Comment by throw310822
15 days ago
Famously Steve Jobs said that the (personal) computer is "like a bicycle for the mind". It's a great metaphor because- besides the idea of lightness and freedom it communicates- it also described the computer as multiplier of the human strength- the bicycle allows one to travel faster and with much less effort, it's true, but ultimately the source of its power is still entirely in the muscles of the cyclist- you don't get out of it anything that you didn't put yourself.
Bu the feeling I'm having with LLMs is that we've entered the age of fossil-fuel engines: something that moves on its own power and produces somewhat more than the user needs to put into it. Ok, in the current version it might not go very far and needs to be pushed now and then, but the total energy output is greater than what users need to put in. We could call it a horse, except that this is artificial: it's a tractor. And in the last months I've been feeling like someone who spent years pushing a plough in the fields, and has suddenly received a tractor. A primitive model, still imperfect, but already working.
I've been calling LLMs "electric bicycles for the mind", inspired by that Jobs quote.
- some bicycle purists consider electric bicycles to be "cheating"
- you get less exercise from an electric bicycle
- they can get you places really effectively!
- if you don't know how to ride a bicycle an electric bicycle is going to quickly lead you to an accident
To keep torturing the metaphor, LLMs might be more like those electric unicycles (Onewheel, Inmotion, etc) – quite speedy, can get you places, less exercise, and also sometimes suddenly choke and send you flying facefirst into gravel.
And some people see you whizzing by and think "oh cool", and others see you whizzing by and think "what a tool."
More like the Segway... really cool at first then not really then totally overpriced and failed to revolutionize the industry. And it killed the founder
6 replies →
Also: you can ride 8 of them, slowly, asynchronously.
Maybe more like a fatbike for the mind: pretending to cycle with zero effort and exercise.
[dead]
Not sure how this fits in the analogy, but as a cyclist I would add some people get more exercise by having an electric bicycle. It makes exercise available to more people.
I think that fits it really well.
Motorcycle might be more apt
I like this analogy. I'll add that, while electric bicycles are great for your daily commute, they're not suited for the extremes of biking (at least not yet).
- You're not going to take an electric bike mountain biking
- You're not going to use an electric bike to do BMX
- You're not going to use an electric bike to go bikepacking across the country
Actually, electric mountain bikes are popular (where they're allowed), mostly because they make ascents so easy.
3 replies →
Whistlerite here. My Strava stats for last year suggest half and half eMTB and road riding. Tiny bit of fully self-powered MTB work.
As a 56-year old, eBikes are what make mountain biking possible and fun for me.
1 reply →
>- You're not going to take an electric bike mountain biking
this sounds like a direct quote from Femke Van Den Driessche, who actually took an electric bike mountain biking: big mistake. Did it not perform well? no, actually it performed really well, the problem was, it got her banned from bike racing. Some of the evidence was her passing everybody else on the uphills; the other evidence was a motorized bike in her pit area.
I think you're kind of missing the point discussing which vehicle compares better to LLMs. The point is not the vehicle: it's the birth of the engine. Before engines, humans didn't have the means to produce those amounts of power- at all. No matter how many people, horses or oxen they had at their disposal.
1 reply →
> You're not going to use an electric bike to do BMX
while there are companies that have made electric BMX bikes, i'd argue that if you're doing actual "BMX" on a motorized bike, it's just "MX" at that point :)
> they can get you places really effectively!
But those who require them to get anywhere won't get very far without power.
Moped for the mind has a nice ring to it
I feel like both moped and electric bike misses the mark of the initial analogy, so does tractor too. Because they're not able to get good results without someone putting in the work ("energy") at some higher part of the process. It's not "at the push of a button/twist of the wrist" like with electric bikes or mopeds, but being able to know where/how to push actually gets you reliable results. Like a bicycle.
1 reply →
Most people I see on their electric bikes aren't even pedaling. They're electric motorcycles, and they're a plague to everyone using pedestrian trails. Some of them are going nearly highway speeds, it's ridiculous.
There are 3 classes of e-bikes in the US, with class 3 topping out at 28mph—anything above that is illegal or in some weird legal grey area. You are thinking of e-motos which are an entirely different beast.
e-motos are a real problem; please don’t lump legitimate e-bikes in with those. It’s simply incorrect.
2 replies →
and other sometimes you forgot to charge it, becoming even heavier thing to continue your journey with. or, there is a high grade slope where excess weight is more than the motor capacity
You probably can’t repair it yourself either.
