If AI was going to advance exponentially I'd have expected it to take off by now.

AI LLMs have been pretty shit, but the advancement in voice, image generation, and video generation in the last two years has been unbelievable.
We went from the infamous Will Smith eating spaghetti to videos that are convincing enough to fool most people... and it only took 2-3 years to get there.
But LLMs will have a long way to go because of how they create content. It's very easy to poison LLM datasets, and they get worse learning from other generated content.

LOL... you did make me chuckle.
Aren't we 18months until developers get replaced by AI... for like few years now?
Of course "AI" even loosely defined progressed a lot and it is genuinely impressive (even though the actual use case for most hype, i.e. LLM and GenAI, is mostly lazier search, more efficient spam&scam personalized text or impersonation) but exponential is not sustainable. It's a marketing term to keep on fueling the hype.
That's despite so much resources, namely R&D and data centers, being poured in... and yet there is not "GPT5" or anything that most people use on a daily basis for anything "productive" except unreliable summarization or STT (which both had plenty of tools for decades).
So... yeah, it's a slow take off, as expected. shrug

how do you grow zero exponentially

I'd argue it has. Things like ChatGPT shouldn't be possible, maybe it's unpopular to admit it but as someone who has been programming for over a decade, it's amazing that LLMs and "AI" has come as far as it has over the past 5 years.
That doesn't mean we have AGI of course, and we may never have AGI, but it's really impressive what has been done so far IMO.
- If you've been paying attention to the field, you'd see it's been a slow steady march. The technology that LLMs are based in were first published in 2016/2017, ChatGPT was the third iteration of the same base model.
  Thats not even accounting for all the work done with RNNs and LSTMs prior to that, and even more prior.
  Its definitely a major breakthrough, and very similar to what CNNs did for computer vision further back. But like computer vision, advancements have been made in other areas (like the generative space) and haven't followed a linear path of progress.
- Agreed. I never thought it would happen in my lifetime, but it looks like we’re going to have Star Trek computers pretty soon.

Things just don't impend like they used to!
- Nobody wants to portend anymore.

It has slowed exponentially because the models get exponentially more complicated the more you expect it to do.
- The exponential problem has always been there. We keep finding tricks and optimizations in hardware and software to get by it but they're only occasional.
  The pruned models keep getting better so now You're seeing them running on local hardware and cell phones and crap like that.
  I don't think they're out of tricks yet, but God knows when we'll see the next advance. And I don't think there's anything that'll take this current path into AGI I think that's going to be something else.

I think we might not be seeing all the advancements as they are made.
Google just showed off AI video with sound. You can use it if you subscribe to thier $250/month plan. That is quite expensive.
But if you have strong enough hardware, you can generate your own without sound.
I think that is a pretty huge advancement in the past year or so.
I think that focus is being put on optimizing these current things and making small improvements to quality.
Just give it a few years and you will not even need your webcam to be on. You could just use an AI avatar that look and sounds just like you running locally on your own computer. You could just type what you want to say or pass through audio. I think the tech to do this kind of stuff is basically there, it just needs to be refined and optimized. Computers in the coming years will offer more and more power to let you run this stuff.
- How is that an advance ? Computers have been able to speak since the 1970s. It was already producing text.

It has taken off exponentially. It’s exponentially annoying that’s it’s being added to literally everything
- Humanity may achieve an annoyance singularity within six months

How do you know it hasn't and us just laying low? I for one welcome our benevolent and merciful machine overlord.
- Duly noted. 🤭 🤫

Computers are still advancing roughly exponentially, as they have been for the last 40 years (Moore's law). AI is being carried with that and still making many occasional gains on top of that. The thing with exponential growth is that it doesn't necessarily need to feel fast. It's always growing at the same rate percentage wise, definitionally.
- Moore's law is kinda still in effect, depending on your definition of Moore's law. However, Dennard Scaling is not so computer performance isn't advancing like it used to.
  
