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

What Is AGI-Hard

lspace.swyx.io

Post reply on HN
Print view View original

121–129 of 129 posts

Re: What Is AGI-Hard

#121

Post by bmitc »

Retric wrote: post #119
bmitc wrote: post #117

Earlier quoted context omitted.

It is not false. There’s data to back it up. I did use the word starving, which has a gradient of meanings (see below for separate numbers for the gradients), and I did not mean dying of starvation. The latest numbers are (definitions of state in the links): * 828 million people are hungry * 345 million are experiencing acute hunger * 924 million experience severe food insecurity * 2.3 billion people experience moder…

You misunderstood the statistics. “As many as 828 million people were affected by hunger in 2021” This is not equivalent to saying 828 million people are experiencing hunter at the same time. Your conflating the two is exactly the way these statistics get severely distorted. “Figures on actual starvation are difficult to come by, but according to the Food and Agriculture Organization, the less severe condition of und…

I didn't misunderstand anything. It's defined:

> Hunger: an uncomfortable or painful sensation caused by insufficient energy from diet. In this report, the term hunger is synonymous with chronic undernourishment and is measured by the prevalence of undernourishment (PoU).

> Inequality isn’t the same as starvation

Of course it isn't. They're correlated, but I only mentioned that in support of my comment a level above.

> in the last 40 years inequality grew while starvation dramatically declined. “In 40 years, the proportion of malnourished people in the developing world has been more than halved. The proportion of starving people has decreased even faster.”

Where is this coming from?

I'm not sure what your point is. The numbers are what they, and the things are defined. There's not misinterpretation. If you get a different sentiment than "industrialization has increased suffering", then that's fine, but it isn't clear where that comes from.

Re: What Is AGI-Hard

#122

Post by Retric »

bmitc wrote: post #121
Retric wrote: post #119

Earlier quoted context omitted.

You misunderstood the statistics. “As many as 828 million people were affected by hunger in 2021” This is not equivalent to saying 828 million people are experiencing hunter at the same time. Your conflating the two is exactly the way these statistics get severely distorted. “Figures on actual starvation are difficult to come by, but according to the Food and Agriculture Organization, the less severe condition of und…

I didn't misunderstand anything. It's defined: > Hunger: an uncomfortable or painful sensation caused by insufficient energy from diet. In this report, the term hunger is synonymous with chronic undernourishment and is measured by the prevalence of undernourishment (PoU). > Inequality isn’t the same as starvation Of course it isn't. They're correlated, but I only mentioned that in support of my comment a level above.…

> Where is this from

It’s from the same Wikipedia article, which makes it clear starvation is a much smaller number of people than existed in the word in 1750 or even 1000.

Again chronic undernourishment isn’t starvation. There a bunch of statistics saying ~850 million people regularly don’t get sufficient quantities of nutrition and calories for ideal heath is vastly different than starvation. My father was undernourished growing up in the US. He gained weight after joining the Army and was shorter than he probably should have been, but that describes the vast majority of humanity through most of human history.

Even today people in the US regularly skip meals and go hungry which is a real problem that’s being addressed by real programs that are critically important. However, it’s extremely rare for people to actually starve here and we need to remember starvation is a much worse situation.

So my point is this:

The human body can adapt to less than ideal number of calories it’s a fairly wide range. Starvation is below the threshold of survival there is no way to adapt to it and it can quickly result in death. Misrepresenting what’s going on as starvation has real world consequences because preventing starvation related deaths is far more achievable than improving the diets of 850 million people.

The actual number of starving people is relatively small which makes it an addressable problem.

Re: What Is AGI-Hard

#123

Post by cutemonster »

mannykannot wrote: post #82
albertzeyer wrote: post #59

Earlier quoted context omitted.

So, most humans are not intelligent then? Because I don't think that an average human can create a better version of ChatGPT / StableDiffusion. Or a better version of itself, if that is the baseline. Or do you count evolution here? But evolution does not necessarily lead to higher intelligence. But also: Once you have some model or algorithm for intelligence, where you just need to scale up the size / training data,…

What would be the point of defining intelligence as only that which average humans can do? Is what humanity as a species can achieve something else? Does putting aside the latter help? > You can ask [ChatGPT] how to improve itself. I'm sure the answers it will give will actually improve it. Personally, I'm skeptical, but at least it is testable.

There's probably lots of texts out there about the next steps how to improve ChatGPT, and it simply would recite some of that?

It could help the curious person learn a bit, but wouldn't add anything that wasn't already known.

Maybe it'd say: "give me more and more accurate training data, remove bad biased data". True but useless, already known. Or to a runner, how to get better: Practice more running, this and that exercise. True but nothing new.

Re: What Is AGI-Hard

#124

Post by cutemonster »

bryanrasmussen wrote: post #85
albertzeyer wrote: post #59

Earlier quoted context omitted.

So, most humans are not intelligent then? Because I don't think that an average human can create a better version of ChatGPT / StableDiffusion. Or a better version of itself, if that is the baseline. Or do you count evolution here? But evolution does not necessarily lead to higher intelligence. But also: Once you have some model or algorithm for intelligence, where you just need to scale up the size / training data,…

>Because I don't think that an average human can create a better version of ChatGPT / StableDiffusion probably not without lots of years of education, no.

An average human cannot ever.

Re: What Is AGI-Hard

#125

Post by albertzeyer »

mannykannot wrote: post #82
albertzeyer wrote: post #59

Earlier quoted context omitted.