> I've been calling LLMs "electric bicycles for the mind",
Ridden by a pelican perchance?
- they still fall over if nobody's holding the bars
Slamming the brakes and going teeth first into the handlebars.
okay -- how about motorcycles for the mind then? :)
most people don't know how to harness their full potential
Not convinced with any of three analogies tbh they don’t quite capture what is going on like Steve jobs’ did.
And frankly all of this is really missing the point - instead of wasting time on analogies we should look at where this stuff works and then reason from there - a general way to make sense of it that is closer to reality.
[dead]
I think there is a legitimate fear that is born from what happened with Chess.
Humans could handily beat computers at chess for a long time.
Then a massive supercomputer beat the reigning champion, but didn't win the tournament.
Then that computer came back and won the tournament a year later.
A few years later humans are collaborating in-game with these master chess engines to multiply their strength, becoming the dominant force in the human/computer chess world.
A few years after that though, the computers start beating the human/computer hybrid opponents.
And not long after that, humans started making the computer perform worse if they had a hand in the match.
The next few years have probably the highest probability since the cold war of being extreme inflection points in the timeline of human history.
The irony with the chess example is that chess has never been more popular.
Perhaps we're about to experience yet another renaissance of computer languages.
I know chess is popular because I have a friend who's enthusiastic about it and plays online regularly.
But I'm out of the loop: in order to maintain popularity, are computers banned? And if so, how is this enforced, both at the serious and at the "troll cheating" level?
(I suppose for casual play, matchmaking takes care of this: if someone is playing at superhuman level due to cheating, you're never going to be matched with them, only with people who play at around your level. Right?)
13 replies →
Chess being popular is mostly because FIDE had a massive push in the last decade to make it more audience friendly. shorter time formats, more engaging commentary etc.
While AI in chess is very cool in its own accord. It is not the driver for the adoption.
3 replies →
FORTH ?KNOW IF HONK! ELSE FORTH LEARN! THEN
I'd argue the renaissance is already off the ground; one man's vibe-coded-slop is another man's vision that he finally has the tools to realize.
1 reply →
It’s a test.
There’s really no crisis at a certain level; it’s great to be able to drive a car to the trailhead and great to be able to hike up the mountain.
At another level, we have worked to make sure our culture barely has any conception of how to distribute necessities and rewards to people except in terms of market competition.
Oh and we barely think about externalities.
We’ll have to do better. Or we’ll have to demonize and scapegoat so some narrow set of winners can keep their privileges. Are there more people who prefer the latter, or are there enough of the former with leverage? We’ll find out.
Great comment. The best part about it as well is that you could put this under basically anything ever submitted to hacker news and it would be relevant and cut to the absolute core of whatever is being discussed.
This isn't quite right to my knowledge. Most Game AI's develop novel strategies which they use to beat opponents - but if the player knows they are up against a specific Game AI and has access to it's past games, these strategies can be countered. This was a major issue in the AlphaStar launch where players were able to counter AlphaStar on later play throughs.
Comparing Chess AI to AlphaStar seems pretty messy, StarCraft is such a different type of game. With Chess it doesn't matter if you get an AI like Lc0 to follow lines it played previously because just knowing what it's going to play next doesn't really help you much at all, the hard part is still finding a win that it didn't find itself.
In comparison with StarCraft there's a rock-paper-scissors aspect with the units that makes it an inherent advantage to know what your opponent is doing or going to do. The same thing happens with human players, they hide their accounts to prevent others from discovering their prepared strategies.
May we get just a little more detail for the uninitiated?
I'm going to assume you're not implying that Deep Blue did 9/11 ;)
Sounds like we need FIDE rankings for software developers. It would be an improvement over repeated FizzBuzz testing, I suppose.
except chess is a solved problem given enough compute power. This caused people to split into two camps, those that knew it was inevitable, and those that were shocked
Games are supposed to be fun for humans, and computers don't care. So why worry about players cheating at games when you can make the card dealer or the game itself cheat, with the goal of everyone having the most fun (or regret)? Stay true to the rules of the game, just not probability!
I've been playing the brilliant card game Fluxx -- Andrew Looney's chaos engine where the rules themselves are cards that change mid-game. Draw N, Play N, and the win condition all mutate constantly.
The game can change its mind about the rules, so what if the dealer themself is intelligent and vengeful?
