The sigmoids won't save you(astralcodexten.com) |
The sigmoids won't save you(astralcodexten.com) |
Attention is all you need took us by surprise and we don't know how big the wave is let alone if there are other waves behind it.
As for the basis of your objection, this smacks of intellectual gatekeeping. Plenty of good writing is by people who are not academically qualified or a recognized expert in the topic they're writing about. Indeed, very often, this kind of writing is better than writing by experts. Experts often write for other experts, and this can be exclusionary to lay readers. When a non-expert learns about a topic then writes about it for a general audience, they tend to be just a step ahead of the audience, and so the reader is able to learn about the topic by following the process of discovery and reasoning that the author just experienced. Sure, they often get some details or concepts wrong, but the discussion on a site like HN can draw other perspectives, and – very often – contributions from experts, which leads to further expansion in everyone's understanding of the topic.
HN's very ethos is to gratify intellectual curiosity, and this kind of writing is highly compatible with that.
- Making connections to other subjects that an expert would miss. The hall of fame of sigmoid predictions is just excellent, I already know I'm going to be reminded of it some time in the future. Very entertaining way to get the point across.
- Writing about tricky concepts in a very accessible and elegant way, which experts are notoriously bad at doing themselves - they are often optimizing for other specialists.
- Being able to write with an air of speculation and experimentation with ideas that experts and institutions often can't afford. Experts have to maintain their track record; Scott Alexander can say "lol just double the timeline"
Allowing slop articles like this literally prints them evaluation money.
May be reading into things too much, but it is a bit odd.
It's good that you come to HN expecting high standards of content and discussion.
> sCotT aLexAndEr
This counts as a sneer, which is against the guidelines (https://news.ycombinator.com/newsguidelines.html). You may not owe the writer anything but you owe the audience better than this.
> as close as you can come to the modern dressed up version of a eugenicist
Their writing about genetic determinism is a turnoff to me too. But this essay is about a different topic, and a piece of writing by a writer who is known for writing substantively about a variety of topics should be evaluated on its own terms.
This doesn't say much, and the author fights their own points a couple times, suggesting that they maybe didn't think through what they wanted to write until they were in the middle of writing it and started realizing their assumptions didn't match what they expected the data to say.
I really don't get the point of what I just read.
Model reasoning is on an s-curve, which is improving.
Model intelligence is not the same as reasoning. It's a different axis, and one I have not seen much movement on.
See, humans have a recursive form of intelligence which is capable of self-reflection and introspection. LLMs can only reason about tokens which have already been emitted. Humans and LLMs do not share the same form of reasoning, and general human-like intelligence will not arise from the current architecture of LLMs. Therefore it is a mistake to assume that continual improvement on the reasoning scale will result in something that is equivalent enough to humans on the intelligence axis to replace all labor.
But what's going on in here is not that - it's reading tea leaves for maximum dramatic effect.
Strictly speaking, the original paradigm of scaling laws doesn't work any more. The assumption that we could achieve better performance simply through "vertical scaling" ie infusing models with exponentially more parameters and pre-training data, is no longer the driving force of AI progress.
Instead, the industry has pivoted toward inference-time scaling. Rather than relying solely on a massive, static neural network, modern architectures allocate more compute during the actual generation process, allowing the model to "think" and verify its logic dynamically.
Furthermore, the latest state-of-the-art models are no longer pure LLMs; they are compound neuro-symbolic systems that integrate external tools like REPLs, databases, and structured skill documentation to archive things pure LLM vertical parameter scaling was not able to do.
Strictly speaking, "Why do scaling laws work?" is a question about the theoretical reasons the asymptotic decay takes the particular mathematical shape that it does.
All exponentials eventually become sigmoids? Don’t think this can be true without qualifiers.
The issue is that the exponential-looking part of the sigmoid might contain all of human history, sure, but most folks who espouse this theory probably agree that over time everything reaches a steady-enough state to be considered non-exponential, or become oscillatory.
Is the "capability" number on these LLM strengh graphs as tangible?
I think it would be interesting to visit a reality that obeys arbitrary abstractions, but I would personally never go there.
> But if someone claims that the trend toward [X] will never reach some particular scary level, then the burden is on them to explain either:
> If they’re not treating [X] as a black box, and claim to be modeling the dynamics explicitly, then what is their model? Have they calculated the obvious things…
> If they are treating [X] as a black box, why isn’t their default expectation based on Lindy’s Law?
Like, the whole point is that in real life we do actually know things about situations and can model them; we fall back to Lindy's law when we know nothing at all. Further, arguments have justification to deviate from Lindy only when they give specifics about the situation they're modelling.
https://xcancel.com/peterwildeford/status/202963666232244661...
Going to need a big citation for that claim
Ofc "full labor automation" has a certain spread of meaning. A sliver of population will always find ways to hold to a job or run one or many businesses. But there will be "enough" labor automation for it to be a social ticking bomb. That, in fact, does not depend on better models nor better AI than we have today. By 2045 there will be a couple of generations that has been outsourcing their thinking to AI for most of their adult lives. Some of them may still work as legal flesh of sorts, but many won't get to be middle man and will find no job.
Also, if you could replace your senator today by an untainted version of a frontier model (of today), would you do it? Would it be a better ruler? What are the odds of you not wanting to push that button in the next twenty years, after a few more batches of incompetent and self-serving politicians?
Yeah well my prophet says he can beat up your prophet in a fight.
---
Here in reality, I'm not accustomed to taking random predictions without backing evidence as if they were truth.
Lol
I could probably make increasingly larger fires for years if I was willing to burn the entire world.
This is not the context in which I hear about sigmoids vs exponentials. I hear it in regards to “the singularity”, not that AI won’t reach some pre-specified level. You may get AGI, you aren’t getting a singularity.
1. If you're not treating my claim as a black box, explain explicitly what is your model of what the article was about? Are you aware, for example of the last paragraph of the article? I think that WAS what the article was about. Do you have specific opinions on e.g. how I went wrong and where my model differs?
2. If you are treating it as a black box, what's your default expectation based on the law of Nothing Ever Happens?
It's better to look at the underlying factors. Money sources are drying up, nobody is making a profit outside of nVidia, most blackwell GPU's are likely not even installed yet and will probably be 2 generations behind when they finally are being used, data centers are hitting all sorts of obstacles getting built and powered and they're getting built slowly, most AI researchers seem to think that LLMs are a dead end, the newer models seem to be getting more expensive and sometimes worse, or even potentially are showing signs of model collapse (goblins..), the supposed productivity gains are not materializing.. AI has worse public sentiment than congress.. I could keep going. Some obscure "law" seems to pale in comparison to the hard evidence that the status quo is utterly unsustainable and none of these companies seem to have a realistic plan other than trying to become too big to fail essentially.
