zlacker

[parent] [thread] 19 comments
1. tsimio+(OP)[view] [source] 2024-05-15 12:34:44
The simplest answer, without adding any extraordinary capabilities to the AGI that veer into magical intelligence, is to have AI assistants that can seemlessly interact with technology the way a human assistant would.

So, if you want to meet with someone, instead of opening you calendar app and looking for an opening, you'd ask your AGI assistant to talk to their AGI assistant and set up a 1h meeting soon. Or, instead of going on Google to find plane tickets, you'd ask you AGI assistant to find the most reasonable tickets for a certain date range.

This would not require any special intelligence more advanced than a human's, but it does require a very general understanding of the human world that is miles beyond what LLMs can achieve today.

Going only slightly further with assumptions about how smart an AGI would be, it could revolutionize education, at any level, by acting as a true personalized tutor for a single student, or even for a small group of students. The single biggest problem in education is that it's impossible to scale the highest quality education - and an AGI with capabilities similar to a college professor would entirely solve that.

replies(4): >>Zambyt+43 >>sebast+Yq >>duped+eH >>JohnFe+bu1
2. Zambyt+43[view] [source] 2024-05-15 12:53:47
>>tsimio+(OP)
This is definitely an interesting way to look at it. My initial reaction is to consider that I can enhance the capabilities of a system without increasing its inteligence. For example, if I give a monkey a hammer, it can do more than it could do when it didn't have the hammer, but it is not more intelligent (though it could probably learn things by interacting with the world with the hammer). That leads me to think: can we enhance the capabilities of what we call "AI systems" to do these things, without increasing their intelligence? It seems like you can glue GPT-4o to some calendar APIs to do exactly this. This seems more like an issue of tooling rather than an issue of intelligence to me.

I guess the issue here is: can a system be "generally intelligent" if it doesn't have access to general tools to act on that intelligence? I think so, but I also can see how the line is very fuzzy between an AI system and the tools it can leverage, as really they both do information processing of some sort.

Thanks for the insight.

replies(1): >>tsimio+Ol2
3. sebast+Yq[view] [source] 2024-05-15 14:53:16
>>tsimio+(OP)
The examples you're providing seem to have been thoroughly solved already.

I'm at the European AI Conference for our startup tomorrow, and they use a platform that just booked me 3 meetings automatically with other people there based on our availability... It's not rocket science.

And you don't even need those narrow tools. You could easily ask GPT-4o (or lesser versions) something along the lines of :

> "you're going to interact with another AI assistant to book meetings for me: [here would be the details about the meeting]. Come up with a protocol that you'll send to the other assistant so it can understand what the meetings are about, communicate you their availability, etc. I want you to come up with the entire protocol, send it, and communicate with the other assistant end-to-end. I won't be available to provide any more context ; I just want the meeting to be booked. Go."

replies(1): >>tsimio+GQ1
4. duped+eH[view] [source] 2024-05-15 16:05:15
>>tsimio+(OP)
> The single biggest problem in education is that it's impossible to scale the highest quality education

Do you work in education? Because I don't think many who do would agree with this take.

Where I live, the single biggest problem in education is that we can't scale staffing without increasing property taxes, and people don't want to pay higher property taxes. And no, AGI does not fix this problem, because you need staff to be physically present in schools to deal with children.

Even if we had an AGI that could do actual presentation of coursework and grading, you need a human being in there to make sure they behave and to meet the physical needs of the students. Humans aren't software to program around.

replies(1): >>tsimio+DR1
5. JohnFe+bu1[view] [source] 2024-05-15 20:12:35
>>tsimio+(OP)
> This would not require any special intelligence more advanced than a human's, but it does require a very general understanding of the human world that is miles beyond what LLMs can achieve today.

Does it? I am quite certain those things are achievable right now without anything like AI in the sense being discussed here.

replies(1): >>tsimio+UR1
◧◩
6. tsimio+GQ1[view] [source] [discussion] 2024-05-15 22:30:12
>>sebast+Yq
GPT-4(o) lacks the ability to connect to any of the tools needed to achieve what I'm describing. Sure, it maybe could give instructions about how this could be done, but it can't actually do it. It can't send an email to your email account, and it can't check your incoming emails to see if any arrived asking for a meeting. It can't then check your calendar, and propose another email, or book a time if the time is available. It doesn't know that you normally take your lunch at some time, so that even though the spot is free, you wouldn't want a meeting at that time. And even if you did take the considerable amount of effort to hook it up with all of these systems, it's failure rate is still far too high to rely on it for such a thing.

