This is interesting stuff. I wonder if these sorts of tricks are already in use at the big labs.
Incidentally, I would recommend trying implementing speculative decoding yourself if you really want to understand LLM inference internals (that, and KV caching of course). I tried it over the Christmas holidays and it was a wonderful learning experience. (And hard work, especially because I forced myself to do it by hand without coding agent assistance.)
i think this matters more for lower batch sizes (local llm and private enterprise deployment where there wont be big user at specific time for big batch size) going from mem Io bottleneck to compute.
Oh, boy. The Shrike. That thing still haunts me in a way that no other monster or alien across all of Sci-fi or fantasy really does. It's something about the inscrutability of it, especially in the first novel (still my favorite) where its purpose and backstory haven't been revealed. Sure, it's scary, but I think the mystery of its motives - and its ability to unpredictably act apparently benevolently sometimes - is where the real terror lies.
Ex historian here, now engineer. I would gently suggest you’re underestimating the magnitude of some of the transformations wrought by the technologies that OP mentioned for the people that lived through them. Particularly for the steam engine and the broader Industrial Revolution around 1800: not for nothing have historians called that the greatest transformation in human life recorded in written documents.
If you think, hey but people had a “job” in 1700, and they had a “job” in 1900, think again. Being a peasant (majority of people in Europe in 1700) and being an urban factory worker in 1900 were fundamentally different ways of life. They only look superficially similar because we did not live the changes ourselves. But read the historical sources enough and you will see.
I would go as far as to say that the peasant in 1700 did not have a “job” at all in the sense that we now understand; they did not work for wages and their relationship to the wider economy was fundamentally different. In some sense industrialization created the era of the “job” as a way for most working-age people to participate in economic life. It’s not an eternal and unchanging condition of things, and it could one day come to an end.
It’s too early to say if AI will be a technology like this, I think. But it may be. Sometimes technologies do transform the texture of human life. And it is not possible to be sure what those will be in the early stages: the first steam engines were extremely inefficient and had very few uses. It took decades for it to be clear that they had, in fact, changed everything. That may be true of AI, or it may not. It is best to be openminded about this.
Not at all, I fully appreciate that these inventions transformed life. I’m skeptical because so much of the breathless AI chatter claims AI will eclipse all these inventions. It is the breathless AI commentators, not I, who have lost all perspective on the magnitude and sweep of history.
It’s not AI per se, but rather ai enabled robotics that can change the world in ways that are different in kind, not just degrees, to earlier changes.
No other change has had the potential to generate value for capital without delivering any value whatsoever to the broader world.
Intelligent robotic agents enable an abandonment of traditional economic structures to build empires that are purely extractive and only deliver value to themselves.
They need not manufacture products for sale, and they will not need money. Automated general purpose labor is power, in the same way that commanding the mongol hordes was power. They didn’t need to have customers or the endorsement of governments to project and multiply that power.
Of course commanding robotic hordes is the steelman of this argument, but the fact that a steelman even exists for this argument, and the unique case that it requests and requires actually zero external or internal cooperation from people makes it fundamentally distinct in character.
Humans will always have some kind of economic system, but it very well may become separate from -and competing for resources with- industrial society, in which humans may become a vanishing minority.
You think an artificial intelligence would have less impact on the world than the steam engine?
The AI commentators are not saying that ELIZA will change the world, they’re saying that one of the big companies is moments away from an AGI. Sam Altman called a recent ChatGPT model a “PhD level expert”; wouldn’t infinite PhDs for $20/month or $200/month be transformative?
That is, your objection isn’t the usual “LLMs aren’t going to be AGI”, you’re saying “even if they do, it won’t be a big deal”?
>You think an artificial intelligence would have less impact on the world than the steam engine?
Not op, but yes, 100%. Steam backs nearly all development of technology of the last 150+ years. Where do you think the power come from to make things? More than half of the world's power *still* runs on steam, as will many of the systems running AI.
If steam power never existed, not only would you not exist but there's a good chance the country you live in wouldn't either. If you don't believe the effect is large, go to the farthest uncontacted place on earth and take out a CO2 meter.
