As of now yes. But we are still in day 0.1 of GenAI. Do you think this will be the case when o3 models are 10x better and 100x cheaper? There will be a turning point but it’s not happened yet.
I basically agree with you, and I think the thing that is missing from a bunch of responses that disagree is that it seems fairly apparent now that AI has largely hit a brick wall in terms of the benefits of scaling. That is, most folks were pretty astounded by the gains you could get from just stuffing more training data into these models, but like someone who argues a 15 year old will be 50 feet tall based on the last 5 years' growth rate, people who are still arguing that past growth rates will continue apace don't seem to be honest (or aware) to me.
I'm not at all saying that it's impossible some improvement will be discovered in the future that allows AI progress to continue at a breakneck speed, but I am saying that the "progress will only accelerate" conclusion, based primarily on the progress since 2017 or so, is faulty reasoning.
> it seems fairly apparent now that AI has largely hit a brick wall in terms of the benefits of scaling
What's annoying is plenty of us (researchers) predicted this and got laughed at. Now that it's happening, it's just quiet.
I don't know about the rest, but I spoke up because I didn't want to hit a brick wall, I want to keep going! I still want to keep going! But if accurate predictions (with good explanations) aren't a reason to shift resource allocation then we just keep making the same mistake over and over. We let the conmen come in and people who get too excited by success that they get blind to pitfalls.
And hey, I'm not saying give me money. This account is (mostly) anonymous. There's plenty of people that made accurate predictions and tried working in other directions but never got funding to test how methods scale up. We say there's no alternatives but there's been nothing else that's been given a tenth of the effort. Apples and oranges...
> What's annoying is plenty of us (researchers) predicted this and got laughed at. Now that it's happening, it's just quiet.
You need to model the business world and management more like a flock of sheep being herded by forces that mostly don't have to do with what actually is going to happen in future. It makes a lot more sense.
> mostly don't have to do with what actually is going to happen
Yet I'm talking about what did happen.
I'm saying we should have memory. Look at predictions people make. Reward accurate ones, don't reward failures. Right now we reward whoever makes the craziest predictions. It hasn't always been this way, so we should go back to less crazy
Practically no one is herded by what is actually going to happen, hardly even by what is expected to happen. Business pretends that it is driven by expectations, but is mostly driven by the past, as in financial statements. What is the bonus we can get this year? There is of course the strategic thinking, I don't want to discount that part of business, but it is not the thing that will drive most of these, AI as a cost saving measure, decisions. This is the unimaginative part of AI application and as such relegated to the unimaginative managers.
> It is difficult to get a man to understand something, when his salary depends on his not understanding it.”
It's all a big hype bubble and not only is no one in the industry willing to pop it, they actively defend against popping a bubble that is clearly rupturing on its own. It's endemic of how modern businesses no longer care about a proper 10 year portfolio and more about how to make the next quarter look good.
There's just no skin in the game, and everyone's ransacking before the inevitable fire instead of figuring out how to prevent the fire to begin with.
I dont see any wall. Gemini 2.5 and o3/o4 are incredible improvements. Gen AI is miles ahead of where it was a year ago which was miles ahead of where it was 2 years ago.
The actual LLM part isn't much better than a year ago. What's better is that they've added additional logic and made it possible to intertwine traditional, expert-system style AI plus the power of the internet to augment LLMs so that they're actually useful.
This is an improvement for sure, but LLMs themselves are definitely hitting a wall. It was predicted that scaling alone would allow them to reach AGI level.
There were definitely people "at the top" who were essentially arguing that more scale would get you to AGI - Ilya Sutskever of OpenAI comes to mind (e.g. "next-token prediction is enough for AGI").
There were definitely many other prominent researchers who vehemently disagreed, e.g. Yann LeCun. But it's very hard for a layperson (or, for that matter, another expert) to determine who is or would be "right" in this situation - most of these people have strong personalities to put it mildly, and they often have vested interests in pushing their preferred approach and view of how AI does/should work.
The improvements have less to do with scaling than adding new techniques like better fine tuning and reinforcement learning. The infinite scaling we were promised, that only required more content and more compute to reach god tier has indeed hit a wall.
People were originally very surprised that you could get so much functionality by just pumping more data and adding more parameters to models. What made OpenAI initially so successful is that they were the first company willing to make big bets on these huge training runs.