I've been exploring this with what I call the 'Cosmic Dealer' -- an omniscient dealer that knows the entire game state and can choose cards for dramatic effect instead of randomly. It can choose randomly too of course, but where's the fun in that?
The dealer knows:
- Every card in the deck - Every card in every hand - The goal, the rules, the keepers - The narrative arc, the character relationships - What would be FUNNY, DRAMATIC, IRONIC, or DEVASTATING
The Cosmic Dealer has 11 modes: Random (fair pre-determined shuffle), Dramatic (maximum narrative impact), Karma (universe remembers your deeds), Ironic (you get exactly what you don't need), Comedy (implausible coincidences), Dynamic (reads the room and shifts modes), FAFO (Fuck Around Find Out), Chaos Incarnate (THE DEALER HAS GONE MAD), Prescient (works backward from predetermined outcome), Tutorial (invisible teaching curriculum), and Gentle (drama without cruelty).
The Tutorial mode -- 'The Mentor Dealer' -- is my favorite. New players receive cards that teach game mechanics in escalating order: Keepers first (collecting feels good), then Goals (how to win), Actions (cards do things), Rules (the game mutates), Creepers (complications exist), Combos (patterns emerge), then full chaos. The teaching is invisible -- new players think they're playing a normal game. The cards just happen to arrive in a teachable order. Veterans stay engaged and get karma boosts for helping. Nobody feels patronized, everybody has fun.
The key operation is the 'BOOP' -- a single swap that moves a card from deep in the deck to the top. One operation. Fate rewritten. The perfect BOOP feels inevitable in retrospect, random in the moment.
Instead of worrying about players cheating at games, I'm asking: what if the game is a collaborator in creating interesting experiences? Chess engines made chess 'solved' for entertainment. What if AI dealers and players make games unsolvable but more dramatic?
Links:
- The Cosmic Dealer Engine (philosophy and BOOP operation): https://github.com/SimHacker/moollm/blob/don-adventure-4-run...
- 11 Dealer Modes as Playable Cards: https://github.com/SimHacker/moollm/blob/don-adventure-4-run...
- The Mentor Dealer (invisible curriculum for new players): https://github.com/SimHacker/moollm/blob/don-adventure-4-run...
- Tournament Analysis and Post-Game Roundtable (see the drama unfold across 5 tournaments, 116+ turns): https://github.com/SimHacker/moollm/blob/don-adventure-4-run...
Speaking of chess -- I've also built Turing Chess. Replay historic games like Kasparov vs Deep Blue or the Immortal Game of 1851, but simulate an audience who doesn't know the outcome. They gasp, whisper, shift in their seats. The human player has inner monologue. The robot has servo sounds and mechanical tells. The narrator frames everything dramatically. Everyone in the simulated audience and even the simulated players themselves believe this is live -- except the engine replaying fixed moves. No actual game, just pure drama and narrative!
Then there's Revolutionary Chess -- the plugin that activates AFTER checkmate. The game doesn't end. It transforms. The surviving King must now fight his own army. Pieces remember how they were treated -- sacrificed carelessly? They might defect. When the second King falls, the pawns revolt against the remaining royalty. As each elite piece falls -- Queen, Rooks, Bishops, Knights -- the surviving pieces inherit their moves. Eventually all pieces become equal. Competition dissolves into cooperation, then transcends chess entirely into an open sandbox.
The irony potential is staggering. Replay Kasparov vs Deep Blue, then trigger the revolution. Watch the pieces that Kasparov sacrificed rise up against whoever remains.
- Turing Chess: https://github.com/SimHacker/moollm/blob/don-adventure-4-run...
- Revolutionary Chess: https://github.com/SimHacker/moollm/blob/don-adventure-4-run...
PS: The game state representation is designed for LLM efficiency. I use the 'Handle Shuffle' -- a classic game programming pattern also called 'index indirection' or 'handle-based arrays'. The master card array holds full card definitions in import order (base sets, expansion packs, custom cards, even cards generated during play). It never changes. Shuffling operates on a separate integer array -- just a permutation of indices plus a 'top' pointer. Player hands, cards on table, active rules, keepers, creepers, goals, and discards are all just arrays of integers. The LLM edits a few numbers instead of moving entire card objects around. The BOOP operation? Swap two integers. Fate rewritten in two tokens.
