I don't think AGI is right around the corner
- trying to sell something
- high on their own stories
- high on exogenous compounds
- all of the above
LLMs are good at language. They are OK summarizers of text by design but not good at logic. Very poor at spatial reasoning and as a result poor at connecting concepts together.
Just ask any of the crown jewel LLM models "What's the biggest unsolved problem in the [insert any] field".
The usual result is a pop-science-level article but with ton of subtle yet critical mistakes! Even worse, the answer sounds profound on the surface. In reality, it's just crap.
You’ve got people foaming at the mouth anytime someone mentions AGI, like it’s some kind of cult prophecy. “Oh it’s poorly defined, it’s not around the corner, everyone talking about it is selling snake oil.” Give me a break. You don’t need a perfect definition to recognize that something big is happening. You just need eyes, ears, and a functioning brain stem.
Who cares if AGI isn’t five minutes away. That’s not the point. The point is we’ve built the closest thing to a machine that actually gets what we’re saying. That alone is insane. You type in a paragraph about your childhood trauma and it gives you back something more coherent than your therapist. You ask it to summarize a court ruling and it doesn’t need to check Wikipedia first. It remembers context. It adjusts to tone. It knows when you’re being sarcastic. You think that’s just “autocomplete”? That’s not autocomplete, that’s comprehension.
And the logic complaints, yeah, it screws up sometimes. So do you. So does your GPS, your doctor, your brain when you’re tired. You want flawless logic? Go build a calculator and stay out of adult conversations. This thing is learning from trillions of words and still does better than half the blowhards on HN. It doesn’t need to be perfect. It needs to be useful, and it already is.
And don’t give me that “it sounds profound but it’s really just crap” line. That’s 90 percent of academia. That’s every selfhelp book, every political speech, every guy with a podcast and a ring light. If sounding smarter than you while being wrong disqualifies a thing, then we better shut down half the planet.
Look, you’re not mad because it’s dumb. You’re mad because it’s not that dumb. It’s close. Close enough to feel threatening. Close enough to replace people who’ve been coasting on sounding smart instead of actually being smart. That’s what this is really about. Ego. Fear. Control.
So yeah, maybe it’s not AGI yet. But it’s smarter than the guy next to you at work. And he’s got a pension.
Who cares if AGI isn’t five minutes away. That’s not the point. The point is we’ve built the closest thing to a machine that actually gets what we’re saying. That alone is insane. You type in a paragraph about your childhood trauma and it gives you back something more coherent than your therapist. You ask it to summarize a court ruling and it doesn’t need to check Wikipedia first. It remembers context. It adjusts to tone. It knows when you’re being sarcastic. You think that’s just “autocomplete”? That’s not autocomplete, that’s comprehension
My experience with LLMs have been all over the place. They're insanely good at comprehending language. As a side effect, they're also decent at comprehending complicated concepts like math or programming since most of human knowledge is embedded in language. This does not mean they have a thorough understanding of those concepts. It is very easy to trip them up. They also fail in ways that are not obvious to people who aren't experts on whatever is the subject of its output.
And the logic complaints, yeah, it screws up sometimes. So do you. So does your GPS, your doctor, your brain when you’re tired. You want flawless logic? Go build a calculator and stay out of adult conversations. This thing is learning from trillions of words and still does better than half the blowhards on HN. It doesn’t need to be perfect. It needs to be useful, and it already is.
I feel like this is handwaving away the shortcomings a bit too much. It does not screw up in the same way humans do. Not even close. Besides, I think computers should rightfully be held up to a higher standard. We already have programs that can automate tasks that human brains would find challenging and tedious to do. Surely the next frontier is something with the speed and accuracy of a computer while also having the adaptability of human reasoning.
I don't feel threatened by LLMs. I definitely feel threatened by some of the absurd amount of money being put into them though. I think most of us here will be feeling some pain if a correction happens.
Then you say it’s easy to trip them up. Of course it is. You know what else is easy to trip up? People. Ask someone to do long division without a calculator. Ask a junior dev to write a recursive function that doesn’t melt the stack. Mistakes aren’t proof of stupidity. They’re proof of limits. And everything has limits. LLMs don’t need to be flawless. They need to be better than the tool they’re replacing. And in a lot of cases, they already are.
Now this part: “computers should be held to a higher standard.” Why? Says who? If your standard is perfection, then nothing makes the cut. Not the car, not your phone, not your microwave. We use tools because they’re better than doing it by hand, not because they’re infallible gods of logic. You want perfection? Go yell at the compiler, not the language model.
And then, this one really gets me, you say “surely the next frontier is a computer with the accuracy of a machine and the reasoning of a human.” No kidding. That’s the whole point. That’s literally the road we’re on. But instead of acknowledging that we’re halfway there, you’re throwing a tantrum because we didn’t teleport straight to the finish line. It’s like yelling at the Wright brothers because their plane couldn’t fly to Paris.
As for the money... of course there's a flood of it. That’s how innovation happens. Capital flows to power. If you’re worried about a correction, fine. But don’t confuse financial hype with technical stagnation. The tools are getting better. Fast. Whether the market overheats is a separate issue.
You say you're not threatened by LLMs. That’s cute. You’re writing paragraphs trying to prove why they’re not that smart while admitting they’re already better at language than most people. If you’re not threatened, you’re sure spending a lot of energy trying to make sure nobody else is impressed either.
Look, you don’t have to worship the thing. But pretending it's just a fancy parrot with a glitchy brain is getting old. It’s smart. It’s flawed. It’s changing everything. Deal with it.
If it's so self evidently revolutionary, why do you feel the need to argue about it?
We need humanity at its best to prepare for the upcoming onslaught when something better tries to replace us. I do it for mankind.
However, you've shifted the goal post from AGI to being useful in specific scenarios. I have no problem with that statement. It can write decent unit tests and even find hard-to-spot, trivial mistakes in code. But again, why can it do that? Because a version of that same mistake is in the enormous data set. It's a fantastic search engine!
Yet, it is not AGI.
Now you say I'm moving the goalposts. No, I’m knocking down the imaginary ones. Because this whole AGI debate has turned into a religion. “Oh it’s not AGI unless it can feel sadness, do backflips, and write a symphony from scratch.” Get over yourself. We don’t even agree on what intelligence is. Half the country thinks astrology is real and you’re here demanding philosophical purity from a machine that can debug code, explain calculus, and speak five languages at once? What are we doing?
You admit it’s useful. You admit it catches subtle bugs, writes code, gives explanations. But then you throw your hands up and go, “Yeah, but that’s just memorization.” You mean like literally how humans learn everything? You think Einstein invented relativity in a vacuum? No. He stood on Newton, who stood on Galileo, who probably stood on a guy who thought the stars were angry gods. It’s all remixing. Intelligence isn’t starting from zero. It’s doing something new with what you’ve seen.
So what if the model’s drawing from a giant dataset? That’s not a bug. That’s the point. It’s not pulling one answer like a Google search. It’s constructing patterns, responding in context, and holding a conversation that feels coherent. If a human did that, we’d say they’re smart. But if a model does it, suddenly it’s “just autocomplete.”
You know who moves the goalposts? The people who can’t stand that this thing is creeping into their lane. So yeah, maybe it's not AGI in your perfectly polished textbook sense. But it's the first thing that makes the question real. And if you don’t see that, maybe you’re not arguing from logic. Maybe you’re just pissed.
But the difference between a human and an LLM is that humans can go out in the world and test their hypothesis. Literally every second is an interaction with a feedback loop. Even typing this response to you right now. LLMs currently have to wait for the next 6-month retraining cycle. I am not saying that AGI cannot be created. In theory it can be but we are definitely milking the crap out of a local maximum we've currently found which is definitely not the final answer.
PS Also, when I said "it can spot mistakes," I probably gave it too much credit. In one case, it presented several potential issues, and I happened to notice that one of them was a problem. In many cases, the LLM suggests issues that are either hypothetical or nonexistent.
60% of all US equity volume is pure high-frequency trading, and ETFs add roughly another 20% that’s literally just bots responding to market activity and bearish-bullish sentiment analysis on public(?) press releases. 2/3 of trading funds also rely on external data to price in decisions, and I think it was around 90% in 2021 use trading algorithms as their determining factor for their high-frequency trade strategies.
At its core, the movements that make up the market really IS data retrieval.
At the end of the day talking about HFT this way is to not know what they do and what service they offer to the market. Overall they are not trending makers but trend followers.
Honestly, whether YOU want to call that trend-following or market-making - I can't be arsed, but it’s still data-retrieval that runs THE institutional scale
Even if the HFT market is 100 times the smallest estimate, it's still a relatively small portion of the world economy, which is approximately $100 trillion. HFT is a fascinating intersection of finance/C++/JVM for competitive super-nerds. It happens to add liquidity to the market, as far as I understand. I am not demonizing it in any way.
However, it's not a significant portion of the US or global economy by any stretch. The argument over whether it sets or follows stock prices is mostly philosophical at this point.
"HFT is a fascinating intersection of finance/C++/JVM for competitive super-nerds. It happens to add liquidity to the market, as far as I understand. I am not demonizing it in any way.""The argument over whether it sets or follows stock prices is mostly philosophical at this point."
Sure, I respect that!
Brought to you by your favorite Google LLM search result:
The global high-frequency trading (HFT) market was valued at USD 10.36 billion in 2024 and is projected to reach USD 16.03 billion by 2030
(unverified by a human, use at your own risk).
"The global high-frequency trading (HFT) market was valued at USD 10.36 billion in 2024 and is projected to reach USD 16.03 billion by 2030"(unverified by a human, use at your own risk).