After their success, I definitely saw a ton of blog posts and general "AI chatter" that to get to AGI all you really needed to do (obviously I'm simplifying things a bit here) was get more data and add more parameters, more "experts", etc. Heck, OpenAI had to scale back it's pronouncements (GPT 5 essentially became 4.5) when they found that they weren't getting the performance/functionality advances they expected after massively scaling up their model.
I basically agree with you also, but I have a somewhat contrarian view of scaling -> brick wall. I feel like applications of powerful local models is stagnating, perhaps because Apple has not done a good job so far with Apple Intelligence.
A year ago I expected a golden age of local model intelligence integrated into most software tools, and more powerful commercial tools like Google Jules to be something used perhaps 2 or 3 times a week for specific difficult tasks.
That said, my view of the future is probably now wrong, I am just saying what I expected.
No, the hype cycle started around 2019, slowly at first. The technology this is built with is more like 20 years old, so no, we are not 2.5 years at most really.
They are self driving the same way a tram or subway can be self driving. They traffic a tightly bounded designated area. They're not competing with human drivers. Still a marvel of human engineering, just quite expensive compared with other forms of public transport. It just doesn't compete in the same space and likely never will.
They are literally competing with human uber drivers in the area they operate and also having a much lower crash and injury rate.
I admit they don't operate everywhere - only certain routes. Still they are undoubtedly cars that drive themselves.
I imagine it'll be the same with AGI. We'll have robots / AIs that are much smarter than the average human and people will be saying they don't count because humans win X Factor or something.
They're driving, but not well in my (limited) interactions with them. I had a waymo run me completely out of my lane a couple months ago as it interpreted 2 lanes of left turn as an extra wide lane instead (or, worse, changed lanes during the turn without a blinker or checking its sensors, though that seems unlikely).
The argument that self-driving cars should be allowed on public roads as long as they are statistically as safe as human drivers (on average) seems valid, but of course none of these cars have AGI... they perform well in the anticipated simulator conditions in which they were trained (as long as they have the necessary sensors, e.g. Waymo's lidar, to read the environment in reliable fashion), but will not perform well in emergency/unanticipated conditions they were not trained on. Even outside of emergencies, Waymos still sometimes need to "phone home" for remote assistance in knowing what to do.
So, yes, they are out there, perhaps as safe on average as a human (I'd be interested to see a breakdown of the stats), but I'd not personally be comfortable riding in one since I'm not senile, drunk, teenager, hothead, distracted (using phone while driving), etc - not part of the class that are dragging the human safety stats down. I'd also not trust a Tesla where penny pinching, or just arrogant stupidity, has resulted in a sensor-poor design liable to failure modes like running into parked trucks.
I'd not personally be comfortable riding in one since I'm not senile, drunk, teenager, hothead, distracted (using phone while driving), etc - not part of the class that are dragging the human safety stats down.
The challenge is that most people think they’re better than average drivers.
I'm not sure what the "challenge" is there, but certainly true in terms of human psychology.
My point was that if you are part of one of these accident-prone groups, you are certainly worse than average, and are probably safer (both for yourself, and everyone around you) in a Waymo. However, if you are an intelligent non-impaired experienced driver, then maybe not, and almost certainly not if we're talking about emergency and dangerous situations which is where it really matters.
Sure, you don't know how well any specific driver is going to react in an emergency situation, and some are going to be far worse than others (e.g. panicking, or not thinking quickly enough), but the human has the advantage of general intelligence and therefore NOT having to rely on having had practice at the specific circumstance they find themselves in.
A recent example - a few weeks ago I was following another car in making a turn down a side road, when suddenly that car stops dead (for no externally apparent reason), and starts backing up fast about to hit me. I immediately hit my horn and prepare to back up myself to get out of the way, since it was obvious to me - as a human - that they didn't realize I was there, and without intervention would hit me.
Driving away I watch the car in my rear view mirror and see it pull a U-turn to get back out of the side road, making it apparent why they had stopped before. I learned something, but of course the driverless car is incapable of learning, and certainly has no theory of mind, and would behave same as last time - good or bad - if something similar happened again.
