The claimed connections here fall apart for me pretty quickly.
CPU instructions, caches, memory access, etc. are debated, tested, hardened, and documented to a degree that's orders of magnitude greater than the LLM-generated code we're deploying these days. Those fundamental computing abstractions aren't nearly as leaky or nearly as in need of refactoring tomorrow.
I think the concern is not that "people don't know how everything works" - people never needed to know how to "make their own food" by understanding all the cellular mechanisms and all the intricacies of the chemistry & physics involved in cooking. BUT, when you stop understanding the basics - when you no longer know how to fry an egg because you just get it already prepared from the shop/ from delivery - that's a whole different level of ignorance, that's much more dangerous.
Yes, it may be fine & completely non-concerning if agricultural corporations produce your wheat and your meat; but if the corporation starts producing standardized cooked food for everyone, is it really the same - is it a good evolution, or not? That's the debate here.
Most people have no idea how to hunt, make a fire, or grow food. If all grocery stores and restaurants run out of food for a long enough time people will starve. This isn't a problem in practice though, because there are so many grocery stores and restaurants and supply chains source from multiple areas that the redundant and decentralized nature makes it not a problem. Thus it is the same with making your own food. Eventually if you have enough robots or food replicators around knowing how to make food becomes irrelevant, because you always will be able to find one even if yours is broken. (Note: we are not there yet)
That is short hand. The problem exists of course, but it is improbable that it will actually occur in our lifetimes. An asteroid could slam into the earth or a gamma ray burst could sanitize the planet of all life. We could also experience nuclear war. These are problems that exist, yet we all just blissfully go on about our lives, b/c there is basically nothing that can be done to stop these things if they do happen and they likely won't. Basically we should only worry about these problems in so much as we as a species are able to actually do something about them.
> that's a whole different level of ignorance, that's much more dangerous.
Why? Is it more dangerous to not know how to fry an egg in a teflon pan, or on a stone over a wood fire? Is it acceptable to know the former but not the latter? Do I need to understand materials science so I can understand how to make something nonstick so I’m not dependant on teflon vendors?
You do need to be able to understand nonstick coating is unhealthy and not magic. You do need to understand your options for pan frying for not sticking are a film of water or an ice cube if you don't want to add an oil into the mix. Then it really depends what you are cooking on how sticky it will be and what the end product will look like. That's why there are people that can't fry an egg, people that cook, chefs, and Michelin chefs. Because nuance matters, it's just that the domain where each person wants to apply it is different. I dont care about nuance in hockey picks but probably some people do. But some domains should concern everyone.
At what point is the threshold between fine and concerning? Seems like the one you put is from your point of view. I’m sure not everyone would agree and is subjective.
Being an AI skeptic more than not, I don't think the article's conclusion is true.
What LLM's can potentially do for us is exactly the opposite: because they are trained on pretty much everything there is, if you ask the AI how the telephone works, or what happens when you enter a URL in the browser, they can actually answer and break it down for you nicely (and that would be a dissertation-sized text). Accuracy and hallucinations aside, it's already better than a human who has no clue about how the telephone works or where to even begin if the said human wanted to understand it.
Human brains have a pretty serious gap in the "I don't know what I don't know" area, whereas language models have such a vast scope of knowledge that makes them somewhat superior, albeit at a price of, well, being literally quite expensive and power hungry. But that's technical details.
LLMs are knowledge machines that are good at precisely that: knowing everything about everything on all levels as long as it is described in human language somewhere on the Internet.
LLMs consolidate our knowledge in ways that were impossible before. They are pretty bad at reasoning or e.g. generating code, but where they excel so far is answering arbitrary questions about pretty much anything.
I am no CS major, nor do I fully understand the inner workings of a computer beyond "we tricked a rock into thinking by shocking it."
I'd love to better understand it, and I hope that through my journey of working with computers, i'll better learn about these underlying concepts registers, bus's, memory, assembly etc
Practically however, I write scripts that solve real world problems, be that from automating the coffee machine, to managing infrastructure at scale.
