Replit intrigues me. Other than Paul Graham talking it up a lot on Twitter - I've never once heard of anyone using it, seen any chatter about it in any or our internal slack forums. Every other AI company I get barraged with- no matter where I go, even antigravity/gemini carbon/argon/models/etc... - which nobody I know uses either - but at least I hear about them a lot.
I wonder what ecosystem replit exists in that seems to completely not overlap with any of the other AI technologies and companies....
Replit used to be a great platform for sharing code in a way which made it easy for anyone to run or modify. It was great for students, and kids learning to code. But then, in 2024, they removed all social features, and reduced the maximum number of repls for free users from practically unlimited to 3. I stopped using it after this, and shortly after, they implemented a strict ratelimit for downloading your own repls.
I'm now taking a look at Replit for the first time in a while, and it looks like they completely redid their interface: Their whole web IDE appears to be gone or vastly simplified, and it seems like half of of everything I click on fails with error 429, after just a tiny bit of looking around. Their old IDE that functioned as a code editor, and gave you basic tools such as a file tree viewer, is no more. Replit turned an awesome product into a glorified LLM chat which is vastly dysfunctional for any sort of hand coding.
I did some contract work on a project that was produced with replit, extracting it to github and setting up its production and staging environments, auditing for issues, resolving some of them.
There was so much that wasn't done, it was basically an Airbnb-like service and it lacked any capacity to register users or hosts, pay for or disburse payments, create or manage listings, notifications, administration. They went from expecting to launch in a couple weeks to looking at my roadmap and quote of about 3 months more development work just for a launchable MVP, which wildly exceeded their budget or commitment to the project.
I don't know how much of this was Replit's fault but some of it certainly was, like despite being a monorepo the frontend had no connectivity with the backend, there was no barriers between user data and most endpoints were not secured, the server had zero hardening, some endpoints would take the user's id as a posted value.
I have been helping a non-profit migrate off Replit. What an awful experience using the platform has been. Their LLMs produce very bad code compared to Claude Code or Codex with the exact same prompt. Ultimately you end up having to go back and forth to get to a better solution while they nickel and dime you on tokens. It feels questionable, like it is intentional to squeeze money out of you.
Having used lovable.dev as well, I must say Lovable is a much better product.
Replit is easily one of the worst products and platforms I've use. I wouldn't even recommend it to my enemies.
I've been picking up bits of fractional CTO work over the last year or so and getting more involved with other founders... and I'm seeing plenty of people using it (and the other systems like it). They're just not generally people from a software engineering background, as you'd probably expect. So it's not a massive surprise that you're not crossing paths with them.
Generally it's entrepreneurial types who would have gone to an agency to get something built, or would have wanted to go to an agency but couldn't afford the costs involved. These platforms are far cheaper and quicker, and if we're really honest not necessarily worse quality than what they would have gotten from the vast majority of those agencies, especially the cheaper ones.
This is not a comment on whether this is a good thing or not, and I have all sorts of concerns, but it's undeniably something I'm seeing, and the nearshore/offshore agencies are certainly starting to feel the impact.
My read on this is they re-invented the OTP supervision tree, process isolation, hibernation, and actor-to-actor message passing, but had to push the isolation boundary all the way down into Ring -1 (KVM/Cloud Hypervisor) and customized Linux tinyconfig kernels because their "actors" are actively trying to jailbreak the hypervisor.
No, that's pretty overwritten. The simple story is: like most neocloud providers, they were using Docker and/or Linux containers as a workload isolation primitive, those have never been safe, and they've now worked out how to scale actual KVM isolation to their users.
A thing I wish more people would do: ask prospective cloud hosting companies "is there a mode of this offering where my workloads might share a kernel with other tenants I don't trust, as would be the case with Docker?", and if the answer to that question is ever "yes", disqualify the provider.
How do people feel about the hardware isolation provided by microVMs versus strong software isolation techniques like gvisor?
Right now, we put a huge amount of stock in "hardware isolation". If that starts falling to agents to any substantial degree, I wonder whether we might end up seeing a transition toward patchable software isolation, similar to the rise of software RAID that occurred a while back.
