yeck

There is a poll of AI researchers floating. A lot are concerned with existential risks. I don't want to give a number because I can't exactly remember it.

I think there might be a reference to it in this article. Also, linking to a section talking about this proposal, since that might answer other questions you have.

https://thezvi.substack.com/p/ai-6-agents-of-change#%C2%A7ad...

How so?

General intelligence just means a model that can seek an objective in broader problem spaces. GPT-4, for instance, is able to predict text for many things, like poetry, tutoring, trivia, coding, how-to guides and in different languages and subject matters. This makes it more general than a model that is limits to a single subject, or language, etc. So general intelligence is more a spectrum than a simple boolean.

Consciousness, while philosophically interesting, doesn't necessarily correlate with the risk a general intelligences causing harm.

It would give the world 6 months to trial, test, integrate the currently released models and provide insights into capabilities and maybe even trickier issues like alignment. Could also give researchers, experts and government leaders time to communicate and start figuring out what a short term plan might be.

How so? Could it not be compared to restrictions on activities like enriching uranium? Or research on biohazardous materials?

How come?

Just a personal anecdote, but at my place of employment we had a wave of layoffs late last year. As part of the process 4 top-level execs "left", although I think saying they were "dismissed" is more accurate.

I am not certain, but it might have been nearly half the people at this level let go. And then there was a trail of layoffs all the way down the leadership chain. Of course the most people let go in absolute numbers are the individual contributors.

It isn't anything revolutionary, just helpful. I think of it like a lightweight framework for writing shell scripts.

Parents actually teach their children a lot before they potentially head out for more formal educating, so even if we assume parents don't have teaching experience from their professions, I don't think it is very accurate to suggest they haven't taught anything.

I do question that having a single teacher for 20+ students teaching all subjects for the entire school year is obviously going to work better than motivated homeschooling parent(s). While teachers do have formal education on the matter, it also includes how to manage many children in a classroom setting and how to interface with schools and their processes. The part of the education pertaining to the subjects teachers teach is actually pretty incomplete. To teach elementary students, you have a major, but still teach them subjects that you might only have an average understanding of (like a history major teaching a music class).

Given all that, it doesn't seem hard to believe that parents could do better for a single child than pre-school and elementary school teachers.

As a child gets older their needs evolve such that no one person has everything required to best support that child. That is an advantage of schools (usually high school), but could also be solved with tutoring and other programs available to homeschoolers.

Enjoying the article, but feel the "Economies of scale" section makes an incorrect comparison between a Spotify business model to a theoretical business using OpenAI's API. The author suggests that since Spotify pays royalties per song played, getting more users doesn't mean more money for them and then claims a business using GPT-3 would have a similar limitation.

There are a couple things I think is wrong with this. First, depending on the sort of users acquired by Spotify it does directly translate to more earnings. What doesn't scale well for Spotify seems to be how active the subscription-paying users are. To which I mean that a user who listens to 50 songs a day will cost more than a user only listening to 10 since the subscription price is static and common across users despite usage.

That last point is where the author gets the next thing wrong: assuming that services employing GPT-3 will be fixed subscriptions instead of a pay-as-you-go model (like AWS). I am sure there will be business using fixed sub prices that are independent of usage, but we shouldn't assume that there is anything about GPT-3 that makes that more likely or even very different from other cases where fixed subs are used. There will always be some costs per user, be it the raw cost of electricity or cloud infrastructure. GPT-3's API would just be one more cost per request to consider.