  Moore’s law is kinda still in effect, depending on your definition of Moore’s law.
  Sounds like the goal post is moving faster than the number of transistors in an integrated circuit.
- We once again congratulate software engineers for nullifying 40 years of hardware improvements.

It has definitely plateaued.

Well, the thing is that we're hitting diminishing returns with current approaches. There's a growing suspicion that LLMs simply won't be able to bring us to AGI, but that they could be a part of or stepping stone to it. The quality of the outputs are pretty good for AI, and sometimes even just pretty good without the qualifier, but the only reason it's being used so aggressively right now is that it's being subsidized with investor money in the hopes that it will be too heavily adopted and too hard to walk away from by the time it's time to start charging full price. I'm not seeing that. I work in comp sci, I use AI coding assistants and so do my co-workers. The general consensus is that it's good for boilerplate and tests, but even that needs to be double checked and the AI gets it wrong a decent enough amount. If it actually involves real reasoning to satisfy requirements, the AI's going to shit its pants. If we were paying the real cost of these coding assistants, there is NO WAY leadership would agree to pay for those licenses.
- Yeah, I don't think AGI = an advanced LLM. But I think it's very likely that a transformer style LLM will be part of some future AGI. Just like human brains have different regions that can do different tasks, an LLM is probably the language part of the "AGI brain".
- What are the "real costs" though? It's free to run a half decent LLM locally on a mid tier gaming PC.
  Perhaps a bigger problem for the big AI companies rather then the open source approach.
  
  Sure, but ChatGPT costs MONEY. Money to run, and MONEY to train, and then they still have to make money back for their investors after everything's said and done. More than likely, the final tally is going to look like whole cents per token once those investor subsidies run out, and a lot of businesses are going to be looking to hire humans back quick and in a hurry.

Iirc there are mathematical reason why AI can't actually become exponentially more intelligent? There are hard limits on how much work (in the sense of information processing) can be done by a given piece of hardware and we're already pretty close to that theoretical limit. For an AI to go singulaity we would have to build it with enough initial intelligence that it could aquire both the resources and information with which to improve itself and start the exponential cycle.

That's only if the exponent is greater than 1.

A few years ago I remember people being amazed that prompts like "Markiplier drinking a glass of milk" could give them some blobs that looked vaguely like the thing asked for occasionally. Now there is near photorealistic video output. Same kind of deal with ability to write correct computer code and answer questions. Most of the concrete predictions/bets people made along the lines of "AI will never be able to do ______" have been lost.
What reason is there to think it's not taking off, aside from bias or dislike of what's happening? There are still flaws and limitations for what it can do, but I feel like you have to have your head in the sand to not acknowledge the crazy level of progress.
- It could do that 3 years ago.
- It's absolutely taking off in some areas. But there's also an unsustainable bubble because AI of the large language model variety is being hyped like crazy for absolutely everything when there are plenty of things it's not only not ready for yet, but that it fundamentally cannot do.
  You don't have to dig very deeply to find reports of companies that tried to replace significant chunks of their workforces with AI, only to find out middle managers giving ChatGPT vague commands weren't capable of replicating the work of someone who actually knows what they're doing.
  That's been particularly common with technology companies that moved very quickly to replace developers, and then ended up hiring them back because developers can think about the entire project and how it fits together, while AI can't - and never will as long as the AI everyone's using is built around large language models.
  Inevitably, being able to work with and use AI is going to be a job requirement in a lot of industries going forward. Software development is already changing to include a lot of work with Copilot. But any actual developer knows that you don't just deploy whatever Copilot comes up with, because - let's be blunt - it's going to be very bad code. It won't be DRY, it will be bloated, it will implement things in nonsensical ways, it will hallucinate... You use it as a starting point, and then sculpt it into shape.
  It will make you faster, especially as you get good at the emerging software development technique of "programming" the AI assistant via carefully structured commands.
  And there's no doubt that this speed will result in some permanent job losses eventually. But AI is still leagues away from being able to perform the joined-up thinking that allows actual human developers to come up with those structured commands in the first place, as a lot of companies that tried to do away with humans have discovered.
  Every few years, something comes along that non-developers declare will replace developers. AI is the closest yet, but until it can do joined-up thinking, it's still just a pipe-dream for MBAs.
  