It's not that I "don't believe the effect is large", but the changes from pre-intelligence planet Earth to post-intelligence planet Earth are larger because they include the invention of steam, and literally everything else too: language, writing, irrigation, cities, trade, numbers, currency, mathematics, chemistry, engineering, nations, governments, supply chains, steam, etc.
An AGI that can solve the problems we think are solvable, but we can't solve, would be huge. Any sci-fi idea that isn't ruled out by the laws of physics, but that we haven't got the brains to solve, any breakthrough that we think should be there but we haven't found, any problem that requires too much time to learn, or too many parts to hold in one human mind, any coordination that is too big for one team, any funding problem, any scarcity problem, any disease or illness problem, any long timeframe problem, are all on the table as possibilities.
There's potential there (with the pocket-PhDs), the question is whether it'll actually make a measurable difference in the long run. I mean I'm sure it will make a difference, the question is whether it's what they say it will be, and whether it'll be financially viable. At the current burn rate of the AI companies, it isn't - before long the first ones will have to give up. They won't die, they'll be subsumed into their competitors.
Anyway, the challenge is making a difference. Current-day LLMs can, for example, generate stories and books; one tweet said "this can generate 1000 screenplays a day". Which sounds impressive by the numbers, but books, screenplays, etc were never about volume.
Same with PhDs - is there a shortage of them? Does adding potentially infinite PhDs (whatever they are) to a project make it better, or does it just make... more?
This is the main difference with the industrial revolution - it, for example, introduced machines that turned 10 people jobs into 1 person jobs. I don't think LLMs will do something like that, it'll just output 10 people's worth of Stuff that will need some use.
I don't think anyone ever asked for 1000 screenplays a day, or infinite PhD's for $20. But then, nobody asked for a riderless carriage yet here we are.
> Same with PhDs - is there a shortage of them? Does adding potentially infinite PhDs (whatever they are) to a project make it better, or does it just make... more?
Yes, there is still a large demand for people with analytical thinking, a deep knowledge base, and good problem-solving skills. This demand shows up broadly across STEM fields, and it's a major reason that these fields pay relatively high.
Even just thinking of R&D, there is an immense amount of work left to be done in basic science. Research is throttled partly by a lack of cheap graduate lab labor. (If that physical + mental labor became much cheaper, the costs of research would shift - what does it take to get reagants? What does it take to build more lab space, and provide water and light? Etc.)
The present issue is that current AI does not really offer the same capabilities as a good grad student or PhD. Not just physically, as in, we don't have good robotics yet, but mentally. LLMs do not exhibit good judgment or problem-solving skills, like a good PhD does. And they don't exhibit continual learning.
No clue on when these will change, but yes, a cheap AI with solid problem-solving skills and good judgment would absolutely upend our economy.
> "I don't think LLMs will do something like that, it'll just output 10 people's worth of Stuff that will need some use."
This is why I said "isn’t the usual “LLMs aren’t going to be AGI”", but you still went straight for "LLMs aren't AGI", which was not in question.
AGI is what OpenAI says they are going for. That's the goal of all this trillion dollar investment, not to output 1000 screenplays a day, but to takeover the world, basically. What would infinite PhDs discover if they could hold all of Arxive in their 'heads' at once and see patterns in every experiment that's ever been done? What could they engineer and manufacture if they could 'concentrate' on millions of steps of a manufacturing process at once without getting fatigued or bored? What ideas could they test if they could be PhD level in a dozen subjects all at once?
A PhD generating knowledge has a cumulative effect that an equivalent intelligence generating prose purely for entertainment does not. And a whole bunch of that work isn’t really about novel insights, it’s about filling in gaps and doing knowledge work that assists people who are capable of having those insights. AI doing this enables them, also making it possible for more people to do the same.
Another interesting thing about the steam engine is much of science in the 1800s was dedicated to figuring out how steam engines actually worked to improve their efficiency. That may be similar for AI, or it may not!