I'm not waiting to pick up a book on x86 assembly first before I write some python however. (I wish it were that easy.)
To the greybeards that do have a grasp of these concepts though? It's your responsibility to share that wealth of knowledge. It's a bitter ask, I know.
I'll hold up my end of the bargain by doing the same when I get to your position and everywhere in between.
Strange article. The problem isn’t that everyone doesn’t know how everything works, it’s that AI coding could mean there is no one who knows how a system works.
No I think the problem is AI coding removes intentionality. And that introduces artifacts and connections and dependencies that shouldn’t be there if one had designed the system with intent. And that makes it eventually harder to reason about.
There is a difference in qualia in it happens to work and it was made for a purpose.
Business logic will strive more for it happens to work as a good enough.
Excellent point. The intention of business is profit, how it arrives there is considered incidental. Any product no matter what, as long as it sells. Compounding effects in computing, the internet and miniaturisation, have enabled large profit margins that further compound these effects. They think of this as a machine that can keep on printing more money and subsuming more and more as software and computers are pervasive.
Including the AI, which generated it once and forgot.
This is going to be a big problem. How do people using Claude-like code generation systems do this? What artifacts other than the generated code are left behind for reuse when modifications are needed? Comments in the code? The entire history of the inputs and outputs to the LLM? Is there any record of the design?
I have experimented with telling Claude Code to keep a historical record of the work it is performing. It did work (though I didn't assess the accuracy of the record) but I decided it was a waste of tokens and now direct it to analyze the history in ~/.claude when necessary. The real problem I was solving was making sure it didn't leave work unfinished between autocompacts (eg crucial parts of the work weren't performed and instead there are only TODO comments). But I ended up solving that with better instructions about how to break down the plan into bite-sized units that are more friendly to the todo list tool.
I have prompting in AGENTS.md that instructs the agent to update the relevant parts of the project documentation for a given change. The project has a spec, and as features get added or reworked the spec gets updated. If you commit after each session then the git history of the spec captures how the design evolves. I do read the spec, and the errors I've seen so far are pretty minor.
Is that documentation useful? I haven't seen a well-documented codebase by AI so far.
To be fair - humans also fail at that. Just look at the GTK documentation as an example. When you point that out, ebassi may ignore you because criticism is unwanted; and the documentation will never improve, meaning they don't want new developers.
I for one I save all conversations in the codebase. Includes both human prompts and outputs. But I’m using a modified codex to do so.
Not sure why it’s not default as it’s useful to have this info.
This happens even today. If a knowledgeable person leaves a company and no KT (or more likely, poor KT) takes place, then there will be no one left to understand how certain systems work. This means the company will have to have a new developer go in and study the code and then deduce how it works. In our new LLM world, the developer could even have an LLM construct an overview for him/her to come up to speed more quickly.
Yes but every time the "why" is obscured perhaps not completely because there's no finished overview or because the original reason cannot be derived any longer from the current state of affairs. Its like the movie memento: you're trying to piece together a story from fragments that seem incoherent.
Does anyone on the planet actually know all of the subtleties and idiosyncrasies of the entire tax code? Perhaps the one inhabitant of Sealand and the Sentinelese but no-one in any western society.
I take a fairly optimistic view to the adoption of AI assistants in our line of work. We begin to work and reason at a higher level and let the agents worry about the lower level details. Know where else this happens? Any human organization that existed, exists, and will exist. Hierarchies form because no one person can do everything and hold all the details in their mind, especially as the complexity of what they intend to accomplish goes up.
One can continue to perfect and exercise their craft the old school way, and that’s totally fine, but don’t count on that to put food on the table. Some genius probably can, but I certainly am not one.
Yeah it’s all trade offs. If it means I get to where I want to be faster, even if it’s imperfect, so be it.
