced the data it was trained on. The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. I'm not sure where you are seeing daylight between this and your own definition of intelligence. FWIW GPT-…

> The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. Increasing training data doesn't increase consistency. Each data point acts as a potential new axiom, and each axiom decreases consistency. GPT-4 is trained to satisfy humans, and humans are wildly inconsistent. Even if humans…

These models don't output sequences - which is where you'd use beam search - they output a single word at a time. The output is a set of probabilities (from a SoftMax) which is then sampled at a given sampling "temperature" (degree of randomness).

There's no point discussing it when you obviously don't have clue how these models work, won't listen when you're told, and just prefer to make stuff up.

Re: How AI knows things no one told it

#113

Post by mike741 »

HarHarVeryFunny wrote: post #112
mike741 wrote: post #111

Earlier quoted context omitted.

> The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. Increasing training data doesn't increase consistency. Each data point acts as a potential new axiom, and each axiom decreases consistency. GPT-4 is trained to satisfy humans, and humans are wildly inconsistent. Even if humans…

These models don't output sequences - which is where you'd use beam search - they output a single word at a time. The output is a set of probabilities (from a SoftMax) which is then sampled at a given sampling "temperature" (degree of randomness). There's no point discussing it when you obviously don't have clue how these models work, won't listen when you're told, and just prefer to make stuff up.

Beamsearch is just one example, the same applies to top sampling and greedy search. Focusing on one approach suggests you've missed the actual point: if you know how a given output is synthesized, then its not autonomous. You're trying to nitpick as an excuse to avoid a substantive response. If you want me to agree with you, you'll need to offer a counterpoint. Saying things like "It's driven by floating point values not 1's and 0's." and conflating pseudo randomness with actual randomness does not inspire confidence.

Re: How AI knows things no one told it

#114

Post by bheadmaster »

byyyy wrote: post #108
bheadmaster wrote: post #107

Earlier quoted context omitted.

Arguments should be evaluated by their logical soundness, not by who said them.

True, but the volume of information in the world today is too overwhelming for one person to evaluate correctly. We live in the information age after all. Most people don't have the faculties to do a proper evaluation that includes both you and me. Would you trust your self to logically evaluate the capabilities of a nuclear reactor design if you yourself aren't an expert? No. You would hire an expert. Therefore when…

A dangerous myth of modern society is that science is beyond the capacity of ordinary men. I strongly believe it's not.

The practice of science itself may be, as it takes years of research to get to the point where one can produce a new result. However, things that are already known can be taught, and iteratively simplified in a way that abstracts away details while keeping the core argument intact.

Take for example the claim that everything in the universe is made of atoms. It's not a trivial thing to understand, yet everyone accepts it nowdays because we've had so much time and effort put towards simplifying the theory and presenting it to people in a way that is easy to grasp.

If those LLM "experts" were truly experts, they could explain their point clearly without the dark-ages-church "trust the priests, peasant" act.

Re: How AI knows things no one told it

#115

Post by HarHarVeryFunny »

mike741 wrote: post #113
HarHarVeryFunny wrote: post #112

Earlier quoted context omitted.

These models don't output sequences - which is where you'd use beam search - they output a single word at a time. The output is a set of probabilities (from a SoftMax) which is then sampled at a given sampling "temperature" (degree of randomness). There's no point discussing it when you obviously don't have clue how these models work, won't listen when you're told, and just prefer to make stuff up.

Beamsearch is just one example, the same applies to top sampling and greedy search. Focusing on one approach suggests you've missed the actual point: if you know how a given output is synthesized, then its not autonomous. You're trying to nitpick as an excuse to avoid a substantive response. If you want me to agree with you, you'll need to offer a counterpoint. Saying things like "It's driven by floating point values…

I was curious what your definition of intelligence was such that you thought GPT-4 doesn't exhibit it. It's your opinion - doesn't have to agree with mine.

You seem to place an importance on whether the models are entirely predictable or not, which is why I pointed out that the output is randomly sampled.

Re: How AI knows things no one told it

#116

Post by sharemywin »

1000thVisitor wrote: post #53
sharemywin wrote: post #48

Earlier quoted context omitted.

This is the answer I got. with this system prompt "You are an intelligent PhD processor explaining topics to your PhD students." In the context of optimization, does the generalized assignment problem satisfy total unimodularity? Assistant: The generalized assignment problem (GAP) is an NP-hard optimization problem that can be formulated as an integer linear program. The problem involves assigning a set of tasks to a…

That is interesting, but i still think the reasoning is weak as it is not mentioning the range of values that the a_ij and b_i are allowed to take on.

I prompted it some more to see if it got any better. I also increased the number of tokens so it could have time to answer. I cut it off before.

USER In the context of optimization, does the generalized assignment problem satisfy total unimodularity?