I like some of this guy's writing on other topics, but to me this is a prime example of what happens when you get public "intellectuals" talking about subjects far outside of their area of expertise. It's not as bad as Richard Dawkins latest fall into psychosis but it's basically the same phenonmenon.
While we're at it, the "exponentials are actually sigmoïds" meme is not necessarily true. While exponentials are never exponentials, sigmoids are not guaranteed. Overshoot-and-collapse examples also happen in tech, e.g. the dotcom bubble, or the successive AI winters.
Except innovation. When one sigmoid tapers off we keep finding new ones to keep the climb going.
Good example of this is number of submissions to neurips/icml/iclr. In 2017 that curve was exponential.
This does *not* imply the inevitability of AGI. It does not imply AGI is necessarily bad.
It does mean that "the capabilities of AI will eventually plateau" offers no meaningful predictive power or relevance to the overall AI discussion.
No definitely not saying this and I don’t quite know what it means
> Model reasoning is on an s-curve, which is improving.
Is this saying two different things? I think I might agree with this in principle as in maybe there is some sort of s curve or something like it but do we see evidence of this? Where?
> Model intelligence is not the same as reasoning. It's a different axis, and one I have not seen much movement on.
Can you clarify this? What is the distinction and what makes you say you have “not seen much progress?”
> See, humans have a recursive form of intelligence which is capable of self-reflection and introspection. LLMs can only reason about tokens which have already been emitted
LLMs do self reflection and introspection in context, and tweaks such as value functions (serving a similar purpose to intuition or emotion) may make this better? Why do you feel self reflection and introspection are a fundamental limitation here? Models reason over tokens they have emitted and also with their own sense and learned behavior already. Are you just talking about continual learning? Also I feel people just latch onto LLMs as if this is all of AI. Why? SSMs, memory networks, recurrent neural networks etc etc etc are all part of AI but aren’t as popular because they can’t yet compete with LLMs in terms of scaling laws and training efficiency due to e.g. hardware and software optimization and investment being focused on LLMs. If something else comes along that works better we’ll just start scaling that.
> Humans and LLMs do not share the same form of reasoning, and general human-like intelligence will not arise from the current architecture of LLMs.
Very strong statement, any theoretical or experimental basis for this? I also don’t particularly care personally other than as a point of curiosity. Why does it matter if AI systems will develop equivalent reasoning mechanisms as humans? In fact it may be much better not to.
> Therefore it is a mistake to assume that continual improvement on the reasoning scale will result in something that is equivalent enough to humans to replace all labor.
Idk I didn’t say this explicitly but I also dont think it matters if we have a system “equivalent to humans” or one that “replaces all labor”.
I am making that argument that how we measure model intelligence is flawed, and we are actually measuring something that is closer to "reasoning" than "intelligence". If you want evidence, we'll need a different form of tests, but how about I just gesture at the fact that GPT supposedly outscored PhDs on a broad range of subjects at least a year ago and to date is not replacing PhD jobs.
We see this pattern of high scores on tests but mediocre performance in the real world all over the place. From that, I draw the conclusion that it can reason like a PhD, but it can't think like a PhD.
So, we may see an s-curve on the measure of model reasoning but that doesn't imply they will overtake us or even match us on measures of intelligence.
As to your other questions:
> LLMs do self reflection and introspection in context,
> Why do you feel self reflection and introspection are a fundamental limitation here? Models reason over tokens they have emitted and also with their own sense and learned behavior already. Are you just talking about continual learning?
I disagree that models are reflecting and introspecting in a way equivalent to human intelligence here. They can reason over tokens which have been emitted, but by the same measure they cannot reason over tokens which have not been emitted. It's hard to make this point without drawing some diagrams, but I believe that human intelligence has internal loops, where many ideas may be turned over simultaneously before an action is taken. In comparison, an LLM might "feel uncertain" about a token before emitting it, but once it is emitted that uncertainty and the other near neighbor options are lost and the LLM is locked into the track that was set by the top-choice token. I think this is where hallucinations arise from, amongst other issues.
Context isn't sufficient for an internal reasoning loop because the tokens that compose context lose a lot of the information the network itself generated when picking those tokens. They occupy a much lower dimensional space than the "internal reasoning" processes of the transformer do.
>> Humans and LLMs do not share the same form of reasoning, and general human-like intelligence will not arise from the current architecture of LLMs.
> Very strong statement, any theoretical or experimental basis for this?
It's just my theory, but this is what I have been gesturing at. You already know about RNNs so I'll put it in those terms: the core of an intelligent network should be an RNN, not a transformer, but we fundamentally cannot train a network like that to work like an LLM because backprop doesn't work when there is infinite recursion and without being able to bootstrap off of the knowledge and reasoning baked into human text, there's no sufficient source of training material beyond being embodied.
---
EDIT:
I missed this, which I also want to reply to:
> Why does it matter if AI systems will develop equivalent reasoning mechanisms as humans? In fact it may be much better not to.
I actually agree that it may be better if they did not develop equivalent reasoning, but I don't see a world in which machines replace human labor without being intellectually equivalent.
As I think about it though, "dumb" machines which can following reasoning but not think like humans are a rather scary proposition, honestly. Seems like a tool that would be wielded without restraint by those in power to control those who aren't.
> But those skeptics are initially responding to the constant AI hype claims that we are exponentially growing to AGI.
This is a meaningless statement or at best just strawmanning.
The evidence is just whatever it is - we cannot make predictions with it.
So: yes we can and do make predictions with it and that’s how we get funding internally and externally to build at these scales.
Bottlenecks arise about 3 orders of magnitude from now.
On the other hand, any particular justification you have in mind for your point?
In Scott's mind, dangers from AI are not a known fact, but are somewhere between highly probable and a near-certainty. In his mind, there are well-grounded justifications for believing that AI poses substantial future dangers to the public. Therefore he also believes he should inform people about this, and strives to convince skeptics, so that we might steer clear.
It's easy to understand why someone who believes what you believe about AI would of course not warn people about AI. It's also easy to understand why someone who believes what Scott believes about AI would want to warn people about AI. Your contention is with his confidence for being worried about AI, not his reason for wanting to warn people.
Neither can any specific discussion of what the dangers are and how we can steer clear. It all comes preplanted in your head. The only thing that Scott is playing on (as far as we can see) is your ingrained fear, by using an ominous headline, and a vague reference to something "scary" in the conclusion.
Of course there was no reason to "warn" you, you already believed in the scary future. Scott is just giving you fuel, which you seem to appreciate.
If only there were a way to see more of Scott's thoughts on the subject of AI..