The general vibe I hear from security practitioners is that they (gvisor and Firecracker/Cloud-Hypervisor) are comparable. You generally do Firecracker if you can get away with it, because it's simpler and faster, but to do it well you generally need bare-metal hosts.
I think "hardware" vs. "software" is the wrong way to frame this. The big difference is between shared-kernel multitenant and guest kernels. gVisor (in some configurations) is like an 80% solution; a partially emulated guest kernel. I think the verdict is in on shared-kernel vs. not: don't share kernels between tenants.
The hardware isolation itself --- AMD-V, VT-X, &c --- has held up for almost 2 decades now. You never know, of course, but if you were going to place a bet it probably wouldn't be on hardware isolation failing before software did.
gVisor and Firecracker are nothing alike. gVisor is an application kernel written in Go. Application kernels will make your system much slower and can only solve a subset of security issues that would affect the kernel. They're like a kernel proxy of sorts.
Firecracker is a KVM virtual machine but made much more efficient. It's the most desirable of the two. Firecracker and KVM don't require bare metal, but in a virtualized environment they will require nested virtualization support. Nested virtualization is supported by just about every hyperscaler.
The reason KVM is preferable is because it provides strong isolation of the host kernel interface and independent emulation of host devices.
I've built systems out of both and they are quite similar, using different mechanisms as vectors towards the same outcome, which is to minimize shared kernel attack surface in untrusted workloads.
Firecracker doesn't require bare metal, but (1) you want it for performance and efficiency at scale and (2) you take a security hit for nesting.
Replit intrigues me. Other than Paul Graham talking it up a lot on Twitter - I've never once heard of anyone using it, seen any chatter about it in any or our internal slack forums. Every other AI company I get barraged with- no matter where I go, even antigravity/gemini carbon/argon/models/etc... - which nobody I know uses either - but at least I hear about them a lot.
I wonder what ecosystem replit exists in that seems to completely not overlap with any of the other AI technologies and companies....
Replit used to be a great platform for sharing code in a way which made it easy for anyone to run or modify. It was great for students, and kids learning to code. But then, in 2024, they removed all social features, and reduced the maximum number of repls for free users from practically unlimited to 3. I stopped using it after this, and shortly after, they implemented a strict ratelimit for downloading your own repls. I'm now taking a look at Replit for the first time in a while, and it looks like they completely redid their interface: Their whole web IDE appears to be gone or vastly simplified, and it seems like half of of everything I click on fails with error 429, after just a tiny bit of looking around. Their old IDE that functioned as a code editor, and gave you basic tools such as a file tree viewer, is no more. Replit turned an awesome product into a glorified LLM chat which is vastly dysfunctional for any sort of hand coding.
I did some contract work on a project that was produced with replit, extracting it to github and setting up its production and staging environments, auditing for issues, resolving some of them.
There was so much that wasn't done, it was basically an Airbnb-like service and it lacked any capacity to register users or hosts, pay for or disburse payments, create or manage listings, notifications, administration. They went from expecting to launch in a couple weeks to looking at my roadmap and quote of about 3 months more development work just for a launchable MVP, which wildly exceeded their budget or commitment to the project.
I don't know how much of this was Replit's fault but some of it certainly was, like despite being a monorepo the frontend had no connectivity with the backend, there was no barriers between user data and most endpoints were not secured, the server had zero hardening, some endpoints would take the user's id as a posted value.
I have been helping a non-profit migrate off Replit. What an awful experience using the platform has been. Their LLMs produce very bad code compared to Claude Code or Codex with the exact same prompt. Ultimately you end up having to go back and forth to get to a better solution while they nickel and dime you on tokens. It feels questionable, like it is intentional to squeeze money out of you.
Having used lovable.dev as well, I must say Lovable is a much better product.
Replit is easily one of the worst products and platforms I've use. I wouldn't even recommend it to my enemies.