Humans aren’t without flaws; prior to coding assistants, I’ve lost count of the times my PM telling me to rush things at the expense of engineering rigor. We validate or falsify the need for a feature sooner and move on to other things. Sometimes it works sometimes a bug blows up in our faces, but things still chug along.
This point will become increasingly moot as AI gets better at generating good code, and faster, too.
I don't think there is anything new here and the metaphor holds up perfectly. There have always been bugs we don't understand in compilers or libraries or implementations beyond that, that make the path we chose unavailable to us at a certain level. The responsibility is to create a working solution, sure, but there is nothing that would prevent us from getting there by typing "Hey LLM, this is not working, let's try a different approach", even though it might not feel great.
To be fair, I don't know how a living human individual work, let alone how they actually work in society. I suspect I'm not alone in this case.
So nothing new under the sun, often the practices come first, then only can some theory emerge, from which point it can be leverage on to go further than present practice and so on. Sometime practice and theory are more entengled in how they are created on the go, obviously.
"It's not slop. It's not forgetting first principles. It's a shift in how the craft work, and it's already happened."
It actually really is slop. He may wish to ignore it but that does not change anything. AI comes with slop - that is undeniable. You only need to look at the content generated via AI.
He may wish to focus merely on "AI for use in software engineering", but even there he is wrong, since AI makes mistakes too and not everything it creates is great. People often have no clue how that AI reaches any decision, so they also lose being able to reason about the code or code changes. I think people have a hard time trying to sell AI as "only good things, the craft will become better". It seems everyone is on the AI hype train - eventually it'll either crash or slow down massively.
We keep delegating knowledge of the natural, physical world for temporary, rapidly-changing knowledge of abstractions and software tools, which we do not control (now LLM cloud tools).
The lack of comprehensive, practical, multi-disciplinary knowledge creates a DEEP DEPENDENCY on the few multinational companies and countries that UNDERSTAND things and can BUILD things. If you don't understand it, if you can't build it, they OWN you.
why does the author imply not knowing everything is a bad thing? If you have clear protocol and interfaces, not knowing everything enables you to make bigger innovations. If everything is a complex mess, then no.
Not knowing everything never "enables" you to do anything. Knowing how something works is always better than not knowing, assuming you want to use it or make changes to it.
The issue with frameworks is not the magic. We feel like it's magic because the interfaces are not stable. If the interfaces were stable we'd consider them just a real component of building whatever
You don't need to know anything about hardware to properly use a CPU isa.
The difference is the cpu isa is documented, well tested and stable. We can build systems that offer stability and are formally verified as an industry. We just choose not to.
Sure, we have complex systems that we don't know how everything works (car, computer, cellphone, etc.) . However, we do expect that those systems behave deterministically in their interface to us. And when they don't, we consider them broken.
For example, why is the HP-12C still the dominant business calculator? Because using other calculators for certain financial calculations were non-deterministically wrong. The HP-12C may not have even been strictly "correct", but it was deterministic in the ways in wasn't.
Financial people didn't know or care about guard digits or numerical instability. They very much did care that their financial calculations were consistent and predictable.
The claimed connections here fall apart for me pretty quickly.
CPU instructions, caches, memory access, etc. are debated, tested, hardened, and documented to a degree that's orders of magnitude greater than the LLM-generated code we're deploying these days. Those fundamental computing abstractions aren't nearly as leaky or nearly as in need of refactoring tomorrow.
That's not how things work in practice.
I think the concern is not that "people don't know how everything works" - people never needed to know how to "make their own food" by understanding all the cellular mechanisms and all the intricacies of the chemistry & physics involved in cooking. BUT, when you stop understanding the basics - when you no longer know how to fry an egg because you just get it already prepared from the shop/ from delivery - that's a whole different level of ignorance, that's much more dangerous.
Yes, it may be fine & completely non-concerning if agricultural corporations produce your wheat and your meat; but if the corporation starts producing standardized cooked food for everyone, is it really the same - is it a good evolution, or not? That's the debate here.