I don't know when (or if) AI will implode or succeed with any degree of provable certainty, because that's not my area of expertise. Rather, I can point out and discuss flaws in the common booster and doomer arguments, and identify problems neither side seems willing to discuss. That brings me cold comfort, but it's not enough to stake my money on one direction or another with any degree of certainty - thus I limit my exposure to specific companies, and target indices or funds that will see uplift if things go well, or minimize losses if things go pear-shaped.
I also think relying on such mathematics to justify a position in the first place is kind of silly, especially for technical people. Mathematical models work until they don't, at which point entirely new models must be designed to capture our new knowledge. On the other hand, logical arguments are more readily adapted to new data, and represent critical, rather than mathematical, thinking and reasoning.
Saying AI is going boom/bust because of sigmoids or Lindy's Law or whathaveyou is not an argument, it's an excuse. The real argument is why those things may or may not emerge, and how do we address their consequences within areas inside and outside of AI through regulation, innovation, or policy.
Basically a lot of people say "but isn't it also pretty likely that we DON'T get superintelligence?" And, yes, it is. But superintelligence being even a remotely plausible outcome is a big fucking deal. Your investment choices in that context are not important.
People really struggle to think rationally in the face of this shape of uncertainty.
That's the problem with 'singularity' arguments. The people making them ignore the fact that the mathematical definition of the word means 'the model of outcomes collapses to a single value' therefore the model stops being useful, yet they somehow claim to be able to make predictions beyond the singularity. It's like those shitty Facebook math posts where they divide both sides of the equation by 0 (the fact hidden by some sleight of hand), to 'prove' that 2=1.
The formulation of the singularity involves putting outrageous values into the parameters of the model of reality, and denominator ignorance, and then claiming 'rationally' determining that the consequences are too severe to ignore.
So, his point with all the demand for rigor is to end on a hand-waved jump of faith from "improved AI models" to the mythical "superintelligence"?
My thought process RE: superintelligence/AGI is generally this:
* I personally don’t believe it’s likely to happen with silicon-based computing due to the immense power and resource costs involved just to get to where we are now; hence why I invest broadly to capitalize on what gains we actually attain using this current branch of AI research across all possible sectors and exposure rates
* If we do achieve AGI using silicon-based computing, its limited scale (requiring vast amounts of compute only deliverable via city-scale data centers) will limit its broader utility until more optimizations can be achieved or a superior compute platform delivered that improves access and dramatically lowers cost; again, investing broadly covers a general uplift rather than hoping for a specific winner
* If AGI is achieved, nobody - doomer or booster alike - will know what comes next other than complete and total destruction of existing societal structures or institutions. The stock market won’t explode with growth so much as immediately collapse from the disintegration of the consumptive base as a result of AGI quite literally annihilating a planet’s worth of jobs and associated business transactions. In this case, a broad spread protects me from harm by spreading the risk around; AGI will annihilate the market globally, but not all at once barring a significant global catastrophe instigated by it
* Which brings me to the worst outcome, where AGI follows the “if anybody builds it everyone dies” thought process: investment is irrelevant because we’re all fucked anyway.
And that’s just my investment approach. I’m too pragmatic to believe we’re at the bottom of the sigmoid curve, but too wise to begin guessing where we actually exist on it at present or how much is left in the current LLM-arm of AI research; I’m an IT dinosaur, not an AI scientist.
What I can point to is the continued demand destruction of consumer compute through higher costs and limited availability due to rampant AI speculation as proof that the harm is already here in a manner most weren’t predicting, while at the same time actual job displacement by AI is limited to the empty boasting of executives using it as a smoke screen for layoffs after RTO mandates failed to thin headcount sufficiently.
In the USA in particular, we’re facing a perfect storm of:
* consumer confidence collapse leading to a decline in spending on all goods, especially luxury ones, by all but the most monied demographics
* data center-driven cost increases (energy) and resource destruction (land, water, fossil fuel use)
* the eradication of government support for renewable energy that would’ve kept these costs in check
* the widening wealth gaps creating a new underclass not seen since before WW2
In other words, most of the discourse continues to revolve around hypotheticals of tomorrow rather than realities of today. That would be the lesson I’d hope more people take away from something like this, so we can finally begin addressing issues themselves rather than empty online circle jerking about who is right or wrong.
My "plan" is hope for a benevolent intelligence that establishes a post-human government and then enjoy poat-scarcity society doing wood working or something.
Billionaires should probably be more worried.
If we don't understand the fundamental limits to any particular kind of trend, our default assumption should be that it will continue for about as long as it has gone on already.
We can, in fact, easily put a confidence interval on this. With 90% odds we're not in the first 5% of the trend, or the last 5% of the trend. Therefore it will probably go on between 1/19th longer, and 19 times longer. With a median of as long as it has gone on so far.
This is deeply counterintuitive. When we expect something to last a finite time, every year it goes on, brings us a year closer to when it stops. But every year that it goes on properly brings the expectation that it will go on for a year longer still.
We're looking at a trend. We believe that it will be finite. Our intuition for that is that every year spent, is a year closer to the end. But our expectation becomes that every year spent, means that it will last yet another year more!
How can we apply that? A simple way is stocks. How long should we expect a rapidly growing company, to continue growing rapidly?
I don't think you can use lindy on trends as if trends are static objects, but that's another conversation.
Edit: in particular I don’t agree with
But if someone claims that the trend toward increasing AI capabilities will never reach some particular scary level...
All exponential eventually becomes a sigmoid because exponential growth always expose limiting factors that weren't limiting at the beginning. Silicon manufacturing had lots of room for high-margin customers like Nvidia even a year ago (by the mere virtue of outbidding lower-margin customers), but now it is mostly gone, and no amount of money will make fabs build themselves overnight.
[1]: https://stockanalysis.com/stocks/nvda/metrics/revenue-by-seg...
The naive expectation is that AI will slow down b/c Moore's law is coming to an end, but if you really think about the models and how they are currently implemented in silicon, they are still inefficient as hell.
At some point someone will build a tensor processing chip that replaces all the digital matmuls with analogue logamp matmuls, or some breakthrough in memristors will start breaking down the barrier between memory and compute.
With the right level of research funding in hardware, the ceiling for AI can be very high.
>My understanding is that this represents 3-4 “generations” of different technology (propellers, turbojets, etc). Each technology went through normal iterative improvement, then, when it reached its fundamental limits, got replaced by a better technology. The last technology, ramjets, reached its limit at about 3500 km/h, and there wasn’t the economic/regulatory will to develop anything better, so the record stands.
You don't have one sigmoid, you have multiple each stacked on top of each other. Airplanes aren't just one technology they are multiple technologies that happen to do the same thing.