In my lens, as long as companies don't want to be held liable for an accident, the shouldn't be on roads. They need to be extremely confident to the point of putting their money where their mouths are. That's true "safety".
That's the main difference with a human driver. If I take an Uber and we crash, that driver is liable. Waymo would fight tooth and nail to blame anything else.
I don't care about SF. I care about what I can but as a typical American. Not as an enthusiast in one of the most technologically advanced cities on the planet
As far as I've seen we appear to already have self driving vehicles, the main barriers are legal and regulatory concerns rather than the tech. If a company wanted to put a car on the road that beetles around by itself there aren't any crazy technical challenges to doing that - the issue is even if it was safer than a human driver the company would have a lot of liability problems.
They are only self-driving in a very controlled environments of few very good mapped out cities with good roads in good weather.
And it took what like 2 decades to get there.
So no, we don't have self-driving even close. Those examples look more like hard-coded solution for custom test cases.
As prototypes, yes. But that's like pointing to Japanese robots in the 80's and expecting robot butlers any day now. Or maybe Boston dynamics 10 years ago. Or when OpenAI was into robotics.
There's a big gap between seeing something work in the lab and being ready for real world use. I know we do this in software, but that's a very abnormal thing (and honestly, maybe not the best)
When people say “we'll have self-driving cars next year”, I understand that self-driving cars will be widespread in the developed world and accessible to those who pay a premium. Given the status quo, I find it pointless to discuss the semantics of whether they exist or not.
I don't think that is a reasonable generalisation. A lot of people would have been talking about the first person to take a real trip in a car that drives itself. A record that is in the past.
Not to mention that HN gets really tetchy about achieving specifically SAE Level 6 when in practice some pretty basic driver assist tools are probably closer to what people meant. It reminds me of a gentlemen I ran into who was convinced that the OpenAI DoTA bot with a >99% win rate couldn't really be said to be playing the game. If someone can take their hands off the wheel for 10 minutes we're there in a common language sense; the human in the car isn't actively in control.
Yeah, it’s a big grid with wide streets. Did it fail there? If so I imagine it’s just due to lack of business—there are almost no taxis in Phoenix. Mostly just from the airport.
100% this. I always argue that groundbreaking technologies are clearly groundbreaking from the start. It is almost a bit like a film, if you have to struggle to get into it in the first few minutes, you may as well spare yourself watching the rest.
I consulted a radiologist more than 5 years after Hinton said that it was completely obvious that radiologists would be replaced by AI in 5 years. I strongly suspect they were not an AI.
Why do I think this?
1) They smelled slightly funny.
2) They got the diagnosis wrong.
OK maybe #2 is a red herring. But I stand by the other reason.
I know a radiologist and talk a decent bit about AI usage in the field. Every radiologist today is making heavy use of AI. They pre screen everything and from what I understand it has led to massive productivity gains. It hasnt led to job losses yet but theres so much money on the line it really feels to me like we're just waiting for the straw that broke the camels back. No one wants to be the first to fully get rid of radiologists but once one hospital does the rest will quickly follow suit.
The quote appears to be “We should stop training radiologists now, it’s just completely obvious within five years deep learning is going to do better than radiologists.”
So there's some room for interpretation, the weaker interpretation is less radical (that AI could beat humans in radiology tasks in 5 years).
You're missing the big picture. Helion can still make their goal. Once they have a working fusion reactor they can use the energy to build a time machine.
They didn't reach that goal. Why would they bother reaching an easier goal when they could shoot for a bigger one? /s Their new goal is to build a fusion plant by 2028 [0].
We’re already heading toward the sigmoid plateau. The GPT 3 to 4 shift was massive. Nothing since had touched that. I could easily go back to the models I was using 1-2 years ago with little impact on my work.
I don’t use RAG, and have no doubt the infrastructure for integrating AI into a large codebase has improved. But the base model powering the whole operation seems stuck.
> I don’t use RAG, and have no doubt the infrastructure for integrating AI into a large codebase has improved
It really hasn't.
The problem is that a GenAI system needs to not only understand the large codebase but also the latest stable version of every transitive dependency it depends on. Which is typically in the order of hundreds or thousands.
Having it build a component with 10 year old, deprecated, CVE-riddled libraries is of limited use especially when libraries tend to be upgraded in interconnected waves. And so that component will likely not even work anyway.