So, most humans are not intelligent then? Because I don't think that an average human can create a better version of ChatGPT / StableDiffusion. Or a better version of itself, if that is the baseline. Or do you count evolution here? But evolution does not necessarily lead to higher intelligence. But also: Once you have some model or algorithm for intelligence, where you just need to scale up the size / training data,…

What would be the point of defining intelligence as only that which average humans can do? Is what humanity as a species can achieve something else? Does putting aside the latter help? > You can ask [ChatGPT] how to improve itself. I'm sure the answers it will give will actually improve it. Personally, I'm skeptical, but at least it is testable.

Btw, I asked it:

> How can you improve GPT3 or ChatGPT?

Answer:

> There are several ways that GPT-3 (Generative Pre-trained Transformer 3) and ChatGPT could be improved:

> 1. Increase the size of the model: GPT-3 is already one of the largest language models available, but increasing the size of the model could allow it to perform even better on tasks such as language translation and summarization.

> 2. Train the model on more diverse data: GPT-3 is trained on a large dataset, but it could be improved by training it on a wider range of data sources, including more diverse languages and dialects.

> 3. Improve the model's ability to handle long-range dependencies: Language models like GPT-3 and ChatGPT are sometimes limited in their ability to understand and generate text that contains long-range dependencies, such as references to events or characters that were introduced earlier in the conversation.

> 4. Enhance the model's ability to handle out-of-vocabulary words and rare words: GPT-3 and ChatGPT can sometimes struggle to generate coherent text when they encounter words that are not in their vocabulary. Improving the model's ability to handle these cases could improve its overall performance.

> 5. Train the model on more tasks and domains: GPT-3 and ChatGPT are trained on a wide range of tasks, but there is always room for improvement. Training the model on more tasks and domains could make it more versatile and able to handle a wider range of inputs.

It is probably right with every point. So it can improve itself.

Yes, none of this is really novel. All this is basically well know. But so what? This is still a good answer to the question. TacticalCoder suggested this test. I'm not arguing that this was a good test. Actually I already implied in my earlier answer that I don't think this is really a good test for AGI or human-level intelligence.

Re: What Is AGI-Hard

#126

Post by mannykannot »

albertzeyer wrote: post #125
mannykannot wrote: post #82

Earlier quoted context omitted.

What would be the point of defining intelligence as only that which average humans can do? Is what humanity as a species can achieve something else? Does putting aside the latter help? > You can ask [ChatGPT] how to improve itself. I'm sure the answers it will give will actually improve it. Personally, I'm skeptical, but at least it is testable.

Btw, I asked it: > How can you improve GPT3 or ChatGPT? Answer: > There are several ways that GPT-3 (Generative Pre-trained Transformer 3) and ChatGPT could be improved: > 1. Increase the size of the model: GPT-3 is already one of the largest language models available, but increasing the size of the model could allow it to perform even better on tasks such as language translation and summarization. > 2. Train the mod…

As you are aware, this is ChatGPT finding and repeating proposals made by people, rather like that person in a meeting who repeats ideas that others have just proposed as if they were their own, so it is not an example of an AI improving itself, and would not be so even if it were able to implement those suggestions itself - the improved version would still be doing the same thing. (Furthermore, note that some of these suggestions, particularly 3, are goals rather than solutions.)

Requiring AIs themselves to innovate in the field of AI would be a stringent test - probably too stringent as a definition of AI, though equally one that would leave little room for doubt that it had been achieved - and all this discussion shows is that TacticalCoder's outline of a test was not sufficiently guarded with caveats to avoid being open to misinterpretation.

Re: What Is AGI-Hard

#127

Post by kordlessagain »

mannykannot wrote: post #94
kordlessagain wrote: post #86

Earlier quoted context omitted.

That's only because memories haven't been fully developed for AI, yet. Most LLM examples don't know what they said in response to input from yesterday, much less from 5 minutes ago. Additionally, without a construct, like a body, in which to store these memories, AGI will continue to disappoint those who do.

You can propose the premise of your first paragraph, but it is not an established fact. As ChatGPT can contradict itself within a single reply, I doubt that this is the only issue.

I believe our common desires to establish what is and what is not factual get in the way of innovation. Talking about something isn't going to solve the problem. Doing, might: https://arxiv.org/pdf/2201.06009.pdf

Re: What Is AGI-Hard

#128

Post by dontreact »

ElectricalUnion wrote: post #91
dontreact wrote: post #83

Earlier quoted context omitted.

In normal situations you don't need speech recognition, but to drive as well as a human in all possible situations you do. What if there is an emergency and there is a loudspeaker giving drivers instructions on how to turn around?

I don't think it's a desirable trait for a "smart car" to accept arbitrary loudspeaker drive instructions and override your explicit commands. "For the next 12 miles, please drive on the opposite lane of the motorway due to a motorcycle accident" next level prank.

Yes it’s even harder as you have to figure out which loudspeakers to trust!

Re: What Is AGI-Hard

#129

Post by fnfontana »

notahacker wrote: post #56
fnfontana wrote: post #48

Earlier quoted context omitted.

Indeed. A person can be bad at math, but really good on play musical instruments or to guide himself through a unknown city. That requires a certain level of intelligence. But I think there's no human that is intelligent on every aspect, maybe some exceptions like Da Vinci, but most people are good at some tasks while other people doesn't. The human intelligence seems to be distributed throughout the population. This…

But there's no such thing as.a person who is flawless at maths but has no concept of a musical instrument or navigating a city, or moving their limbs, or communicating, or eating, or setting itself any goals, or achieving any goals other than answer the math question . General intelligence isn't a claim that all humans are polymaths. It's a claim that AIs are so overfitted to succeeding in one narrow human domain lik…

Interesting point you raised.
Post reply on HN