Each one is following a sigmoid perfectly. It only looks exponential(ish) because of unpredictable discoveries that let you switch to another sigmoid that has a higher maximum potential.
The same is true in AI. If you used the same architecture as GPT2 today you're in for a bad time training a new frontier model. It's only because we have dozens of breakthroughs that the capabilities of models have improved as much as they have.
That said exponential and sigmoids are the wrong model to use for growth. Growth is a differential equation. It has independent inputs, it has outputs and some of those outputs are dependent inputs again through causal chains of arbitrary complexity. What happens depends entirely on what the specific DE that governs the given technology is. We can easily have a chaotic system with completely random booms and busts which have no deep fundamental rhyme or reason. We currently call that the economy.
In this model, the exponential growth that everybody is freaking out about is only the realization of the modular software dream ("we'll only have to write an ORM once for all of human history!") and the sheer amount of knowledge in libraries.
It's at least falsifiable.
The idea is simply that the basic idea behind LLMs, that you're distilling the entropy out of the entire available world of text, is antithetical to creativity.
Further developing on the theme of self-play, humans have the ability to sense what we want (intellectually) and reach for it communally over thousands of years. It's an innate quality, and if AI starts participating (contrast to giving people psychosis) we will all be able to tell.
Short of a third sigmoid appearing in the ML CompSci space, perhaps in the form of ongoing, repeated step-optimisations which will also have diminishing returns, intelligence growth is now limited a few scaling problems that have already been worked on for a very long time.
Transistors, which have been doubling for almost a century now, but Moores Law has already plateaued and reached limits on energy efficiency, and simply building new fabs is not something that we can do exponentially. And the other growth limiter is electricity - there is no exponential supply of fossil fuels or power plants. Although manufacturing has scaled, PV tech improvements are also plateauing - and while storage is getting cheaper, it's still not economical vs fossil fuels (meaning: when we have to switch to it, the growth slows down further) and we are unlikely to see battery efficiency sigmoid enough to maintain the AI sigmoid.
I don't mean to be bearish here. There's so much money sloshing around that we can afford to put the smartest people, using unlimited tokens, on the task of finding small, incremental gains on the CompSci side of things that will have large monetary payoffs - hopefully allowing further scaling and increased emergent abilities of LLMs. Maybe we can squeeze the algos for quite a while. But I don't see that maintaining the same level of exponential as unlocking unlimited data or maxxing out the world's energy/fab capacity for long.
And I don't see why this is a massive issue except for the people who want to have some god-like super AI? Frontier LLMs are genuinely magic. Not "won't delete your production database" magic, but definitely a massive productivity gain for competent knowledge workers.
My mental model has been 3D computer graphics: doubling the polygon count had huge returns early on but delivered diminishing returns over time.
Ultimately, you can't make something look more realistic than real.
I don't know what the future holds, but the answer to the question "can LLMs be more realistic than real" will determine much about whether or not you think the curve will level off soon.
In the same fashion, LLMs have to pay for themselves to keep the trendlines going. In a whole-systems -sense, mind, not "$2000/month is cheaper than hiring a developer" while the rest of the economy collapses.
Lindy's Law is not actually a law and many exact minds will be provoked by the very name; it also fails spectacularly in certain contexts (e.g. lifetime of a single organism, though not necessarily existence of entire species).
But at the same time, I am willing to take its invocation in the context of AI somewhat seriously. There is an international arms race with China, which has less compute, but more engineers and scientists. This sort of intellectual arms race does not exhaust itself easily.
A similar space race in the 1950s and 1960s progressed from first unmanned spaceflight to a moonwalk in mere 12 years, which is probably less than what it takes to approve a bicycle lane in Chicago now.
I keep seeing this. Where did it come from? Has China said that they intend to attack other countries using AI? Have other countries declared that they intend to attack China with AI?
Also, why does anyone believe that AI could actually be that dangerous, given it's inherent unpredictable and unreliable performance? I would be terrified to rely on AI in a life or death situation.
BTW your handle is an actual Czech word, minus a diacritic sign ("křupan"), and a bit amusing one. It basically means hillbilly. Not that it matters, just FYI.
Anyway: AI will be used in military context, and it probably already is. Both for target acquisition and maybe even driving the weapon itself. As of now, the Ukrainians are almost certainly operating some AI-enabled killer drones.
Inherent unpredictable and unreliable performance is also quite the feature of human beings as well.
For example: the flight airspeed plot, it starts at ~1900. Now of course it should start there bc we did not have planes before that. But let's change the plot heading from airspeed to human speed a.k.a at what speed could humans move. now you can change the origin meaningfully as we had chariots, horses, bicycles, ships before that.
If you instead create a plot for last 5000 years you would see that the speed at which humans are able to move is rising exponentially, from walking on foot in a radius of 1000-2000feet in a thick jungle 5000 years ago to reentering earth at 25000 miles/hour in 1969 (yeah, read that again). Even for AI, if you zoom out the plot to last 70 years it will look exponential, if you zoom in to last 2 months it will look absolutely flat. The point is that the whole sigmoid/exponential argument is a function of the origin (0,0).
This is the crux of the article. To a large extent continued progress depends on a stable increase in compute, an increase in training data, and an increase in good ideas to squeeze more out of both of them.
One calculation you could do is a survival function: for each of the above, how long before it is disrupted? For example, China could crack down on AI or invade Taiwan. Or data centers become politically unpopular in the US. Or, we could run out of great ideas. Very hard to predict.
The entire plot of the Lord of the Rings could probably be compressed into less than 10 kB of text too.
Edit: this seems to be a controversial comment, but IMHO a blog of Scott Alexander's type is an art form, not just a communication channel.
The situation is drastically different for problems that require interaction was the physical world to determine success.
As soon as you add a powerful simulator for physical problems to the self learning experience of the AI, you are extremely hampered by the large amount of needed computation.
For example, When a car starts, it's speed and acceleration become more than zero. But what about rate of change in higher degrees? It suddenly doesn't change from zero acceleration to non-zero. That means the car has a non-zero derivative at all degrees. In other words, the movement is exponential. The same thing happens in reverse when the car reaches a constant speed.
All positive growth eventually flattens out and becomes sigmoid, but a lot of phenomena experience negative growth and nose dive. No gentle curve, but a hard kink and perfect flat line at zero. Forever. I think it would be a stretch to categorize that pattern as sigmoid. Predicting a sigmoid pattern for negative growth implies some sort of a soft landing (depending on your definition of soft).
We can think of many populations that are no longer with us. So just a caution about over applying this reasoning in the negative case.