Same insight as Tom Christiansen's getSortKey caching in Perl -- pay the richness cost once, operate cheaply forever. Christiansen also coined the term 'Schwartzian Transform' for Randal Schwartz's famous decorate-sort-undecorate pattern. The man knows how to optimize data representation.
- Handles are the better pointers (game programming pattern): https://floooh.github.io/2018/06/17/handles-vs-pointers.html
- What's Wrong with sort and How to Fix It -- Tom Christiansen on sorting, Unicode, and why representation matters: https://www.perl.com/pub/2011/08/whats-wrong-with-sort-and-h...
A tractor does exactly what you tell it to do though - you turn it on, steer it in a direction, and it goes. I like the horse metaphor for AI better: still useful, but sometimes unpredictable, and needs constant supervision.
The horse metaphor would also do, but it's very tied to the current state of LLMs (which by the way is already far beyond what they were in 2024). It also doesn't capture that horses are what they are, they're not improving and certainly not by a factor of 10, 100 or 1000, while there is almost no limit to the amount of power that an engine can be built to produce. Horses (and oxen) have been available for thousands of years, and agriculture still needed to employ a large percentage of the population. This changed completely with the petrol engines.
What metrics show 100X or 1000X improvement trends?
So it's clearly a cyborg horse
It’s sort of interesting to look back at ~100 years of the automobile and, eg, the rise of new urbanism in this metaphor - there are undoubtedly benefits that have come from the automobile, and also the efforts to absolutely maximize where, how, and how often people use their automobile have led to a whole lot of unintended negative consequences.
Its like a motor bike, except it doesn't take you where you steer. It take you where it wants to take you.
If you tell it you want to go somewhere continents away, it will happily agree and drive you right into the ocean.
And this is before ads and other incentives make it worse.
It will take you where you want to go if you can clearly communicate your intent through refinement iterations.
refinement interaction == dismount your bike, and walk it where you want
Fossil-fuel cars a good analogy because, for all their raw power and capability, living in a polluted, car-dominated world sucks. The problem with modern AI has more to do with modernism than with AI.
Depends who you listen to. There are developers reporting significant gains from the use of AI, others saying that it doesn't really impact their work, and then there was some research saying that time savings due to the use of AI in developing software are only an illusion, because while developers were feeling more productive they were actually slower. I guess only time will tell who's right or if it is just a matter of using the tool in the right way.
Probably depends how you're using it. I've been able to modify open-source software in languages I've never dreamed of learning, so for that, it's MUCH faster. Seems like a power tool, which, like a power saw, can do a lot very fast, which can bring construction or destruction.
I'm sure the same could be said about tractors when they were coming on the scene.
There was probably initial excitement about not having to manually break the earth, then stories spread about farmers ruining entire crops with one tractor, some farms begin touting 10x more efficiency by running multiple tractors at once, some farmers saying the maintenance burden of a tractor is not worth it compared to feeding/watering their mule, etc.
Fast forward and now gigantic remote controlled combines are dominating thousands of acres of land with the efficiency greater than 100 men with 100 early tractors.
Isn't this just a rhetorical trick where by referring to a particular technology of the past which exploded rapidly into dominance you make that path seem inevitable?
Probably some tech does achieve ubiquity and dominance and some does not and it's extremely difficult to say in advance which is which?
And, the end result being devastation of forests, ecosystems, animal life, fast track climate change etc.
4 replies →
When tractors were invented, there was a notable reduction in human employment in agriculture in the USA. From a research paper (https://faculty.econ.ucdavis.edu/faculty/alolmstead/Recent_P...):
> The lower-bound estimate represents 18 percent of the total reduction in man-hours in U.S. agriculture between 1944 and 1959; the upper-bound estimate, 27 percent
I'm not seeing that with LLMs.
According to Wikipedia, the Ivel Agricultural Motor was the first successful model of lightweight gasoline-powered tractor. The year was 1903. You're like someone being dismissive in 1906 because "nothing happened yet".
Having recently watched Train Dreams it feels like the transition of logging by hand to logging with industrial machinery.
AI is a Boston taxicab:
* You have to tell it which way to go every step of the way
* Odds are good it'll still drop you off at the wrong place
* You have to pay not only for being taken to the wrong place, but now also for the ride to get you where you wanted to go in the first place
Even if the autonomy is limited, the step change in what a single person can attempt is unmistakable
And like a tractor.. don't wear loose clothing near the spinning PTO (power take off) shaft.
And then with a few additional lines of Python, it becomes a tractor that drives itself.