  But any actual developer knows that you don’t just deploy whatever Copilot comes up with, because - let’s be blunt - it’s going to be very bad code. It won’t be DRY, it will be bloated, it will implement things in nonsensical ways, it will hallucinate… You use it as a starting point, and then sculpt it into shape.
  Yeah, but I don't know where you're getting the "never will" or "fundamentally cannot do" from. LLMs used to be only useful for coding if you ask for simple self-contained functions in the most popular languages, and now we're here; most requests with small scope, I'm getting a result that is better written than I could have done myself by spending way more time, it makes way fewer mistakes than before and can often correct them. That's with only using local models which became actually viable for me less than a year ago. So why won't it keep going?
  From what I can tell there is not very much actually standing in the way of sensible holistic consideration of a larger problem or codebase here, just context size limits and being more likely to forget things in the context window the longer it is, which afaik are problems being actively worked on where there's no reason they would be guaranteed to remain unsolved. This also seems to be what is holding back agentic AI from being actually useful. If that stuff gets cracked, I think it's going to mean things will start changing even faster.
- Agreed. LLM Ai has gotten insanely good insanely fast, and an LLM of course isn’t going to magically turn into an AGI. That’s a whole different ball game.
- Yes, the goal posts keep moving, but they do so for a rather solid reason: We humans are famously bad at understanding intelligence and at understanding the differences between human and computer intelligence.
  100 years ago, doing complex calculations was seen as something very complex that only reasonably smart humans could do. Computers could easily outcompete humans, because calculations are inherently easy for computers while very difficult for humans.
  30 years ago we thought that high-level chess was something reserved only to the smartest of humans, and that it was a decent benchmark for intelligence. Turns out, playing chess is something that benefits greatly from large memory and fast computations, so again, it was easy for computers while really hard for humans.
  Nowadays AI can do a lot of things we thought would be really hard to do, but that computers can actually do. But there's hardly any task performed by LLMs where it's actually better than a moderately proficient human being. (Apart from tasks like "Do homework task X", where again LLMs benefit from large memory since they can just regurgitate stuff from the training set.)
- Linear growth can be faster than exponential growth. Exponential implys tomorrow we will see it advance faster then it did the day before so every day we would see even crazier shit.

When people talk about AI taking off exponentially, usually they are talking about the AI using its intelligence to make intelligence-enhancing modifications to itself. We are very much not there yet, and need human coaching most of the way.
At the same time, no technology ever really follows a particular trend line. It advances in starts and stops with the ebbs and flows of interest, funding, novel ideas, and the discovered limits of nature. We can try to make projections - but these are very often very wrong, because the thing about the future is that it hasn't happened yet.
- And at that point, we wouldnt ever know anyway that it did.
- Although i agree with the general idea, AI (as in llms) is a pipe dream. Its a non product, another digital product that hypes investors up and produces "value" instead of value.
  
  Not true. Not entirely false, but not true.
  Large language models have their legitimate uses. I'm currently in the middle of a project I'm building with assistance from Copilot for VS Code, for example.
  The problem is that people think LLMs are actual AI. They're not.
  My favorite example - and the reason I often cite for why companies that try to fire all their developers are run by idiots - is the capacity for joined up thinking.
  Consider these two facts:
  Humans are mammals.
  Humans build dams.
  Those two facts are unrelated except insofar as both involve humans, but if I were to say "Can you list all the dam-building mammals for me," you would first think of beavers, then - given a moment's thought - could accurately answer that humans do as well.
  Here's how it goes with Gemini right now:
  Now Gemini clearly has the information that humans are mammals somewhere in its model. It also clearly has the information that humans build dams somewhere in its model. But it has no means of joining those two tidbits together.
  Some LLMs do better on this simple test of joined-up thinking, and worse on other similar tests. It's kind of a crapshoot, and doesn't instill confidence that LLMs are up for the task of complex thought.
  And of course, the information-scraping bots that feed LLMs like Gemini and ChatGPT will find conversations like this one, and update their models accordingly. In a few months, Gemini will probably include humans in its list. But that's not a sign of being able to engage in novel joined-up thinking, it's just an increase in the size and complexity of the dataset.
- I do expect advancement to hit a period of exponential growth that quickly surpasses human intelligence. Given it adapts the drive to autonmously advance. Whether that is possible is yet to be seen and that's kinda my point.
  
