TLDR: It’s actually more about kernels changing with batch sizes, and you can solve it by making these kernels not depend on batch sizes. It took their inference time from 26s to 42s.
Perhaps this is a form of Gell-Mann Amnesia (but kinda inverted) where everyone views AI as too inaccurate for their own niche, but perfectly fine for every other field that they know comparably little about.
I’ve been thinking about this, or a variant of it.
Hypothetically, I’m scared and sad that AI can replace me (it currently can’t, not literally, but a lot of my skill and expertise, built up over say 30 years, that used to be valuable and rare is now cheap to get from an AI).
Let’s try to see the upside. How ‘powerful’ would it make me if, at the cost of my own edge being dulled, can access everyone else’s edge?
I am still my own expert. Now with AI I have a minor expert in everything else as well. What is the best way to use that? don’t have an answer but it’s an aspect I haven’t seen discussed much and I think it is worth bringing up.
As of today, "minor expert" is the wrong way to phrase it. The flagship LLMs today are at best at an initiate or apprentice level in every field.[1] This is not meant as a derisive remark – having something at hand that is initiate level in every field is remarkable and useful. But it's nowhere near expertise anywhere.
You seriously think current LLM is just at apprentice level in programming? It can write stuff one shot that I’d expected even some experts to struggle to do even with ample more time allowed.
They will one shot with bunch of duplicated code, somehow they will just omit basic things, like security middleware. It will all kind of work. But then you make additional passes to do review and clean up, suddenly there is a lot more work to make it half decent.
You think every person only thinks of their own job and no one dreams of anything bigger from humanity’s perspective? I’m never going to be in some kind of space colony but do I want to see them happen? You betcha.
> You think every person only thinks of their own job and no one dreams of anything bigger from humanity’s perspective?
Yes, the vast majority of people care about their jobs first rather than a theoretical future where mostly rich people colonize space. It's much easier to imagine yourself becoming poor than living in Mars.
What I was 9, I lay awake thinking about how humanity will one day explore space and discover all sorts of new species, make new friends.. I imagined music concerts with humans and animals and all sorts of creatures.. and then it hit me, I won't be alive by then. I basically realized I'm mortal, and cried for hours.
I want that type of sadness back, please. Right now I'm relieved if anything to know I won't see 2126 and beyond, but that's not because I'm inherently pessimistic, I came to this world optimistic as can be.
> I want that type of sadness back, please. Right now I'm relieved if anything to know I won't see 2126 and beyond, but that's not because I'm inherently pessimistic, I came to this world optimistic as can be.
This is normal I think. You grew up. Not wanting to live forever is totally sensible.
I didn't think about "forever", then or now. It's more about seeing the next few hundred years as an exciting time of progress and adventures, of ideas -- rather than just a snuff movie I'd rather pass on.
And "growing up" kind of can mean everything and nothing. I noticed that humans prey on humans, destroy and poison animal species and each other for nothing, for entertainment or to compensate for psychological issues. It is normal for a child to reject that, and a moral imperative for an adult to act on that, not to "get over it".
Come to think of it, kids aren't "naive" that they don't know these things. Adults are correctly ashamed of them, so they only discuss them with people who will accept shrugs, i.e. other adults. If we were honest to children about the world, the things in it we accept, they would freak the fuck out and probably be traumatized. We tell them to not kick a dog, but oh by the way, Abu Ghraib. If they say "shit" they get scolded, if a strange man scares them you should tell their parents and they'll be mad at that man, but if people murder thousands of little children, just evaporate them, that's just how it is, nothing can be done. Some people are insane and rule over us, starve us out, and that's why mommy and daddy have to work 5 jobs to have nothing compared to what the grandparents had, but here's that cartoon you like.
I'd say what is called "rational" and "grown up" is often just diminished and broken, barely surviving via constant distraction from one's own state.
I think it’s this but also that we all see the value in our own niche because it’s ours, and have more trouble seeing the values in other niches. So it becomes a self perpetuating positive reinforcement thing.
I don't see why that should be the case. The only reason software is getting focused on first is:
1. Software devs are obviously going to have a better idea how to apply AI to software development compared to other fields. So of course the coding tools are going to be the first things made.
2. Formal verification makes the problem easier by allowing for iterative feedback (compilers, proofs, etc.)
The second argument is, I think, somewhat valid, but ignores that a lot of other professions also have similar verification systems even if they're a bit less rigorous. The first argument just explains why things are the way they are now, it's not indicative of the future. I don't want to fall into the trap of thinking that other jobs than mine require less cognitive horsepower or whatever, but I don't see what's particularly special about other jobs if it can do hard STEM stuff.