Most people have no idea how to hunt, make a fire, or grow food. If all grocery stores and restaurants run out of food for a long enough time people will starve. This isn't a problem in practice though, because there are so many grocery stores and restaurants and supply chains source from multiple areas that the redundant and decentralized nature makes it not a problem. Thus it is the same with making your own food. Eventually if you have enough robots or food replicators around knowing how to make food becomes irrelevant, because you always will be able to find one even if yours is broken. (Note: we are not there yet)
>If all grocery stores and restaurants run out of food for a long enough time people will starve. This isn't a problem in practice though...
I fail to see how this isn't a problem? Grid failures happen? So do wars and natural disasters which can cause grids and supply chains to fail.
That is short hand. The problem exists of course, but it is improbable that it will actually occur in our lifetimes. An asteroid could slam into the earth or a gamma ray burst could sanitize the planet of all life. We could also experience nuclear war. These are problems that exist, yet we all just blissfully go on about our lives, b/c there is basically nothing that can be done to stop these things if they do happen and they likely won't. Basically we should only worry about these problems in so much as we as a species are able to actually do something about them.
In Star Trek they just 3D printed everything via light.
> that's a whole different level of ignorance, that's much more dangerous.
Why? Is it more dangerous to not know how to fry an egg in a teflon pan, or on a stone over a wood fire? Is it acceptable to know the former but not the latter? Do I need to understand materials science so I can understand how to make something nonstick so I’m not dependant on teflon vendors?
You do need to be able to understand nonstick coating is unhealthy and not magic. You do need to understand your options for pan frying for not sticking are a film of water or an ice cube if you don't want to add an oil into the mix. Then it really depends what you are cooking on how sticky it will be and what the end product will look like. That's why there are people that can't fry an egg, people that cook, chefs, and Michelin chefs. Because nuance matters, it's just that the domain where each person wants to apply it is different. I dont care about nuance in hockey picks but probably some people do. But some domains should concern everyone.
At what point is the threshold between fine and concerning? Seems like the one you put is from your point of view. I’m sure not everyone would agree and is subjective.
> AI will make this situation worse.
Being an AI skeptic more than not, I don't think the article's conclusion is true.
What LLM's can potentially do for us is exactly the opposite: because they are trained on pretty much everything there is, if you ask the AI how the telephone works, or what happens when you enter a URL in the browser, they can actually answer and break it down for you nicely (and that would be a dissertation-sized text). Accuracy and hallucinations aside, it's already better than a human who has no clue about how the telephone works or where to even begin if the said human wanted to understand it.
Human brains have a pretty serious gap in the "I don't know what I don't know" area, whereas language models have such a vast scope of knowledge that makes them somewhat superior, albeit at a price of, well, being literally quite expensive and power hungry. But that's technical details.
LLMs are knowledge machines that are good at precisely that: knowing everything about everything on all levels as long as it is described in human language somewhere on the Internet.
LLMs consolidate our knowledge in ways that were impossible before. They are pretty bad at reasoning or e.g. generating code, but where they excel so far is answering arbitrary questions about pretty much anything.
Perhaps a dose of pragmatism is needed here?
I am no CS major, nor do I fully understand the inner workings of a computer beyond "we tricked a rock into thinking by shocking it."
I'd love to better understand it, and I hope that through my journey of working with computers, i'll better learn about these underlying concepts registers, bus's, memory, assembly etc
Practically however, I write scripts that solve real world problems, be that from automating the coffee machine, to managing infrastructure at scale.
I'm not waiting to pick up a book on x86 assembly first before I write some python however. (I wish it were that easy.)
To the greybeards that do have a grasp of these concepts though? It's your responsibility to share that wealth of knowledge. It's a bitter ask, I know.
I'll hold up my end of the bargain by doing the same when I get to your position and everywhere in between.