  They've been saying "AGI in 18 months" for years now.
  
  Ask ChatGPT to list every U.S. state that has the letter 'o' in its name.

What do you consider having "taken off"?
It's been integrated with just about everything or is in the works. A lot of people still don't like it, but that's not an unusual phase of tech adoption.
From where I sit I'm seeing it everywhere I look compared to last year or the year before where pretty much only the early adopters were actually using it.
- What do you mean when you say AI has been integrated with everything? Very broad statement that's obviously not literally true.
  
  True, I tried to qualify it with just about or on the way.
  From the perspective of my desk, my core business apps have AI auto suggest in key fields (software IDEs, ad buying tools, marketing content preparation such as Canva). My Whatsapp and Facebook messenger apps now have an "Ask meta AI" feature front and center. Making a post on Instagram, it asks if I want AI assistance to write the caption.
  I use an app to track my sleeping rhythm and it has an AI sleep analysis feature built in. The photo gallery on my phone includes AI photo editing like background removal, editing things out (or in).
  That's what I mean when I say it's in just about everything, at least relative to where we were just a short bit of time ago.
  You're definitely right that it's not literally in everything.

This is precisely a property of exponential growth, that it can take (seemingly) very long until it starts exploding.
- What are you talking about it asymptoped at 5 units. It cant be described as exponential until it is exponential otherwise its better described as linear or polynomial if you must.
  
  Exponential growth is always exponential, not just if it suddenly starts to drastically increase in the arbitrarily choosen view scale.
  A simple way, to check wether data is exponential, is to visualize it in loc-scale, and if it shows there a linear behavior, it has a exponential relation.
  Exponential growth means, that the values change by a constant ratio, contrary to linear growth where the data changes by a constant rate.
  
  It's exponential along its entire range, even all the way back to negative infinity.
  
  The derivative of an exponential is exponential. The relative difference between -1 and -2 is the same as 1 and 2.
  I’d say the development is exponential. Compare what we had 4 years ago, 2 years ago and now. 4 years ago it was inconceivable that an AI model could generate any convincing video at all. 2 years ago we laughed at Will Smith eating pasta. Today we have Veo 3 which generates videos with sound that are near indistinguishable from real life.
  It’s not going to be long until you regularly see AI generated videos without realizing it’s AI.

That's exactly what AI would say. Hmmm...

A major bottleneck is power capacity. Is is very difficult to find 50Mwatts+ (sometime hundreds) of capacity available at any site. It has to be built out. That involves a lot of red tape, government contracts, large transformers, contractors, etc. the current backlog on new transformers at that scale is years. Even Google and Microsoft can't build, so they come to my company for infrastructure - as we already have 400MW in use and triple that already on contract. Further, Nvidia only makes so many chips a month. You can't install them faster than they make them.
- And the single biggest bottleneck is that none of the current AIs "think".
  They. Are. Statistical. Engines.
  
  Maybe we are statistical engines too.
  When I heard people talk they are also repeating the most common sentences that they heard elsewhere anyway.
  
  How closely do you need to model a thought before it becomes the real thing?
  
  And it's pretty great at it.
  AI's greatest use case is not LLM and people treat it like that because it's the only thing we can relate to.
  AI is so much better and many other tasks.
  