I was assured that MCP was going to solve all of this but nope.
Those large number of outdated dependencies are in the LLM "index" which can't be rapidly refreshed because of the training costs.
MCP would allow it to instead get this information at run-time from language servers, dependency repositories etc. But it hasn't proven to be effective.
I use LLM’s daily and live them but at the current rate of progress it’s just not really something worth worrying about. Those that are hysterical about AI seem to think LLM’s are getting exponentially better when in fact diminishing returns are hitting hard. Could some new innovation change that? It’s possible but it’s not inevitable or at least not necessarily imminent.
I agree that the core models are only going to see slow progression from here on out, until something revolutionary happens... which might be a year from now, or maybe twenty years. Who knows.
But we are going to see a huge explosion in how those models are integrated into the rest of the tech ecosystem. Things that a current model could do right now, if only your car/watch/videogame/heart monitor/stuffed animal had a good working interface into an AI.
Not necessarily looking forward to that, but that's where the growth will come.
How are we in 0.1 of GenAI ? It's been developed for nearly a decade now.
And each successive model that has been released has done nothing to fundamentally change the use cases that the technology can be applied to i.e. those which are tolerant of a large percentage of incoherent mistakes. Which isn't all that many.
So you can keep your 10x better and 100x cheaper models because they are of limited usefulness let alone being a turning point for anything.
Funding is behind the curve. Social networks existed in 2003 and Facebook became a billion dollar company a decade later. AI horror fantasies from the 90’s still haven’t come true. There is no god, there is no Skynet.
AlphaGo beating the top human player was in 2016. To my memory, that was one of the first public breakthroughs of the new era of machine learning.
Around 2010 when I was at university, a friend did their undergraduate thesis on neural networks. Among our cohort it was seen as a weird choice and a bit of a dead-end from the last AI winter.
How does it work if they get 10x better in 10 years ? Everything else will have already moved on and the actual technology shift will come from elsewhere.
Basically, what if GenAI is the Minitel and what we want is the internet.
No doubt from me that it’s a sigmoid, but how high is the plateau? That’s also hard to know from early in the process, but it would be surprising if there’s not a fair bit of progress left to go.
Human brains seem like an existence proof for what’s possible, but it would be surprising if humans also represent the farthest physical limits of what’s technologically possible without the constraints of biology (hip size, energy budget etc).
Biological muscles are proof that you can make incredibly small and forceful actuators. But the state of robotics is nowhere near them, because the fundamental construction of every robotic actuator is completely different.
We’ve been building actuators for 100s of years and we still haven’t got anything comparable to a muscle. And even if you build a better hydraulic ram or brushless motor driven linear actuator you will still never achieve the same kind of behaviour, because the technologies are fundamentally different.
I don’t know where the ceiling of LLM performance will be, but as the building blocks are fundamentally different to those of biological computers, it seems unlikely that the limits will be in any way linked to those of the human brain. In much the same way the best hydraulic ram has completely different qualities to a human arm. In some dimensions it’s many orders of magnitudes better, but in others it’s much much worse.
Biological muscles come with a lot of baggage, very constrained operating environments, and limited endurance.
It’s not just that ‘we don’t know how to build them’, it’s that the actuators aren’t a standalone part - and we don’t know how to build (or maintain/run in industrial enviroments!) the ‘other stuff’ economically either.
I don’t think it’s hard to know. We’re already seeing several signs of being near the plateau in terms of capabilities. Most big breakthrough these days seems to be in areas where we haven’t spent the effort in training and model engineering. Like recent improvements in video generation. So of course we could get improvements in areas where we haven’t tried to use ML yet.
For text generation, it seems like the fast progress was mainly due to feeding the models exponentially more data and exponentially more compute power. But we know that the growth in data is over. The growth in compute has a shifted from a steep curve (just buy more chips) to a slow curve (have to make exponentially more factories if we want exponentially more chips)
Im sure we will have big improvements in efficiency. Im sure nearly everyone will use good LLMs to support them in their work, and they may even be able to do all they need to do on-device. But that doesn’t make the models significantly smarter.
