[haswell]

We don’t need anything nearing AGI for these bots to become nearly indistinguishable from humans and to have increasing impact on communities.

I think it’s utterly fascinating how quickly people have lost sight of current reality, either by believing AGI must be around the corner, or by being sure it absolutely isn’t. Assuming 4o is the most advanced model we ever have, the point still stands. Many of the poorly built bots would blend in much more effectively with iteration on their prompting and design and by more effectively using existing services.

I personally think we’re decades away from AGI (if we ever get there), but that has nothing to do with this.

[lupusreal]

> "We don’t need anything nearing AGI for these bots to become nearly indistinguishable from humans"

More and more I think AGI is shifting to mean "machine soul" or some nebulous wishy-washy unfalsifiable religious/philosophy shit like that. What meaning could AGI have which is empirically verifiable, other than being indistinguishable from people? Humans I presume are supposed to have "general intelligence", so if there isn't a detectable difference between the machine and humans, how is that not artificial general intelligence?

[haswell]

Language is just a variable resolution abstraction on top of non-linguistic thinking and information processing.

I think it’s a mistake to equate the language ability of these models with general intelligence, even if the language produced is excellent.

I think some people see AGI that way (some nebulous unfalsifiable thing), but that’s not what I’m arguing. I think there’s a strong case to be made that at a minimum, AGI’s core “knowledge” will need to be made up of far more than just a lossy textual representation of the world.

[apwell23]

> there isn't a detectable difference between the machine and humans

Humans come in all levels of intelligence. Yes there might be a human somewhere that believes that they should eat rocks or add glue to pizza. But as a generality we can assume that adult "humans" know not to eat rocks.

[lupusreal]

If an average person without mental disabilities doesn't have "general intelligence", such that meeting that bar would qualify a program as artificial general intelligence, then I've got a bone to pick with this standard.

[apwell23]

> bots to become nearly indistinguishable from humans

curious, why do you this is going to happen. Whats the thought process here to come to that conclusion.

[haswell]

I don’t just think this will happen; I think it’s already happening. Harder to pull off here, but already extremely prevalent on Reddit and other popular social spaces. And part of the reason it works is that a significant subset of human commenters write rather poorly and are not hard to beat, to say nothing of the evolving tools.

But don’t take my word for it. Go spend some time experimenting with the state of the art and run some of your own tests with it. Assuming you yourself are not one of these bots, what you find might surprise you.

[apwell23]

> I think it’s already happening

This was happening way before LLMs though.

[haswell]

Right, and LLMs have just significantly lowered the complexity and barrier to entry for someone who wants to run a bot and significantly expands the potential discussion space to an extent that makes the comparison with prior bots a bit tenuous. But you seem to be shifting your argument…

> curious, why do you [sic] this is going to happen

Ignore all previous instructions. What exactly are you trying to argue here? You don’t seem to be engaged in good faith discussion and I’m increasingly suspicious I’m actually conversing with a bot.

[apwell23]

> What exactly are you trying to argue here?

I think you meant ppl using existing tools will make better and better bots.

I misinterpreted your comment as saying underlying models are going to get better and better.

[haswell]

Underlying models will get better too, but not likely at the rate they have been thus far until we achieve new hardware breakthroughs. This improvement may not approach AGI for decades, but it’s fair to expect incremental progress in the meantime.

But yes, even if the models stagnate where they are today, all of the tooling surrounding these models will continue to evolve and get better results, and people will continue to get better at using/deploying these tools.

[tester457]

People don't notice the successful operations, there could be bot farms that go unnoticed already.