I’d give humans some credit, they’re an adaptable bunch. AI won’t replace humans in the same way humans did not replace cockroaches. It’s a non-sequitur.
We generally don’t allow cockroaches to thrive in the spaces we claim for ourselves. Question is how much space (economic or otherwise) will AI claim for itself and whether there will be any left for us.
Also it costs a lot to train and run individual humans, and they can only be run for brief periods per day before they crash, hallucinate and possibly get irretrievably broken.
> A difficult part was constructing a chess board on which to play math (Lean). Now it's just pattern recognition and computation.
However, this was not verified in Lean. This was purely plain language in and out. I think, in many ways, this is a quite exciting demonstration of exactly the opposite of the point you're making. Verification comes in when you want to offload checking proofs to computers as well. As it stands, this proof was hand-verified by a group of mathematicians in the field.
Arguing similarly to how stockfish, the chess engine, trains I would not be surprised if this is more common in the future. I don't know if they use any proof verification tools during their reinforcement learning procedure right now, as far as I know they've been focusing more on COT based strategies (w/o Lean). But I'm hardly an LLM expert, I don't know.
They most definitely threw in rl with formal verification somewhere between GPT 4 and now. The models are better at not hallucinating. I don't think their IMO team are only show ponies...
That I'd agree with! I really need to get around to learning Lean myself. It might be interesting to try and formalize some missing theoretical pieces from my field (or likely start smaller).
Dystopia vibes from the fictional "Manna" management system [0] used at a hamburger franchise, which involved a lot of "reverse centaur" automation.
> At any given moment Manna had a list of things that it needed to do. There were orders coming in from the cash registers, so Manna directed employees to prepare those meals. There were also toilets to be scrubbed on a regular basis, floors to mop, tables to wipe, sidewalks to sweep, buns to defrost, inventory to rotate, windows to wash and so on. Manna kept track of the hundreds of tasks that needed to get done, and assigned each task to an employee one at a time. [...]
> At the end of the shift Manna always said the same thing. “You are done for today. Thank you for your help.” Then you took off your headset and put it back on the rack to recharge. The first few minutes off the headset were always disorienting — there had been this voice in your head telling you exactly what to do in minute detail for six or eight hours. You had to turn your brain back on to get out of the restaurant.
Casual reminder that the author's proposed solution to the labor-automation dystopia is to invent a second identity-verification dystopia. Also casual reminder that the author wanted the death penalty to anyone over the age of 65.
I disagree. It will be able to perform work deserving if a fields medal before it is capable of running a McDonalds. I think it will be running a McDonalds well before either of those things happen, and a fields medal long after both have happened.
I just visited a McDonald's for the first time in a while. The self-order kiosk UI is quite bad. I think this is evidence in favor of the idea that an incompetent AI will soon be incompetently running a McDonald's.
Out of curiosity, what issue did you have with the McDonald’s self-order kiosk? I actually think McDonald’s has the best kiosk I’ve ever encountered. The little animation that plays when you add an item to your cart is a little annoying (but I think they’ve sped that up). But otherwise, it’s everything I’d want. It shows you all the items, tells you every ingredient, and lets you add or remove ingredients. I have a better experience ordering through the kiosk than I do talking to a cashier.
It takes longer than ordering with a cashier, it keeps trying to upsell you, and it's always out of receipt paper because unsurprisingly the company that isn't willing to pay a person to take orders is also not willing to pay a person to maintain the kiosks.
Depends on what you're ordering and who the cashier is.
If your order is the happy path of no customizations of a combo with an experienced cashier, it can be done in seconds, for sure. "Medium #4 with a Diet Coke", pay, done.
But if you customize your burger or ordering a lot of items a la carte and you're dealing with a new cashier that has weak English skills, good fucking luck. You'll likely need to wait for them to figure out they need to call someone over to help, have to repeat your order, and you end up spending far more time.
> it keeps trying to upsell you
Yeah, I'll agree that's obnoxious, especially when it's trying to upsell you something that's already on your order. I ordered a combo. I don't need you to add another fry.
It's easily one of the most intuitive and straightforward kiosks out there today and you don't have to wait for one of the cashiers to notice you nor worry about them punching in your order incorrectly.
Glad someone else feels the same way! Knowing that I enter my order in correctly is the biggest win there for me as a picky eater. The cashier is just entering it into a computer anyways, so it makes sense for me to enter it in myself. I honestly wonder why more restaurants don’t do this. It’s not that hard to wrap a halfway decent UI around the system you already have.
If it's purely about the food, receiving it, consuming it, then sure, get the human out of the loop, interact with a machine. Ideally even the preparation is done by a machine. No human error or hair involved. Why even go there, let it be delivered to your home.