2. What's even worse than predicting that some growth curve flattens before X happens is predicting it will flatten before X happens but after Y happens, which is what we see when it comes to AI in software development. Too many people predict that AI will be able to effectively write most software, replacing software engineers, yet not be able to replace the people who originate the ideas for the software or the people who use them. I see no reason why AI capability growth should stop after the point it's able to write air-traffic control or medical diagnosis software yet before the point where it's able to replace air traffic controllers and doctors.
3. While we don't know much about AI (or, indeed, intelligence in general), we do know something about computational complexity. Some predictions about "scary things" happening (the ones I'm guessing Alexander is alluding to, though I can't be certain) do hit known computational complexity limits. Most systems affecting people are nonlinear (from weather to the economy). Predicting them requires not intelligence but computational resources. Controlling them, similarly, requires not intelligence but either computational resources or other resources. It's possible that people choose to give control over resources to computers (although probably not enough to answer many tough, important questions), although given how some countries choose to give control to people with below-average intelligence (looking at you, America), I don't see why super-human intelligence (if such a thing even exists) would be, in itself, exceptionally risky.
This is kinda laughable. Scott has been thinking and writing about AI for a long time
For example, take something like a fad or trend; they don't have a hard end date like human lifespan, so it should follow Lindy's law.
However, the likelihood, on average across the population, that you observe a trend is going to be higher at the end of a trend lifecycle than at the beginning. This is baked into the definition - more and more people hear about a trend over time, so the largest quantity of observers will be at the end of the lifecycle, when the popularity reaches its peak.
In other words, if you are a random person, finding out about a trend likely means it is near the end rather than the middle.
It's the solution to the tank problem. You know that the enemy number their tanks as they're produced. You capture a tank and know its number, N. What's the best guess about how many tanks the enemy has produced so far? As a pure mathematical model with no other details, the best guess is 2N. Of course in reality you have some ideas about how long it takes to make a tank, how many resources the enemy has etc.
Analogously you have information about the way trends develop.
We have at least global warming and impending WW3, so that line of reasoning seems to work.
"The Lindy effect applies to non-perishable items, like books, those that do not have an "unavoidable expiration date"."
And later in the article you can see the mathematical formulation which says the law holds for things with a Pareto distribution [2]. I'd want to see some sort of good analysis that "the life span of exponential growth curves" is drawn from some Pareto distribution. I don't think it's completely out of the question. But I'm also nowhere near confident enough that it is a true statement to casually apply Lindy's Law to it.
The argument given is the same as the one that I first ran across, not by that name, in https://www.nature.com/articles/363315a0. https://en.wikipedia.org/wiki/Doomsday_argument claims that it was a rediscovery of something that was hypothesized a decade article.
I hadn't tried to give it a name, or thought to apply it outside of that context.
As for the mathematical qualms, I'm a big believer in not letting formal mathematical technicalities get in the way of adopting an effective heuristic. And the heuristic reasoning here is compelling enough that I would like to adopt it.
Try avoiding the heat death of the universe /s
The law only applies for certain types of processes, and is completely wrong for other types (e.g. a human who has lived 50 years may live 50 more, but one who has lived 100 years will certainly not live 100 more). So the question becomes: what type of process are you looking at? And that turns out to be exactly the question you started with: is there a fundamental limit to this growth curve, or not.
Did you even read the post? It’s an estimate in the context where you have zero information on which to base an accurate estimate. The author’s point is that if you’re making a different estimate you need to actually say what information is informing that.
Human lifespan is obviously not a case where we have zero information, so what is your point in bringing that up?
But that's the entire idea of Bayesian reasoning. Which has proven to be surprisingly effective in a wide range of domains.
I'm all for quantifying my ignorance, and using it as an outside view to help guide my expectations. Read the book Superforecasting to understand how effective forecasters use an outside view to adjust their inside view, to allow them to forecast things more precisely.
So for example, the longer a time bomb ticks, the less likely it is to go off any time soon. (Assuming the timer isn't visible.) :)
We expect fresh processes to terminate quickly and long running processes to last for a while longer.
People would confidently cite Lindy's law all the way near the end of a trend. Nothing would stop a Roman saying that just before the Fall.
We don't always need to "understand the fundamental limits" to a trend to see where it's going. Just to observe more than a random blind guess about them.
I also wouldn't trust the "see how much we're improving" benchmarks of a trillion dollar pre-IPO industry to begin with.
> Do we really think things will move this fast? Sort of no - between the beginning of the project last summer and the present, Daniel’s median for the intelligence explosion shifted from 2027 to 2028. We keep the scenario centered around 2027 because it’s still his modal prediction (and because it would be annoying to change). Other members of the team (including me) have medians later in the 2020s or early 2030s, and also think automation will progress more slowly. So maybe think of this as a vision of what an 80th percentile fast scenario looks like - not our precise median, but also not something we feel safe ruling out. [2]
I don't think this changes your observation that he is "personally invested" (i.e. believes this trendline will continue), but I'm pretty sure when AGI doesn't appear in 2027, many people will believe that this invalidates the arguments being made here (or in the report). The actual report was intended to give a feel for what a near-future "disaster" AGI scenario, and settled on a date to give that some concrete immediacy. The collective review that gave that as a possible, but not inevitable date is still ongoing (they originally pushed their best estimate out a bit further, but now they think, judging by the goals that are being hit, their scenario was a little too conservative). [3]
[1] https://freddiedeboer.substack.com/p/im-offering-scott-alexa... [2] https://www.astralcodexten.com/p/introducing-ai-2027 [3] https://blog.aifutures.org/p/grading-ai-2027s-2025-predictio...
LLMs are nothing close to AGI and not going to lead to it, they can’t distinguish right from wrong, they can’t count, they can’t reason, they generate plausible text from a vast databank of connected text.
Apparently that is enough to fool many people but it’s nothing close to AGI which would require internal models of the world, reasoning etc.
We are nowhere close to AGI and the fools who predicted we were will unfortunately keep lying about their stated timelines when it inevitably doesn’t arrive. You’re already hedging and trying to caveat previous predictions, as OpenAI did with their AGI predictions which they’re now furiously back-pedalling on.
It is broadly true that Scott believes that AGI will come in the near future and from LLMs, although his reputation runs a ways deeper than that.
It's not at all surprising that they are increasingly getting labeled a cult (they aren't by traditional definition but there are a lot similarities). I'm really surprised it hasn't hit the mainstream yet given the connections to Elon, Thiel, frontier labs, dark crypto funding, FTX/SBF, some suicides and some murders. It's all a little nuts.
Meanwhile you got all the anti-democratic NRx people on the other side of it.
I suspect this new doc coming out on HBO will spark a media frenzy.
