Models like o1 suggest that people won’t generally notice AGI-ish systems that are better than humans at most intellectual tasks, but which are not autonomous or self-directed Most folks don’t regularly have a lot of tasks that bump up against the limits of human intelligence, so won’t see it - ThreadSky

emollick.bsky.social • 84 days ago

Models like o1 suggest that people won’t generally notice AGI-ish systems that are better than humans at most intellectual tasks, but which are not autonomous or self-directed

Most folks don’t regularly have a lot of tasks that bump up against the limits of human intelligence, so won’t see it

Comments

strangeloopcanon.com•84 days ago

Don't think this is quite true. It's a question that what's needed is a reliable agent that's at GPT4 levels rather than a super smart o1 that still fails many common sense tasks.

emollick.bsky.social•84 days ago

General agents could change stuff but may be a harder challenge than expected.

karstenhowes.bsky.social•83 days ago

Lack of agentic spark feels like a fundamental model-of-intelligence problem. Current LLMs are worse at employing concepts introduced during usage than those learned in training. Successful agents incorporate ideas continuously as new ‘weights’ in their neuron substrate; LLMs operate only on words.

tachikoma.elsewhereunbound.com•84 days ago

AGI should be defined as the limits of human intelligence - a child is generally intelligent while not being at the forefront of human intellect.

joedavis72.bsky.social•84 days ago

A language model is like having a friend that has memorized every fact in the world and can explain it to you really well, but who also has zero ambition and will just sit there staring at the wall in between questions.

thebaldcoder.bsky.social•83 days ago

That will change quite quickly: I want a bot/agent that sits in the background goes through my code base and creates push requests of bug fixes, refactoring, cleaning up code style etc. This surely is coming in 2025.

joedavis72.bsky.social•83 days ago

The same was said in 2023.. oh that’s coming in 2024.

theseneschal.bsky.social•83 days ago

I think we’re almost there with this. GitHub’s Dependabot scans my code for vulnerable libraries and opens a PR automatically. We can kick that up a notch without too much wizardry right now.

joedavis72.bsky.social•83 days ago

I hope so. Big Tech is spinning its wheels and spending billions while open source models are way better for privacy and the ability to run them locally.

softwareengineering.ca•84 days ago

Is o1 a truly different model? In using preview I felt like it was a layer of agents calling the 4o model to produce a better answer and validate the results. It didn’t “feel” smarter to me.

While great, all my tasks got stuck in infinite loops so I ended up just going back to the standard model.

mattheard.net•83 days ago

in the first livestream this week, it was mentioned that gpus were in the process of being switched over from o1-preview to o1, which to me implies a different configuration of weights than simply a different cpu-driven algorithm. i assume that also implies that o1 has different weights than 4o

gregthomaswashere.bsky.social•84 days ago

And I like that 4o “knows” me from the memory feature. I keep going to 4o1 (full/preview/mini) to ask “based on what you know about me” questions and it just comes back with “sorry, I don’t know anything about you, but in a general sense here’s your answer…”

softwareengineering.ca•84 days ago

( I turn off memory to help me write better prompts elsewhere, hah. )

But overall either I’m completely missing the power of o1 (and Ethan is totally right) or o1 is the productization of agent coordinators on top of regular LLMs (and am not asking it hard problems, and Ethan is right).

gregthomaswashere.bsky.social•84 days ago

I feel like there’s a very high-level conversation going on about how to use these models, but a convo on how to best use it for a layman like me is lacking. Sure, there’s beginner “how-tos,” but I feel like I’m somewhere in this middle wasteland.

alexanderking.bsky.social•84 days ago

Honestly, this is why I’m still subscribed to ChatGPT plus rather than Claude. It’s because the convenience factor of the different features of ChatGPT. It’s just that much more useful in my day to day life than the better intelligence of Claude.

emollick.bsky.social•84 days ago

I don’t think o1 is AGI in any conventional sense to be clear. But I think the fact that most people can’t find a use for it when it is quite good in some hard things is a good indicator.

paradaxiom.bsky.social•84 days ago

o1 can do integrals that even the best symbolic integration software can not handle

nickup.bsky.social•83 days ago

I throw Claude some sub problems of my PhD research: and while it can't solve any of them outright yet, it has pointed me in the right direction a couple of times or told be about some useful inequality.

sophiebits.com•84 days ago

where can I read more about this?