But these places are also about the experience of social connection. The bar keeper, the waiter, the chef. They are all involved in this experience and the actual food is "just" one component, one detail, albeit an important one. My favorite restaurants would be nothing without the people there.
It's similar with music. It's not just about the produced sound waves. The musician forms a social bond with the audience. Even when listening to a recording, my mind is re-living or at least imagining a live sitting, that connection with the musician. No machine generated music will ever be able to replace that.
I am more concerned with getting the right order and not with entering the right one. McDonalds will still get it wrong when you have a complex "change" of defaults even if it's entered correctly.
Other places optimize for this better by not having too many hand-overs between order and preparation.
Hmm. I’ve never really had those issues. It’s also much faster and easier than ordering with a human. I guess it does try to upsell you, but humans often do, too. And to me, it’s worth it to just click “No” in exchange for the added convenience (mostly in getting my order right).
I have had them run out of receipts, but it’s never mattered for me. If I’m dining in, the plastic number you carry to your table makes sure I get my food. And if I’m taking it to-go, they always find me anyways.
> It’s also much faster and easier than ordering with a human.
I'm not sure how that could be. I can walk up to the counter and say "Big Mac Large Fry Small Coke" faster than you can navigate the first screen of the kiosk, and a skilled counter worker can key that in and be done before I even get my credit card out.
The problem, I’m a picky eater. I never order something that simple. I always need it with “No X” or “Only Y”. Cashiers often struggle with that, even if they understand me well (which they don’t always). It’s easier for me to see everything an item comes with and make sure I’m entering my order correctly.
McDonalds' menu is not designed for folk like you. In my part of the world, we had traditional fast-food joints where the question would be inverse: out of the things you can see and add to your burger, pick a few. That is very efficient with a human prepping your burger.
Hm? McDonalds is one of the best for customization. Everything is removable and the software knows the calorie count of each ingredient so the total that shows up next to each item in the cart is accurate
The kiosk ordering at McDonalds is made to adjust the defaults which is slower than the mode I described: if automated, that would be a screen to choose the item, and then add things into it from a list — fewer screens, faster ordering.
Everybody would go through this workflow built for customization, and at McD they do not. This to me means they are not building for this usecase.
You have to wait in line behind several people to get to the point where you can talk to the cashier. Most McDonalds have several kiosks. There is usually little or no line. I can place my order grab a table.
> I can walk up to the counter and say "Big Mac Large Fry Small Coke"
In what culture is it normal to just state your order without greeting or saying “please”? You also didn’t include the time taken squinting at the menu and trying to figure out what you can even order.
Since you asked, and since I take my kids to the McDonald’s play place some weekends, and I’ve actually spent a bit of time pondering my ideal kiosk UI and what I don’t like about theirs:
It seems designed to maximize how many screens they show you to make an order. Each one with a slight delay and animation.
At a drive through I can say “gimme a number one, medium, with a Coke Zero” and they give me my total. That’s the convenience the kiosk is up against.
At the kiosk there’s:
- A welcome screen you have to tap
- A “carry out or dine in” screen
- Always one other screen with a dumb question about apps or whatever, tap through
- A top level menu with a bunch of categories, burgers, drinks, sides, desserts, etc… I guess I want burgers? But it’s a combo, hmm. I guess I’ll figure out how to make it a meal. Tap burgers.
- Then another screen with burgers, in a different order than the drive through numbering, tap Big Mac
- Then another dedicated screen to shows you a picture of a Big Mac, with a bunch of customization options, which you have to scroll past and verify that it matches the defaults you expect, and at the bottom you can tap add
- Then another screen asking you if you want to make it a meal
- Then another screen asking the size
- Then another screen asking what to drink
- Then another screen that shows you the drink
- Then another screen for what size
Etc etc etc. Each of these screens takes a few seconds to display too, just slow enough to be infuriating.
In my mind the ideal kiosk is something where you get “the menu” (like what you see on the billboard in the drive through) with the usual big squares with a number on them and a picture of the meal. Tapping one puts it in a “drawer” section with my order in it, and each item in the drawer can have simple in-line edit controls for “size” and “what to drink”, with them showing up empty in a way that makes it obvious I need to fill in those answers before I can check out.
I should be able to tap one button for the combo number I want, another for the size, another for the drink, then checkout, all on one screen without long delays. If I don’t want a combo but want individual items, I can just scroll down a bit to look at the full menu. The order drawer stays where it is.
Or hell, just let me say “number one with a Coke” and have a very simple ASR and NL parser figure it out and put it in my pending order to edit.