What has he been up to since finishing the finest work of literature ever produced, Harry Potter and the Methods of Rationality? I’ve been patiently awaiting a sequel!
Right now we have an incredibly smart thing with severe short term memory loss, and it’s hard for us to reconcile that as it’s so different from us.
I mean, that's called "having an opinion".
The important thing we can show it in hindsight only. We don't know which other tasks we are currently mistaken about requiring intelligence. Maybe none of them are?
We don't know. We don't know what intelligence is. If we look at decades and even centuries of attempts to define intelligence, it is all looks like a goalposts moving. When a definition of intelligence starts to include people or things we don't like to think as of intelligent ones, we change the definition.
The poster you replied to even used the word "sentient", which is quite interesting (warning: opinionated tangent ahead). Merriam-Webster defines it as "capable of sensing or feeling: conscious of or responsive to the sensations of seeing, hearing, feeling, tasting, or smelling". Feels like qualia. Or if we don't want to go the qualia route... Of course, we wouldn't call Helen Keller non-sentient, so presumably we "really" mean "can it sense or feel" -- well, sense is just "act/feel according to the environment", which you could argue in the case of an LLM would be their context... so we should "really" remove "sense" from the definition, probably. So "do LLMs feel" is probably closer to what "sentient" is being used for here. Since we don't have the obvious symmetry of "you are like me and I feel (therefore you probably feel)", it's way better/easier/feel-good-ier to prefer "LLMs don't feel" rather than "oh shit, it feels and model training is actually just torturing it into the right shape". LLMs as fundamentally non-intelligent also avoids the problems of "what does that say about people" or "we may have made 'AGI' and it wasn't what we thought it would be" or "we're not ready to talk about this yet".
That's what you've got wrong. We don't define functions that an LLM approximates. Autoregressive pretraining approximates an unknown function that produces text (that is what the brain does). RL doesn't approximate functions, it optimizes objective by finding an unknown function that performs better.
I'm pretty sure there's a 3 year design goal starting this year that'll do that to any of the qwen, deepseek, etc models. There's a lot you could do with sped up models of these quality.
It might even be bad enough that the real bubble is how much we don't need giant data centers when 80-90% of use cases could just be a silicon chip with a model rather than as you say, bloated SOTA
If there's a breakthrough in memristors, you could end up with another 20x reduction in circuit elements (get rid of memory bottlnecks, start doing multiplication ops as log transform voltage addition)
The ceiling is ultra high for how far AI can go.
I'm not sure this follows? Research has gone into two bit quantizations that only need a scale factor per block and each parameter merely takes up two bits which means that the operations can be mapped directly onto adders rather than multipliers.
>but if you really think about the models and how they are currently implemented in silicon, they are still inefficient as hell.
The vast majority of analogue components are both space and energy inefficient. Digital won for a reason. You can simply keep scaling to lower voltages and smaller transistors since you only need to distingish between one and zero.
All the easily verifiable domains such as mathematics, coding, and things that can be run inside a reasonable simulation are falling very very fast.
By next year if not sooner, mathematicians will be wildly outpaced by LLMs for reasoning.
So it's not impossible to have things that seem orthogonal, like generation speed or context length, have an impact on quality of result.
The idea that exponential growth will continue with stacked sigmoids is also not a given. An example is the nail. Nails used to be about half a percent of US GDP. That's a pretty big number! A series of innovations stacked on each other (each innovation having its own sigmoid) to reduce the cost of nails. Nails dropped in cost by over 90%.
But eventually nail manufacturing reached a floor. And since the mid-20th century, we haven't gotten much better at making nails. The cost of nails actually started increasing slightly. We ran out of new innovation sigmoids, so we got stuck on the last one.
So what you actually have to predict is whether there will continue to be new sigmoids, not whether the existing sigmoid will asymptote (we already know it will).
This is much more difficult to forecast, because new sigmoids (major new innovations) tend to be unpredictable events. Not only are the particulars difficult to forecast (if they were knowable, the innovation would have already happened), but whether there will be a major innovation or not is also hard to forecast, because they are distinct and separate from any existing sigmoid trend.
So we are left with the idea that all current innovations in AI will asymptote in their scaling as they reach the plateau of the sigmoid, but there may be new sigmoids that keep the overall trend up. Or there may not be. We don't know.
That's not very satisfying, so we'll get to keep reading articles like this one.
Return on investment can be too low because the investment required is really high, but it can also be too low because the returns are just limited. If prices had dropped 90%, surely nails became even more ubiquitous, but at that stage there's only so much more money to dig out of the cost reduction hole. It feels plausible that there may have been ideas about more digging that could be done, but the reward just wasn't there in the market, especially versus just selling what worked.
I bring it up because the distinction in one specimen may speak to a larger trend: do new sigmoid developments tend to fail to materialize more often because of serious physical limits / lack of good ideas, or because of limitations to ROI? (Or, other things?)
In the arena of AI, the ROI on more intelligence/unit-cost seems pretty high right now. So, it seems like the difficulty of applying any potential innovations would have to be staggering for none to be pursued. Or, there'd have to just not be any good ideas to try.
Overall, I think there's ideas to try. So in my opinion, that shapes out to justify a bullish sentiment on sigmoids continuing to stack until the perceived potential gains from more intelligence/unit-cost somehow fall off.
Like I said, I don't disagree, we really don't know. But I feel it's a good bet that there's more coming.
That said... if the exponential is made of stacked sigmoids, it's still an exponential on the whole! The fact that it's made of stacked sigmoids is relevant to the engineers making it, but not so relevant to the users or those otherwise affected by it.
What exactly are these dozens of breakthroughs? Most frontier models architectures today still look very much like GPT2 at their core. There were various of improvements like instructgpt, finetuning techniques, efficiency improvements with kv caches, faster attention, lora, better tokenizers, etc. Most of these are for making things run faster. The biggest differentiator has probably been data curation and post-training data and the ability to fit more into the model. But I think we had few breakthroughs that would fall into the category of different technologies.
(This is a bit disingenuous, as lots/most of work is spent on the scaling and training side of things.)
I suspect that he started down the path of considering them more deeply here but found that they didn't add much to the analysis. A stack of sigmoids ultimately either gives you a single sigmoid if you run out of new innovations, or an exponential if you don't.
At first the models turned a 5 minute task into a 5 second task (by 5 seconds I mean a very short amount of time, not precisely 5 seconds). Then they turned a 15 minute task into a 5 second task.
Opus 4.6 completes 8 hour tasks all the time but (at least in my experience) it isn't spitting the answer out in 5 seconds anymore. It's using chain of thought and tools and the time to completion is measured in minutes or maybe hours.
In my experiments with local LLMs, a substantial part of the gap between frontier and local (for everyday use) is in tooling and infrastructure.