I thought the same but i dont think so anymore. My wife is a senior manager at a big 4 consultancy gig and she says copilot became freaking good at understanding tax, multinational company structure etc etc. Even if you need a few partners and experts at the top to validate things you can cut huge amounts of workers.
Exactly. Regarding software, it is trained on a massive corpus of code and the feedback loop can be very fast (playing well into LLM's upsides) and results are ... mediocre.
Recently I had to go through some building regulations and Claude's advices were catastrophic.
The real trouble isn't that it can replace us. Instead, consider that when there have been two comparable technologies in the market, the market has invariably chosen the lowest-quality/worst one. Why? That's easy to understand... while the chosen option is objectively the worst, it's always the cheapest. And cheapest wins. It's not "can AI do his job?" so much as "is the AI cheaper than a human?". And I think we all know the answer to that... even if the silicon's expensive now, volume pricing, data center buildouts, and other economic forces will soon make it cheaper.
The thing that is truly mysterious to the managerial and ruling classes though... when everyone is unemployed, who will be able to afford to buy your junk? Whatever industry you're in, whatever it is you're selling, the people buying that have the money to buy it because they still have jobs. If you're cutting jobs at your company, that helps the bottom line, but every other company is doing the same thing. And they're laying off your customers.
The idea of the social contract impacting perceptions of AI is interesting to me. I hate to use the words “permanent underclass”, but perhaps the main difference is a fear of that permanent underclass actually materializing. In the US, it seems that that would be the logical endpoint of the capitalist system and many people predict AI simply replacing them permanently. Of course, China is not completely communist, but since their social contract is much less individualistic and more collectivist, maybe that makes people see AI as much more likely to uplift society as a whole or at least “trickle down”.
I think this might be a bigger reason as China’s economy for the youth isn’t looking the brightest right now.
> Non-technical middle managers who have not written a line of code in their lives, now feel that the biggest obstacle between them and greatness has lifted.
I find it interesting how this is almost the “democratization” you mentioned that AI provides. While AI “democratizes” certain technical ability, in some ways the democratization of things can actually be bad, in that this “democratization” pushes us towards a system in which people are completely fungible, and so lose their individual bargaining capability. By democratizing this ability to the non-technical middle manager, the junior software engineer ends up losing their unique contribution and hence vote.
I read a while ago about boycotting AI if you can, and I would love to, but this issue makes me wonder if that could even be effective. If the goal is to remove every unique contribution you provide, what can you take away with a boycott?
I'm thinking of some similarity between the non-technical middle manager using agents to avoid the biggest obstacle of their developers, and a person installing a wall socket at home.
It's not too difficult to install a wall socket, but many things can go wrong, so people would normally call an electrician. In some jurisdictions you are not allowed to install it without a license, or probably you can install it only for yourself but not for anyone else if you don't have a license, and you need to call an inspector and check your work when you're done. Some other jurisdictions truly don't care, you can do whatever, and all the possible damage is on you.
Electricians are somewhat fungible, because you often don't care who will do the job for you, but the profession exists and they can pay their bills.
I wonder how far we are until, at least in some jurisdictions, a person won't be legally allowed to update the website that stores customers' data or processes payments without being a licensed software developer, and how rules, should they be adopted, will change our profession.
> in that this “democratization” pushes us towards a system in which people are completely fungible, and so lose their individual bargaining capability.
Hasn't this been the case since the power loom put hand weavers out of a job?
This time it is different. Ai promised to be the power loom for all things. Knowledge work and all physical work through robotics. And there will be no more jobs to turn to. Why might an AI company allow for humanity to live in this future? They’d have no utility. No economic purpose. They’d have no unique skill. Supporting them over the datacenter amounts to investing in waste heat. And they may even rise up and destabilize the horror you’ve created for planet earth. Game theory suggests keeping humans on earth is increasingly a liability.
I hate the term "democratization". It's putting a respectable face on something that shouldn't be respectable at all. In the age of the internet, coding and other creative skills have always been largely democratized to people that care enough to learn. Nothing is being "democratized" by AI, there's simply (an attempt) at driving the value of actual skill to zero so the skill-less and stupid can purchase their way to mediocrity (without the benefit of transferring that money to someone who has worked to be skilled). There is NOTHING "democratic" about that.
It’s not coding that’s being democratized, it’s the ability to create games, tools, etc, that require coding.
Suppose it became possible to buy a Ferrari (not that new hideous one) for $5k. That would be democratizing Ferrari ownership: far more people could do it. I’m sure there would be investment bankers who would complain it devalues their hard work (I am not a fan of those people, but it is notoriously long hours). Does that make a $5k Ferrari less of a democratization?