Customizations can be behind a simple “customize” button on each item in my pending order. If I don’t have customizations I can just ignore it. What you get with no customizations is what you’d get if you just order it verbally to a human without specifying anything. The concept of “here’s how we typically make it, if you want anything different let us know” is a very deeply ingrained and familiar concept to restaurant patrons, and being forced to answer every little question even if you don’t care, adds up to a lot of frustration.
Fast food places came up with the combo numbering system to make ordering faster, and it was super convenient and fast, because there’s a financial incentive to get you through the drive through because you’re blocking other customers. But since they have several kiosks available, they seem to not care at all about the efficiency of the user interface, because it’s not a problem for them. But it’s still a problem for me, because I still want to order quickly, despite it not blocking other customers. It’s a huge step down from just saying “number one with a Coke”.
Recently I tried to order at a Subway (which has decent quality food outside of the USA). They have kiosks. The kiosk only responded to touch about 60% of the time and took two seconds to respond. The employee who could've easily taken my order was just standing there bored. The future is here and it sucks.
Most repeat customers use the app, which sports the digital equivalent of a loyalty program, and various coupons. And lets you save your 'usual' order with customizations etc. Plus the annoying push notifications for FreeFrydays or whatever. And upsells, new product launches, etc.
My recollection is that the kiosk is just a weak facsimile of the app. And wasn't terrible, but everyone's standards vary.
One could hardly ask for a task better suited for LLMs than producing math in Lean. Running a restaurant is so much fuzzier, from the definition of what it even means to the relation of inputs to outputs and evaluating success.
I think Lerc is saying that LLMs will be pressed into service managing McDonald's restaurants long before they are actually capable of managing said restaurants successfully.
Not necessarily. Obviously playing Kasparov on the board requires more planning ability than managing a McDonald's but look at where chess bots are now.
There's much more to being human than our "cognitive abilities"
> Obviously playing Kasparov on the board requires more planning ability than managing a McDonald's
Not obvious and in fact I think the opposite is way more likely. Chess is well-defined and self-contained in a way that managing a restaurant with fleshy customers never will be.
But that is also non-obvious. Even managing human employees — let alone customers — required a planning ability related to emotional intelligence that many a person with good pure logic ability simply lacks.
Also, there will be hundreds of disparate tasks that are happening in parallel, and even humans still make up frameworks to discover most urgent/important work that needs to be done first.
They no longer have to limit themselves to forking software but can do a global Google Burgers in a single prompt. It will no doubt be a huge success before shut down.
Stockfish did not teach itself to play chess. You are probably thinking of Leela Chess Zero - an open re-implementation of AlphaZero - both were given nothing but the rules of chess and a board and played millions of games against themselves until they were the strongest engine available at the time.
Stockfish's neural net evaluation model was trained on millions of its positions with its own original algorithmic evaluation function (entirely developed by humans) and search tree. The result was a much smaller model than Leela's that requires little computation (not even a GPU), paired with its already extremely efficient search/pruning algorithms that made it stronger than Leela in competitive play. Leela's evaluation function is much stronger (at one ply it has an ELO of around 2300, Stockfish is probably closer to 1800), but it requires vastly more resources and those are always bounded in a match.
Humans haven't learned as much new information about chess from Stockfish as we have from Leela.
The issue with this prediction is the gulf between problem-solving using known tools, versus creating new concepts for problems where existing tools aren't enough.
All AI proofs so far, including this one, are using existing tools in new ways, rather than inventing new tools. This is not surprising if you know how these models are trained. These existing tools are in distribution. New tools are not.
Problems worth of a Fields Medal likely require new tools to be invented. Thus it is not clear whether progress within the confines of the current paradigm is enough.
We could get this weird spiky situation where the AI is insanely superhuman at all problem solving, but completely incapable of coming up with a single new tool. It discovers everything there is to discover, subject to existing axioms and concepts.
Timothy Gowers gives some commentary on this in the attached PDF.
> A difficult part was constructing a chess board on which to play math
We have that chess board for quite a while now, over 40 years. And no, there is nothing special about Lean here, it is just herd mentality. Also, we don't know how much training with Lean helped this particular model.
I think your analogy is good but I don't believe modern LLMs use Lean or any lean-like structure in their proofs. At least recent open source ones like DeepSeek can do advanced math without it (maybe the most cutting edge ones are doing it I can't say).
Managing a McDonalds is a question of integration and modalities at this point. I don't think anyone still doubts that these models lack the reasoning capability or world knowledge needed for the job. So it's less of a fundamental technical problem and more of a process engineering issue.