Assistant:

The generalized assignment problem (GAP) is an NP-hard optimization problem that can be formulated as an integer linear program. The problem involves assigning a set of tasks to a set of agents, subject to capacity constraints, with the goal of maximizing the total profit or minimizing the total cost. The integer linear programming formulation of GAP is as follows:

Objective function:

Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij

Subject to:

1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent)

2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent)

3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)

Total unimodularity (TU) is a property of a matrix that guarantees that the linear programming relaxation of an integer linear program will result in integer optimal solutions. If the constraint matrix is totally unimodular, you can solve the linear programming relaxation and obtain an integer solution without using specialized integer programming algorithms. ASSISTANT what are some weaknesses to the this argument? ASSISTANT In the context of the original answer, the main weakness is that it does not directly answer the question of whether the generalized assignment problem (GAP) satisfies total unimodularity. The answer provides a description of GAP and the concept of total unimodularity, but it does not provide a conclusion on whether the constraint matrix of GAP is totally unimodular or not.

Another weakness is that the answer does not provide any evidence or examples to support the conclusion. To strengthen the argument, one could analyze the constraint matrix of GAP and determine whether it satisfies the conditions for total unimodularity, such as all subdeterminants being 0, 1, or -1. Additionally, providing examples or counterexamples of instances where GAP exhibits total unimodularity or not would help to clarify the relationship between the problem and this property.

In general, the answer could be improved by directly addressing the question and providing evidence or reasoning to support the conclusion. USER Please craft a better argument. ASSISTANT In order to determine whether the generalized assignment problem (GAP) satisfies total unimodularity, we need to analyze its constraint matrix and check if it satisfies the conditions for total unimodularity.

Recall the integer linear programming formulation of GAP:

Objective function:

Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij

Subject to:

1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent) 2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent) 3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)

The constraint matrix A of GAP can be written as:

      [ a_11   a_12  ... a_1n  0  0  ...  0 ]
      [ a_21   a_22  ... a_2n  0  0  ...  0 ]
A = [ ... ... ... ... ... ... ... ... ] [ a_m1 a_m2 ... a_mn 0 0 ... 0 ] [ 1 0 ... 0 1 0 ... 0 ] [ 0 1 ... 0 0 1 ... 0 ] [ ... ... ... ... ... ... ... ... ] [ 0 0 ... 1 0 0 ... 1 ]

The matrix A is not totally unimodular. To see why, consider a submatrix formed by the first m rows and n columns:

      [ a_11   a_12  ... a_1n ]
      [ a_21   a_22  ... a_2n ]
A' = [ ... ... ... ... ] [ a_m1 a_m2 ... a_mn ]

The determinant of A' can be an arbitrary integer, depending on the values of a_ij. Therefore, the condition for total unimodularity, which requires all subdeterminants to be 0, 1, or -1, is not satisfied.

This implies that the constraint matrix of GAP does not satisfy total unimodularity, and as a result, the linear programming relaxation of GAP does not guarantee integer optimal solutions. Solving GAP requires specialized integer programming algorithms or approximation algorithms to obtain optimal or near-optimal solutions.

Re: How AI knows things no one told it

#117

Post by c22 »

unsupp0rted wrote: post #82
version_five wrote: post #68

Earlier quoted context omitted.

Also want to add: ELIZA: How do you do. Please tell me your problem. YOU: I'm worried about shmerple ELIZA: Do you believe it is normal to be worried about shmerple ? The fact that it can insert your made up word into patterns has no significance in itself.

ChatGPT answers the way a human might: > I'm sorry to hear that you're worried about someone named "shmerple." Can you please provide me with more information about who shmerple is and why you're worried about them?

Sorry, but this doesn't sound like any human I've ever spoken to.

Re: How AI knows things no one told it

#118

Post by unsupp0rted »

c22 wrote: post #117
unsupp0rted wrote: post #82

Earlier quoted context omitted.

ChatGPT answers the way a human might: > I'm sorry to hear that you're worried about someone named "shmerple." Can you please provide me with more information about who shmerple is and why you're worried about them?

Sorry, but this doesn't sound like any human I've ever spoken to.

How would a human, say the receptionist at your dentist's office who doesn't know you at all, respond to something like this?

First question would be: is Shmerple a person or a concept, and assume it's a person, since most people don't start conversations about worrying about concepts.

Therefore: who is Shmerple and why are you worried about them?

Re: How AI knows things no one told it

#119

Post by mike741 »

HarHarVeryFunny wrote: post #115
mike741 wrote: post #113

Earlier quoted context omitted.

Beamsearch is just one example, the same applies to top sampling and greedy search. Focusing on one approach suggests you've missed the actual point: if you know how a given output is synthesized, then its not autonomous. You're trying to nitpick as an excuse to avoid a substantive response. If you want me to agree with you, you'll need to offer a counterpoint. Saying things like "It's driven by floating point values…

I was curious what your definition of intelligence was such that you thought GPT-4 doesn't exhibit it. It's your opinion - doesn't have to agree with mine. You seem to place an importance on whether the models are entirely predictable or not, which is why I pointed out that the output is randomly sampled.

It's not random, its pseudo-random. Those are not the same thing.

Re: How AI knows things no one told it

#120

Post by c22 »

unsupp0rted wrote: post #118
c22 wrote: post #117

Earlier quoted context omitted.

Sorry, but this doesn't sound like any human I've ever spoken to.

How would a human, say the receptionist at your dentist's office who doesn't know you at all, respond to something like this? First question would be: is Shmerple a person or a concept, and assume it's a person, since most people don't start conversations about worrying about concepts. Therefore: who is Shmerple and why are you worried about them?

Sure, your sentence sounded human to me. It's not the content I'm looking at, but the delivery.
Post reply on HN