I’d rather talk about the history of steam engines than AI today, so: let’s just say it sounds like at some time in the past you saw a clunky inefficient Newcomen steam engine pumping water out of a coal mine, and you hated it, and now you think that’s all steam engines are or can be or can do: they’re loud and annoying and they’re just for pumping coal mines. Then one day someone tells you they’re powering mechanized looms in cotton mills and you flat out deny it and you don’t even want to go into the mill to take a look, because you hated that first steam engine so much.
It’s right there. You can go and see it any time, doing the things you don’t think it’s capable of doing. Just a little curiosity is all you need.
No no, an intelligent person looking at a crude steam engine could see what potential it has. This is not hindsight.
It is generating large amount of power on demand.
From that one can imagine what it could do. But more importantly in this context, one could also imagine what it could NEVER do. If someone say "Oh, the mighty steam engine! It lets us print 100x more books than we were doing before. Who knows, may be some day it will even start writing new books!"
And at that point, if you understand anything about the steam engine, or writing, you can call bluff. But if you don't understand what the steam engine is doing, and if you don't actually know what it takes to come up with a story, one could take a look at the engine printing the books, and blunder into the conclusion that it printing an entirely new book is only a question of time.
So in short, it is not "hate", just the acknowledgement about what it is not.
It does take a lot of imagination and creativity to come up with new and better ways to use an already existing idea. We're currently just scratching the surface of what LLMs are going to do for us
> The aeolipile is considered to be the first recorded steam engine or reaction steam turbine, but it is neither a practical source of power nor a direct predecessor of the type of steam engine invented during the Industrial Revolution.
The ancient Greeks surely would have realised that an aeolipile could be used as a source of power, if they'd had abundant combustible fuel, a need for rotary motion, and no better source of it.
Newcomen engines are mere curiosities today, because we have better sources of power (better engines). In the past, they had better sources of power too (donkeys, wind, water, or human slaves). Newcomen engines, like all technologies, are only viable in certain economic environments. In all others they are curiosities.
Early steam engines did not produce large amounts of power on demand, though. They produced small amounts of power, were a hassle to fuel and maintain, and broke often. It was reasonable that the engineers of the 1700s said "well, until someone improves on this, it's not worth using"..
.. which is not far off from what people said about ChatGPT in 2022.
I don't know how long it'll take for AI to be as broadly impactful as the steam engine was, but.. it's definitely coming. I expect the world to look radically different in 50 years.
Thank you for your post. Very informative. Why is it too early for AI? It’s clearly an emergent cultural evolutionary byproduct that’s been many years in the making and quite mature. Perhaps your own bias is limiting you to imagine what AI is truly capable of?
I have come to think “predict the next token” is not a useful way to explain how LLMs work to people unfamiliar with LLM training and internals. It’s technically correct, but at this point saying that and not talking about things like RLVR training and mechanistic interpretability is about as useful as framing talking with a person as “engaging with a human brain generating tokens” and ignoring psychology.
At least AI-haters don’t seem to be talking about “stochastic parrots” quite so much now. Maybe they finally got the memo.
I mean that's really just a comparison to how silicon circuits work though isn't it.
"Thinking rocks" vs "thinking meat sacks" isn't much of a distinction really.
Conversely if you approach conversations the same way an LLM does and just repeat what you've heard other people say a lot without actually knowing what it means then you're also likely to be compared to a feathery chatterbox.
I think talking to people unfamiliar with LLM training using words like "RLVR training and mechanistic interpretability" is about as useful as a grave robber in a crematorium.
Obviously you don’t just say those words and leave it at that. Both those things can be explained in understandable terms. And even having a superficial sense of what they are gives people a better picture of what modern LLMs are all about than tired tropes from three years ago like “they’re just trained to predict the next token in the training data, therefore…”
Must one be an "AI-hater" to use the term "stochastic parrot"? Which is probably in response to all the emergent AGI claims and pointless discussions about LLMs being conscious.
Sampling over a probability distribution is not as catchy as "stochastic parrot" but I have personally stopped telling believers that their imagined event horizon of transistor scale is not going to deliver them to their wished for automated utopia b/c one can not reason w/ people who did not reach their conclusions by reasoning.