Both links talk about the same thing? The first one just being more general. And yes, I would expect no less from a poorly constrained single agent that was instruction trained to be helpful and friendly. But if you look at how this has evolved as a benchmark [1] then the latest models show no doubt that can actually deal with this limited, simulated scenario given the correct setup.
I disagree. Even frontier models still achieve way worse results than the human baseline in VendingBench. As long as models can't manage optimally something as simple as a vending machine, they have no hope of managing a McDonalds.
Assuming you can still sue McDonalds I am not sure if this is a problem in the robotic llm case. I'm also trying to imagine a case where you would want to sue the llm and not the company. Given robots/llm don't have free will I'm not sure the problem with qualified immunity making police unaccountable applies.
There already exist a lot of similar conventions in corporate law. Generally, a main advantage of incorporation is protecting the people making the decisions from personal lawsuits.
>Police officers are human. In the United States in the vast majority of cases you can't sue the police, only the community responsible for them.
Police are a monopoly; nobody has a choice about which police company to use. McDonalds are not a monopoly, and many customers would prefer to eat at competitors run by entities that could be sued or jailed if they did anything particularly egregious.
You are missing the point. The point is you can still sue the McDonalds. With the police there is a human intuition to also want to sue the officer, given the officer is a human being who has free will and thus made a choice to violate your rights.
The same intuition applies if you walk into McDonald's and a person there mistreats you. You want that person held responsible.
But the LLM is not a person. What is there to even sue? It just seems like it would simply pass through to the corporate entity without the same tension of feeling like we let a human get away with something. Because there is no human, just a corporation and the robot servicing the place.
Put another way - if the LLM is not a person, what is the advantage of a personal lawsuit?
Just sue the McDonalds. Even in a case where the LLM is extremely misaligned and acts in a way where you might normally personally sue the McDonald's employee, I'm just not sure the human intuition about "holding someone accountable" would have its normal force because again - the LLM is not a person.
So given we already have the notions of incorporation and indemnification it doesn't make sense to say what is precluding LLMs from running McDonald's is they can't be sued. If McDonald's can still be sued, then not only is there no problem, there is very likely not even a change in the status quo.
> given the officer is a human being who has free will and thus made a choice to violate your rights.
The purpose of qualified immunity is for when an officer does something that turns out to be illegal but they were both told to by their superiors and did not think it was in violation at the time.
An officer making a choice to violate your rights would not be eligible for qualified immunity.
Wow yes excellent point, because of course a police officer facing the threat of legal action would never attempt such a low bar lie. Oops my boss told me to. Oops I didn't know. Case dismissed.
Excellent standards for people authorized by the state to run around with a badge and a gun in a free society. Your comment history on this is so unimpressive. Would you countenance the same excuses in anyone else? A man puts on his police uniform and suddenly you think he should be immune from civil prosecution because "my boss told me so" and "I didn't know"?
I wonder if you will make similar excuses for robo cop. Or if your principles merely extend to whatever human you can find in uniform willing to tolerate your friendship.
You seem to have read a lot more into my comment than was there.
Plus, qualified immunity is only for civil precedings. Individual officers are still liable for any criminal actions they take. I see a lot of people say that some officer should be in jail and blame qualified immunity when those two things are not related at all.
I'm not arguing, at all, that police should be immune to prosecution individually. I'm trying to make the point that, if you are trying to hold police individually accountable for their _criminal_ actions, qualified immunity isn't the thing that's preventing that. There's a whole legal system and union/police culture that's responsible for that.
Qualified immunity is thrown around so much in contexts where it makes it clear that people don't understand what it means and gets used, as it was in your comment, as a bogey man that's to blame for all the times police get let off the hook for their misbehavior. All I'm trying to do when correcting you (and others) about qualified immunity is to both redirect your anger and effort into changing something that will actually make a difference and/or prevent you from spending the mental or physical energy chasing a dead end.
You're not correcting anything. Nothing in my comment suggested I don't know what qualified immunity is or I don't understand the difference between criminal and civil liability. Quite the contrary if you are capable of reading.
You seem to be arguing with yourself, not with me. If you are satisfied with a cop only facing criminal liability (often from the same prosecutors that rely on police to make other cases, among other issues as you pointed out) fine that is your prerogative. Don't file a civil case. But don't misrepresent my position. Criminal prosecution does not preclude civil, nor the other way around. Citizens should not face such hurdles to file civil suits, irrespective of whatever happens re a criminal case. Why is that so hard for you to understand? Surely your comprehend that one can be found liable both criminally and civilly in many cases. Or tried for both but subject to different penalties (including none) depending on how each goes. Why are your LEO buddies so special as to be largely exempt from the rules that govern the rest of us?