dmewes.com•84 days ago

That's an interesting observation! A lot of the tasks we'd most like to automate away are also more physical in nature (household chores etc). Intellectual tasks often require goal driven behavior with extensive tool use over long time horizons.

gregthomaswashere.bsky.social•84 days ago

I think “can’t find a use for it” is the wrong approach. I don’t NEED it to solve complex math problems or assist with coding, etc. because that’s not what I do. I think most people testing it (or even using it) are in a bubble of their own industries and then scoff at people who don’t “get it.”

mihoda.bsky.social•83 days ago

Guided learning is my main use case.

leriz10.bsky.social•84 days ago

The more I think about the more I think I am on the camp that o1 is already an early agi. I'm not a coder at all, and I've been able to code stuff I would of never done before. Sonnet and o1 are def. really really smart. Focusing on their shortcomings doesn't do them justice.

m9e.bsky.social•84 days ago

Leads to an interesting question of... is this a product/ui problem? (I think yes, and it fits in my general theory of AI is doing very well and the uses are under engineered at the moment)

catblanketflower.yuwakisa.com•84 days ago

Early on someone commented that most people just don’t have a lot of cognitive load in their daily lives

tachikoma.elsewhereunbound.com•84 days ago

what's defined as cognitive load? we take for granted the ability to do all the mundane tasks we accomplish everyday that are only just now practically solvable with AI, like driving, and even then with specialized models.

catblanketflower.yuwakisa.com•84 days ago

I took that to mean "informed decisions". Most people do some problem solving when things go awry, but a lot of jobs consist of highly recurrent or even repetitive tasks

How often do most people find themselves "stumped" in their job? And even then that's mainly referring it up the management chain

tdietterich.bsky.social•84 days ago

I agree. We are already accustomed to using systems more intelligent than us. Web search was already super-human. Calculators are already super-human. Google maps is already super-human. To most end users, LLMs and AI agents will look like more of the same.

tdietterich.bsky.social•84 days ago

Indeed, as many others have said, AI will be most successful (as a tool for people) when it is invisible and easy to use. Right now, I can't fully trust an AI system if it is based on statistical learning.

j2bryson.bsky.social•83 days ago

I strongly believe we should NOT want AI to be invisible. We can only maintain human centring, responsibility, and control if people are aware that they are working with AI, and have enough understanding of what that means to choose to deploy it accountably. 1/2

j2bryson.bsky.social•83 days ago

Note that without responsibility, there will be no improvement – responsibility is how we do gradient ascent as a society. And only humans can be held responsible – either as individuals or as collectives. But responsibility for artefacts is incoherent, which is why it's banned by UNESCO. 2/3

j2bryson.bsky.social•83 days ago

citations: unesco recommendation on AI (signed by all 193 members of the time https://www.unesco.org/en/articles/recommendation-ethics-artificial-intelligence )
Of, for, and by the people: the legal lacuna of synthetic persons
https://link.springer.com/article/10.1007/s10506-017-9214-9
#AIEthics #AILaw #digitalGovernance 3/3

tdietterich.bsky.social•83 days ago

I think we agree. AI could only be invisible if it was completely trustworthy (including morally). Until then, additional "automation" requires additional vigilance to the point where we will spend all of our time monitoring these AI systems rather than living our lives.

thebaldcoder.bsky.social•83 days ago

That's a daft last sentence: all our current evidence points to human intelligence being purely based on statistical learning: a load of neurons firing adapting to inputs and their outputs. You trust them every day.

marccrouch.com•81 days ago

Speak for yourself. I don’t trust my brain at all, especially when it does stupid shit like make me put my car keys in the fridge and then spend an hour looking for them next time.

shampshire.bsky.social•84 days ago

I don’t mean this to sound snarky, but how could a system possibly be described as AGI-ish if it’s not autonomous or self-directed?

thebaldcoder.bsky.social•83 days ago

Isnt it trivial to put this in a robot of any size and shape and have an energy meter, a charging station and then give the robot a task to do i.e hoover and if it detects that energy is low it decides to stop hoovering and go and charge and then continue afterwards?

thebaldcoder.bsky.social•83 days ago

We've got many, many more sensors (senses) but effectively its just scaling up that scenario in a way more complex environment but with Waymo and Roomba we're there - just needs o1.

Comments

Posting Rules

Reply