Hacker News Reader

Comment by mach1ne | Hacker News Reader

I wonder if we may have crossed the threshold into singularity. LLMs will almost inevitably be on pretty much everything in 5 years. Looking at how well current GPTs are able to write calls to external APIs, their integration to systems will be surprisingly deep. No apparent end in sight to their advancement.

It may well be that this is an unexpected vector from which to enter the singularity.

View on HN

> Looking at how well current GPTs are able to write calls to external APIs

Are you talking about the behavior of ChatGPT constantly inventing API parameters that would be kind of useful to solve the task at hand, but which unfortunately aren't part of the external API?

I would say that understanding that there's a difference between your own code, where you have full freedom to design your contracts, and "code of others" where you have to adhere to some already-defined contract, would be an absolute minimum bar that must be passed in order to call someone (or something) proficient at writing calls to external APIs.

davinci has been pretty good about using APIs that I've shown it, I can certainly see it hallucinating parameters but I also misremember interfaces. It seems that you could easily solve that a number of ways e.g. analysis, tests, submitting the code along with the spec and a prompt that asks for bug checking with the temperature set to 0.

Just because you can't one shot a perfect completion every time with 0 engineering investment doesn't mean you can't use this technology to build some cool things.

The point of the singularity is that at that point it's not us making the further technological advances, it's AIs, so the rate and type of advance is both opaque to us, and beyond our control. That's why there's a horizon we can't see beyond. An AI helping us write chunks of code still has us in the loop.

No it's not: the point of the singularity is that you can't meaningfully speculate past it.

We're not there yet, but it's reasonable to think we might've crossed an event horizon.

I have no clue where LLMs will lead in 6 months, so we're pretty much there already.

Exactly, we're seeing the time horizon shrink rapidly. There have already been numerous threads ruminating on the fact that even a few months of not paying attention will result in you lagging behind in this technology. I can't wait to see what happens when state space models are developed further[0]. If the prompt and output length can be increased significantly then large chunks of the work of building software could be offloaded to these tools.

[0] https://arxiv.org/abs/2212.14052

Plenty of people have not known where technology would take us in six months, for the last 200 years or more. AI researchers have been predicting systems like ChatGPT for many years now, and even successfully predicting their failure modes and pathologies.

You really don't see a difference between today and 200 years ago?

>AI researchers have been predicting systems like ChatGPT for many years now, and even successfully predicting their failure modes and pathologies.

Any particular examples you are thinking of here that you could point to? I’d be interested to read.

The reason it's a singularity is not because we can't predict past it: it's because it represents a point of rapidly accelerating growth in capability of AI, enabled by the fact that the AI itself is either modifying its own code, or creating the next generation of AI, which will be smarter, and able to create a smarter next generation, etc.

Humans not being able to accurately predict what comes next technologically isn't the singularity; it's just a basic truism, and has been for decades.

Yes, I said as much, but I also said why can't we meaningfully speculate past it.

Debatable whether crossing the event horizon is distinguishable from beeing there

I know it's popular to mention singularity in the context of AI, but i am afraid it doesn't do justice to the "real" singularity which is a mystery of both scientifical and mystical proportions.

Just wait until efforts to improve the accuracy and usefulness of LLMs by implementing conceptual understanding in the models rather than pure language construction, maybe including physical interactions with the world as part of the training process, trigger a legitimate debate as to whether these things are conscious. Now that will make justice to the real singularity.

"trigger a legitimate debate as to whether these things are conscious"

That debate has been raging at least since The Mechanical Turk, and even now philosophers can't even agree on what consciousness is or if humans themselves are really conscious or some kind of automata.

Chatgpt simply brought such speculation to the masses, but it's been raging in academia and science fiction for quite some time.

The same debate has been going on since ELIZA was first introduced. The difference between arguing about a text generating model and one that inhabits a virtual world or can interact with the physical world is huge.

How is this unexpected? AI agents that interact via language and are able to write code which improves itself is classic singularity. It basically describes HAL from 2001.

> AI agents that interact via language

Via English. I'll be inclined to believe we have general AI when it teaches itself language that is not widely available on the internet.

The goal posts are always moving.

How would anyone teach themselves a language without having access to it?

There was no singularity in 2001. Or in the entire series.

2001 is prior to the concept of the singularity being well articulated, but the HAL artificial intelligence clearly fed ideas for what an advanced AI could do for future fiction.

Maybe the progenitors had a singularity. It's off-camera and not even imagined in the story but after the fact we can see that it would at least not be inconsistent. But almost anything can be said not to conflict with almost anything off-camera.

All I think 2001 depicts is a transcendence. A level up.

And actually that was all Dave being levelled up. I don't remember Hal being all that important at all. Some transcended elders or their left-behind machines picked up Dave and gave him the VGER treatment (give god powers to the insect) and afterwards, we mere humans can still follow his train of thought as he comes back home.

To me a singularity is something you can't even imagine. If you can make any sort of sense of something at all, then it's still contemporary and not a singularity. It's just a faster car than last years. AI that becomes smarter than us and take over is not a singularity. That's just the Europeans wiping out the native Americans.

Compute power could probably produce a singularity. What happens when Jupiter is converted into solid compute? So it's merely adjacent. But a smart ai is not the definition of a singularity.

Reply on news.ycombinator.com