You don't have to extrapolate. There's a frenzy of talent being applied to this problem, it's drawing more brainpower the more progress that is made. Young people see this as one of the most interesting, prestigious, and best-paying fields to work in. A lot of these researchers are really talented, and are doing more than just scaling up. They're pushing at the frontiers in every direction, and finding methods that work. The progress is broadening; it's not just LLMs, it's diffusion models, it's SLAM, it's computer vision, it's inverse problems, it's locomotion. The tooling is constantly improving and being shared, lowering the barrier to entry. And classic "hard problems" are yielding in the process. It's getting hard to even find hard problems any more.
I'm not saying this as someone cheering this on; I'm alarmed by it. But I can't pretend that it's running out of steam. It's possible it will run out of money, but even if so, only for a while.
The AI bubble is already starting to burst. They Sam Altmans' of the world over-sold their product and over-played their hand by suggesting AGI is coming. It's not. What they have is far, far, far from AGI. "AI" is not going to be as important as you think it is in the near future, it's just the current tech-buzz and there will be something else that takes its place, just like when "web 2.0" was the new hotness.
It's gonna be massive because companies love to replace humans at any opportunity and they don't care at all about quality in a lot of places.
For example, why hire any call center workers? They already outsourced the jobs to the lowest bidder and their customers absolutely hate it. Fire those people and get some AI in there so it can provide shitty service for even cheaper.
In other words, it will just make things a bit worse for everyone but those at the very top. usual shit.
This getting too abstract. The core issue of LLMs that others have pointed out is the lack of accuracy; Which is how they are supposed to work because they should be paired with a knowledge representation system in a proper chatbot system.
We've been trying to build a knowledge representation system powerful enough to capture the world for decades, but this is something that goes more into the foundations of mathematics and philosophy that it has to do with the majority of engineering research. You need a literal genius to figure that out. The majority of those "talented" people and funding aren't doing that.
> There's a frenzy of talent being applied to this problem, it's drawing more brainpower the more progress that is made. Young people see this as one of the most interesting, prestigious, and best-paying fields to work in. A lot of these researchers are really talented, and are doing more than just scaling up. They're pushing at the frontiers in every direction, and finding methods that work.
You could have seen this exact kind of thing written 5 years ago in a thread about blockchains.
Yes, but I didn't write that about blockchain five years ago. Blockchains are the exact opposite of AI in that the technology worked fine from the start and did exactly what it said on the tin, but the demand for that turned out to be very limited outside of money laundering. There's no doubt about the market potential for AI; it's virtually the entire market for mental labor. The only question is whether the tech can actually do it. So in that sense, the fact that these researchers are finding methods that work matters much more for AI than for blockchain.
Really, cause I remember an endless stream of people pointing out problems with blockchain and crypto and being constantly assured that it was being worked on and would be solved and crypto is inevitable.
For example, transaction costs/latency/throughput.
I realize the conversation is about blockchain, but I say my point still stands.
With blockchain the main problem was always "why do I need this?" and that's why it died without being the world changing zero trust amazing technology we were promised and constantly told we need.
With LLMs the problem is they don't actually know anything.
Amount of effort applied to a problem does not equal guarantee of problem being solved. If a frenzy of talent was applied to breaking the speed of light barrier it would still never get broken.
I mean, a frenzy of talent was applied to breaking the sound barrier, and it broke, within a very short time. A frenzy of talent was applied to landing on the moon and that happened too, relatively quickly. Supersonic travel also happens to be physically possible under the laws of our universe. We know with confidence that human-level intelligence is also physically possible within the laws of our universe, and we can even estimate some reasonable upper bounds on the hardware requirements that implement it.
So in that sense, if we're playing reference class tennis, this looks a lot more like a project to break the sound barrier than a project to break the light barrier. Is there a stronger case you can make that these people, who are demonstrating quite tangible progress every month (if you follow the literature rather than just product launches), are working on a hopelessly unsolvable problem?
I do think the Digital realm, where the cost of failure and iteration is quite low, will proceed rapidly. We can brute force with a lot of compute to success, and the cost of each failed attempt is low. Most of these models are just large brute force probabilistic models in any event - efficient AI has not yet been achieved but maybe that doesn't matter.
