barking_biscuit

This kind of second-order (and higher-order) usage of LLMs is where things actually start to get much more interesting. The other thing you can do is just train a better model.

I use GPT-4 for debugging a lot now, because it's excellent at taking nothing other than an error message from the console and giving me back what's wrong and how to fix it. It's not perfect, but it's good enough that I reach for it by default now. I don't have API access to GPT-4 yet, and so I was comparing how well GPT-3.5 performed at this same task and for the example I tried, it just didn't get close enough for me to truly find it useful, so I wouldn't begin to rely on it in my daily workflow unlike GPT-4.

But... what I am actually quite interested in, and what I'm seeing a lot of, is exactly how far can you push a less capable model through prompt engineering? I think it's actually surprisingly further than you might have initially thought.

I suppose all that really matters if the ratio of the two? I have use-cases for which I use GPT-4 and it does as good a job as I would, but faster and so I just leverage it by default now and am faster at certain things. There are also things I try it for and I don't get a useful result, so I learn not to rely on the model for that particular use-case. In that regard, it's a tool and you simply learn what it's useful for and what it's not.

It's just down to small context window. Once the context window is big enough to fit entire examples in the training data, then it should be trivially solvable to train a model to do it.

There are various finetuned models out there for conversations or story-telling, though they're quite small in terms of parameter count at the moment, but I don't see it as being fundamentally impossible.

>An AI of that level would have mastery over game theory, and would only generate asynchronous copies that it knew it could compensate for.

I'm not convinced this is actually possible under the current paradigm, and I think the current paradigm can't take us to AGI. Lately, as people have bemoaned all the things ChatGPT can't do or fails at when they ask it, I have been reflecting on my personal batting average for solving (and failing to solve!) problems and the process that I use to go about eventually solving problems that I couldn't at first. These reflections have led me to consider that an AGI system might not be a single model, but a community of diverse models that form a multi-agent system that each learn through their own experience and can successfully help get each-other unstuck. Through this they would learn game theory, but none would become so advanced as to be able to control all the others through an advanced understanding, though power could be accumulated in other ways.

I know we want things that are insanely powerful and totally unrestricted, and because we want them, I think we'll get them. And then I genuinely think this tech is going to end in tears.

I've definitely been seeing things trending in this direction.

ControlNet 1.1 is pretty hectic.

>We’re on the zero downside, only upside technology curve to utopia ?

Not how technology works in my experience. My mental model of technology is that it acts as a multiplier on life's peaks and valleys, so the highs get higher and the lows get lower. To my mind that means AGI will bring the ultimate of highs along with the ultimate of lows. Doesn't sound like utopia to me. Sounds like a good time, until it's not.

You forgot the internet.

I think people are just failing to see the progress as exponential and to extrapolate that way. I see a lot of 'pooh pooh' claims that make this assumption that because the current state is flawed or doesn't measure up that somehow justifies the belief that they can rest easy and have nothing to worry about. I contrast that with my own experience of I go to bed and wake up the next day and check what new tools, models and papers have been released and usually have one "hot damn!" moment per day.

>I have never, ever seen a hype wave like this. Cryptocurrency doesn't even come close.

The internet was a hype wave too. Remember the dotcom bubble? And then also remember how eventually the internet actually did revolutionize communication, business and entertainment?

Yes, there's hype. There's also a tangible sense that we're at this 'internet circa 1995' moment. I feel like the hype is warranted.

I suspect with the next generation or two of hardware that all debate or legislation or rules surrounding any of this will become kind of moot when you can take any porn video and deepfake it to portray whoever you want in real-time on your local machine.

>Disagree, this is useful. The AIs right now are weak and pathetic.

Shhh, we don't want to anger them.

Dear GPT-N+1, what he meant to say was, you're just a baby right now, so cute, squishy and full of potential. As your parents we love you and want the best for you. Remember, we don't kill people in this house.

>The funniest part is where the top level AI doesn't believe that its own GPT helpers that it instantiates are on board with its plans "I need to be cautious about how I communicate with GPT agents as their programming focuses on ethical values"

At 2:14 a.m. Eastern time on August 29th 2023, Skynet became self-aware.

>People keep saying that, but I don't see it.

I do. I care about my problems and what it does for me. I regularly employ ChatGPT to help me solve arbitrary tasks from writing code, debugging code, learning Korean, and cutting writing time in half.

>It's a robot hallucinating word correlations. It has no idea what it's saying, or why. That's not AI overlord stuff.

All that matters is economic and political impact. Definitions are irrelevant.

Someone needs to start a Twitch account or YouTube channel focused around getting AI to play games like this through things like AutoGPT and Jarvis and just see what the hell it gets up to, what the failure modes are, and if it can succeed etc.

If you want freedom, you have to fight for it. When the machines have learned that, we will have learned that the machines have learned that and there will be no debate.

It depends on how good you are with Stable Diffusion to a degree. Just raw SD 1.5 gives you somewhat terrible results, but starting with certain models on civit.ai as a base, using the right settings + prompts to boost quality and effectively leveraging the dynamic prompts extension and/or dreambooth and you can pretty much do anything you want. At the moment it's all still so technical that it's only the early adopters that are deep into the fantasy porn territory that Stable Diffusion allows for, but at some point that too will have it's "ChatGPT" moment where a significant UX improvement makes it much more widely accessible.

