There could be some interesting actual OS design innovations for new permissions, access control, and sandboxing systems that are designed from the ground up for these "conjurable" experiences (or even OS components themself).
Finding some balance between access and safety for vibe-conjuring parts of the OS in a way that's still powerful but won't accidentally nuke or corrupt itself.
I do think there's some interesting space to be explored in capability-based systems, but even then I think it would mostly end up looking similar to how existing systems are, just with one more element added
Oh, ok. I guess computer science is finished, then. I should have known. It's not like there's any evidence that it's been blown wide open recently.
(The irony is that all the reading that's gone into our design is like, old CSAIL papers. Maybe it's true! Maybe everything we're doing is just crate-digging through old settled computer science. That would be on brand. I should find a way to work Athena into it. Maybe we'll use Zephyr for messaging.)
I think operating systems have changed. They don't interact directly with hardware much anymore and computers are filled with SoCs all running systems of their own, sometimes with entire network stacks, communicating with devices and each other, collecting and processing data and that activity can be totally invisible to the operating system the user gets to work with.
That makes a lot of sense in a world where we’re importing all our software from strangers. It makes less sense in the world we’re heading to, where most of the software we’re carving up fiefdoms for has the same provenance.
If I'm running software written by an LLM, even if I was the one who prompted the LLM, I would still want my OS to treat it as if written by a stranger.
I know he's not necessarily talking about relaxing security models here, but he's not being very specific about his vision otherwise.
Nope. There has never been a "good" post written in this genre ("I'm leaving my last team, here's my new project") and I'm not going to waste interesting bits on my pro-forma transition post.
I didn't submit this! Happy to chat about it, but I feel like I was pretty clear in the post that I wasn't bidding for the front page.
I think OP may be onto something real, while you are definitely right if we translate concepts directly. But maybe the challenge, with current platforms included, is because we try to cram identities by reference. If a phone had its own identity, with which you transact as with any party then lots of the security concerns would dissolve. To tone down crankiness, I refer to identity in the technical sense, not personality etc. This doesn’t solve the challenge but it might point to a different foundation for a security model.
I love the idea that normal people building software to solve real problems in their life is, necessarily, a toy; the only real software is blessed by our sacred priesthood of PLT experts.
Consider: what does computing look like in a world where many (maybe most) applications have an audience of just 1-2 people? How is software distributed? What does it run on?
The same. The slop apps remain just that. They may do something useful for you, but that's it. A bit like a branch in a repository.
It's the same thing as looking down your nose at people who use Excel as a database because it's not a "real" database.
Yes, I'm sure we all have horror stories about spreadsheets that got out of hand, but that ignores the thousands (or more) spreadsheets that are silently productive everyday.
I agree with the gist of how weird everything is going to get, and how we are living in a time where people haven't come close to internalizing the shift.
IMO even this AI phone is doomed in the slightly longer run.
One more step into the weirdness and people will just be working with AI to build their own bespoke phone with the specs sent over to a manufacturer that is "like PCBWay, but spits out whole devices".
I have no idea what this does to society, the economy, etc when we reach a point where nobody really "needs" any kind of middleman for almost anything (well, except for things they actually really need, like food and shelter), but it looks like I'll see it happen for a bit before I die.
Taking it at face value, it seems to me this will also collapse and get folded into the model.
Unclear what will be left next to it. It will just live on your phone and take care of everything. The apps will be markdown files describing what they do, probably compiled ahead of time for latency.
But the model will be in control, so what else is left to implement besides giving it bash, a compiler and a display?
I guess there is value in the year before that happens.
In my opinion, this post misses the forest for the trees. Yes, the model of an OS running various off-the-shelf apps for various tasks will soon become obsolete. But it's not the idea of the OS that's outdated - it's the idea of apps. Why would you ever want to ask an AI to make an app for you to complete some task, if you can instead ask the AI to complete the task directly? The future isn't a bunch of extremely personalized apps - the future is a single AI assistant doing all the work.
An AI assistant can give different answers for the same inputs. An application written by an AI assistant, if it doesn't just call an LLM at runtime, will mostly likely produce the same answer for the same inputs. That is, unless there's something intrinsic to its business logic that makes it nondeterministic.
Sometimes you really want the latter even if the former is less effort.
I want an agent studio and execution sandbox where my LLM driven assistant builds and maintains my apps with a receipt or SBOM like history of what it did to create each app. Determinism built per use case with non deterministic assistance orchestrating the build and management layer. Extra credit if I can trivially share and collaborate with others per “deterministic silo” via a shared link.
What it has always been. Nothing has changed in this regard.
This is by and for people who do not distinguish between the UI and the OS, like how "MSIE" is "The Internet" for them.
There could be some interesting actual OS design innovations for new permissions, access control, and sandboxing systems that are designed from the ground up for these "conjurable" experiences (or even OS components themself).
Finding some balance between access and safety for vibe-conjuring parts of the OS in a way that's still powerful but won't accidentally nuke or corrupt itself.
I do think there's some interesting space to be explored in capability-based systems, but even then I think it would mostly end up looking similar to how existing systems are, just with one more element added
I've heard very competent people claiming similar things. I think there's an abstraction I'm missing. I know they don't mean it literally
I mean it literally.
