What would convince you to take AI seriously?

Significantly cheaper labor. If it's displacing real thought work then the results should be self-evident.

Do you take it seriously? Why are you asking

My thought is that some people don't want to adapt, but will be forced to adapt when AI is used everywhere. I'm not a fan of AI either, but it does have its uses and if you don't use it now, you'll have to learn later on.

When OpenAI claimed to have solved Dota, it did not do so by outwitting humans. It had superhuman reaction time (which directly caused them to win one game against OG they wouldn't have otherwise), and they drastically restricted the hero pool, completely transforming the game by restricting most of the strategic counterplay.

When DeepMind claimed to have solved StarCraft II, it did not do so by outwitting humans. Instead, it took advantage of its inhuman control to eke wins out of losing strategies. Another pro, Harstem, beat it and commented on his game (https://www.youtube.com/watch?v=3F9VKAArwPE):

> "It feels more like [he's] praying for mistakes of his opponent rather than actually playing properly or playing well [...] I hope in a couple of months he will be a lot stronger and can finally show some strategic innovation"

Generative art and text still fall short of skilled human work, so much so that it's often possible to tell when someone's used AI. The details are good, and can fool non-artists, but the overall composition is always nonsensical. If you study art, you will see that artists often structure around an idea, like "make this person the focal point, and subtly direct the viewer's eye towards it". None of that intentionality is present in AI work.

OpenAI's AtCoder bot this weekend seems to continue this trend (https://www.tomshardware.com/tech-industry/artificial-intell...):

> The AI model, branded OpenAIAHC, was widely expected to dominate the market. Still, Dębiak’s innovative, heuristic-driven approach -- using problem-solving shortcuts and educated guesses instead of brute-force calculation—secured the win. Contest administrator Yoichi Iwata praised his distinctive method, noting that while the AI excelled at raw optimization, it “fell short of human creativity.

For me, I would want to see some decision that makes me go, "Ah, that's really clever. It feels quite human. I would have never thought of that." -- but of course it's difficult to codify that rigorously. Still, I have seen absolutely zero progress towards this sort of creativity. That's not to say that AI itself doesn't excite me; it's very cool to see how far these models have come in such a short time.

In my opinion, that's a silly question.

Why do I even need to make up my mind about it?

It isn’t about how well AI can answer solved problems. Can it invent the future?

And let’s say AI does automate all human labor… what’s the plan? That happening, will lead to chaos, without some massive changes in how society is organized and functions. That change itself with be chaotic and massively disruptive and painful for millions. If someone can’t answer that question, they have no business hyping up the end of human involvement in civilization.

Solving benchmarks isn't intelligence and doesn't mean anything. It would have to display intelligence, not solve a constrained problem that someone made up.

More fundamentally I'd look for an architecture that could actually support intelligence and not LLMs like we currently have.

Convince me what? That it’s ready to replace thought work?

The moment it can do that is the moment we each singularity.

We are not there yet. Everyone will know when we hit that point.

For the now it’s a great tool for helping with thought work and that’s about it.

Have it admit it doesn’t know instead of sounding like a Reddit thread full of “experts” trying to one up each other

When you ask it to do you something and it tells you to fuck off and do it yourself.

To start with, AI should first exist.

For development work: something like DHH’s “build a blog in 15 minutes” demo.

For original work: solving some well known but unsolvable problem, like the Collatz conjecture.

Four things: do meaningful novel work, learn over the course of a long conversation without poisoning its context, handle lies and accidental untruths, and generally be able to onboard itself and begin doing useful work for a business given the same resources as a new hire.

This isn't an exhaustive list, it's just intended to illustrate when I'd start seriously thinking AGI was possible.

I take AI seriously in the sense that it's useful, can solve problems, and represents a lot of value for a lot of businesses, but not in the sense that I think the current methods are headed for AGI without further breakthroughs. I'm also not an expert.

I take it seriously, but I'd take it more seriously if I could use it to solve real problems without encountering issues, or see others use it solve real problems without issues.

There are many good tasks improved by AI, e.g. generic writing, brainstorming, research and simple tasks. However, I keep finding it struggle with more complicated or open-ended problems, making obvious mistakes. If it stops making obvious mistakes, or even if we simply discover automated ways to correct them, I'd take it more seriously.

AI is also not creative IMO: I find AI-generated art noticeably lower-quality than real art, even though it looks better on the surface, because it doesn't have as much semantic detail. If I find an AI-generated art that would be almost as good as the human-talent equivalent, or an especially impressive AI-generated art that would be very hard to generate without AI, I'd also take it more seriously.

I use LLMs daily as a software engineer and they save me dozens of hours a week and I can’t imagine going back to a time without them.

But call me when you can load all human knowledge circa 1905 and have them spit out the Theory of Relativity.

And even then I might shift my goalposts.

I will take AI seriously when the data used for training is gathered with consent from its authors.

Other people taking it seriously, honestly. It's hard to take it seriously when everyone is thinking it will dethrone God or put everyone out of a job and if you're not using it you are going back to the Stone Age while it embarrasses them repeatedly.

I would start worrying if AI models can understand, reason, learn and incorporate new information on-the-fly without retraining or just stuffing information in context windows, RAG, etc. The worry would also depend on the economics of the entire model lifecycle, as well as the current state of mechanical automation.

We aren't getting that with next-token generators. I don't think we'll get there by throwing shit at the wall and seeing what sticks, either, I think we'll need a deeper understanding of the mind and what intelligence actually is before we can implement it on our own, virtually or otherwise.

Similarly, we're pretty good at creating purpose-built machines, but when it comes to general/universal purpose, it's still in its infancy. The hand is still the most useful universal tool we have. It's hard to compete with the human mind + body when it comes to adapting to and manipulating the environment with purpose. There's quite literally billions of them, they create themselves and their labor is really cheap, too.

There's my serious answer.

Write something fictional that doesn't suck and doesn't sound like it was written by a computer.

Idk though, I'm not sure you could ever convince me to agree that anything can replace all human labor and cognitive output

Once it stops making obvious mistakes on a regular basis, then I'll take it seriously.

That's like asking what I want to hear in a song.

You just know when it's good.

When it starts curing disease for real, like one or two sessions: done.

Because then I know: they have been using it for real human benefit without trying to get humans hooked on re-occurring costs.

When it starts solving actual human problems like climate or start filling in our gaps of knowledge in science. When it starts lifting humans up to higher grounds instead of replacing them to make a buck.

It is not a big deal because OpenAI has been known to cheat on LLM benchmarks before and I have no reason to believe that the AI actually solved the problems by itself without training on the solutions. I’ll be more impressed if a similar performance is obtained by an open-source model that can be independently verified.