Strange article. The problem isn’t that everyone doesn’t know how everything works, it’s that AI coding could mean there is no one who knows how a system works.
No I think the problem is AI coding removes intentionality. And that introduces artifacts and connections and dependencies that shouldn’t be there if one had designed the system with intent. And that makes it eventually harder to reason about.
There is a difference in qualia in it happens to work and it was made for a purpose.
Business logic will strive more for it happens to work as a good enough.
Excellent point. The intention of business is profit, how it arrives there is considered incidental. Any product no matter what, as long as it sells. Compounding effects in computing, the internet and miniaturisation, have enabled large profit margins that further compound these effects. They think of this as a machine that can keep on printing more money and subsuming more and more as software and computers are pervasive.
Including the AI, which generated it once and forgot.
This is going to be a big problem. How do people using Claude-like code generation systems do this? What artifacts other than the generated code are left behind for reuse when modifications are needed? Comments in the code? The entire history of the inputs and outputs to the LLM? Is there any record of the design?
I have experimented with telling Claude Code to keep a historical record of the work it is performing. It did work (though I didn't assess the accuracy of the record) but I decided it was a waste of tokens and now direct it to analyze the history in ~/.claude when necessary. The real problem I was solving was making sure it didn't leave work unfinished between autocompacts (eg crucial parts of the work weren't performed and instead there are only TODO comments). But I ended up solving that with better instructions about how to break down the plan into bite-sized units that are more friendly to the todo list tool.
I have prompting in AGENTS.md that instructs the agent to update the relevant parts of the project documentation for a given change. The project has a spec, and as features get added or reworked the spec gets updated. If you commit after each session then the git history of the spec captures how the design evolves. I do read the spec, and the errors I've seen so far are pretty minor.
Is this an actual problem? Takes minutes for an AI to explore and document a codebase. Sounds like a non problem.
Is that documentation useful? I haven't seen a well-documented codebase by AI so far.
To be fair - humans also fail at that. Just look at the GTK documentation as an example. When you point that out, ebassi may ignore you because criticism is unwanted; and the documentation will never improve, meaning they don't want new developers.
Yes, exactly my point as well. It cuts both ways.
I for one I save all conversations in the codebase. Includes both human prompts and outputs. But I’m using a modified codex to do so. Not sure why it’s not default as it’s useful to have this info.
This happens even today. If a knowledgeable person leaves a company and no KT (or more likely, poor KT) takes place, then there will be no one left to understand how certain systems work. This means the company will have to have a new developer go in and study the code and then deduce how it works. In our new LLM world, the developer could even have an LLM construct an overview for him/her to come up to speed more quickly.
Yes but every time the "why" is obscured perhaps not completely because there's no finished overview or because the original reason cannot be derived any longer from the current state of affairs. Its like the movie memento: you're trying to piece together a story from fragments that seem incoherent.
Not just tech.
Does anyone on the planet actually know all of the subtleties and idiosyncrasies of the entire tax code? Perhaps the one inhabitant of Sealand and the Sentinelese but no-one in any western society.
I take a fairly optimistic view to the adoption of AI assistants in our line of work. We begin to work and reason at a higher level and let the agents worry about the lower level details. Know where else this happens? Any human organization that existed, exists, and will exist. Hierarchies form because no one person can do everything and hold all the details in their mind, especially as the complexity of what they intend to accomplish goes up.
One can continue to perfect and exercise their craft the old school way, and that’s totally fine, but don’t count on that to put food on the table. Some genius probably can, but I certainly am not one.
But what if the AI agent has a 5% chance of adding a bug to that feature? Surely before any feature was completely bug free
Yeah it’s all trade offs. If it means I get to where I want to be faster, even if it’s imperfect, so be it.
Humans aren’t without flaws; prior to coding assistants, I’ve lost count of the times my PM telling me to rush things at the expense of engineering rigor. We validate or falsify the need for a feature sooner and move on to other things. Sometimes it works sometimes a bug blows up in our faces, but things still chug along.