Oh, ok. I guess computer science is finished, then. I should have known. It's not like there's any evidence that it's been blown wide open recently.
(The irony is that all the reading that's gone into our design is like, old CSAIL papers. Maybe it's true! Maybe everything we're doing is just crate-digging through old settled computer science. That would be on brand. I should find a way to work Athena into it. Maybe we'll use Zephyr for messaging.)
I think operating systems have changed. They don't interact directly with hardware much anymore and computers are filled with SoCs all running systems of their own, sometimes with entire network stacks, communicating with devices and each other, collecting and processing data and that activity can be totally invisible to the operating system the user gets to work with.
If I'm running software written by an LLM, even if I was the one who prompted the LLM, I would still want my OS to treat it as if written by a stranger.
I know he's not necessarily talking about relaxing security models here, but he's not being very specific about his vision otherwise.
Nope. There has never been a "good" post written in this genre ("I'm leaving my last team, here's my new project") and I'm not going to waste interesting bits on my pro-forma transition post.
I didn't submit this! Happy to chat about it, but I feel like I was pretty clear in the post that I wasn't bidding for the front page.
I think you got your comment replies mixed up? I don't see the relevance of this (https://news.ycombinator.com/item?id=49851040) to that (https://news.ycombinator.com/item?id=49850998).
Sorry, I'm responding to the (apt) complaint that there aren't a lot of details in the post. They're right, and I'm just adding why.
I think OP may be onto something real, while you are definitely right if we translate concepts directly. But maybe the challenge, with current platforms included, is because we try to cram identities by reference. If a phone had its own identity, with which you transact as with any party then lots of the security concerns would dissolve. To tone down crankiness, I refer to identity in the technical sense, not personality etc. This doesn’t solve the challenge but it might point to a different foundation for a security model.
If the phone injures or defrauds you, does it have to make restitution? If so, how?
'Transacting' requires not only identity but accountability, a completely controlled-by-you phone cannot meaningfully transact separately from you.
Kernel + Shell + Slop
Don't forget systemd :-\
Adult Roblox
I love the idea that normal people building software to solve real problems in their life is, necessarily, a toy; the only real software is blessed by our sacred priesthood of PLT experts.
Didn’t mean any judgement just the metaphor that came to mind
Roblox is a game where you make the game. This sounds like an OS where you make the OS.
Probably could have spelled it out more, being too pithy perhaps
No, you're fine, and I might be a little hair trigger; sorry about that.
not just startup nerd, he also make gas town and drunken rant
steve yegge is not tptacek (thanks for cryptopals, tptacek!)
You've misunderstood; the quoted line from the post is most likely subtweeting Steve Yegge (in particular https://steve-yegge.medium.com/why-i-left-google-to-join-gra...).
Why does this article have this title? I could see nothing in it at all that relates to the question posed by the title.
The same. The slop apps remain just that. They may do something useful for you, but that's it. A bit like a branch in a repository.
Have you thought carefully about this? What does it mean to call a piece of software that solves a real problem you have in your life "slop"?
It's the same thing as looking down your nose at people who use Excel as a database because it's not a "real" database.
Yes, I'm sure we all have horror stories about spreadsheets that got out of hand, but that ignores the thousands (or more) spreadsheets that are silently productive everyday.
I agree with the gist of how weird everything is going to get, and how we are living in a time where people haven't come close to internalizing the shift.
IMO even this AI phone is doomed in the slightly longer run.
One more step into the weirdness and people will just be working with AI to build their own bespoke phone with the specs sent over to a manufacturer that is "like PCBWay, but spits out whole devices".
I have no idea what this does to society, the economy, etc when we reach a point where nobody really "needs" any kind of middleman for almost anything (well, except for things they actually really need, like food and shelter), but it looks like I'll see it happen for a bit before I die.
Taking it at face value, it seems to me this will also collapse and get folded into the model.
Unclear what will be left next to it. It will just live on your phone and take care of everything. The apps will be markdown files describing what they do, probably compiled ahead of time for latency.
But the model will be in control, so what else is left to implement besides giving it bash, a compiler and a display?
I guess there is value in the year before that happens.
In my opinion, this post misses the forest for the trees. Yes, the model of an OS running various off-the-shelf apps for various tasks will soon become obsolete. But it's not the idea of the OS that's outdated - it's the idea of apps. Why would you ever want to ask an AI to make an app for you to complete some task, if you can instead ask the AI to complete the task directly? The future isn't a bunch of extremely personalized apps - the future is a single AI assistant doing all the work.
An AI assistant can give different answers for the same inputs. An application written by an AI assistant, if it doesn't just call an LLM at runtime, will mostly likely produce the same answer for the same inputs. That is, unless there's something intrinsic to its business logic that makes it nondeterministic.
Sometimes you really want the latter even if the former is less effort.
I want an agent studio and execution sandbox where my LLM driven assistant builds and maintains my apps with a receipt or SBOM like history of what it did to create each app. Determinism built per use case with non deterministic assistance orchestrating the build and management layer. Extra credit if I can trivially share and collaborate with others per “deterministic silo” via a shared link.
[delayed]
A bunch of dot files, obviously.