  Same
  
  Humans don’t actually think either, we’re just electricity jumping to nearby neural connections that formed based on repeated association. Add to that there’s no free will, and you start to see how “think” is a immeasurable metric.
  
  Markov chains with extra steps
  
  You're not going to get an argument from me.
- Is this the AI?

Consider age of the planet, eons and eras. This is the age of exponential growth.
- That's ridiculous

We humans always underestimate the time it actually takes for a tech to change the world. We should travel in self-flying flying cars and on hoverboards already but we're not.
The disseminators of so-called AI have a vested interest in making it seem it's the magical solution to all our problems. The tech press seems to have had a good swig from the koolaid as well overall. We have such a warped perception of new tech, we always see it as magical beans. The internet will democratize the world - hasn't happened; I think we've regressed actually as a planet. Fully self-drving cars will happen by 2020 - looks at calendar. Blockchain will revolutionize everything - it really only provided a way for fraudsters, ransomware dicks, and drug dealers to get paid. Now it's so-called AI.
I think the history books will at some point summarize the introduction of so-called AI as OpenAI taking a gamble with half-baked tech, provoking its panicked competitors into a half-baked game of oneupmanship. We arrived at the plateau in the hockey stick graph in record time burning an incredible amount of resources, both fiscal and earthly. Despite massive influences on the labor market and creative industries, it turned out to be a fart in the wind because skynet happened a 100 years later. I'm guessing 100 so it's probably much later.
- AI has been advancing exponentially, it's just a very small exponent.
  In the 1980s, it was "five years out" - and it more or less has been that until the past 5-10 years. It's moving much faster now, but still much slower than people expect.
  They think because they saw HAL in the 2001 movie back in 1968, that should have been reality by the 1970s, or certainly by 2010.
  Some things move faster than people expect, like the death of newspapers and the first class letter, but most move slower.

It's not anytime soon. It can get like 90% of the way there but those final 10% are the real bitch.
- So logarithmic then.
- The AI we know is missing the I. It does not understand anything. All it does is find patterns in 1's and 0's. It has no concept of anything but the 1's and 0's in its input data. It has no concept of correlation vs causation, that's why it just hallucinates (presents erroneously illogical patterns) constantly.
  Turns out finding patterns in 1's and 0's can do some really cool shit, but it's not intelligence.
  
  This is why I hate calling it AI.
  
  Humans are just nurons, we don't "understand either" until so many stack on top of each other than we have a sort of consciousness. The it seems like we CAN understand but do we? Or are we just a bunch of meat computers? Also, llms handle language or correlations of words, don't humans just do that (with maybe body language too) but we're all just communicating. If llms can communicate isn't that enough conceptually to do anything? If llms can program and talk to other llms what can't they do?
  
  This is not necessarily true. While it's using pattern recognition on a surface level, we're not entirely sure how AI comes up with it's output.
  But beyond that, a lot of talk has been centered around a threshold when AI begins training other AI & can improve through iterations. Once that happens, people believe AI will not only improve extremely rapidly, but we will understand even less of what is happening when an AI black boxes train other AI black boxes.
  
  Distill intelligence - what is it, really? Predicting what comes next based on... patterns. Patterns you learn in life, from experience, from books, from genetic memories, but that's all your intelligence is too: pattern recognition / prediction.
  As massive as current AI systems are, consider that you have ~86 Billion neurons in your head, devices that evolved over the span of billions of years ultimately enabling you to survive in a competitive world with trillions of other living creatures, eating without being eaten at least long enough to reproduce, back and back and back for millions of generations.
  Current AI is a bunch of highly simplified computers with up to hundreds of thousands of cores. Like planes fly faster than birds, AI can do some tricks better than human brains, but mostly: not.
- It can get like 90% of the way there
  I'm still waiting for the first 10%

It "took off" by companies forcing it into everything

Pretty much like everyone expected flying cars to have taken off by now (pun intended)
- The biggest thing holding back flying cars is that everyone calls them helicopters.
  
