As Rocks May Think
evjang.com
evjang.com
I thought they meant the plural of ASRock as in "ASRocks May think" and thought this was about ASRock motherboards getting a BIOS/UEFI with an integrated LLM or something.
https://cmpr.ai/hutter/archive/
Claude does the research, like having an army of very fast but forgetful grad students working for free.
The individual paper doesn't make much incremental improvement but once you get a pipeline of results building in a particular direction it's powerful.
Alternatively, we are entering a dark age where the billionaires who control most of the world's capital will no longer need to suffer the indignity of paying wages to humans in order to generate more revenue from information products and all of the data they've hoarded over the past couple of decades.
> the real kicker is that we now have general-purpose thinking machines that can use computers and tackle just about any short digital problem.
We already have those thinking machines. They're called people. Why haven't people solved many of the world's problems already? Largely because the people who can afford to pay them to do so have chosen not to.
I don't see any evidence that the selfishness, avarice, and short-term thinking of the elites will be improved by them being able to replace their employees with a bot army.
Like every previous invention that improves productivity (cf. copiers, steam power, the wheel), this wave of AI is making certain forms of labor redundant, creating or further enriching a class of industrialists, and enabling individuals to become even more productive.
This could create a golden age, or a dark age -- most likely, it will create both. The industrial revolution created Dickensian London, the Luddite rebellion & ensuing massacres, and Blake's "dark satanic mills," but it also gave me my wardrobe of cool $30 band T-shirts and my beloved Amtrak train service.
Now is the time to talk about how we predict incentive structures will cause this technology to be used, and what levers we have at our disposal to tilt it toward "golden age."
Capitalists have openly gloated in public about wanting to replace at least one profession. That was months or years ago. What are people doing in response? Discussing incentive structures?
SC coders paid hundreds of thousands a year are just letting this happen to them. “Nothing to be done about another 15K round of layoffs, onlookers say”
Great, let them try. They'll find out that AI makes the human SC coder more productive not less. Everyone knows that AI has little to nothing to do with the layoffs, it's just a silly excuse to give their investors better optics. Nobody wants to admit that maybe they've overhired a bit after the whole COVID mess.
Coding may be a limited exception, but even then the AI's job is to be basically a dumb (if sometimes knowledgeable) code monkey. You still need to do all the architecture and detailed design work if you want something maintainable at the end of the day.
I mean, I know what you are getting at. I agree with you on the current state of the art. But advancements beyond this point threaten everyone's job. I don't see a moat for 95% of human labor.
There's no reason why you couldn't figure out an AI to assemble "the architecture and detailed design work". I mean I hope it's the case that the state of the art stays like this forever, I'm just not counting on it.
> There's no reason why you couldn't figure out an AI to assemble "the architecture and detailed design work".
I'd like to see that because it would mean that AI's have managed to stay at least somewhat coherent over longer work contexts.
The closest you get to this (AIUI) is with AI's trying to prove complex math theorems, where the proof checking system itself enforces the presence of effective large-scale structure. But that's an outside system keeping the AI on a very tight leash with immediate feedback, and not letting it go off-track.
Even the most pointless bullshit job accomplishes a societal function by transferring wages from a likely wealthy large corporation to a individual worker who has bills to pay.
Eliminating bullshit jobs might be good from an economic efficiency perspective, but people still gotta eat.
Seems to me like that's still a worthy goal.
Seems to me like that's still a worthy target if chasing it fights that outcome.
Of course that comes with its own set of problems, e.g. that you will lose training, connections, the ability to exert influence through the job or any hope of building a career.
But one is well-advised to inflate and test the new lifeboat before jumping out of the current one, not after.
Also, these productivity gains arent used to reduce working time for the same number of people, but instead to reduce the number of people needed to do the same amount of work. Working people get to see the productivity benefits via worsening material conditions.
https://www.independent.co.uk/news/world/americas/donald-tru...
At an individual level, we have a lot of psychological plasticity and can work to overcome our limitations. At societal scale, though, we are social primates and any system that takes advantage of natural social primate behavior is likely to succeed indefinitely.
I see literally zero people doing the equivalent of “breaking the factories” like the luddites attempted
Do you not see the overwhelmingly negative response to AI produced goods and services from the average westerner?
At a certain point it’s too late.
What if it’s over 10 years?
AI ghosts can do a lot of things, but they're limited by being non-physical.
> The entire global economy is re-organizing around the scale-up of AI models.
> Software engineering is just the beginning; ...
> Air conditioning currently consumes 10% of global electricity production, while datacenter compute less than 1%. We will have rocks thinking all the time to further the interests of their owners. Every corporation with GPUs to spare will have ambient thinkers constantly re-planning deadlines, reducing tech debt, and trawling for more information that helps the business make its decisions in a dynamic world.