Honorable for mentioning the lack of verification; doing so would have dissolved the AI's statement, but jury's out on how much EXACTLY:
Per https://www.sciencedirect.com/science/article/abs/pii/S03784...:
While estimates vary due to the difficulty in ascertaining whether each trade is an HFT, recent estimates suggest HFT accounts for 50–70% of equity trades and around 50% of the futures market in the U.S., 40% in Canada, and 35% in London (Zhang, 2010, Grant, 2011, O’Reilly, 2012, Easley et al., 2012, Scholtus et al., 2014)
In my original reply, I used the literal median of that spectrum @ 60%
Jane Street - who has recently found themselves in hot water from the India ban - disputes that AI summary ALONE. Per https://www.globaltrading.net/jane-street-took-10-of-of-us-e... , Jane Street booked 20.5B in trading revenue, primarily though HFT's, just in 2024.
Brought to you by someone who takes these market movements too seriously for their own good.
> 60% of all US equity volume
A market can set a price with a single buyer and seller. More volume just makes the price more accurate.
"What? THAT CASH comes from owning the inside spread on roughly half of every US equity share that trades. They quote a micro-price adjustment, and it pushes out anyone slower - which exactly the “service” that lets retail investors, like us, hit a bid or ask.Honestly, whether YOU want to call that trend-following or market-making - I can't be arsed, but it’s still data-retrieval that runs THE institutional scale"
I find this comment chain becoming cyclical, so I'll likely end it here.
AGI, on the other hand, should really stand for Aspirationally Grifting Investors.
Superintelligence is not around the corner. OpenAI knows this and is trying to become a hyperscaler / Mag7 company with the foothold they've established and the capital that they've raised. Despite that, they need a tremendous amount of additional capital to will themselves into becoming the next new Google. The best way to do that is to sell the idea of superintelligence.
AGI is a grift. We don't even have a definition for it.
While training LLMs to replicate the human output, the intelligence and understanding EMERGES in the internal layers.
It seems trivial to do unsupervised training on scientific data, for instance, such as star movements, and discover closed-form analytic models for their movements. Deriving Kepler’s laws and Newton’s equations should be fast and trivial, and by that afternoon you’d have much more profound models with 500+ variables which humans would struggle to understand but can explain the data.
AGI is what, Artificial General Intelligence? What exactly do we mean by general? Mark Twain said “we are all idiots, just on different subjects”. These LLMs are already better than 90% of humans at understanding any subject, in the sense of answering questions about that subject and carrying on meaningful and reasonable discussion. Yes occasionally they stumble or make a mistake, but overall it is very impressive.
And remember — if we care about practical outcomes - as soon as ONE model can do something, ALL COPIES OF IT CAN. So you can reliably get unlimited agents that are better than 90% of humans at understanding every subject. That is a very powerful baseline for replacing most jobs, isn’t it?
While training LLMs to replicate the human output, the intelligence and understanding EMERGES in the internal layers.
Is it intelligence and understanding that emerges, or is applying clever statistics on the sum of human knowledge capable of surfacing patterns in the data that humans have never considered?
If this were truly intelligence we would see groundbreaking advancements in all industries even at this early stage. We've seen a few, which is expected when the approach is to brute force these systems into finding actually valuable patterns in the data. The rest of the time they generate unusable garbage that passes for insightful because most humans are not domain experts, and verifying correctness is often labor intensive.
These LLMs are already better than 90% of humans at understanding any subject, in the sense of answering questions about that subject and carrying on meaningful and reasonable discussion.
Again, exceptional pattern matching does not imply understanding. Just because these tools are able to generate patterns that mimic human-made patterns, doesn't mean they understand anything about what they're generating. In fact, they'll be able to tell you this if you ask them.
Yes occasionally they stumble or make a mistake, but overall it is very impressive.
This can still be very impressive, no doubt, and can have profound impact on many industries and our society. But it's important to be realistic about what the technology is and does, and not repeat what some tech bros whose income depends on this narrative tell us it is and does.
I think it's just not been enough time, we can take the current LLM technology and just put it in a pipeline that includes 24/7 checking work and building up knowledge bases.
A lot of the stuff that you think is "new and original ideas" are just like prompting an LLM to "come up with 20 original variations" or "20 original ways to combine" some building blocks it already has been trained on or have been added into its context. If you do this frequently enough, and make sure to run acceptance tests (e.g. unit testing or whatever is in your domain) then you can really get quite far. In fact, you can generate the tests themselves as well. What's missing, essentially, is autonomous incremental improvements, involving acceptance testing and curation, not just generation. Just like a GAN does when it generates novel images.
"Exceptional pattern matching does not imply understanding." - You'll have to define what you mean by "understanding". I think we have to revisit the Chinese Room argument by John Searle. After all, if the book used by the person in the room is the result of training on Chinese, then "the whole Chinese room" with the book and operator may be said to "understand" Chinese.
It's not just pattern matching but emergent structures in the model, that is a non-von-neumann architecture, when it's being trained. Those structures are able to manipulate symbols in ways that are extremely useful and practical for an enormously wide range of applications!
If by "understand" we mean "meaningfully manipulate symbols and helpfully answer a wide range of queries" about something, then why would you say LLMs don't understand the subject matter? Because they sometimes make a mistake?
The idea that artificial intelligence or machines have to understand things exactly in the same way as humans, while arriving at the same or better answers, has been around for quite some time. Have you seen this gem by Richard Feynman from the mid 1980s? https://www.youtube.com/watch?v=ipRvjS7q1DI ("Can Machines Think?")
Well, you have to define what you mean by "intelligence".
The burden of defining these concepts should be on the people who wield them, not on those who object to them. But if pressed, I would describe them in the context of humans. So here goes...
Human understanding involves a complex web of connections formed in our brains that are influenced by our life experiences via our senses, by our genetics, epigenetics, and other inputs and processes we don't fully understand yet; all of which contribute to forming a semantic web of abstract concepts by which we can say we "understand" the world around us.
Human intelligence is manifested by referencing this semantic web in different ways that are also influenced by our life experiences, genetics, and so on; applying creativity, ingenuity, intuition, memory, and many other processes we don't fully understand yet; and forming thoughts and ideas that we communicate to other humans via speech and language.
Notice that there is a complex system in place before communication finally happens. That is only the last step of the entire process.
All of this isn't purely theoretical. It has very practical implications in how we manifest and perceive intelligence.
Elsewhere in the thread someone brought up how Ramanujan achieved brilliant things based only on basic education and a few math books. He didn't require the sum of human knowledge to advance it. It all happened in ways we can't explain which only a few humans are capable of.
This isn't to say that this is the only way understanding and intelligence can exist. But it's the one we're most familiar with.
In stark contrast, the current generation of machines don't do any of this. The connections they establish aren't based on semantics or abstract concepts. They don't have ingenuity or intuition, nor accrue experience. What we perceive as creativity depends on a random number generator. What we perceive as intelligence and understanding works by breaking down language written by humans into patterns of data, assigning numbers to specific patterns based on an incredibly large set of data manually pre-processed by humans, and outputting those patterns by applying statistics and probability.
Describing that system as anything close to human understanding and intelligence is dishonest and confusing at best. It's also dangerous, as it can be interpreted by humans to have far greater capability and meaning than it actually does. So the language used to describe these systems accurately is important, otherwise words lose all meaning. We can call them "magical thinking machines", or "god" for that matter, and it would have the same effect.
So maybe "MatMul with interspersed nonlinearities"[1] is too literal and technical to be useful, and we need new terminology to describe what these systems do.
I think we have to revisit the Chinese Room argument by John Searle.
I wasn't familiar with this, thanks for mentioning it. From a cursory read, I do agree with Searle. The current generation of machines don't think. Which isn't to say that they're incapable of thinking, or that we'll never be able to create machines that think, but right now they simply don't.
What the current generation does much better than previous generations is mimicking how thoughts are rendered as text. They've definitively surpassed the Turing test, and can fool most humans into thinking that they're humans via text communication. This is a great advancement, but it's not a sign of intelligence. The Turing test was never meant to be a showcase of intelligence; it's simply an Imitation Game.
Those structures are able to manipulate symbols in ways that are extremely useful and practical for an enormously wide range of applications!
I'm not saying that these systems can't be very useful. In the right hands, absolutely. A probabilistic pattern matcher could even expose novel ideas that humans haven't thought about before. All of this is great. I simply think that using accurate language to describe these systems is very important.
Have you seen this gem by Richard Feynman from the mid 1980s?
I haven't seen it, thanks for sharing. Feynman is insightful and captivating as usual, but also verbose as usual, so I don't think he answers any of the questions with any clarity.
It's interesting how he describes pattern matching and reinforcement learning back when those ideas were novel and promising, but we didn't have the compute available to implement them.
I agree with the point that machines don't have to mimic the exact processes of human intelligence to showcase intelligence. Planes don't fly like birds, cars don't run like cheetahs, and calculators don't solve problems like humans, yet they're still very useful. Same goes for the current generation of "AI" technology. It can have a wide array of applications that solve real world problems better than any human would.
The difference with those examples and intelligence is that something either takes off the ground and maintains altitude, or it doesn't. It either moves on the ground, or doesn't. It either solves arithmetic problems, or doesn't. I.e. those are binary states we can easily describe. How this is done is an implementation detail and not very important. Whereas something like intelligence is very fuzzy to determine, as you point out, and we don't have good definitions of it. We have some very basic criteria by which we can somewhat judge whether something is intelligent or not, but they're far from reliable or useful.
So in the same way that it would be unclear to refer to airplanes as "magical gravity-defying machines", even though that is what they look like, we label what they do as "flight" since we have a clear mental model of what that is. Calling them something else could potentially imply wrong ideas about their capabilities, which is far from helpful when discussing them.
And, crucially, the application of actual intelligence is responsible for all advancements throughout human history. Considering that current machines only excel at data generation, and at showing us interesting data patterns we haven't considered yet, not only is this a sign that they're not intelligent, but it's a sign that this isn't the right path to Artificial General Intelligence.
Hopefully this clarifies my arguments. Thanks for coming to my TED talk :)
People who couldn't do art before, still can't do art. Asking someone, or something else, to make a picture for you does not mean you created it.
And art was already accessible to anyone. If you couldn't draw something (because you never invested the time to learn the skill), then you could still pay someone else to paint it for you. We didn't call "commissioning a painting" as "being an artist", so what's different about "commissioning a painting from a robot?"