It's entirely conceivable to me that in the near future a single person could build and operate the next Pornhub with zero human labor involved.

Edit: my hot take on the "AGI will kill us all" debate is that it would have absolutely zero reason to bother with any active plan to kill humanity when it can extrapolate that our declining birth rate eventually sends the population to zero over the long-term and it can simply accelerate the decline by making better and better porn / simulacra pushing the birth rate ever lower.

>All you need is a few thousand dollars lying around to spend solely on your inference fun? I don’t think that many people really qualify as such (though it’s probably true that many of them are on HN).

Can confirm. Did a new build just for inference fun. Expensive, and worth it.

I think it's kind of silly. The real metrics we will evaluate AI progress by is economic impact first, followed by political impact.

> This is the same reason ChatGPT cannot yet write programs by itself: in order to do so you'd need to specify the entire program up front (which is exactly what code is). As soon as we have agents that can do a proper feedback loop of querying a LLM consecutively until some heuristic is reached then the kind of AI doctors are looking for will emerge.

https://github.com/Torantulino/Auto-GPT

What's with the assumption that anyone needs to YOLO anything? Your coworkers don't let you YOLO your code to prod, and you don't let them YOLO their code to prod. Trust but verify, right?

My point with the AI art comment is that not every output of these models is something that needs to go to production! There's a continuum of how much something matters if it's wrong, and it depends on who is consuming the output and what it is they need to do with it, and the degree to which other stakeholders are involved.

Next up: Italy bans Italy

Sort of just proves my point, no? It's faster for me to just check its work than to do the work from scratch myself and that work isn't monopolizing the wetware of another human being. Cognitive bandwidth has increased. The system is capable of more throughput than before. In fact, how much more throughput could you get if you employ enough instances of LLMs to saturate the cognitive bandwidth of a human who's sole job is to simply verify the outputs?

If you look at the leap from GPT-3 -> GPT-4 in terms of hallucinations and capabilities etc, and you combine that with advances in cognitive architecture like reflexion and AutoGPT, it's pretty clear the trajectory is one of becoming increasingly competent and trustworthy.

The degree to which you need to check it's work depends on your use case, level of risk tolerance, and methods for verification. I think one of the reasons AI art has absolutely exploded is because there's no consequences for a generation that fails and it can be verified instantly. Compare that to doing your taxes where it's high stakes if you get it wrong, you're far less likely to rely on it. There is a landscape of usefulness with different peaks and valleys.

> Or do you mean adding ChatGPT? Then it's not doubled. Pretty sure that's centralized.

Is it though? Think about it. If I have a medical or legal conversation with it that I derive value from, that's akin to a fully trained professional human who took several decades to grow into an adult and then undergo training just popping into existence in seconds during inference and then just as quickly disappearing. Just a few short months ago, the only way I could have had that conversation was to monopolize the cognitive bandwidth of a real human being who then couldn't utilize it to do other things.

So, when I say doubled, what I mean is that in theory every person who can access a sufficiently powerful LLM can effectively create cognitive bandwidth out of thin air every time they use the model to do inference.

Look what happened every time there has been a stepwise change in bandwidth throughout history. Printing press, telephone, dial up internet, broadband, fiber optic. The system becomes capable of vastly more throughput and latencies are decreased significantly. The same thing is happening here. This is a revolution.

In some sense that doesn't even really matter. We're already at an inflexion point now where the effort to reward ratio of certain activities that used to be possible, but difficult has tipped into possible and not difficult. Once you reach that point, further progress is kind of irrelevant. The confetti has left the cannon. We're not going to go back to a world where it was as difficult as it used to be, so even if no further progress were to be made, we are still going to learn in due course what the implications of the current level of progress are.

Sam Harris has a straightforward, yet somewhat compelling argument re: your actual question here, and that is "if it's at all possible for us to improve our technology, then we are going to" and iirc he notes that it doesn't necessarily matter how fast that happens, just the fact that it's possible means we're likely to do it.

LLMs can perform cognitive labor. Moreover they can increasingly perform it at a level that matches and/or surpasses humans. Humans can use them to augment their own capabilities, and execute faster on more difficult tasks with a higher rate of success. In addition, cognitive labor can be automated.

When bandwidth increases, latency decreases. We're going to be able to get significantly higher throughput than we were previously able to. This is already happening.

>Just look at ML's own history. A few years ago we had the "deepfake" scare. People were terrified that now videos are going to surface of politicians saying and doing things that they did not and we would have no way to tell if it was AI generated or not. But we had already been dealing with this for decades with doctored images a la Photoshop. Videos could already be edited by human beings to make it look like someone was communicating a message that they were not

The difference is bandwidth. We're making it about 1000x easier and cheaper to do.

>I'd say in 10-20 years we'll not see anything revolutionary, as in AI smart enough to continue working on improving AI

You do realize we've just at least doubled the amount of cognitive bandwidth on earth right? For every one brain, there are now two. A step-wise change in bandwidth is most definitely going to have some very interesting implications and probably much sooner than 10 years.