> Militaries will scramble every FLOP they can find to play out wargames, like rollouts in a MCTS search. What will happen when the first decisive war is won not by guns and drones, but by compute and information advantage? Stockpile your thinking tokens, for thinking begets better thinking.
So he is extending this to more than just computer science.
Almost all of their wealth is ultimately derived from people.
The rich get richer by taking a massive cut of the economy, and the economy is basically people providing and paying for services and goods. If all the employees are replaced and can earn no money, there is no economy. Now the elite have two major problems:
a) What do they take a cut of to keep getting richer?
b) How long will they be safe when the resentment eventually boils over? (There's a reason the doomsday bunker industry is booming.)
My hunch is, after a period of turmoil, we'll end up in the usual equilibrium where the rest of the world is kept economically growing just enough to keep (a) us stable enough not to revolt and (b) them getting richer. I don't know what that looks like, could be UBI or something. But we'll figure it out because our incentives are aligned: we all want to stay alive and get richer (for varying definitions of "richer" of course.)
However, I suspect a lot will change quickly, because a ton of things that made up the old world order is going to be upended. Like, you'd need millions in funding to hire a team to launch any major software product; this ultimately kept the power in the hands of those with capital. Now a single person with an AI agent and a cloud platform can do it themselves for pocket change. This pattern will repeat across industries.
The power of capital is being disintermediated, and it's not clear what the repercussions will be.
Anyway, like data centers in space there are lots of material limitations that all of these exuberant "ZOMG rocks can think now" essays all sweep under the rug to drive a very biased narrative about what is actually happening & the fact that all those binary bits are produced by real materials that have lifecycles & production limits not visible in the digital artifacts.
> In order to predict where thinking and reasoning capabilities are going, it's important to understand the trail of thought that went into today's thinking LLMs.
No. You don't understand at all. They don't think. They don't reason. They are statistical word generators. They are very impressive at doing things like writing code, but they don't work the way that is being inferred here.
People get the idea that just because LLMs are initially trained to predict word sequences, that's all they can do. This is not the case.
Transformers are general-purpose learning mechanisms. Word sequences are just the first thing we teach them. Then we train them more with human feedback. Then we sometimes hook them up to math and logic engines to train them some more, so they learn logical and mathematical reasoning. The article describes this process in a bit of detail.
With "reasoning models" we also let the model have some internal monologue before generating output, so it can "think through" the problem.
We didn't do all that with early LLMs. Those were just word predictors. But now we do those things, and that's why our lowly LLMs are writing huge software projects that actually work, and solving famous math problems that have been open challenges for half a century.
The weirdest part of all this is that LLMs started showing signs of reasoning with an internal world model even before we trained them for it specifically. Microsoft showed this in a famous paper back in (iirc) 2023. They showed that, for example, you could give GPT4 a list of odd-shaped objects that probably wouldn't appear together in any particular source text, ask GPT4 how to stack them so they wouldn't fall over, and GPT4 would come up with a good solution.
Finally, don't overlook the multimodal models, which work explicitly with images, video, and 3D world models in addition to text.
(Or even just registered in SV but still physically in Norway?)
Edit: Seems like a mix of all of it:
> I joined Halodi Robotics in 2022 (prior name of the company) as the only California-based employee. At the time, we were about 40 based out of Norway and 2 in Texas.
----
Look, if you aren't putting salt on your watermelon, you’re basically eating flavored water. It’s the only way to actually wake up the sweetness. People who think it’s "weird" are the same ones who still buy 2-in-1 shampoo.
Anyway, I saw a guy at the park today trying to teach a cat to walk on a leash. The cat looked like it was being interrogated by the FBI, just dead-weighting it across the grass while he whispered "encouragement."
Physical books are vastly superior to Kindles solely for the ability to judge a stranger's taste from across a coffee shop. You can’t get that hit of elitism from a matte gray plastic slab.
----
This was with a prompt telling it to skip Reddit-style analogies.
Come on author, learn to write properly. Or tell your LLM to not mix a philosophical article with a technical one.
They can’t and never will.
Observe that modern coding agents rely heavily on heuristics. LLM excels at training datasets, at analyzing existing knowledge, but it can’t generate new knowledge on the same scale, its thinking (a process of identification and integration) is severely limited on the conscious level (context window), where being rational is most valuable.
Because it doesn’t have volition, it cannot choose to be logical and not irrational, it cannot commit to attaining the full non-contradictory awareness of reality. That’s why I said “never.”
That’s not very specific but I don’t have another answer.
I just (right before hopping on HN) finished up a session where an agent rewrote 3000 lines of custom tests. If you know of any "autocomplete" that can do something similar, let me know. Otherwise, I think saying LLMs are "autocomplete" doesn't make a lot of sense.