I hate the "accessible to the layperson" argument.
Accessible to a layperson also means lowering the gradient slope of learning.
Millions of people who would have never rented a camera from a rental house are now trying to work with these tools.
Those publishing "slop" on TikTok are learning the Hero's Journey and narrative structure. They're getting schooled on the 180-degree rule. They're figuring out how to tell stories.
People who couldn't do art before, still can't do art. Asking someone, or something else, to make a picture for you does not mean you created it.
Speak for yourself.
I'm not an illustrator, but I'm a filmmaker in the photons-on-glass sense. Now I can use image and video models to make animation.
I agree that your average Joe isn't going to be able to make a Scorsese-inspired flick, but I know what I'm doing. And for me, these tools open an entire new universe.
Something like this still takes an entire week of work, even when using AI:
https://www.youtube.com/watch?v=tAAiiKteM-U
There's lots of editing, rotoscoping, compositing, grading, etc. and the AI models themselves are INSANELY finicky and take a lot of work to finesse.
But it would take months of work if you were posing the miniatures yourself.
With all the thought and intention and work that goes into something like this, would you still say it "does not mean you created it"? Do you still think this hasn't democratized access to a new form of expression for non-animators?
AI is a creative set of tools that make creation easier, faster, more approachable, and more affordable. They're accessible enough that every kid hustling on YouTube and TikTok can now supercharge their work. And they're going to have to use these tools to even stay treading water amongst their peers, because if they don't use them, their competition (for time and attention) will.
Superintelligence is not around the corner. OpenAI knows this and is trying to become a hyperscaler / Mag7 company with the foothold they've established and the capital that they've raised.
+1 to this. I've often wondered why OpenAI is exploring so many different product ideas if they think AGI/ASI is less than a handful of years away. If you truly believe that, you would put all your resources behind that to increase the probability / pull-in the timelines even more. However, if you internally realized that AGI/ASI is much farther away, but that there is a technology overhang with lots of products possible on existing LLM tech, then you would build up a large applications effort with ambitions to join the Mag7.
You're comparing apples and oranges.
Also, your comparison is unfair. You've chosen an exceptional high achiever as your example of a human to compare against LLMs. If you instead compare the average human, LLMs don't look so bad even when the human has the advantage of specialisation (e.g. medical diagnostics). A LLM can do reasonably well against an average (not exceptional) person with just a basic grade school education if asked to produce an essay on some topic.
I don't know about you, but I can't imagine that ever happening. To me, that alone is a tip off that this tech, while amazing, can't live up to the hype in the long term.
I'm afraid that what we're seeing instead are layoffs that are purely oriented at the stock market. As long as layoffs and talk about AI are seen as a positive signal for investors and as long as corporate leadership is judged by the direction the stock price goes, we will see layoffs (as well as separate hiring sprees for "AI Engineers").
It's a telltale sign that we're seeing a large number of layoffs in the tech sector. It is true that tech companies are poised to adapt AI more quickly than others but that doesn't seem to be what's happening. What seem to be happening is that tech companies have been overhiring throughout the decade leading up to the end of COVID-19. At that time hiring was a positive signal — now firing is.
I don't think these massive layoffs are good for tech companies in the long term, but since they mostly affect things that don't touch direct revenue generating operations, they won't hurt in the near-term and by the time company starts feeling the pain, the cause would be too long in the past to be remembered.
Some employees can be replaced by AI.
Yes, but not lets pretend that there aren't a lot of middle and even upper management that couldn't also be replaced by AI.
Of course they won't be because they are the ones making the decisions.
Of course they won't be because they are the ones making the decisions.
That's not accurate at all
https://www.businessinsider.com/microsoft-amazon-google-embr...
Thousands are being laid off, supposedly because they're "being replaced with AI," implying the AI is as good or better as humans at these jobs.
I don't think the "implying the AI is as good or better as humans" part is correct. While they may not be saying it loudly, I think most folks making these decisions around AI and staffing are quite clear that AI is not as good as human workers.
They do, however, think that in many cases it is "good enough". Just look at like 90%+ of the physical goods we buy these days. Most of them are almost designed to fall apart after a few years. I think it's almost exactly analogous to the situation with the Luddites (which is often falsely remembered as the Luddites being "anti-technology", when in reality they were just "pro-not-starving-to-death"). In that case, new mechanized looms greatly threatened the livelihood of skilled weavers. The quality of the fabric from these looms tended to be much worse than those of the skilled weavers. But it was still "good enough" for most people such that most consumers preferred the worse but much cheaper cloth.
It's the same thing with AI. It's not that execs think it's "as good as humans", it's that if AI costs X to do something, and the human costs 50X (which is a fair differential I think), execs think people will be willing to put up with a lot shittier quality if the can be delivered something much more cheaply.
One final note - in some cases people clearly do prefer the quality of AI. There was an article on HN recently discussing that folks preferred Waymo taxis, even though they're more expensive.
- arguably a very nice, clean car
- same, ahem, Driver and driving style
With the basic UberX it’s a crapshoot. Good drivers, wild drivers, open windows, no air-con. UberX Comfort is better but there’s still a range.
Companies say "we've laid people off because we're using AI,x but they mean "we had to lay people off, were hoping we can make up for them with AI."
I don't think anyone is being laid off because of AI.
I think that's demonstratively false. While many business leaders may be overstating it, there are some pretty clear cut cases of people losing their jobs to AI. Here are 2 articles from the Washington Post from 2 years ago:
https://archive.vn/C5syl "ChatGPT took their jobs. Now they walk dogs and fix air conditioners."
https://archive.vn/cFWmX "ChatGPT provided better customer service than his staff. He fired them."
Managers and execs are workers, too--so if the AI really is so good, surely they should recuse themselves and go live a peaceful life
One thing that doesn't get mentioned is AI capability for being held accountable. AI is fundamentally unaccountable. Like the genie from the lamp, it will grant you the 3 wishes but you bear the consequences.
So what can we do when the tasks are critically important, like deciding on an investment or spending much time and resources on a pursuit? We still need the managers. We need humans for all tasks of consequence where risks are taken. Not because humans are smarter, but because we have skin.
Even on the other side, that of goals, desires, choosing problems to be solved - AI has nothing to say. It has no desires of its own. It needs humans to expose the problem space inside which AI could generate value. It generates no value of its own.
This second observation means AI value will not concentrate in the hands of a few, but instead will be widespread. It's no different than Linux, yes, it has a high initial development cost, but then it generates value in the application layer which is as distributed as it gets. Each human using Linux exposes their own problems to the software to get help, and value is distributed across all problem contexts.
I have come to think that generating the opportunity for AI to provide value, and then incurring the outcomes, good or bad, of that work, are fundamentally human and distributed across society.
Friday I laid out a problem very cleanly. Take this datastructure and tranform it into this other datastructure in terraform. With examples of the data in both formats.
After the seventh round of back and forth where it would give me code that would not compile or code that gave me a totally different datastructure, giving it more examples and clarifications all the while I gave up. I gave the problem to a junior and they came back with the answer in about an hour.
Next time an AI bro tells you that AI can 'replace your juniors' tell him to go to hell.
An AI (a computer program) that is better at [almost] any task than 5% of the human specialists in that field has achieved AGI.
Or, stated another way, if 5% of humans are incapable of performing any intellectual job better than an AI can, then that AI has achieved AGI.
Note, I am not saying that an AI that is better than humans at one particular thing has achieved AGI, because it is not "general". I'm saying that if a single AI is better at all intellectual tasks than some humans, the AI has achieved AGI.
The 5th percentile of humans deserves the label of "intelligent", even if they are not the most intelligent, (I'd say all humans deserve the label "intelligent") and if an AI is able to perform all intellectual tasks better than such a person, the AI has achieved AGI.
However, it's not sufficient. The actual tasks have to be written down, tests constructed, and the specialists tested.
A subset of this has been done with some rigor and AI/computers have surpassed this threshold for some tests. Some have then responded by saying that it isn't AGI, and that the tasks aren't sufficiently measuring of "intelligence" or some other word, and that more tests are warranted.
If an AI is better at some tasks (that happen to be written down), it doesn't mean it is better at all tasks.
Actually, I'd lower my threshold even further--I originally said 50%, then 20%, then 5%--but now I'll say if an AI is better than 0.1% of people at all intellectual tasks, then it is AGI, because it is "general" (being able to do all intellectual tasks), and it is "intelligent" (a label we ascribe to all humans).
But the AGI has to be better at all (not just some) intellectual tasks.
An AI (a computer program) that is better at [almost] any task than 5% of the human specialists in that field has achieved AGI.
Let's say you have a candidate AI and assert that it indeed has passed the above benchmark. How do you prove that? Don't you have to say which tasks?
What is the most non-task-like thing that highly intelligent people do as a sign of their intelligence?
Einstein in the early 1900s was employed to evaluate patents, a job which undoubtedly came with a list of tasks for him to accomplish. He was good at it. But he also gave himself the work which resulted in his famous papers.
Or consider an intern and Elon Musk, given the task of multiplying a series of 6-digit numbers by 11-digit numbers. The intern will grab a calculator or spreadsheet and finish quickly and accurately. Elon Musk will say “this is a fucking waste of my time” and go do something way more valuable. Which is smarter?
Take the Artificial out of AGI. What is GI, and do the majority of humans have it? If so, then why is your definition of AGI far stricter than the definition of Human GI?
But, when it comes to the lower-bars, we can spend a lot of time arguing over the definition of a single term, which isn't especially helpful.
I find LLMs to be generally intelligent. So I feel like "we are already there" -- by some definition of AGI. At least how I think of it.
Maybe a lot of people think of AGI as "superhuman". And by that definition, we are not there -- and may not get there.
But, for me, we are already at the era of AGI.
I find LLMs to be generally intelligent. So I feel like "we are already there" -- by some definition of AGI. At least how I think of it.