This point will become increasingly moot as AI gets better at generating good code, and faster, too.
Let me make it worse. Much worse. :)
https://youtu.be/36myc8wQhLo (USENIX ATC '21/OSDI '21 Joint Keynote Address-It's Time for Operating Systems to Rediscover Hardware)
But people are expected to understand the part of the system they are responsible for at the level of abstraction they are being paid to operate.
This new arrangement would be perfectly fine if they aren't responsible when/if it breaks.
I don't think there is anything new here and the metaphor holds up perfectly. There have always been bugs we don't understand in compilers or libraries or implementations beyond that, that make the path we chose unavailable to us at a certain level. The responsibility is to create a working solution, sure, but there is nothing that would prevent us from getting there by typing "Hey LLM, this is not working, let's try a different approach", even though it might not feel great.
> Nobody knows how the whole system works
True.
But in all systems up to now, for each part of the system, somebody knew how it worked.
That paradigm is slowly eroding. Maybe that's ok, maybe not, hard to say.
> But in all systems up to now, for each part of the system, somebody knew how it worked.
If the project is legacy or the people just left the company that’s just not true.
> If the project is legacy or the people just left the company that’s just not true.
Yeah, that's why I said "knew" instead of "knows".
Wikipedia knows how it all works, and that's good enough in case we need to reboot civilization.
To be fair, I don't know how a living human individual work, let alone how they actually work in society. I suspect I'm not alone in this case.
So nothing new under the sun, often the practices come first, then only can some theory emerge, from which point it can be leverage on to go further than present practice and so on. Sometime practice and theory are more entengled in how they are created on the go, obviously.
Adam Jacob's quote is this:
"It's not slop. It's not forgetting first principles. It's a shift in how the craft work, and it's already happened."
It actually really is slop. He may wish to ignore it but that does not change anything. AI comes with slop - that is undeniable. You only need to look at the content generated via AI.
He may wish to focus merely on "AI for use in software engineering", but even there he is wrong, since AI makes mistakes too and not everything it creates is great. People often have no clue how that AI reaches any decision, so they also lose being able to reason about the code or code changes. I think people have a hard time trying to sell AI as "only good things, the craft will become better". It seems everyone is on the AI hype train - eventually it'll either crash or slow down massively.
We keep delegating knowledge of the natural, physical world for temporary, rapidly-changing knowledge of abstractions and software tools, which we do not control (now LLM cloud tools).
The lack of comprehensive, practical, multi-disciplinary knowledge creates a DEEP DEPENDENCY on the few multinational companies and countries that UNDERSTAND things and can BUILD things. If you don't understand it, if you can't build it, they OWN you.
why does the author imply not knowing everything is a bad thing? If you have clear protocol and interfaces, not knowing everything enables you to make bigger innovations. If everything is a complex mess, then no.
Not knowing everything never "enables" you to do anything. Knowing how something works is always better than not knowing, assuming you want to use it or make changes to it.
It is the same with the global financial system
I don't like this thing where we dislike 'magic'
The issue with frameworks is not the magic. We feel like it's magic because the interfaces are not stable. If the interfaces were stable we'd consider them just a real component of building whatever
You don't need to know anything about hardware to properly use a CPU isa.
The difference is the cpu isa is documented, well tested and stable. We can build systems that offer stability and are formally verified as an industry. We just choose not to.
Sure, we have complex systems that we don't know how everything works (car, computer, cellphone, etc.) . However, we do expect that those systems behave deterministically in their interface to us. And when they don't, we consider them broken.
For example, why is the HP-12C still the dominant business calculator? Because using other calculators for certain financial calculations were non-deterministically wrong. The HP-12C may not have even been strictly "correct", but it was deterministic in the ways in wasn't.
Financial people didn't know or care about guard digits or numerical instability. They very much did care that their financial calculations were consistent and predictable.
The question is: Who will build the HP-12C of AI?