The wonderful thing about a sigmoid is that, just as it seems like it's going exponential, it goes back to linear. So I'd guess we're not going to see 1000x from here - I could be wrong, but I think the low hanging fruit has been picked. I would be surprised in 10 years if AI were 100x better than it is now (per watt, maybe, since energy devoted to computing is essentially the limiting factor)
The thing about the latter 1/3rd of a sigmoid curve is, you're still making good progress, it's just not easy any more. The returns have begun to diminish, and I do think you could argue that's already happening for LLMs.
Progress so far has been half and half technique and brute force. Overall technique has now settled for a few years, so that's mostly in the tweaking phase. Brute force doesn't scale by itself and semiconductors have been running into a wall for the last few years. Those (plus stagnating outcomes) seem decent reasons to suspect the plateau is neigh.
with autonomous vehicles, the narrative of imperceptibly slow incremental change about chasing 9's is still the zeitgeist despite an actual 10x improvement in homicidality compared to humans already existing.
There is a lag in how humans are reacting to AI which is probably a reflexive aspect of human nature. There are so many strategies being employed to minimize progress in a technology which 3 years ago did not exist and now represents a frontier of countless individual disciplines.
This is my favorite thing to point out from the day we started talking about autonomous vehicles on tech sites.
If you took a Tesla or a Waymo and dropped into into a tier 2 city in India, it will stop moving.
Driving data is cultural data, not data about pure physics.
You will never get to full self driving, even with more processing power, because the underlying assumptions are incorrect. Doing more of the same thing, will not achieve the stated goal of full self driving.
You would need to have something like networked driving, or government supported networks of driving information, to deal with the cultural factor.
Same with GenAI - the tooling factor will not magically solve the people, process, power and economic factors.
> You would need to have something like networked driving, or government supported networks of driving information, to deal with the cultural factor.
Or actual intelligence. That observes its surroundings and learns what's going on. That can solve generic problems. Which is the definition of intelligence. One of the obvious proofs that what everybody is calling "AI" is fundamentally not intelligent, so it's a blatant misnomer.
One of my favorite things to question about autonomous driving is the goalposts. What do you mean the “stated goal of full self driving”, which is unachievable? Any vehicle, anywhere in the world, in any conditions? That seems an absurd goal that ignores the very real value in having vehicles that do not require drivers and are safer than humans but are limited to certain regions.
Absolutely driving is cultural (all things people do are cultural) but given 10’s of millions of miles driven by Waymo, clearly it has managed the cultural factor in the places they have been deployed. Modern autonomous driving is about how people drive far more than the rules of the road, even on the highly regulated streets of western countries. Absolutely the constraints of driving in Chennai are different, but what is fundamentally different? What leads to an impossible leap in processing power to operate there?
> What do you mean the “stated goal of full self driving”, which is unachievable? Any vehicle, anywhere in the world, in any conditions? That seems an absurd goal that ignores the very real value in having vehicles that do not require drivers and are safer than humans but are limited to certain regions.
I definitely recall reading some thinkpieces along the lines of "In the year 203X, there will be no more human drivers in America!" which was and still is clearly absurd. Just about any stupidly high goalpost you can think of has been uttered by someone in the world early on.
Anyway, I'd be interested in a breakdown on reliability figures in urban vs. suburban vs. rural environments, if there is such a thing, and not just the shallow take of "everything outside cities is trivial!" I sometimes see. Waymo is very heavily skewed toward (a short list of) cities, so I'd question whether that's just a matter of policy, or whether there are distinct challenges outside of them. Self-driving cars that only work in cities would be useful to people living there, but they wouldn't displace the majority of human driving-miles like some want them to.
I mean, even assuming the technical challenges to self-driving can be solved, it is obvious that there will still be human drivers because some humans enjoy driving, just as there are still people who enjoy riding horses even after cars replaced horses for normal transport purposes. Although as with horses, it is possible that human driving will be seen as secondary and limited to minor roads in the future.
I'd apprecite that we dont hurry past the acknowledgement that self driving will be a cultural artifact. Its been championed as a purely technical one, and pointing this out has been unpopular since day 1, because it didn't gel with the zeitgeist.
As others will attest, when adherence to driving rules is spotty, behavior is highly variable and unpredictable. You need to have a degree of straight up agression, if you want to be able to handle an auto driver who is cheating the laws of physics.