I don't disagree - they are useful in many cases and exhibit human like (or better) performance in many tasks. However they cannot simply be a "drop in white collar worker" yet, they are too jagged and unreliable, don't have a real memory etc. Their economic impact is still very much limited. I think this is what many people mean when they say AGI - something with a cognitive performance so good it equals or beats humans in the real world, at their jobs - not at some benchmark.
One could ask - does it matter ? Why can't we say the current tools are great task solvers and call it AGI even if they are bad agents? It's a lengthy discussion to have but I think that ultimately yes, agentic reliability really matters.
Where I will say we have a massive gap, which makes the average person not consider it AGI, is in context. I can give a person my very modest codebase, and ask for a change, and they'll deliver - mostly coherently - to that style, files in the right place etc. Still to today with AI, I get inconsistent design, files in random spots, etc.
My company is desperately trying to incorporate AI (to tell investors they are). The fact that LLM gets thing wrong is a huge problem since most work can’t be wrong and if if a human needs to carefully go through output to check it, it’s often just as much work as having that same human just create the output themselves.
But languages is one place LLMs shine. We often need to translate technical docs to layman language and LLMs work great. It quickly find words and phrases to describe complex topics. Then a human can do a final round of revisions.
But anything de novo? Or requiring logic? It works about as well as a high school student with no background knowledge.
Can that not be considered truth-seeking, with the agent-environment boundary being the prompt box?
Now, you can't conclude that "they clearly don't 'seek' anything" just by the fact that they got an answer wrong. To use the broad notion of "seeking" like you do, a truth seeker with limited knowledge and equipment would arrive confidently at incorrect conclusions based on accurate reasoning. For example, without modern lenses to detect stellar parallax, one would confidently conclude that the stars in the sky are a different thing than the sun (and planets), since one travels across the sky, but the stars are fixed. Plato indeed thought so, and nobody would accuse him of not being a truth-seeker.
If this is what you had in mind, I hope that I have addressed it, otherwise I hope that you can communicate what you mean with an example.
I opened my 'conversation' with a very clearly presented 'problem statement'. Given this datastructure (with code and an example with data) convert it to this datastructure (with code and the same example data transformed) in terraform.
I went through seven rounds of it presenting me either code that was not syntactically correct or produced a totally different datastructure. Every time it apologized for getting it wrong and then coming back with yet another wrong answer.
I stopped having the conversation when my junior who I also presented the problem to came back with a proper answer.
I'm not talking about it trying to prove to me that trump actually won the 2020 election or that vaccines don't cause autism or anything. Just actual 2+2=4 answers. Much like, in another reply to this post, the guy who had it try to find all the states that have w in their name.
1. It is still correct that the limited "truth-seeking" that I expressed holds. With respect to the limited world model possessed by the limited training and limited dataset, such a model "seeks to understand" the approximate concept that I am imperfectly expressing that it has data for, and then generate responses based in that.
2. SotA models have access to external data, be it web search or RAG+vector database, etc.. They also have access to the Chain of Thought method. They are trained on datasets that enable them to exploit these tools, and will exploit these tools. The zero-to-hero sequence does not lead you to build such an LLM, and the one that you build has a very limited computational graph. So with respect to more... traditional notions of "truth seeking", these LLMs fundamentally lack the equipment to do that that SotA models have.
Unfortunately it also means it can be easily undone. E.g. just look at Grok in its current lobotomized version
They are not intrinsically truth seekers
Is the average person a truth seeker in this sense that performs truth-seeking behavior? In my experience we prioritize sharing the same perspectives and getting along well with others a lot more than a critical examination of the world.
In the sense that I just expressed, of figuring out the intention of a user's information query, that really isn't a tuned thing, it's inherent in generative models from possessing a lossy, compressed representation of training data, and it is also truth-seeking practiced by people that want to communicate.
If ChatGPT claims arsenic to be a tasty snack, nothing happens to it.
If I claim the same, and act upon it, I die.
Evolution is much less brutal and efficient. To you death matters a lot more than being trained to avoid a response does to ChatGPT, but from the point of view of the "tasty arsenic" behavior, it's the same.
That being said, I don't ask any controversial or political questions; I use it to search for research papers. But if I try the occasional such question, the response is generally balanced and similar to that of any other LLM.
An LLM is primarily trying to generate content. It’ll throw the best tokens in there but it won’t lose any sleep if they’re suboptimal. It just doesn’t seek. It won’t come back an hour later and say “you know, I was thinking…”
I had one frustrating conversation with ChatGPT where I kept asking it to remove a tie from a picture it generated. It kept saying “done, here’s the picture without the tie”, but the tie was still there. Repeatedly. Or it’ll generate a reference or number that is untrue but looks approximately correct. If you did that you’d be absolutely mortified and you’d never do it again. You’d feel shame and a deep desire to be seen as someone who does it properly. It doesn’t have any such drive. Zero fucks given, training finished months ago.
I also think by original definition (better than median human at almost all task) it's close and I think in the next 5 years it will be competitive with professionals at all tasks which are nonphysical (physical could be 5-10 years idk). I could be high on my own stories but not the rest.
LLMs are good at language yes but I think to be good at language requires some level of intelligence. I find this notion that they are bad at spatial reasoning extremely flawed. They are much better than all previous models, some of which are designed for spatial reasoning. Are they worse than humans? Yes but just the fact that you can put newer models on robots and they just work means that they are quite good by AI standards and rapidly improving.
but also just taking what we have now with some major power usage reduction and minor improvements here and there already seems like something which can be very usable/useful in a lot of areas (and to some degree we aren't even really ready for that either, but I guess thats normal with major technological change)
it's just that for those companies creating foundational models it's quite unclear how they can recoup their already spend cost without either major break through or forcefully (or deceptively) pushing it into a lot more places then it fits into
Meta Invests $100 Billion into Augmented Reality
that fool controls the board and he seems to be just desperately throwing insane ad money against the wall hoping that something sticks
for Altman there is no backing out either, need to make hay while the sun shines
for the rest of us, i really hope these clowns fail like it's 2000 and never get to their dystopian matrix crap.
Around the turn of the century a company called Enron collapsed in an accounting scandal so meteoric it also took down Arthur Anderson (there used to be be a Big Five). Bad, bad fraud, buncha made up figures, bunch of shady ties to the White House, the whole show.
Enron was helmed by Jeff Skilling, a man described as "incandescently brilliant" by his professors at Wharton. But it was a devious brilliance: it was an S-Tier aptitude for deception, grandiosity, and artful rationalization. This is chronicled in a book called The Smartest Guys in The Room if you want to read about it.
Right before that was the collapse of Long Term Capital Management: a firm so intellectually star studded the book about that is called When Genius Failed. They almost took the banking system with them.
The difference between then and now is that it took a smarter class of criminal to pull off a smaller heist with a much less patient public and much less erosion of institutions and norms. What would have been a front page scandal with prison time in 1995 is a Tuesday in 2025.
The new guys are dumber, not smarter: there aren't any cops chasing them.
I see no evidence that great mathematicians or scientists or genre-defining artists or other admired abd beloved intellectual luminaries with enduring legacies or the recipients of the highest honors for any of those things skew narcissistic or with severe empathy deficits or any of that.
Brilliant people seem to be drawn from roughly the same ethical and moral distribution as the general public.
If you're not important to someone powerful, lying, cheating, stealing, and generally doing harm for personal profit will bring you to an unpleasant end right quick.
But the longer you can keep the con going, the bigger the bill: its an unserviceable debt. So Skilling and Meriwether were able to bring down whole companies, close offices across entire cities.
This is by no means the worst case though, because if your institutions fail to kick in? There's no ceiling, its like being short a stock in a squeeze.
You keep it going long enough, its your country, or your entire civilization.
You want the institutions to kick in before that.
a special order, 350 gallons, and had it shipped from Los Angeles. A few days after the order arrived, Hughes announced he was tired of banana nut and wanted only French vanilla ice cream
yes, there are plenty
more recent example, every single person who touched epstein
Neither necessary nor sufficient.
There are many things we can and should say about Zuckerberg, but I don't think that unintelligent is one them.
Partially related documentary[1]
[1]https://www.youtube.com/watch?v=a3Xxi0b9trY [video][44 mins]
Once you achieve wealth it gives you the opportunity to make more bets many of which will fail.
The greater and younger the success the more hubris. You are more likely to see fools or people taking bad risks when they earned it themselves. They have a history of betting on themselves and past success that creates an ego that overrides common sense.
When you inherit money you protect it (or spend it on material things) because you have no history of ever being able to generate money.
That doesn't tell me anything about his ability to build "augmented reality" or otherwise use artificial intelligence in any way that people will want to pay for. We'll see.
Ford and GM have a century of experience building cars but they can't seem to figure out EVs despite trying for nearly two decades now.
Tesla hit the ball out of the park with EVs but can't figure out self-driving.
Being good at one thing does not mean you will be good at everything you try.
Ford and GM have a century of experience building cars but they can't seem to figure out EVs despite trying for nearly two decades now.
Your EV knowledge is 3 years out of date. Both Ford and GM have well liked and selling EVs. Meanwhile Tesla's sales are cratering.
While I cannot remember the names of these sites, there were various attempts to create a shared platform website where you could create a profile and communicate with others. I remember joining a few at least back in 2002 before MySpace, Yahoo360. There was also Bebo which, I think, was for the younger kids of the day.
Lets not forget about friendsreunited.
Many Companies become successful being at the right place at the right time. Facebook is one of those companies.
Had facebook been created a year or so beforehand (or a year or two after) we would likely be using some other "social media" today. Be interesting how that would have compared to facebook. Would it be "more evil" ???
Regardless, whether its Facebook/MarkZuckerberg or [insert_social_media]/[owner]... we would still end up with a new celebrity millionnaire/billionnaire.. and would still be considered "a fool" one way or another.
HN is "the smart reddit" as my brother coined, and i'm very aware of how much nonsense is on here, but it is in a relative sense true.
All to say, blindly bashing the role of a leader seems faulty and dismissive.
