I believe that you are half right. AI is going to radically change the world, but not now, maybe in 10 years. The pace of research is slower than the expectations of investors. This will lead to the perception of a boom and bust and AI winter, but in reality there will be slow steady improvements. It’s just been commercialized too soon.

There is also the other (very high in my opinion) possibility that LLM will hit a wall in regards to performance and a new model is required. It's unfortunate that the current VC landscape encourages this single shot mentality instead of diversification. This will not help in the AI winter since most people will be invested in the same thing.

Wouldn't it stand to reason that LLMs should be obsolete in about a year due to the singularity event and the rapid speed which we're moving towards AGI and then onto exponential technology growth due to ASI?

At this point, as an AI researcher, I guess you'd just have to sit back and watch is all unfold, very soon everything you do is obsolete almost immediately.

This suggests that LLM is the path to take us to AGI. There is no proof for that. That seems to be the current bet, but my thinking (and it is strictly mine) that we are still technologically under-powered to achieve such a leap. Maybe in 10 years, or in one year or in a hundred of years. However, for that leap (and strictly my opinion), we need significant infrastructure leaps. Something like processors are x100 faster, or your laptop is powered by a 5Ghz 200 cores CPU...

As it stands, we can't even get the Mark Metaverse right. You are trying to convince me that we have the infra. for AGI? Not convinced.

Three observations here. Firstly it has been a really eye opening experience watching the innovation around Stable Diffusion and locally run LLMs and seeing that the unoptimized research code that needed such beefy hardware could actually be optimized to run on consumer hardware given sufficient motivation.

Secondly it wasn't obvious that deep learning was going to work as well as it did if you simply threw enough compute at it. Now that this tech has reached critical mass there is a tonne more money being poured into infra to support it.

Lastly, compute power is increasing as always. Nvidia releasing H100 and also their recent work on computational lithography. Also DeepMind finding new state-of-the-art algorithms for doing matrix multiplication with AlphaTensor. You can kinda already see the positive feedback loop in action.

I dunno... at this point I just wouldn't bet against the trajectory that we're on.

What actually is the trajectory we're on and what will we do once there?

So, it seems the trajectory is one of increasing generality and capability of models and increasing reliance on them.

If it's at all possible to improve our technology then we will. If we improve it it increases in utility. If it increases in utility we use it more.

What other thesis is there?

The model architecture has stayed roughly the same since the original AIAYN transformer in 2017. That’s 6 years of nothing fundamental happening.

Now, obviously the models have got hugely better in capabilities since BERT. Everything else has advanced. Tweaking, tuning and scaling have delivered true intelligence, albeit sub-human. But it seems unlikely that transformers are what take us to human-parity AGI and beyond, because the more we optimize these word predictors the more we find their limitations.

The lack of architecture changes over the last 6 years creates a huge amount of “potential energy”. A new model architecture might well push us over the human-parity threshold. It wouldn’t surprise me if I wake up one day to find that transformers are obsolete and Google has trained a human-parity AGI with a new arch.

This could happen tomorrow or in 20 years, transformers had an easy discovery path from RNNs, to RNNs with attention mechanisms, to Transformers. Architecture X seems to have a much more obscure discovery path.

It's certainly going to be very interesting what comes out of the training runs that are going to be done on giant clusters of H100s.

I also believe we might see something unfold but we still don't know.

It would be wrong though not to keep an very close eye on it or also to embrace it because if it will not just happen in the next 20 years you still need to earn money and with expertise in ml you might be better of

You couldn’t be further from the truth.

This is the single greatest leap in productivity we’ve had in the last 100 years.

Last week, I used GPT-4 to write my code. Later, I used it to analyze 100+ websites and come up with a personalized pitch for a relevant plugin/product idea - something that would have taken me 100+ hours.

How did you use it to analyse websites?

Did it go scrape stuff for you or did already have the data or did you paste in the website data?

It has plugins now. It can search the web. It's connected to Wolfram. Zapier. It can do things.

Even before that you could hack something together. I told it how to request an image be included in responses - it includes [fluffy unicorn], I parse that out, feed it to another GPT to get a better description, then feed that to DALL-E to get the image to include

How did you provide the info for those 100 websites to GPT?

Yes, but it hallucinates sometimes. /s

I mean, that’s exactly what it’s being used for — hallucination.

Can you elaborate on this? Or are you just being cheeky?

When you use it for brainstorming, hallucination is not that big of a deal - you need to verify ideas anyway. And it's great at "thinking" outside of the box, if you prompt it right.

If it makes you feel better (or worse?). I am using it daily and extensively. It is a productivity boost but not something that will propel me to god-level and the more I use it (Codex in particular) the more I discover its limitations; and that's where my conclusion has came from: It is too limited beyond the local scope it has. It is not clear (and there are no indications) it can make a jump from that local scope.

It seems then that it's usefulness depends on the intersection of what you want/need to do, what the model is capable of doing, and what you're capable of getting the model to do.

In some cases the impact might be enormous, and in others perhaps less so. One thing is for sure, the models are getting more capable, and along with that people are investing time/effort/money into improving their capability to leverage what the models can do.

Are you using GPT-3.5 or GPT-4?

Ironically you've proven you have no value add over GPT-4 by choosing to use it to complete your tasks.

That seems like it supports his point pretty well. If he can use it to basically replace most of his work, then it doesn't seem like a bubble at all

1000x this.

What it does is it makes you multidisciplinary. You might be a coder who struggles with writing social media posts and blog headlines. GPT fills that shortcoming.

It’s also a massive skill multiplier. It can turn someone with a 3-6 months knowledge of a discipline equivalent to someone with 2-3 years (or even more) in the field.

How did you make it analyse 100+ websites? Scraping html and pasting it to gpt, or some other way?

I'm doing similar things, and wondering how other people handle it.

All current indications says otherwise.

Web3 or crypto didn't disrupt markets.

Chatgpt or let's say ml already did and there is no writing on the wall it will stop. Contrary it shows how much potential there is and how excited a lot of people are.

We have already ml now in office, in bing, in Google workspace. There is midjourney, SD etc.

You can find ml generated porn pictures.

We have constantly news about advantages in multiple ml fields.

This is so different to crypto and stuff.