Another example of something thats obvious based on crimes in India; people can and will come up to your car during a traffic jam, tap your chassis to make it sound like there was an impact, and then snatch your phone from the dashboard when you roll your window down to find out what happened.
This is simply to illustrate and contrast how pared down technical intuitions of "driving" are, when it comes to self driving discussions.
This is why I think level 5 is simply not happening, unless we redefine what self driving is, or the approach to achieving it. I feel theres more to be had from a centralized traffic orchestration network that supplements autonomous traffic, rather than trying to solve it onboard the vehicle.
Why couldn’t an autonomous vehicle adapt to different cultures? American driving culture has specific qualities and elements to learn, same with India or any other country.
Do you really think Waymos in SF operate solely on physics? There are volumes of data on driver behavior, when to pass, change lanes, react to aggressive drivers, etc.
Yeah exactly. It’s kind of absurd to take the position that it’s impossible to have “full self driving” because Indian driving is different than American driving. You can just change the model you’re using. You can have the model learn on the fly. There are so many possibilities.
Because this statement, unfortunately, ends up moving the underlying goal posts about what self driving IS.
And the point that I am making, is that this view was never baked into the original vision of self driving, resulting in predictions of a velocity that was simply impossible.
Physical reality does not have vibes, and is more amenable to prediction, than human behavior. Or Cow behavior, or wildlife if I were to include some other places.
Marketers gonna market. But if we ignore the semantics of what full self driving actually means for a minute, there is still a lot of possibilities for self driving in the future. It takes longer than we perceive initially because we don’t have insight into the nuances needed to achieve these things. It’s like when you plan a software project, you think it’s going to take less time than it does because you don’t have a detailed view until you’re already in the weeds.
To quote someone else, if my grandmother had wheels, she would be a bicycle.
This is a semantic discussion, because it is about what people mean when they talk about self driving.
Just ditching the meaning is unfair, because goddamit, the self driving dream was awesome. I am hoping to be proved wrong, but not because we moved our definition.
Carve a separate category out, which articulates the updated assumptions. Redefining it is a cop out and dare I say it, unbecoming of the original ambition.
Frankly, we don't know. That "turning point" that seemed so close for many tech, never came for some of them. Think 3D-printing that was supposed to take over manufacturing. Or self-driving, that is "just around the corner" for a decade now. And still is probably a decade away. Only time will tell if GenAI/LLMs are color TV or 3D TV.
> Think 3D-printing that was supposed to take over manufacturing.
3D printing is making huge progress in heavy industries. It’s not sexy and does not make headlines but it absolutely is happening. It won’t replace traditional manufacturing at huge scales (either large pieces or very high throughput). But it’s bringing costs way down for fiddly parts or replacements. It is also affecting designs, which can be made simpler by using complex pieces that cannot be produced otherwise. It is not taking over, because it is not a silver bullet, but it is now indispensable in several industries.
You're misunderstanding the parent's complaint and frankly the complaints with AI. Certainly 3D printing is powerful and hasn't changed things. But you forgot that 30 years ago people were saying there would be one in every house because a printer can print a printer and how this would revolutionize everything because you could just print anything at home.
The same thing with AI. You'd be blind or lying if you said it hasn't advanced a lot. People aren't denying that. But people are fed up being constantly being promised the moon and getting a cheap plastic replica instead.
The tech is rapidly advancing and doing good. But it just can't keep up with the bubble of hype. That's the problem. The hype, not the tech.
Frankly, the hype harms the tech too. We can't solve problems with the tech if we're just throwing most of our money at vaporware. I'm upset with the hype BECAUSE I like the tech.
So don't confuse the difference. Make sure you understand what you're arguing against. Because it sounds like we should be on the same team, not arguing against one another. That just helps the people selling vaporware
>Think 3D-printing that was supposed to take over manufacturing
This was never the case, and this is obvious to anyone who has ever been to factories that doing mass-produced plastic
>Or self-driving, that is "just around the corner" for a decade now.
But it is really around the corner, all that remains is to accept it. That is, to start building and modifying the road infrastructure and changing the traffic rules to enable effective integration self-driving cars into road traffic.
There's a lot of "when" people are betting on, and not a lot of action to back it. If "when" is 20 years, then I still got plenty career ahead of me before I need to worry about that.