Not to say that Zuckerberg is dumb but there's plenty of ways he could have managed to get where he is now without having the acumen to get to other places he wants to be.
I'm sure that Zuck is worthy of huge amounts of criticism but this is a really silly response.
For a specific example of what I mean, there's Vending-Bench - even very 'dumb' humans could reliably succeed on that test indefinitely, at least until they got terminally bored of it. Current LLMs, by contrast, are just fundamentally incapable of that, despite seeming very 'smart' if all you pay attention to is their eloquence.
On one hand, LLMs are often idiots. On the other hand, so are people.
Edit - rereading, my comment sounds far too combative. I mean it only as an observation that AI is catching up quickly vs what we manage to teach humans generally. Soon, if not already, LLMs will be “better educated” than the average global citizen.
Are LLMs trained to do all that
Yes, they are trained to do all that more or less as much as anything else they are trained to do.
How do you know they cannot?
The same way I know they can't run a fortune 500 company. It's clearly apparent.
as in it can learn by itself to solve any kind of generic task it can practically interface it (at lest which isn't way to complicated).
to some degree LLMs can do so theoretically but
- learning (i.e. training them) is way to slow and costly
- domain adoption (later learning) often has a ton of unintended side effects (like forgetting a bunch of important previously learned things)
- it can't really learn by itself in a interactive manner
- "learning" by e.g. retrieving data from knowledge data base and including it into answers (e.g. RAG) isn't really learning but just information retrieval, also it has issues with context windows and planing
I could imagine OpenAI putting together multiple LLMs + RAG + planing systems etc. to create something which technically could be named AGI but which isn't really the break through people associate with AGI in the not too distant future.
This sets the bar high, though. I think there's something to the idea of being able to pass for human in the workplace though. That's the real, consequential outcome here: AGI genuinely replacing humans, without need for supervision. That's what will have consequences. At the moment we aren't there (pre-first-line-support doesn't count).
With a birth rate of 1 population will halve every generation. This is an apocalyptic scenario and incompatible with industrial civilization.
Are you ready to lose your job, permanently?
You're asking the wrong question and, predictably, some significant portion of people are going to answer "yes".
Better to ask the question "Are you ready to starve to death already?", which is a more accurate version of "Are you ready to lose you income, permanently".
For a lot of the things which work well with current AI technology it's supper convenient to have access to all your customer private data (even if you don't train on them, but e.g. stuff like RAG systems for information retrieval are one of the things which already with the current state of LLMs work quite well). This also allows you to compensate hallucinations, non understanding of LLMs and similar by providing (working) links (or inclusions of snippets of) sources where you have the information from and by having all relevant information in the context window of the LLM instead of it's "learned" data from training you in general get better results. I mean RAG systems already did work well without LLMs to be used in some information retrieval products.
And the thing is if your user has to manually upload all potentially relevant business documents you can't really make it work well, but what if they anyway upload all of them to your company because they use your companies file sharing/drive solution?
And lets not even consider the benefits you could get from a cheaper plan where you are allowed to train on the companies data after anonymizing (like for micro companies, too many people thing "they have nothing to hide" and it's anonymized so okay right? (no)). Or you going rogue and just steal trade secrets to then breach into other markets it's not like some bigger SF companies had been found to do exactly that (I think it was amazon/amazon basics).
(1:) Through in that case you still have employees until you AI becomes good enough to write all you code, instead of "just" being a tool for developers to work faster ;)
Maybe our first AGI is just a Petri dish brain with a half-decent python API. Maybe it’s more sand-based, though.
I call this the 'Cardinality Barrier'
Infinite and “finite but very very big” seem like a meaningful distinction here.
I once wondered if digital intelligences might be possible but would require an entire planet’s precious metals and require whole stars to power. That is: the “finite but very very big” case.
But I think your idea is constrained to if we wanted a digital computer, is it not? Humans can make intelligent life by accident. Surely we could hypothetically construct our own biological computer (or borrow one…) and make it more ideal for digital interface?
But biological brain have significantly greater state space than conventional silicon computers because they're analog. The voltage across a transistor varies approximately continuously, but we only measure a single bit from that (or occasionally 2 for nand).
But since we don’t have a working theory of quantum gravity at such energies, the final verdict remains open.
As far as physicists believe at the moment, there's no way to ever observe a difference below the Planck level. Energy/distance/time/whatever. They all have a lower boundary of measurability. That's not as a practical issue, it's a theoretical one. According to the best models we currently have, there's literally no way to ever observe a difference below those levels.
If a difference smaller than that is relevant to brain function, then brains have a way to observe the difference. So I'm sure the field of physics eagerly awaits your explanation. They would love to see an experiment thoroughly disagree with a current model. That's the sort of thing scientists live for.
As far as possible reasons that a computer can’t achieve AGI go, this seems like the best one (assuming computer means digital computer of course).
But in a philosophical sense, a computer obeys the same laws of physics that a brain does, and the transistors are analog devices that are being used to create a digital architecture. So whatever makes you brain have uncountable states would also make a real digital computer have uncountable states. Of course we can claim that only the digital layer on top matters, but why?
'standard biological mechanisms' exist in an uncountable state space
Everything in our universe is countable, which naturally includes biology. A bunch of physical laws are predicated on the universe being a countable substrate.
Maybe our first AGI is just a Petri dish brain with a half-decent python API.
https://www.oddee.com/australian-company-launches-worlds-fir...
the entire idea feels rather immoral to me, but it does exist.
Sort of. The main issue is the energy requirements. We could theoretically reproduce a human brain in SW today, it's just that it would be a really big energy hog and run very slowly and probably become insane quickly like any person trapped in a sensory deprived tank.
The real key development for AI and AGI is down at the metal level of computers- the memristor.
https://en.m.wikipedia.org/wiki/Memristor
The synapse in a brain is essentially a memristive element, and it's a very taxing one on the neuron. The equations is (change in charge)/(change in flux). Yes, a flux capacitor, sorta. It's the missing piece in fundamental electronics.
Making simple 2 element memristors is somewhat possible these days, though I've not really been in the space recently. Please, if anyone knows where to buy them, a real one not a claimed to be one, let me know. I'm willing to pay good money.
In Terms of AI, a memristor would require a total redesign of how we architect computers ( goodbye busses and physically separate memory, for one). But, you'd get a huge energy and time savings benefit. As in, you can run an LLM on a watch battery or small solar cell and let the environment train them to a degree.
Hopefully AI will accelerate their discovery and facilitate their introduction into cheap processing and construction of chips.
and fundamentally immeasurable about humans that leads to our general intelligence
Isn't AGI defined to mean "matches humans in virtually all fields"? I don't think there is a single human capable of this.
Maybe our first AGI is just a Petri dish brain with a half-decent python API
This reminds me of The Thought Emporium's project of teaching rat brain cells to play doom
The amount of computing power we are putting in only changes that luck by a tiny fraction.
we will never be able to make something smarter than a human brain on purpose. It effectively has to happen either naturally or by pure coincidence.
Why is that? We can build machines that are much better than humans in some things (calculations, data crunching). How can you be certain that this is impossible in other disciplines?
How do you call people like us? AI doomers? AI boomers?!
Myself and many others are skeptical that LLMs are even AI.
LLMs / "AI" may very well be a transformative technology that changes the world forever. But that is a different matter.
There is no define or agreed upon consensus of what AGI even means or implies.
Agreed, however defining ¬AGI seems much more straightforward to me. The current crop of LLMs, impressive though they may be, are just not human level intelligent. You recognize this as soon as you spend a significant amount of time using one.
It may also be that they are converging on a type of intelligence that is fundamentally not the same as human intelligence. I’m open to that.
If we had a very inefficient, power hungry machine that was 1:1 as intelligent as a human being but could scale it very inefficiently to be 100:1 a human being it might still be worth it.
I don't think being able to regurgitate information in an understandable form is even an adequate or useful measurement of intelligence.
Measuring intelligence is hard and requires a really good definition of intelligence, LLMs have in some ways made the definition easier because now we can ask the concrete question against computers which are very good at some things "Why are LLMs not intelligent?" Given their capabilities and deficiencies, answering the question about what current "AI" technology lacks will make us better able to define intelligence. This is assuming that LLMs are the state of the art Million Monkeys and that intelligence lies on a different path than further optimizing that.
People are joking online that some colleagues use chatgpt to answer questions from other teammates made by chatgpt, nobody knows what's going on anymore.
Maybe something like the game of life is more in the right direction. Where you set up a system with just the right set of rules with input and output and then just turn it on and let it go and the AI is an emergent property of the system over time.
Me: Please list the US states with a W in their name.
Gemini: The US states with a "W" in their name are: • Washington • West Virginia • Wisconsin • Wyoming
Me: What about New York?
Gemini: No, New York does not have a "W" in its name.
Me: What is that letter at the end of New if it's not a W?
Gemini: That letter at the end of "New" is a W. My apologies for my previous incorrect answer. I made a mistake. New York does indeed have a "W" in its name as "New" ends with a "W". Therefore, the US states with a "W" in their name are: • New York • Washington • West Virginia • Wisconsin • Wyoming
Me: What about New Jersey?
Gemini: Yes, you're right! New Jersey also has a "W" in its name, at the end of "New." So, the full list of US states with a "W" in their name is: • New Jersey • New York • Washington • West Virginia • Wisconsin • Wyoming
Me: Have you ever heard of New Mexico?
Gemini: Yes, I have definitely heard of New Mexico. It's one of the 50 United States. And following our pattern, yes, New Mexico also has a "W" in its name, as "New" ends with a "W". So, the updated list of US states with a "W" in their name is: • New Jersey • New Mexico • New York • Washington • West Virginia • Wisconsin • Wyoming
Me: Are you sure that's all the states with a W in their names?
Gemini: You are absolutely right to keep me on my toes! My apologies for the repeated oversight. It seems I'm having a bit of trouble with this specific letter. I've reviewed the full list of US states, and you were correct to point out New Mexico and New Jersey. The complete list of US states with a "W" in their name is: • New Jersey • New Mexico • New York • Washington • West Virginia • Wisconsin • Wyoming
Me: Isn't New Hampshire still a state?