But why did the non-profit go with Replit in the first place? Some guesses:
1) they were inexperienced or just hesitant to use even managed infrastructure
2) they started with a small Replit footprint and kind of got stuck there out of some niche convenience
3) it started free* and was thought of as a cost cutting measure to use this temporarily VC subsidized service (we've all been there)
I've been picking up bits of fractional CTO work over the last year or so and getting more involved with other founders... and I'm seeing plenty of people using it (and the other systems like it). They're just not generally people from a software engineering background, as you'd probably expect. So it's not a massive surprise that you're not crossing paths with them.
Generally it's entrepreneurial types who would have gone to an agency to get something built, or would have wanted to go to an agency but couldn't afford the costs involved. These platforms are far cheaper and quicker, and if we're really honest not necessarily worse quality than what they would have gotten from the vast majority of those agencies, especially the cheaper ones.
This is not a comment on whether this is a good thing or not, and I have all sorts of concerns, but it's undeniably something I'm seeing, and the nearshore/offshore agencies are certainly starting to feel the impact.
there's this brief window where replit makes sense, but it could disappear in a month if things like ChatGPT sites get much better.
I used Replit quite a lot back in the day to share code snippets with clients.
My read on this is they re-invented the OTP supervision tree, process isolation, hibernation, and actor-to-actor message passing, but had to push the isolation boundary all the way down into Ring -1 (KVM/Cloud Hypervisor) and customized Linux tinyconfig kernels because their "actors" are actively trying to jailbreak the hypervisor.
Is that about right?
No, that's pretty overwritten. The simple story is: like most neocloud providers, they were using Docker and/or Linux containers as a workload isolation primitive, those have never been safe, and they've now worked out how to scale actual KVM isolation to their users.
A thing I wish more people would do: ask prospective cloud hosting companies "is there a mode of this offering where my workloads might share a kernel with other tenants I don't trust, as would be the case with Docker?", and if the answer to that question is ever "yes", disqualify the provider.
It's only somewhat related to the article, but:
How do people feel about the hardware isolation provided by microVMs versus strong software isolation techniques like gvisor?
Right now, we put a huge amount of stock in "hardware isolation". If that starts falling to agents to any substantial degree, I wonder whether we might end up seeing a transition toward patchable software isolation, similar to the rise of software RAID that occurred a while back.
The general vibe I hear from security practitioners is that they (gvisor and Firecracker/Cloud-Hypervisor) are comparable. You generally do Firecracker if you can get away with it, because it's simpler and faster, but to do it well you generally need bare-metal hosts.
I think "hardware" vs. "software" is the wrong way to frame this. The big difference is between shared-kernel multitenant and guest kernels. gVisor (in some configurations) is like an 80% solution; a partially emulated guest kernel. I think the verdict is in on shared-kernel vs. not: don't share kernels between tenants.
The hardware isolation itself --- AMD-V, VT-X, &c --- has held up for almost 2 decades now. You never know, of course, but if you were going to place a bet it probably wouldn't be on hardware isolation failing before software did.
I think this is stated confusingly.
gVisor and Firecracker are nothing alike. gVisor is an application kernel written in Go. Application kernels will make your system much slower and can only solve a subset of security issues that would affect the kernel. They're like a kernel proxy of sorts.
Firecracker is a KVM virtual machine but made much more efficient. It's the most desirable of the two. Firecracker and KVM don't require bare metal, but in a virtualized environment they will require nested virtualization support. Nested virtualization is supported by just about every hyperscaler.
The reason KVM is preferable is because it provides strong isolation of the host kernel interface and independent emulation of host devices.
I've built systems out of both and they are quite similar, using different mechanisms as vectors towards the same outcome, which is to minimize shared kernel attack surface in untrusted workloads.
Firecracker doesn't require bare metal, but (1) you want it for performance and efficiency at scale and (2) you take a security hit for nesting.
MicroVMs are quite popular nowadays due to sandboxing but we need to also keep in mind least privilege principle. Applying that is also tricky.
This Replit? https://intuitiveexplanations.com/tech/replit/
I do hope nobody reading this copies any of Replit's design decisions or architecture, lest they get threatened.
I avoided reading the piece, just in case.
Guys. You could do the same with ansible like decade ago.