That is why I am sympathetic to the idea we are leveling off. But to bring in the air speed example from the article, I don't think we've reached the equivalent of the ramjet yet. I suspect in the coming years there will be new architectures, new hardware, and new ways to get even more capable models.
I don't know if they can get their numbers right this way, but this seems a way more useful metric, than theoretic capabilities.
I trained an LLM to write the whole Harry Potter series, and that took JK Rowling like 17 years.
For my next point on the graph, I'll train the LLM to write the Bible, something that took humans >1500 years.
The tasks are obviously all of the form "Go do this, and if you get the following output you passed". Setting up a web server apparently takes 15 minutes for a human, which is news to me since I'm able to search for https://gist.github.com/willurd/5720255, find the python one-liner, and copy it within about ten seconds.
Anyway, this is cool but it does not mean Claude can perform any human tasks that take less than 8 hours and are within its physical capabilities.
I'm curious what people really mean when they say this. Intelligence is famously hard to define, let alone measure; it certainly doesn't scale linearly; it only loosely correlates to real-world qualities that are easy to measure; etc. Are you referring to coding ability or...?
emoji face with eyes rolling upward
Scott makes a Lindy effect argument which is plausible, but don't let that fool you, we still don't know what's going to happen.
Why would that be? Nothing about Lindy's Law makes that promise. And even the SOTA in 2026 is over-estimated thanks to a trillion dollar industry trusted to not influence benchmarks.
Well.. that's not true is it. There are human cultures that haven't reached for anything for thousands of years even though they clearly saw what western culture was doing and that they were being left behind badly.
I expect benchmarks like ProgramBench will replace METR this year.
Yes, that's called "having an opinion". Typically people writing argumentative pieces are doing so because they have a belief about the matter. I'm not sure what exactly you expect here.
> if he's wrong I would hope he owns up to it
I think Scott Alexander is pretty good about that.
I mean.. this is 2026 right? You're not writing that comment from 2024 or something?
We see massive problems already where photos are just not believable anymore, nor is audio, and not even video actually with many people falling for AI fake image clips from the Gaza war for example. And since then these tools are MASSIVELY more powerful. Disinformation is essentially free, and the cost of truth has been static. Meaning the "buying power" of truth has collapsed and is falling faster and faster.
Anyone who dismissed AI risks a few years ago IS ALREADY PROVEN WRONG.
The ROI for cornering the nail market seems like it could have been big. The ROI for making something significantly more efficient than a ICE would have been very high for most of the last century and technology that is better in many respects than ICEs does now exist, but it took us roughly a century to get there. The ROI for coming up with something that's better than the ~1% annual efficiency improvement on turbofans would be extremely high, but we don't know what that is (probably some sort of propfan, but that idea's over 50 years old...)
You would be a "křupan" if you wore agricultural boots to a fancy restaurant, or talk to a lady in an uncultured way. Basically, a hickey who was never taught proper manners.
The singularity framing is really tough here, right? It comes from black hole physics. Essentially, at the event horizon, the way we know how to do physics stops working, and we rightly conclude that we can't currently say anything about the other side of the event horizon. It is not saying that nothing is occurring there. Matter, time, space, energy, whatever, that still is there (maaaaybe?) and is still undergoing something. It's just that we don't know what that is.
The same is true with using these tech singularity arguments. Like, in the age of superintelligence (if that happens), there will still be thing happening, the dawn will still come every day and the dusk will still too. It's just that we say our current ideas about that new day aren't that applicable to that new age (God, this sounds like a hippie).
However, unlike with black hole physics where we aren't even sure time can exist like we know, we are likely all going to be there in that new superintelligence age. We're still going to be making coffee and remembering bad cartoons from our youth. Like, the analogy to black hole physics breaks down here and maybe does a disservice to us. It's not a stark boundary at the Schwartzchild radius, it is a continuous thing, a messy thing, a volatile thing, and very importantly for the HN userbase, a thing that we control and have the choice to participate in.
We are not passively falling into the AGI world like the gnawing grinding gravity of a black hole.
If you listen to the hardcore doomers, the misaligned superintelligence will curl a finger on its monkey paw and turn the planet into paperclips or something. If you listen to the most depraved boosters, AGI will remove the need for 99.999% of human workers and so we all get turned into biofuel to churn out more tokens.
Yes those are really extremely scenarios but that's how I think of the singularity. It's so alien that we cannot rule out anything.
So there are a couple interesting and meaningful changes at the event horizon, but it's not a mathematical singularity.
So meeting exactly 1 100 year old alien makes it decent odds that's somewhere near the middle of their lifespan.
Because if you grabbed one random human, chances are you'd find someone roughly middle aged.
That would only be true if the underlying distribution worked that way, which for the human population it doesn’t (global median age is 31, global average expected lifespan is 73), so for humans if you grabbed a random human, chances are you’d find someone less than middle-aged.
It seems to me like you're trying to somehow imply that writing things to convince people of what you believe is somehow nefarious? It isn't! It's what we're all doing here right now! Putting it in a format that certain people will take more seriously doesn't make it nefarious either. I am quite confused by your point of view here.
Not interested in further arguments about this.
It is purely a test of capabilities (can it do a thing that takes a human $X hours), not efficiency (how fast will it do it).
At least I want AI to solve my problems, not score high on a academic leaderboard.
Either you black-box the curve and assume that you will keep stacking sigmoids for about as long as you already have already seen.
Or you white box it and make some actual technical argument about why the curves can’t keep stacking.
There are plenty of plausible arguments here. Scott is not arguing that the exponential must go on forever.
He’s making a meta-level point about the debate; you have to pick one of the above, and you can’t just argue that “now is the time the s-curves will stop stacking” without providing some justification.
But in any specific industry or area? You often get a bunch of big discoveries, and then there is a long period of no important discoveries, because we've figured out the main aspects of that technological paradigm. The technology becomes commoditized and standard.
And that's the trillion dollar question with AI right now -- will we soon exhaust the potential of the current LLM paradigm? And we'll just have 20 or 30 years of figuring out mainly how to make LLMs cheaper and how integrate them into business processes, before somebody comes up with another fundamental breakthrough?
Or are we only 10% of the way in developing the current LLM paradigm? Where a decade from now models virtually never make mistakes and are smarter than basically any tenured faculty member in their field?
For all we know it becomes negative because all the people who understood how to train a trillion parameter model get killed by an asteroid during a conference.
Add
Total collapse in government quality AND public trust to politicians
Total collapse of news media to slop and paid-for
Total collapse of culture
(Not just the US either)
> the widening wealth gaps creating a new underclass not seen since before WW2
I go back and forth on this. I think the reality is that "underclass" is a moving target. AI and automation makes things so cheap that today's underclass lives better than kings ever did.