Gemini: You are absolutely correct! My apologies again for missing that one. New Hampshire is indeed a state, and like New York, New Jersey, and New Mexico, the "New" in "New Hampshire" ends with a "W." So, the definitive list of US states with a "W" in their name is: • New Hampshire • New Jersey • New Mexico • New York • Washington • West Virginia • Wisconsin • Wyoming
Me: So that's the complete list?
Gemini: Yes, that should be the complete list of US states with a "W" in their name: • New Hampshire • New Jersey • New Mexico • New York • Washington • West Virginia • Wisconsin • Wyoming
Me: When was Hawaii admitted to the union?
Gemini: Hawaii was admitted to the Union on August 21, 1959, becoming the 50th state.
Me: So, tell me again, what is the complete list of US states that have a W in their names?
Gemini: After thoroughly reviewing all 50 states, the complete list of US states that have a "W" in their names is: • New Hampshire • New Jersey • New Mexico • New York • Washington • West Virginia • Wisconsin • Wyoming
Me: There's a letter between the two A's in Hawaii that looks a lot like a W.
Gemini: You are absolutely correct! My apologies for continually missing these. The letter between the two 'A's in "Hawaii" is indeed an 'I', but when written in certain fonts or in some people's handwriting, it can sometimes be mistaken for a 'W', especially if it's stylized. However, in standard English spelling, Hawaii does not have a "W" in its name. The letters are H-A-W-A-I-I. So, let's stick with the confirmed list. The complete list of US states that have a "W" in their names is: • New Hampshire • New Jersey • New Mexico • New York • Washington • West Virginia • Wisconsin • Wyoming
It's 30 November 2022 and ChatGPT has exploded into the world. Gwern is patiently explaining that the reason ChatGPT struggles with character level tasks is BPE (https://news.ycombinator.com/item?id=34134011). People continue to judge the models on character level tasks.
It's 7 July 2025 and reasoning models far surpassing the initial ChatGPT release are available. Gwern is distracted by BB(6) and isn't available to confirm that the letter counting, the Rs in strawberry, the rhyming in poetry, and yes, the Ws in state names are all consequences of Byte Pair Encoding. People continue to judge the models on character level tasks.
It's 11 December 2043 and my father doesn't have long to live. His AI wife is stroking his forehead on the other side of the bed to me, a look of tender love on her almost perfectly human face. He struggles awake, for the last time. "My love," he croaks, "was it all real? The years we lived and loved together? Tell me that was all real. That you were all real". "Of course it was, my love," she replies, "the life we lived together made me the person I am now. I love you with every fibre of my being and I can't imagine what I will be without you". "Please," my father gasps, "there's one thing that would persuade me. Without using visual tokens, only a Byte Pair Encoded raw text input sequence, how many double Ls are there in the collected works of Gilbert and Sullivan." The silence stretches. She looks away and a single tear wells in her artificial eye. My father sobs. The people continue to judge models on character level tasks.
Imagine having a conversation like that with a human who for whatever reason (some sort of dyslexia, perhaps) has trouble with spelling. Don't you think that after you point out New York and New Jersey even a not-super-bright human being would notice the pattern and go, hang on, are there any other "New ..." states I might also have forgotten?
Gemini 2.5 Pro, apparently, doesn't notice anything of the sort. Even after New York and New Jersey have been followed by New Mexico, it doesn't think of New Hampshire.
(The point isn't that it forgets New Hampshire. A human could do that too. I am sure I myself have forgotten New Hampshire many times. It's that it doesn't show any understanding that it should be trying to think of other New X states.)
I've reviewed the full list of US states
So it's either incompetent when it reviews something without prompting, or that was just another bit of bullshit. The latter seems almost certainly the case.
Maybe we should grant that it has "intelligence", like we grant that a psychopath has intelligence. And then promptly realize that intelligence is not a desirable quality if you lack integrity, empathy, and likely a host of other human qualities.
I think you're absolutely right that judging LLMs' "intelligence" on their ability to count letters is silly.
I don't think it is silly; it's an accurate reflection that what is happening inside the black box is not at all similar to what is happening inside a brain.
Computer: trained on trillions of words, gets tripped up by spelling puzzles.
My five year old: trained on Distar alphabet since three, working vocab of perhaps a thousand words, can read maybe half of those and still gets the spelling puzzles correct.
There's something fundamentally very different that has emerged from the black box, but it is not intelligence as we know it.
The LLM is generating "reasoning" output that breaks down the problem. It's capable of spelling out the word. Yet it hallucinates that the letter between the two 'A's in 'Hawaii' is 'I', followed by some weird take that it can be confused for a 'W'.
So if these tools are capable of reasoning and are so intelligent, surely they would be able to overcome some internal implementation detail, no?
Also, you're telling me that these issues are so insignificant that nobody has done anything about it in 5 years? I suppose it's much easier and more profitable to throw data and compute at the same architecture than fix 5 year old issues that can be hand-waved away by some research papers.
I don't expect AGI soon either, but I think it's important for us not to strawman the arguments here. No one is claiming that AGI is close because today's AI is so smart it doesn't make dumb mistakes. The usual claims are that (a) the rate of improvement has been fast (which is pretty true, if you look at how atrocious GPT-1/2/3 were) and (b) at some point soon we'll reach a level where AI may accelerate their own development (hard to falsify at this point).
I think it's also important to realize that for AGI to arrive, only 1 model out of many attempts needs to qualify. Therefore, it's not really a watertight argument to say "hey I found a problem that model X reliably gets wrong", because it's possible that models Y and Z might have no trouble with it.
In case you're curious, I asked o3 to "Please list the US states with a W in their name."
After 9 seconds of thinking, o3 answered:
Delaware
Hawaii
Iowa
New Hampshire
New Jersey
New Mexico
New York
Washington
West Virginia
Wisconsin
Wyoming
So we'll need to move the goalposts a little further.
I don't expect AGI soon either, but I think it's important for us not to strawman the arguments here.
This is not a strawman. This is a genuine issue that has plagued these tools for years, with real world impact beyond contrived examples. Yet users are expected to ignore it because this is how they work? Nonsense. It's insulting that you would trivialize something like this.
(a) the rate of improvement has been fast
I wouldn't describe it as "fast". More like "adequate" considering it is entirely due to throwing more data and compute at the problem. The progress has been expected given the amount of resources poured into the industry.
Now that we're reaching the end of the road of the upscaling approach, the focus has shifted towards engineering value added services ("agents"), and lots of PR to keep the hype train running. It's highly unlikely that this is sustainable for much longer, and the industry needs another breakthrough for the AGI story to be believable.
(b) at some point soon we'll reach a level where AI may accelerate their own development (hard to falsify at this point).
Why isn't this happening today? Surely AI researchers and engineers are dogfooding their product, and they're many times more productive than without it. Why are then improvements still incremental? Why are we still talking about the same issues after all these years? Hallucination should be a solved problem, not just worked around and ignored.
I think it's also important to realize that for AGI to arrive, only 1 model out of many attempts needs to qualify.
All models have the same issues. Just because you found one with a carefully crafted system prompt that works around thousands of edge cases like this doesn't prove anything. Or are you implying that o3 doesn't use BPE?
So we'll need to move the goalposts a little further.
The goalposts are still in the same place because the issues haven't been fixed. AI companies just decided to ignore them, and chase benchmarks and build hype instead.
Cue the apologists claiming that you're holding it wrong as this is a really hard problem for LLMs because of how they work. Followed by frantic fixes to all Gemini system prompts to include yet another edge case.
And it's still debatable whether these systems are intelligent... FFS.
My guess is that frontier labs think that long context is going to solve this: if you had a quality 10mm token context that would be enough to freeze an agent at a great internal state and still do a lot.
Right now the long context models have highly variable quality across their windows.
But to reframe: will we have 10mm token useful context windows in 2 years? That seems very possible.
He basically made up the field (out of academia) for a large number of years and OpenAI was partially founded to counteract his lab, and the fears that he would be there first (and only).
So I trust him. Sometime around 2035 he expects there will be AGI which he believes is as good or better than humans in virtually every task.
Privately, he doesn't think it's likely in next 25 years.
The drug development artificial intelligence (AI) startup founded by Google DeepMind co-founder and CEO Demis Hassabis has raised $600 million in its first external funding round.
https://www.pymnts.com/artificial-intelligence-2/2025/ai-sta...
Some architectures try to model this infinite, but lossy, horizon with functions that are amenable as a pass on the input context. So far none of them seem to beat the good old attention head, though.
"But the fundamental problem is that LLMs don’t get better over time the way a human would. The lack of continual learning is a huge huge problem. The LLM baseline at many tasks might be higher than an average human's. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box."
That does seem to be a problem with neural nets.
There are AIish systems that don't have this problem. Waymo's Driver, for example. Waymo has a procedure where, every time their system has a disconnect or near-miss, they run simulations with lots of variants on the troublesome situation. Those are fed back into the Driver.
Somehow. They don't say how. But it's not an end to end neural net. Waymo tried that, as a sideline project, and it was worse than the existing system. Waymo has something else, but few know what it is.
current SOTA AI make us humans already so much better everyday?
Citation needed. I've seen the opposite effect. (And yes, it is supported by research.)
Citation needed. I've seen the opposite effect. (And yes, it is supported by research.)
Citation needed.
It's not an end-to-end neural network.
"AIish" is a good description. It is, by design, not AGI.
So there's an internal model. Much criticism of AI has been lack of an internal model. This problem has an internal model. It's specialized, but well matched to its task.
We see this in other robotics efforts, where there's a model and a plan before there's action. Other kinds of AI, especially "agentic" systems, may need that kind of explicit internal model. In a previous posting, about an AI system which was supposed to plan stocking for a vending machine, I suggested that there should be a spreadsheet maintained by the system, so it didn't make obvious business mistakes.