Technical concepts can be broken down into ideas anyone can understand if they're interested. Token prediction is at the core of what these tools do, and is a good starting point for more complex topics.
On the other hand, calling these tools "intelligent", capable of "reasoning" and "thought", is not only more confusing and can never be simplified, but dishonest and borderline gaslighting.
How nice for you. But since you totally neglected to say anything about your use-case or schema or query patterns, it’s impossible to know what this even means. Some use cases can trivially be done without any explicit transactions and you’re not giving anything up. For others (usually, something where you need to enforce invariants under high concurrency writes or writes+reads on the same data across multiple tables), transactions are pretty critical. So, it depends.
I actually like this article a lot. I do a bit of teaching, and I imagined the ideal audience for this as a smart junior engineer who knows SQL and has encountered transactions but maybe doesn’t really understand them yet. I think introducing things via examples of isolation anomalies (which most engineers will have seen examples of in bugs, even if they didn’t fully understand them) gives the explanation a lot more concreteness than starting with serializability as a theoretical concept as GP is proposing. Sure, strict serializability is a powerful idea that ties all this together and is more satisfying for an expert who already knows this stuff. But for someone who is just learning, you have to motivate it first.
If anything, I’d say it might be better to start with the lower isolation levels first, highlight the concurrency problems that can arise with them, and gradually introduce higher isolation levels until you get to serializability. That feels a bit more intuitive rather than downward progression from serializability to read uncommitted as presented here.
It also might be nice to see a quick discussion of why people choose particular isolation levels in practice, e.g. why you might make a tradeoff under high concurrency and give up serializability to avoid waits and deadlocks.
But excellent article overall, and great visualizations.
There’s a name for that: the Anthropic principle. And it is deeply unsatisfying as an explanation.
And does it even apply here? If the charge on the electron differed from the charge on the proton at just the 12th decimal place, would that actually prevent complex life from forming. Citation needed for that one.
I agree with OP. The unexplained symmetry points to a deeper level.
I was born to this world at a certain point in time. I look around, and I see environment compatible with me: air, water, food, gravity, time, space. How deep does this go? Why I am not an ant or bacteria?
Slightly tangential, but I discovered recently that the famous literary critic Harold Bloom was a huge fan of Ursula Le Guin and rated her one of the great canonical writers of the 20th century, in all of literature not just sci-fi. Also, they never met but they struck up a polite friendship over email when they were both old and chatted back and forth.
Some might consider this raises the stature of Ursula Le Guin. I consider it rather as raising the stature of Harold Bloom. He recognized how she transcended genre and belongs alongside (or perhaps, above) writers of highbrow literary fiction.
> He recognized how she transcended genre and belongs alongside (or perhaps, above) writers of highbrow literary fiction.
In the 70s and 80s, Le Guin and other SFF authors were very aware of the literary divide that often regarded most science fiction and fantasy as little better than pulp fiction. Gene Wolfe's essays and speeches in Castle of Days touch on this several times.
What changed was the arrival of a new generation of literary critics, researchers, and readers who knew greatness in some of the SFF works of the era.
> Today, I would say that about 90% of my code is authored by Claude Code. The rest of the time, I’m mostly touching up its work or doing routine tasks that it’s slow at, like refactoring or renaming.
> I see a lot of my fellow developers burying their heads in the sand, refusing to acknowledge the truth in front of their eyes, and it breaks my heart because a lot of us are scared, confused, or uncertain, and not enough of us are talking honestly about it. Maybe it’s because the initial tribal battle lines have clouded everybody’s judgment, or maybe it’s because we inhabit different worlds where the technology is either better or worse (I still don’t think LLMs are great at UI for example), but there’s just a lot of patently unhelpful discourse out there, and I’m tired of it.
Incidentally, I would recommend trying implementing speculative decoding yourself if you really want to understand LLM inference internals (that, and KV caching of course). I tried it over the Christmas holidays and it was a wonderful learning experience. (And hard work, especially because I forced myself to do it by hand without coding agent assistance.)
reply