Hacker News Reader

Comment by md_ | Hacker News Reader

Seems like a long-winded way of saying, "no". ;)

It is telling though, that a year ago myself and I think most programmers would have just quickly said "nope definitely not", whereas now you kind of have to get long winded to say no, and many of us are not saying no but maybe

AI has progressed in fits and starts. We could be one more breakthrough away from AGI, maybe many breakthroughs are needed. Nobody knows for sure, or how quickly those will come. I think things could get wild really quicky, given how effective a really primitive "Guess the next word" neural network is by just throwing a ton of data at it.

What will we get with more nuanced ideas and multiple networks wired together with feedback loops and such?

Gonna be fun.

> many of us are not saying no but maybe

I'm not sure if this is a controversial stance, but not all opinions are equal. HN commenters span a wide range of competencies, but very few of them have been in the trenches with ML (not just using, but designing and building the actual architectures, working on theory, or writing cogent philosophy about it) long enough to have good opinions about it worth listening to.

Gene Wilder meme:

Tell me all about how your opinions are better than everyone else's!

You will note I did not proffer any of my own. To avoid Dunning-Kruger, you must know what things you don't have sufficient knowledge about and avoid speaking as if you do.

Also please don't "meme" on HN, it degrades discussion quality very quickly. There are other venues for that.

oh, I beg your pardon. You did say "very few of them have been in the trenches with ML (not just using, but designing and building the actual architectures, working on theory, or writing cogent philosophy about it) long enough to have good opinions about it worth listening to."

I thought you meant that your opinions are worth listening to, unlike the other HN'ers. Are they not?

That's pretty irrelevant within the context of this comment. I said what I said; I didn't say anything else.

I have thoughts about this and related topics in my comment history, but still focus pretty tightly on saying things I can reasonably back up, avoiding speculation about things where my understanding is murkier. So in that wider sense, yes, I am living my values.

I think it's pretty surprising, and counterintuitive, how powerful generative language models are. It turns out that a lot of intellectual tasks are sort of weakly simulatable by just doing a great job at next-token prediction.

I don't think that means that "good next-token prediction" is identical to "AGI". As Yan LeCun says, I think LLMs are an off-ramp on the highway to AGI. But they are powerful at specific tasks.

Unfortunately, instead of evaluating their actual utility at a specific task, people just seem to think "if we throw AI at the problem, we will solve it," which is sort of like every other tech bubble ever.

  It turns out that a lot of intellectual tasks are sort of weakly simulatable by just doing a great job at next-token prediction.

Yes, this is a good way of putting it. I've been saying for years, it's less that we're making big discoveries about what "AI" can do, and more that we're showing that many things humans do that appear complex actually reduce to something pretty simple. But that simple thing is still just fitting a pattern. It's the cases where it doesn't work, even if it only fails 1% of the time, that define the difference between pattern matching and actual human intelligence.

One thing that follows from that (that people don't like) is that we actually need to move goalposts about how intelligence is defined. "Pass a turing test" is not very valuable now. And as mode tasks are shown to be possible with pattern matching / next token prediction, we need to further refine tests away from these tasks to settle on a good definition of what separates human intelligence. It should be obvious that the distinction is there, but it's still tough to nail down. (I'd argue that by defining a "task" you've already done most of the work to solving it, so it's not to exciting to learn that AI can finish the job)

I am not convinced actual human intelligence is all that different. Not all people do the kind of reasoning we talk about when we try to distinguish between GPT and humans, and even those of us humans that CAN do it often don't bother.

We may find that augmenting something as simple or nearly as simple as ChatGPT with a few extra tools to handle step by step reasoning (see the Wolfram plugin) may get you there.

I see these takes a lot, but I struggle to take them seriously.

As a trivial example: LLMs don't learn new skills on the fly. A key aspect of human behavior--implicit memory--simply doesn't exist for these models.

That seems like a pretty huge gap! Like, if ChatGPT is mediocre at your job today, it's going to be mediocre at it tomorrow and the next day, until a new model is trained!

With current computing power limits training a new model takes a long time. Seems like months at this point.

But lets say Nvidia pulls a rabbit out of its magic hat, and comes out with a training chip that is a million times more powerful. Now instead of months training that drops down to 24 hours. Would you still say that is a pretty huge gap? And I ask this because this is Nvidia's goal within a decade.

Where, for you, does a big gap turn into 'that is dangerously close'?

I don't think the difference between human memory and retraining a DNN is purely about the latency to retraining. They seem to be fundamentally different things, no?

We would both agree that birds and planes are fundamentally different things, but we would both agree that they fly, correct?

Both humans and LLMs have 'short term' and 'long term' memory. The process of turning short term to long term is significantly different. In LLMs this is taking the history data given to it and putting it in the training corpus then recalculating the weights. In humans this involves falling to sleep for some number of hours.

I think in the long term here AI will actually have the benefit of forming long term memories. You have to sleep to form them, it just has to run reweighting on another cluster while the primary cluster goes about its business.

Well, my point was that explicit memory and "skill development" (as one kind of what is known as "implicit memory") are fairly different. No matter how many times you retrain GPT4, it isn't going to learn to drive a car.

There are a few problems with your argument here....

You're using just text of GPT-4 and have not used any of the multi-modal capabilities. Trying to assess the total capability of a system when exposed to only one mode of a system is difficult if not impossible. I would assume that total compute capacity and cost of running are the biggest inhibitors of this being pushed out to more (any?) users.

Now lets imagine GPT-4 with image input, and command output. Tell it that it has an accelerator and a brake. Now feed it an image of a wide open road with no obstructions. My assumption would be that it would reason that it can accelerate. If you fed it a second image of a wide open road, would it assume that it can maintain its speed? Then a third image that shows a child in the road some distance in front of it. The system that already exists already has the reasoning capability that it should stop or it would hit a child.

Now you have the problem of "Is this learning to drive a car?". Via just image and text alone it's never going to get very good as we need some more sensory and environmental data passed to the system. But here's the thing about LLMs... Just about any structured information can be considered a language. Passing simulated driving data to an LLM could very likely teach it how to drive.

I think your intuition that gpt4 could function as you described is frankly highly unrealistic.

I'm guessing you're ignoring the evidence that's been presented by different groups chaining tools together already.

Reply on news.ycombinator.com