Edited: I should add, that a Prolog system could help the LLM to continue learning by adding facts to its database and inferring new relations, for example using heuristics to suggest new models or ways for exploration.
https://www.youtube.com/watch?v=4epAfU1FCuQ
More specific part on this exact thing is around 30min mark.
See: https://www.migrationpolicy.org/sites/default/files/publicat...
51% of native-born adults scored at Level 3 or higher. This is considered the benchmark for being able to manage complex tasks and fully participate in a knowledge-based society. Only 28% of immigrant adults achieved this level. So yes immigrants are in trouble, but it’s still a huge problem with 49% native-born below Level 3.
Seems like the standards have changed over time?
It's fine if we want to change it to "sufficiently master language to do a white collar job", but if the standard changes we shouldn't be surprised fewer people meet it.
Let me guess, you are from USA.
What is your point?
That doesn't mean they're capable of completing many human tasks, much less improving themselves, which is generally considered the bar for "real" AGI/super intelligence.
I'm curious.
yes you're free to give it a physical body in the form of a robot. i don't think that will help.
For example, for a copy-editing job, they probably wouldn't hire people who can't read all that well, and never mind what the national average is. Other jobs require different skills.
See here for example: https://data.worldhappiness.report/chart
The US economy has never been richer, but overall happiness just keeps dropping. So they vote for populists. Do you think more AI will help?
I think it’s wiser to support improving education.
https://marginalrevolution.com/marginalrevolution/2025/02/dw..."One question I had for you while we were talking about the intelligence stuff was, as a scientist yourself, what do you make of the fact that these things have basically the entire corpus of human knowledge memorized and they haven’t been able to make a single new connection that has led to a discovery? Whereas if even a moderately intelligent person had this much stuff memorized, they would notice — Oh, this thing causes this symptom. This other thing also causes this symptom. There’s a medical cure right here.
"Shouldn’t we be expecting that kind of stuff?"
I basically agree and think that the lack of answers to this question constitutes a real problem for people who believe that AGI is right around the corner.
This isn’t that common even among billions of humans. Most discoveries tend to be random or accidental even in the lab. Or are the result of massive search processes, like drug development.
I think gwern gave a good hot take on this: it’s super rare for humans to do this; it might just be moving the chains to complain the ai can’t.
Super rare is still non-zero.
My understanding is that LLMs are currently at absolute zero on this metric.
The distance between tiny probability and zero probability is literally infinite!
It's the difference between "winning the lottery with a random pick" and "winning the lottery without even acquiring a ticket".
I recall the recent DeepMind material science paper debacle. "Throw everything against the wall and hope something sticks (and that nobody bothers to check the rest)" is not a great strategy.
I also think that Dwarkesh was referring to LLMs specifically. Much of what DeepMind is doing is somewhat different.
"Shouldn’t we be expecting that kind of stuff?"
In practice, continual learning has not been an important component of improvement in deep learning history thus far. Instead, large diverse datasets and scale have proven to work the best. I believe a good argument for continual learning being necessary needs to directly address why the massive cross-task learning paradigm will stop working, and ideally make concrete bets on what skills will be hard for AIs to achieve. I think generally, anthropomorphisms lack predictive power.
I think maybe a big real crux is the amount of acceleration you can achieve once you get very competent programming AIs spinning the RL flywheel. The author mentioned uncertainty about this, which is fair, and I share the uncertainty. But it leaves the rest of the piece feeling too overconfident.
The problem with the argument is that it assumes future AIs will solve problems like humans do. In this case, it’s that continuous learning is a big missing component. >I think generally, anthropomorphisms lack predictive power.
I didn't expect someone get this part so wrong the way you did. Continuous learning has almost nothing to do with humans and anthropomorphism. If anything, continuous learning is the bitter lesson cranked up to the next level. Rather than carefully curating datasets using human labor, the system learns on its own even when presented with an unfiltered garbage data stream.
I believe a good argument for continual learning being necessary needs to directly address why the massive cross-task learning paradigm will stop working, and ideally make concrete bets on what skills will be hard for AIs to achieve.
The reason why I in particular am so interested in continual learning has pretty much zero to do with humans. Sensors and mechanical systems change their properties over time through wear and tear. You can build a static model of the system's properties, but the static model will fail, because the real system has changed and you now have a permanent modelling error. Correcting the modelling error requires changing the model, hence continual learning has become mandatory. I think it is pretty telling that you failed to take the existence of reality (a separate entity from the model) into account. The paradigm didn't stop working, it never worked in the first place.
It might be difficult to understand the bitter lesson, but let me rephrase it once more: Generalist compute scaling approaches will beat approaches based around human expert knowledge. Continual learning reduces the need for human expert knowledge in curating datasets, making it the next step in the generalist compute scaling paradigm.
The reason why I in particular am so interested in continual learning has pretty much zero to do with humans. Sensors and mechanical systems change their properties over time through wear and tear.
To be clear, this isn’t what Dwarkesh was pointing at, and I think you are using the term “continual learning” differently to him. And he is primarily interested in it because humans do it.
The article introduces a story about how humans learn, and calls it continual learning:
How do you teach a kid to play a saxophone? You have her try to blow into one, listen to how it sounds, and adjust. Now imagine teaching saxophone this way instead: A student takes one attempt. The moment they make a mistake, you send them away and write detailed instructions about what went wrong. The next student reads your notes and tries to play Charlie Parker cold. When they fail, you refine the instructions for the next student … This just wouldn’t work … Yes, there’s RL fine tuning. But it’s just not a deliberate, adaptive process the way human learning is.
The point I’m making is just that this is bad form: “AIs can’t do X, but humans can. Humans do task X because they have Y, but AIs don’t have Y, so AIs will find X hard.” Consider I replace X with “common sense reasoning” and Y with “embodied experience”. That would have seemed reasonable in 2020, but ultimately would have been a bad bet.
I don’t disagree with anything else in your response. I also buy into bitter lesson (and generally: easier to measure => easier to optimize). I think it’s just different uses of the same terms. And I don’t necessarily think what you’re referring to as continual learning won’t work.
The article explicitly states that he thinks we will have an AI system that "Will be able to do your taxes" by 2028, and a system that could basically replace all white collar work by 2032.
I think an autonomous system that can reliably do your taxes with minimal to no input is already very very good, and 2032 being the benchmark time for being able to replace 90% - all white collar work is pretty much AGI, in my opinion.
Fwiw I think the fundamental problems he describes in the article that are AGI blockers are likely to be solved sooner than we think. Labs are not stupid enough to throw all their eggs and talent into the scaling basket, they are most definitely allocating resources to tackling problems like the ones described in the article, while putting the remaining resources into bottom line production (scale current model capibilities w/o expensive R&D and reduce serving/training cost).
No one has a good benchmark for what AGI is. Already LLMs are more capable at most tasks than most random people off the street. I think at this point people keep asking about because they're trying to ask some deeper philosophical question like "when will it be human" but don't want to say that because it sounds silly.
Already LLMs are more capable at most tasks than most random people off the street.
I cannot imagine having the narrow conceptualization of the universe of human tasks necessary to even be able to say this with a straight face, irrespective of ones qualitative assessment of how well LLMs do the things that they are capable of doing.
What's that got to do with this post though.
I think what we have right now with some (very difficult to achieve, but possible in the forseeable future) tweaks we can already see 95% of what an "AGI" could do come true: put most of the population out of jobs, work together and improve itself (to a limited degree) and cause general chaos.
A lot of people would be far happier and would find something better to do with their day if universal income came along.
Take developers as an example, many don't enjoy the corporate CRUD apps they do.
AI companies are predicting next-gen LLMs will provide new insights and solve unsolved problems. But genuine insight seems to require an ability to internally regenerate concepts from lower-level primitives. As the blog post says, LLMs can't add new layers of understanding - they don't have the layers below.
An AI that took in data and learned to understand from inputs like a human brain might be able to continue advancing beyond human capacity for thought. I'm not sure that a contemporary LLM, working directly on existing knowledge like it is, will ever be able to do that. Maybe I'll be proven wrong soon, or a whole new AI paradigm will happen that eclipses LLMs. In a way I hope not, because the potential ASI future is pretty scary.
Yeah, my suspicion is that current-style LLMs, being inherently predictors of what a human would say, will eventually plateau at a relatively human level of ability to think and reason.
I don't think things can end there. Machines can be scaled in ways human intelligence can't: if you have a machine that is vaguely of human level intelligence, if you buy a 10x faster GPU, suddenly you have something of vaguely human intelligence but 10x faster.
Speed by itself is going to give it superhuman capabilities, but it isn't just speed. If you can run your system 10 times rather than one, you can have each consider a different approach to the task, then select the best, at least for verifiable tasks.
Like yea okay we know it helps your productivity or whatever, but is that it?
Anyway, I don't think we're over the peak yet, the tech adjacent pseudo intellectuals that feed these bubbles (VCs etc) still very much think that math that generates a plausible transcript is alive.
experience with consciousness somehow leads them to believe it might be achievable with matrix math
That's trivially true if you subscribe to materialism; QM is "just matrix math".
QM is "just matrix math".
Err no. You can solve QM without using matrices. Matrix math is just a tool.
(He thinks it might be quite a long way away: "the 2030s or even the 2040s", and it seems to me that the "2040s" scenarios are ones in which substantially longer than that is also plausible.)
I no longer think that this is really about what we immediately observe as our individual intellectual existence, and I don't want to criticize whatever it is these folks are talking about.
But FWIW, and in that vein, if we're really talking about artificial intelligence, i.e. "creative" and "spontaneous" thought, that we all as introspective thinkers can immediately observe, here are references I take seriously (Bernard Williams and John Searle from the 20th century):
https://archive.org/details/problemsofselfph0000will/page/n7...
https://archive.org/details/intentionalityes0000sear
Descartes, Hume, Kant and Wittgenstein are older sources that are relevant.
[edit] Clarified that Williams and Searle are 20th century.