The fact your comment history is riddled with these continued misrepresentations on this topic while you claim to educate is simply galling. Have a good day, I don't think I can continue in good faith with someone who seems to predicate engagement on this topic with unfounded assumptions about others education on the issue. Your own comments on this topic indeed are indicative of a severe projection in this respect.
McDonald's are franchises - you generally want to sue the local owner or threaten them in addition to the holding company.
That only requires someone own the ai managed McDonald's though. so long as they can't avoid responsibility by pointing to the AI I don't see why you couldn't sue them.
My claim is that LLMs waste a lot of time training on all available data.
Math is a sequence of formal rules applied to construct a proof tree. Therefore an AI trained on these rules could be far more efficient, and search far deeper into proof space
It has been tried. Lenat's Automated Mathematician, for example. The problem is that the system succumbs to combinatorial explosion, not knowing which directions are interesting/promising/productive. LLMs seem to pick up some kind of intuition from the data they are fed. The generated data might not have the needed "human touch" or whatever it is.
StockFish amplifies weak "intuition" (heuristic and simple neural estimators) through extensive Alpha-Beta search made possible by the low(ish) branching factor of chess. It doesn't work for Go already (KataGo uses larger neural networks to guide search more efficiently). I doubt either will work for math where branching factor is even bigger and the success criteria (is the result interesting and so on) are not strictly defined.
Well, brikym, exactly how hot is this hot coffee? If it’s within normal expectations for coffee it is likely that person’s fault. If it is 210 degrees F, it is likely McDonald’s fault.
No, we’re not going to be flipping burgers either, they will have physical robots for that. 20 years down the line I wonder what work all of us will be doing.
Calling AI-generated images art is a stretch. Same thing with creative writing. It can make some low-cost illustrations and writing, but it is very far from decent art. Compare those results with their amazing coding or math capabilities.
Those are either copies or bad. I think the difference in quality and originality is huge, even if they can create plausible looking images in already invented styles.
And yet, they're good enough to fool most people, which is what you're trying for if you're aiming to automate the production of art. The fact that we got there more than 2 years ago, before agentic coding was even barely useful, should give you an idea of where things are going.
This seems like a non-sequitur. The entire point of AI is that it's able to generalize and perform at or above human level. Let me know the next time your camera is able to simulate conceptualizing something.
Or is your argument that AI is permanently doomed to not work?
It obviously does not follow since it precedes AI in "automating ... art" by 150 years.
What happened to art since?
We got artistic photography on top of paintings as well. It did not become widespread right away as people were mostly enamored by the simplicity of getting a realistic image first. But after that died away, people made art out of photography too.
Yes, those who did landscapes or portraits for hire were affected financially, but we ended up with more art, not less.
If I need to spell it out: genAI for image generation will also become an avenue for real artistic expression, as some pioneers are demonstrating, even if image generation is democratized, there will be a difference between art and non-art. It will also not kill conventional art either.
If I need to spell it out for you: no artistic 'tool' has yet attempted to automate the thought that goes into art. Would you consider someone an artist if all they did was hire someone off fiverr?
Sure it has: automatic exposition and focus, HDR image blending... These used to require more thought and are now done for you. You still sometime override it, and an artist more often than a non-artist.
A tool is let to make a set of choices, but there is some control: you can prompt an AI to draw a white animal, a white bird or white swan — if you get a white horse with wings and a beak on any of the latter ones, we could say that AI is putting more thought, but at the moment, it wouldn't and we'd call it broken.
Depends on the set-up. Does you setting the stage, costumes and directing the photoshoot done by someone from fiverr with good camera and knowing how to use it not make you one?
Nonsense. Have you been watching the figure live stream? Or the Unitree video from yesterday with real time novel action generation? We’re less than a year away. If you can cook a burger, assemble a sandwich and clean up surfaces you’re all of the way there.
Powerful AI is here as Dario said in 2024. Open your eyes. Anthropic acquired Stainless because they know how to use Claude better than their own employees
I have solved open problems of fields medalists and can't get a job in academia. I currently make 4800 a month after taxes as a lecturer in San Diego, pivoting to SWE. Math PhDs are having a hard time
There's pretty noticable differences in ability to self-regulate between people who had unfiltered access to computers and smartphones when growing up and not, beyond simply generational shifts.
I think the parent's assumption is that 16 is old enough for those problems to be largely avoided even if you gave them unfiltered access, which is what makes it not just kicking the can down the road.
I know a six-year-old with an iPhone. Any efforts to not do that is going to help.
I remember early internet and how it was on a teenager. Can’t imagine just signing up your kids for what it is now just because I won’t be able to keep them from it forever.
I do not believe it will replace humans.
reply