If that happened, Farrari's would simply lose their value and something more rare would take it's place. The scarcity is the point. Same with art, if you "democratise" it somehow humans will seek out a new form of artistic expression that is not accessible to people with a $10 midjourney subscription. And that is OK, good actually.
Perhaps to some degree, but there is intrinsic value to a Ferrari if you enjoy driving. That doesn’t change with scarcity.
But I agree, and even celebrate, that a hypothetical future with abundance will be bad for speculators and the luxury-obsessed, and great for people who enjoy the first-order benefits of products.
That's a product, not a skill. We're not talking about people getting cheap cars, we're talking about nobody being able to make a living anymore. And no, this isn't like other revolutions, because there's no "upskilling" into it when it threatens all white collar work. (I'm sorry but I don't think prompting AI is going to be some amazing new profession, it'll be at best a transitory phase). Best case is to buy some durable jeans and learn a trade?
Of course, I don't think this is happening any time soon because I don't believe the hype machine. But if you do believe the hype machine, you should be honest about what you're facing.
Also, for what it's worth as a person that's written (and will continue to write) games and tools etc., there were already plenty of forces democratizing that (free engines, asset stores with reusable code and art, etc.) The difference was that there was still skill involved, and at the end of the day you were paying another human being for their effort if you bought or used pre-existing components. There's no skill in AI, not really.
Also, I'm really not that interested in playing a game created by AI, and judging from the reaction gamers and game developers have had to this technology, I don't think many people are. So, yay, you can make a game that has a tiny audience with zero satisfaction from having done something hard and learned something? Also it's likely going to be extremely derivative, because the AI can only really work with what it's been trained on.
Wait a second, you say its not coding thats getting democratized. It is the "ability to create" software... hm i wonder what would be the ability to write software? Maybe something like software engineering?
Also your example with Ferrari is completely flawed, and I dont give a fuck about what investment bankers think or do. What are you even talking about here?
Please explain the first part more than you hand waved away to make a completely unrelated metaphorical case. What is the ability to create software if not software engineering? How are both different?
The whole point of AI coding is (partial) decoupling of software from human coding.
You might as well ask what the ability is to create arbitrary images to any specification, if not painting? But AI image gen is not painting.
It’s painful and confusing to those who’ve mistaken the tool for the product. Why, you ordered that pizza for delivery, that’s not cooking! No, but it is eating.
Same thing with software. Some people just want the outcome, not the process, and it’s pointless to whinge about how it’s somehow less legitimate.
I am against how its being used as well, in fact Authoritarianism is literally "central power to preserve the political status quo" - and that is what you have folks: OpenAI/Anthropic/Google/Facebook/etc are all central powers.
Coding was always power that you can have if you just read the freely available info but having access to a computer and internet is not free - if we can change this it would make democratized.
>You just had to put in the effort of learning and understanding.
Yeah it's always so weird when people anthropomorphize the harsh physical reality of learning a skill that is accessible to everyone as some sort of secret boys club, where the existing club members do not let you in, when the only difference between the members and outsiders is whether they've paid the entry fee or not.
A 10'000 hour entry fee does rule out a fair bunch of people though, in practice. While there are few artificial barriers to learning to code, there still are some natural ones, like time.
is time a barrier or a requirement? its a bit like saying time is a natural barrier to solving hashes. like yeah, its an inherent part of the process that cannot be skipped. You cannot learn without time, but thats just the laws of the universe not a barrier to learning.
Thinking that "individual bargaining capability" capability is an especially useful thing for society as a whole is one of the things that has got us into the current Hobbesian mess
Unlike a lot of anti-AI writing I have seen which can drift into the territory of reactionist rhetoric, this feels very grounded. As a student entering the job market, I personally am quite conflicted about my usage of AI, due to all the effects you have said. Especially when testing out more recent “agentic” coding tools where they do literally all the work, not only are there the moral qualms but they’re also not enjoyable to use.
However, in the tech bubble I live in to use these tools for as much as possible in order to “make it” in some sense and actually be employable. Like you said it is a game of chicken. Perhaps the best strategy (for me) is to campaign for institutional guardrails on usage while continuing to individually try to be competitive?
Kinda a smart alec response, but how do you know you aren’t going to increment off the end of the array when operating normally? I guess it is twice the danger.
I think in the context of these it’s more of “we’ve discovered a bug” which gives you more information than “there is a bug”. The main difference in information being that the former implies not only there is a bug but that LLMs can find it.
Either way you would still need some check to ensure your digits are digits and not some other type of character. Having zeroed out memory read as a bunch of NUL characters instead of like “00000000” would probably be useful, as “000000” is sometimes a legitimate user input
Also, if your laptop is work provided, your laptop is going to have little to no privacy no matter what.