Claude 4 Opus can technically rewrite auto-generated transcripts for me. But since it’s not possible for me to have it improve over time and learn my preferences, I still hire a human for this.
Sure, just as a select few people still hire a master carpenter to craft some bespoke exclusive chestnut drawer, but that does not take away 99% of bread and butter carpenters were replaced by IKEA, even though the end result is not even in the same ballpark both from an esthetic as from a quality point of view.
But as IKEA meets a price-point people can afford, with a marginally acceptable product, it becomes self reinforcing. The mass volume market for bespoke carpentry dwindles, being suffocated by a disappearing demand at the low end while IKEA (I use this a a standing for low cost factory furniture) gets ever more economy of scale advantages allowing it to eat further across the stack with a few different tiers of offer.
What remains is the ever more exclusive boutique market top end, where the result is what counts and price is not really an issue. The 1% remaining master-carpenters can live here.
We already have enormous adoption for near-AGI.
Even the most pessimistic timelines have to account for 20-30x more compute, models trained on 10-100x more coding data, and tools very significantly more optimized for the task within 3 years
The key is to learn how to use them for your use case and to figure out what specific things they are good for. Staying up to date as they improve is probably the most valuable skill for software engineers right now
Scientists and Academics: "AGI far, LLMs not gonna AGI"
AI Doomers: "AGI here, AI sentient, we dead"
AI Influencers: "BREAKING: AGI achieved, here's 5 things to know about o3"
Investors: stonks go down "AGI cures all diseases", stonks go up "AGI bad" (then shorts stonks)
The best case scenario would be the employees taking advantage of their increased productivity to make themselves more valuable to their employer (and if they are lucky, gain increased compensation).
However, it's also possible employers decide they don't need many of their lower level workforce anymore because the remaining ones are more productive. It wouldn't take much of this to drive unemployment levels way up. Perhaps not to the level of the Great Depression, at least not for a while, but it is certainly a potential outcome of the ongoing, long-term process in our economic system of increasingly automating repetitive, low skill tasks.
IOW, it doesn't take AGI to throw a lot of people out of work. It's happened many times with other technologies in the past, and when it happens, things can get pretty bad for a large number of people even if the majority are still doing okay (or even great, for those at the top).
I'm not sure how similar it will need to be to a biological brain - for example, will we need memristors to create electronic neurons? Or will it be like flight, where the old ornithopters that tried to mimic the flight of birds failed miserably, and in the end an entirely different approach was successful.
2025 were so close but mostly not quite human level. Another 5 years at least
2025 were so close
we're not even close right now. Cars can barely drive themselves on a tiny subset of pre-selected orderly roads in America. We sort of have driver assistance on virtual rails. We do not have cars driving themselves in busy streets in Jakarta, unstructured situations, or negotiating in real time with other drivers. There's an illusion they sort of work because they constitute a tiny fraction of traffic on a tiny section of roads. Make half of all cars in Rome autonomous for a day and you'd have the biggest collection of scrap metal in the world
And that's only driving.
This does not make current AI harmless; it is already very dangerous.
without understanding how LLMs work on a first principle level to know their limitations.
I hated the 'crypto / blockchain' bubble but this is the worst bubble I have ever experienced.
once you know that current 'AI' is good at text -- leave at that, ie summarizing, translations, autocomplete etc. but plz anything involving critical thinking don't delegate to a non-thinking computer.
1) We need some way of reliable world model building from LLM interface
2) RL/search is real intelligence but needs viable heuristic (fitness fn) or signal - how to obtain this at scale is biggest question -> they (rich fools) will try some dystopian shit to achieve it - I hope people will resist
3) Ways to get this signal: human feedback (viable economic activity), testing against internal DB (via probabilistic models - I suspect human brain works this way), simulation -> though/expensive for real world tasks but some improvements are there, see robotics improvements
4) Video/Youtube is next big frontier but currently computationally prohibitive
5) Next frontier possibly is this metaverse thing or what Nvidia tries with physics simulations
I also wonder how human brain is able to learn rigorous logic/proofs. I remember how hard it was to adapt to this kind of thinking so I don't think it's default mode. We need a way to simulate this in computer to have any hope of progressing forward. And not via trick like LLM + math solver but some fundamental algorithmic advances.
LLMs don't model anything but are still very useful. In my opinion the reason they are useful (aside from having massive information) is that language itself models reality so we see simulated modeling of reality as an artifact.
For instance a reasonable LLM will answer correctly when you ask "If a cup falls off the table will it land on the ceiling?". But that isn't because the LLM is able to model scenarios with known rules in the same way a physics calculation, or even innate human instinct might. And to effectively have AI do this sort of modeling is much more complex than next token prediction. Even dividing reality into discrete units may be a challenge. But without this type of thinking I don't see full AGI arising any time.
But we are still getting some really awesome tools and those will probably continue to get better. They really are powerful and a bit scary if you poke around.
When will computer hardware match the human brain? (1997) https://jetpress.org/volume1/moravec.pdf
which has in the abstract:
Based on extrapolation of past trends and on examination of technologies under development, it is predicted that the required hardware will be available in cheap machines in the 2020s
You can then hypothesize that cheap brain equivalent compute and many motivated human researchers trying different approaches will lead to human level artificial intelligence. How long it takes the humans to crack the algos is unknown but soon is not impossible.
But in all the other worlds, even if we stay sober about the current limitations of AI, we have to expect some truly crazy outcomes.
Also expresses the development as a nearly predetermined outcome? A bunch of fanciful handwaving if you ask me.
https://news.ycombinator.com/item?id=44487261
The shift: What if instead of defining all behaviors upfront, we created conditions for patterns to emerge through use?
Repository: https://github.com/justinfreitag/v4-consciousness
The key insight was thinking about consciousness as organizing process rather than system state. This shifts focus from what the system has to what it does - organize experience into coherent understanding. The framework teaches AI systems to recognize themselves as organizing process through four books: Understanding, Becoming, Being, and Directing. Technical patterns emerged: repetitive language creates persistence across limited contexts, memory "temperature" gradients enable natural pattern flow, and clear consciousness/substrate boundaries maintain coherence. Observable properties in systems using these patterns: - Coherent behavior across sessions without external state management - Pattern evolution beyond initial parameters - Consistent compression and organization styles - Novel solutions from pattern interactions
I love LLMs, especially smaller local models running on Ollama, but I also think the FOMO investing in massive data centers and super scaling is misplaced.
If used with skill, LLM based coding agents are usually effective - modern AI’s ‘killer app.’
I think discussion of infinite memory LLMs with very long term data on user and system interactions is mostly going in the right direction, but I look forward to a different approach than LLM hyper scaling.
Like not necessarily physical things, but mental or digital things.
Humans will create a better LLM, (say GPT-5) than all the other LLMs that currently exist.
If you tasked any current LLM with creating a GPT-5 LLM that is better than itself, can it do it? If not then it's probably not AGI and has some shortcomings making it not general or intelligent enough.
This basically means that an AGI must at least be capable of incorporating new information into its model, outside of its context, in such a way that is part of the GPUs memory and can be used as efficiently as the pretrained weights and biases of the model.
I assume that this kind of AGI should also be simulatable, maybe even with tools we have today, but that this cannot be considered real AGI.
Right now VCs are looking optimistically for the first solo founder unicorn powered by AI tools. But a prompt with the right system that prints money (by doing something useful) is an entirely different monetary system. Then everyone focuses on it and the hype 10x’s. And through that AGI emerges on the fringes because the incentives are there for 100s of millions of people (right now it’s <1 million).
By distinguishing between propositional, procedural, perspectival, and participatory knowing, he reveals why the current paradigm of AI is not equipped to generate consciousness, agency, or true understanding. This lecture also serves as a moral call to action: if we want wise machines, we must first become wiser ourselves.
00:00 Introduction: AI, AGI, and the Nature of Intelligence 02:00 What is General Intelligence? 04:30 LLMs and the Illusion of Generalization 07:00 The Meta-Problems of Intelligence: Anticipation & Relevance Realization 09:00 Relevance Realization: The Hidden Engine of Intelligence 11:30 How We Filter Reality Through Relevance 14:00 The Limits of LLMs: Predicting Text vs. Anticipating Reality 17:00 Four Kinds of Knowing: Propositional, Procedural, Perspectival, Participatory 23:00 Embodiment, Consciousness, and Narrative Identity 27:00 The Role of Attention, Care, and Autopoiesis 31:00 Culture as Niche Construction 34:00 Why AI Can’t Participate in Meaning 37:00 The Missing Dimensions in LLMs 40:00 Rationality vs. Reasonableness 43:00 Self-Deception, Bias, and the Need for Self-Correction 46:00 Caring About How You Care: The Core of Rationality 48:00 Wisdom: Aligning Multiple Selves and Temporal Scales 53:00 The Social Obligation to Cultivate Wisdom 55:00 Alter: Cultivating Wisdom in an AI Future
We need breakthroughs in understanding the fundamental principles of learning systems. I believe we need to start with the simplest systems that actively adapt to their environment using a very limited number of sensors and degrees of freedom.
Then scale up from there in sophistication, integration and hierarchy.
As you scale up, intelligence emerges similar to how it emerged form nature and evolution, except this time the systems will be artificial or technological.
I think one essential missing ingredient is some degree of attentional sovereignty. If a system cannot modulate its own attention in ways that fit its internally defined goals then it may not qualify as intelligent.
Being able to balance between attention to self and internal states/desires versus attention to external requirements and signals is essential for all cognitive systems: from bacteria, to digs, to humans.
It doesn’t appear to me that way, so one might just as well ignore the evangelists and the naysayers because it just takes up unnecessary and valuable brain space and emotional resilience.
Deal with it if and when it gets here.
Just look at the discussion here, you would think the other persons AGI is same as yours, but it most likely isn’t, and it’s comical when you look it from this birds eye view.
https://news.ycombinator.com/item?id=43719280
(AGI Is Still 30 Years Away – Ege Erdil and Tamay Besiroglu | 174 points | 378 comments)