And getting it to actually buy stuff like plane tickets on your behalf would be entirely crazy.

Sure, it can be made to do some parts of this for very narrowly defined scenarios, like the specific platform of a single three day conference. But it's nowhere near good enough for dealing with the general case of the messy general world.

replies(2): >>Zambyt+dm3 >>sebast+Wj4
◧◩
7. tsimio+DR1[view] [source] [discussion] 2024-05-15 22:39:18
>>duped+eH
Having individual tutors for each child is not often discussed because it is self-evidently impossible for any cost whatsoever - it would require far too high a percentage of the workforce of a country to be dedicated to education. But it is the most responsible thing for the difference between the education the elites get, especially the elites of the past, and the general education.

Sure, this doesn't mean you could just fire all teachers and dissolve all schools. You still need people to physically be there and interact with the children in various ways. But if you could separate the actual teaching from the child care part, and if you could design individualized courses for each child with something approaching the skill of the best teachers in the whole world, you would get an inconceivably better educational system for the entire population.

And I don't need to work in education for much of this. Like all others, I was intimately acquainted with the educational system (in my country) for 16 years of my life through direct experience, and much more since in increasingly less direct experience. I have very very good and very direct experience of the variance between teachers and the impact that has on how well students understand and interact with the material.

replies(1): >>duped+Ig4
◧◩
8. tsimio+UR1[view] [source] [discussion] 2024-05-15 22:43:23
>>JohnFe+bu1
Show me one product that can offer me an AI assistant that can set up a meeting with you at a time that doesn't contradict any of our plans, given only my and your email address.
replies(1): >>JohnFe+gh3
◧◩
9. tsimio+Ol2[view] [source] [discussion] 2024-05-16 04:11:46
>>Zambyt+43
I'm sure some aspects of this can be achieved by manually programming GPT-4 links to other specific services. And obviously, some interaction tools would have to be written manually even for an AGI.

The difference though is the amount of work. Today if you wanted GPT-4 to work as I describe, you would have to write an integration for Gmail, another one for Office365, another one for Proton etc. You would probably have to create a management interface to give access to your auth tokens for each of these to OpenAI so they can activate these interactions. The person you want to sync with would have to do the same.

In contrast, an AGI that only has average human intelligence, or even below, would just need access to, say, Firefox APIs, and should easily be able to achieve all of this. And it would work regardless if the other side is a different AGI using a different provider, or even if they are just a regular human assistant.

replies(1): >>Zambyt+Lo3
◧◩◪
10. JohnFe+gh3[view] [source] [discussion] 2024-05-16 14:08:15
>>tsimio+UR1
I've never looked into actual products as this isn't something I'm interested in. I'm just saying that accomplishing this can be done without involving AI of the sort being discussed here. I'm not sure what such AI would bring to the table for this sort of task.

> given only my and your email address.

AI or not, such an application would need more than just email addresses. It would need access to our schedules.

replies(1): >>tsimio+tq4
◧◩◪
11. Zambyt+dm3[view] [source] [discussion] 2024-05-16 14:35:26
>>tsimio+GQ1
> Sure, it maybe could give instructions about how this could be done [...]

If you were in a room with no computer, would you consider yourself to be not intelligent enough to send an email? Does the tooling you have access to change your level of intelligence?

◧◩◪
12. Zambyt+Lo3[view] [source] [discussion] 2024-05-16 14:51:37
>>tsimio+Ol2
What if you ask GPT-4 to write the integration between its API and an email provider? You're not really "manually" creating the integration then.
replies(1): >>tsimio+xA4
◧◩◪
13. duped+Ig4[view] [source] [discussion] 2024-05-16 20:00:56
>>tsimio+DR1
That's like claiming you know how to run a restaurant because you like to eat out. Or worse actually, since you're extrapolating your individual experience from a small set of educational systems to education as a whole.

If you're looking for insight into the problems faced in education, speak to educators. I really doubt they would tell you that the quality of individual instructors is their biggest problem.

replies(1): >>tsimio+0r4
◧◩◪
14. sebast+Wj4[view] [source] [discussion] 2024-05-16 20:22:57
>>tsimio+GQ1
Here's what's strange about your argument.