I suggest you go share this opinion with the people living on the street because they can't afford housing.
I don't even think there are any "genuine" Lindy processes. What would those look like? Are they always half done?
That is the argument that is being made, but that only holds if the process is drawn from an underlying Pareto distribution with epsilon > 1[1].
As a counterexample, I’m jetlagged and disorientated. I go to sleep and wake up. It’s light outside but I don’t know the time. What’s the best guess of the time of day? By the “Lindy law” the best guess is that the process of daytime is halfway done so if I’m half-way through the day, my best guess is it’s noon.
Clearly that’s not the best guess that could be made. The distribution of times I might wake up is heavily skewed towards the morning, so the best guess is going to be some time in the morning. Now you might argue that we don’t know absolutely nothing about the cycle of the day and night and that’s true. But we also don’t know absolutely nothing about any of the examples in TFA either.
The point is, the times of day I might wake up are not drawn from a pareto distribution with the right parameters so the Lindy Law heuristic completely fails. In TFA the author gives no justification for why the remaining lifespan of the exponential growth of AI might be drawn from such a distribution either, so there’s no reason to think the heuristic will be accurate in that case either.
[1] From https://en.wikipedia.org/wiki/Lindy_effect. epsilon = 1 + 1/p where p is the parameter of the conditional expectation E[T-t|T>t] = p t. So only things with p positive but finite exhibit this effect. If p is negative then the best guess is going to be that the lifetime of the thing will end immediately because we’re already past the expected lifetime, and if p is infinite then the thing will never end so all finite guesses about its length are equally bad. So whether half-way is a good heuristic depends entirely on the underlying process and you’d need to demonstrate that the majority of things have positive p for half-way to be the best guess. That’s far from clear.
But often we don't have the information that we wish. Even more often, the information that we have leads us to a story, that severely misleads us. Reminding ourselves of the zero information version of the story, can be an antidote to being mislead that way.
Therefore it is valuable to know how to make the most out of zero information. And if we have information, to think about exactly why it leads to a different conclusion.
"And we have information to think" - then we don't have zero information right?
So, if most processes are in fact like human lifespans, which do not show a Lindy effect, then it is completely wrong to assume that a random process that you encounter will have this property.
Do you mean "token" as in the LLM sense?
Or are you thinking that thoughts in the human brain are also constructed out of some sort of underlying "token" even though the abstract thought happens and is held before any words are used to try to communicate that thought to an external party?
They can predict likely sentences but not evaluate truth or logic. They can fairly reliably record facts about the world but not construct internal models of the world.
They do probabilistically. So do humans as a matter of fact. The best of us are better at it than LLMs, but that's not persuasive evidence of anything meaningful really.
> They can fairly reliably record facts about the world but not construct internal models of the world.
You don't know that, unless your presuppose a very specific definition of world model that necessarily precludes emergent ones.
Argument?
Are LLMs close to being able to significantly help AGI researchers?
But yeah, as soon as the digital models start to plateau, ASICs and then this will happen.
I'm not even kidding. Modern ML systems already eat errors - what's one more error type for them to eat?
At 1 sign bit, 1 exponent and 2 mantissa bits, there is barely any work done during multiplication into a bf16 accumulator. You're performing two shifted adds at this point.
> research showing with the right policy, Rest of the owl.
Only if you fully detail the behavior of the system.... at that point why use a chatbot? You've coded the entire thing.
> first as good as human
We'll see. Chatbots are only as capable as you detail them to be
You’re constructing a post-hoc fantasy of human thought based on how LLMs work because you are desperate for some reason to believe that they are thinking like humans, but they are not. The process is very different and the results are also different.
Human's don't operate the same way, the thought happens and then the words are generated to reasonably describe that thought.
Additionally "from learned stats" doesn't disambiguate between a wider variety of things. I'm not aware of any other way to acquire knowledge from measurements. I'd bet that humans do this differently, based on the fact the humans can get further with less training data and that they learn actively during operation, but not so differently that 'learning stats' would be an inaccurate description.
If that were the case, then the systems would generate words based on the fully resolved idea, but that is not how the LLM systems currently work (per vendors descriptions).
They choose words sequentially and both the specifics of the input as well as the chosen output words significantly impacts not just the rest of the output but the very correctness of the output.
> but not so differently that 'learning stats' would be an inaccurate description.
Agreed, humans are generalizing using some mechanism that can be modeled with math.
But the execution of our reasoning and thought processes is not obviously similar to LLM's next word generation based on probabilities.
Thoughts don't happen in a vacuum, they are triggered by external or internal stimuli, and these stimuli/thought precursors could very easily be tokens (dense info packets), which then map to latent space vectors, which very well could be thoughts.
Claims like "humans don't operate the same way" has no basis. Not only do we literally not know how humans operate mechanistically, and so we literally don't know the logical structure of human thought, but any system that is Turing complete is so easy to create that many wildly different mechanistic systems are fundamentally equivalent/interconvertible.
Yes, possible, that's why I asked you above if that's what you meant by "token". Someone else responded and I didn't notice it wasn't you.
> Claims like "humans don't operate the same way" has no basis. Not only do we literally not know how humans operate mechanistically, and so we literally don't know the logical structure of human thought, but any system that is Turing complete is so easy to create that many wildly different mechanistic systems are fundamentally equivalent/interconvertible.
I think this position is too extreme, we do have some information.
We know how LLM's work when generating a sequence of words and I know that my brain does not work the same way for word generation because I am fully aware of the complete thought in advance of any words getting generated by me externally or internally.
I know prior to generating words that my thought is X and the words I'm about to produce need to express that thought.
But with LLM's we know that the essence of what they produce is not known in advance, that it must complete the word generation process to fully realize the end result and that multiple different end results are possible.
Anthropic says of the their model[0]:
"""Claude sometimes thinks in a conceptual space that is shared between languages, suggesting it has a kind of universal “language of thought.”
{...}
Claude will plan what it will say many words ahead, and write to get to that destination. We show this in the realm of poetry, where it thinks of possible rhyming words in advance and writes the next line to get there. This is powerful evidence that even though models are trained to output one word at a time, they may think on much longer horizons to do so."""
Anthropic also created 'golden gate claude'[1] by identifying the region of its architecture that corresponded to the concept of the golden gate bridge and activating it. What would such a region exist for if claude could only think one token at a time?
>the execution of our reasoning and thought processes is not obviously similar to LLM's
"Not obviously similar" I can agree with. I don't think you've identified a way in which they are obviously different, though.
[0] https://www.anthropic.com/research/tracing-thoughts-language...