  Though helicopters are not what was envisioned as flying cars most of the time. Including their usability.

In the spirit of showerthoughts: I feel the typical LLM is reaching a plateau. The "reasoning" type was a big advance though.
Companies are putting a lot of effort on how to handle the big influx of AI requests.
With the huge resources, both academic and operational, going into AI we should expect unexpected jumps in power :)

Already happening https://youtu.be/evSFeqTZdqs

Latest models give until 2043 for the takeoff
AI, the one currently used for actual productive work by scientific researchers, healthcare specialists, energy development, manufacturing, agriculture and such, is poised to be able to handle about 20% of all human related work by 2040.
By 2043, it will be able to handle 100% of any human related work in the fields. The takeoff is merely 3 years
- Yeah, imma need a source for those numbers.
  
  Whats it even matter. The variance between one exponential function and another can be astronomical if the confidence interval isn't extremely tight.
- Ok I'll take your word for it but Ill tell you i never felt the need to build a bunker for y2k

ask the nearest school where nearly all students have smartphones whether it has taken off
- AI is when you can
  ... Checks notes
  Use a dumb terminal to communicate with another dumb terminal?
  
  it's when people use chatGPT to do their homework

Only the uneducated don't see AI "taking off" right now.
Every idiot who says this thinks that ChatGPT encompasses all of "AI". They're the same people who didn't get internet in their household until 2010.
- AI isn't taking off because it took off in the 60s. Heck, they were even working on neural nets back then. Same as in the 90s when they actually got them to be useful in a production environment.
  We got a deep learning craze in the 2010s and then bolted that onto neural nets to get the current wave of "transformers/diffusion models will solve all problems". They're really just today's LISP machines; expected to take over everything but unlikely to actually succeed.
  Notably, deep learning assumes that better results come from a bigger dataset but we already trained our existing models on the sum total of all of humanity's writings. In fact, current training is hampered by the fact that a substantial amount of all new content is already AI-generated.
  Despite how much the current approach is hyped by the tech companies, I can't see it delivering further substantial improvements by just throwing more data (which doesn't exist) or processing power at the problem.
  We need a systemically different approach and while it seems like there's all the money in the world to fund the necessary research, the same seemed true in the 50s, the 60s, the 80s, the 90s, the 10s... In the end, a new AI winter will come as people realize that the current approach won't live up to their unrealistic expectations. Ten to fifteen years later some new approach will come out of underfunded basic research.
  And it's all just a little bit of history repeating.
  
  in the 60s. Heck, they were even working on neural nets back then
  I remember playing with neural nets in the late 1980s. They had optical character recognition going even back then. The thing was, their idea of "big networks" was nowhere near big enough scale to do anything as impressive as categorize images: cats vs birds.
  We've hit the point where supercomputers in your pocket are....
  The Cray-1, a pioneering supercomputer from the 1970s, achieved a peak performance of around 160 MFLOPS, it cost $8 million - or $48 million in today's dollars, it weighed 5 tons
  Modern smartphones, even mid-range models, can perform significantly faster than the Cray-1. For example, a 2019 Google Pixel 3 achieved 19 GFLOPS
  19000/160 = over 100x as powerful as a Cray from the 1970s.
  I just started using a $110 HAILO-8 for image classification, it can perform 26TOPS, that's over 160,000x a 1970s Cray (granted, the image processor is working with 8 bit ints, the Cray worked with 64 bit floats... but still... 20,000x the operational power for 1/436,000th the cost and 1/100,000th the weight.)
  There were around 60 Crays delivered by 1983, HAILO alone is selling on the order of a million chips a year...
  Things have sped up significantly in the last 50 years.
- There are AI experts (I could air quote that, but really, people who work in the industry) who are legitimately skeptical about the speed, power and real impact AI will have in the short term, so it isn’t a case of everyone who “really knows” thinks we’re getting doomsday AGI tomorrow.
- I didnt even imply it hasn't happened but ok..
  What do I know, im just uneducated, right?
  
  Right!