I had a (human) assistant in my previous business, super-smart MBA type, and by your definition she wasn't a general intelligence on the day of onboarding:

- she didn't have access to my email account or calendar

- she didn't know my usual lunch time hours

- she didn't have a company card yet.

All of those points you're raising are logistics, not intelligence.

Intelligence is "When trying to achieve a goal, can you conceive of a plan to get there despite adverse conditions, by understanding them and charting/reviewing a sequence of actions".

You can definitely be an intelligent entity without hands or tools.

replies(2): >>tsimio+su4 >>Infini+Pag
◧◩◪◨
15. tsimio+tq4[view] [source] [discussion] 2024-05-16 21:02:48
>>JohnFe+gh3
My point is that an AGI would give you this use case for free. Currently this kind of product, AI or not, simply doesn't exist. It's in principle doable, but the number of integrations required makes it uneconomical. An AGI assistant could use the same messy interfaces we use, and thus it would be compatible with every email provider and client ever created.

> AI or not, such an application would need more than just email addresses. It would need access to our schedules.

It needs access to my schedule, yes, but it only needs your email address. It can then ask you (or your own AGI assistant) if a particular date and time is convenient. If you then propose another time, it can negotiate appropriately.

◧◩◪◨
16. tsimio+0r4[view] [source] [discussion] 2024-05-16 21:05:13
>>duped+Ig4
Educators don't like to discuss the performance of other educators, as most professionals don't like to diss their colleagues, especially not in front of their customers. But the quality of educators is absolutely a huge problem, so huge that there are even consecrated sayings about it (those who can, do; those who can't, teach). So huge that one of the most well known rock anthems of all time is about the poor quality of educators (Pink Floyd's Another Brick in the Wall Part II).

Educators are the best people to ask about how to make their jobs easier. They are not necessarily the best people to ask about how to make children's education better.

Edit:

> That's like claiming you know how to run a restaurant because you like to eat out.

No, it's like claiming you know some things about the problems of restaurants, and about the difference between good and bad restaurants, after spending 8+ hours a day almost every day, for 16 years, eating out at restaurants. Which I think would be a decent claim.

◧◩◪◨
17. tsimio+su4[view] [source] [discussion] 2024-05-16 21:20:34
>>sebast+Wj4
I'm pretty certain your assistant learned to do all of those things more or less on her own. Of course, you shared your schedule and email with them, and similarly, you'd have to share your schedule and email with an AGI.

But you certainly didn't have to write a special program for your assistant to integrate with your inbox, they just used an existing email/calendar client and looked at their screen.

GPT-4 is nowhere near able to interact with, say, the Gmail web page at this level. And even if you created the proper integrations, it's nowhere near the level that it could read all incoming email and intelligently decide, with high accuracy, which emails necessitate updates to your calendar, which don't, and which necessitate back-and-forth discussions to negotiate a better date for you.

Sure, your assistant didn't know all of this on day one, but they learned how to do it on their own, presumably with a few dozen examples at most. That is the mark of a general intelligence.

replies(1): >>sebast+5wp
◧◩◪◨
18. tsimio+xA4[view] [source] [discussion] 2024-05-16 21:53:07
>>Zambyt+Lo3
You can try that. I don't think it will be as reliable as you'd want for something like this.
◧◩◪◨
19. Infini+Pag[view] [source] [discussion] 2024-05-21 13:29:18
>>sebast+Wj4
I think the intelligence part is to think of any potential logistical obstacles and figure out ways to deal with them with minimal disruption except when necessary because of potential conflicts with other goals.
◧◩◪◨⬒
20. sebast+5wp[view] [source] [discussion] 2024-05-24 08:32:11
>>tsimio+su4
I think we're disagreeing on the current capacity of models, as much as we're disagreeing about the definition of AGI.

I'm pretty sure, from previous interactions with GPT-4o and from their demos, that if you used their desktop app (which enables screensharing) and asked it to tell you where to click, step-by-step, in the Gmail web page, it would be able to do a pretty good job of navigating through it.

Let's remember that the Gmail UI is one of the most heavily documented (in blogs, FAQs, support pages, etc) in the world. I can't see GPT-4o having any difficulty locating elements in there.

[go to top]