Not sure if that same pace applies to the physical realm where costs are high (resources, energy, pollution, etc), and the risk of getting it wrong could mean a lot of negative consequences. e.g. I'm handling construction materials, and the robot trips on a barely noticeable rock leaking paint, petrol, etc onto the ground costing more than just the initial cost of materials but cleanup as well.
This creates a potential future outcome (if I can be so bold as to extrapolate with the dangers that has) that this "frenzy of talent" as you put it will innovate themselves out of a job with some may cash out in the short term closing the gate behind them. What's left is ironically the people that can sell, convince, manipulate and work in the physical world at least for the short and medium term. AI can't fix the scarcity of the physical that easily (e.g. land, nutrients, etc). Those people who still command scarcity will get the main rewards of AI in our capital system as value/economic surplus moves to the resources that are scarce and have advantage via relative price adjustments.
Typically people had three different strengths - physical (strength and dexterity), emotional IQ, and intelligence/problem solving. The new world of AI at least in the medium term (10-20 years) will tilt the value away from the latter into the former (physical) - IMO a reversal of the last century of change. May make more sense to get good at gym class and get a trade rather than study math in the future for example. Intelligence will be in abundance, and become a commodity. This potential outcome does alarm me not just from a job perspective, but in terms of fake content, lack of human connection, lack of value of intelligence in general (you will find people with high IQ's lose respect from society in general), social mobility, etc. I can see a potential to the old world where lords that command scarcity (e.g. landlords) command peasants again - reversing the gains of the industrial revolution as an extreme case depending on general AI progress (not LLMs). For people who's value is more in capital or land vs labor, AI seems like a dream future IMO.
There's potential good here, but sadly I'm alarmed because the likelihood that the human race aligns to achieve it is low (the tragedy of the commons problem). It is much easier, and more likely, certain groups use it and target people of value economically now, but with little power (i.e the middle class). The chance of new weapons, economic displacement, fake news, etc for me trumps a voice/chat bot and a fancy image generator. The "adjustment period" is critical to manage; and I think climate change, and other broader issues tells us sadly IMO our likely success in doing this.
Do you expect the hockeystick graph of technological development since the industrial evolution to slow? Or that it will proceed, only without significant advances in AI?
Seems like the base case here is for the exponential growth to continue, and you'd need a convincing argument to say otherwise.
That's no guarantee that AI continues advancing at the same pace, and no one has been arguing against overall technological progress slowing
Refining technology is easier than the original breakthrough, but it doesn't usually lead to a great leap forward.
LLMs were the result of breakthroughs, but refining them isn't guaranteed to lead to AGI. It's not guaranteed (or likely) to improve at an exponential rate.
Which chart are you referencing exactly? How does it define technological development? It's nearly impossible for me to discuss a chart without knowing what axis refer.
Without specifics all I can say is that I don't acknowledge any measurable benefits of AI (in its' current state) in real world applications. So I'd say I am leaning towards latter.
That's probably what every self-driving car company thought ~10 years ago or so, everything was moving so fast for them back then. Now it doesn't seem like we're getting close to solution for this.
Surely this time it's going to be different, AGI is just around a corner. /s
Would you have predicted in summer of 2022 that gpt4 level conversational agent is a possibility in the next 5 years? People have tried to do it in the past 60 years and failed. How is this time not different?
On a side note, I find this type of critique of what future of tech might look like the most uninteresting one. Since tech by nature inspiries people about the future, all tech get hyped up. all you gotta do then is pick any tech, point out people have been wrong, and ask how likely is it that this time it is different.
Unfortunately, I don't see any relevance in that argument, if you consider GPT-4 to be a breakthrough -- then sure, single breakthroughs happen, I am not arguing with that. Actually, same thing happened with self-driving: I don't think many people expected Tesla to drop FSD publicly back then.
Now, chain of breakthroughs happening in a small timeframe? Good luck with that.
Just to make it clear, I see only 1 breakthrough [0]. Everything that happened afterwards is just application of this breakthrough with different training sets / to different domains / etc.
They are the same breakthrough applied to different domains, I don't see them as different. We will need a new breakthrough, not applying the same solution to new things.
I'm not saying this as someone cheering this on; I'm alarmed by it. But I can't pretend that it's running out of steam. It's possible it will run out of money, but even if so, only for a while.