InsanityBit
@insanitybit.bsky.social
I think the dependency "cooldown" approach is fundamentally flawed and a total distraction from the work that would actually solve supply chain issues - sandboxing and attestation.
insanitybit.github.io/2025/11/22/o...
insanitybit.github.io/2025/11/22/o...
On Dependency Cooldowns -
InsanityBit
insanitybit.github.io
November 24, 2025 at 4:01 PM
I think the dependency "cooldown" approach is fundamentally flawed and a total distraction from the work that would actually solve supply chain issues - sandboxing and attestation.
insanitybit.github.io/2025/11/22/o...
insanitybit.github.io/2025/11/22/o...
Doing a thing with wasm again. I always end up disappointed, let's see how it goes this time!
November 10, 2025 at 1:35 PM
Doing a thing with wasm again. I always end up disappointed, let's see how it goes this time!
news.ycombinator.com/item?id=4556...
Interesting how broken your epistemology has to be to say something like this. "We call it X therefor it is X".
Interesting how broken your epistemology has to be to say something like this. "We call it X therefor it is X".
> we are on the cusp of intelligent machines. Nah, we aren't. There's a reason t... | Hacker News
news.ycombinator.com
October 12, 2025 at 9:08 PM
news.ycombinator.com/item?id=4556...
Interesting how broken your epistemology has to be to say something like this. "We call it X therefor it is X".
Interesting how broken your epistemology has to be to say something like this. "We call it X therefor it is X".
Trying to modularize crates so that the entire API surface and implementation of each can fit into a small context window.
October 12, 2025 at 2:35 PM
Trying to modularize crates so that the entire API surface and implementation of each can fit into a small context window.
Reposted by InsanityBit
I totally get why people think this (due to how rapidly the SOTA has changed), but it's not really accurate. Transformer models operate in high-dimensional embedding spaces where semantic relationships differ from surface form. It's not just about the text. It's the meanings, the relationships.
As an example: AI doesn't understand "no." Because the statements "no ketchup on my burger" and "ketchup on my burger" are almost identical to a machine that does not and cannot actually reason. It's only a 2 letter difference.
AI doesn't know 'no' – and that's a huge problem for medical bots
Many AI models fail to recognise negation words such as “no” and “not”, which means they can’t easily distinguish between medical images labelled as showing a disease and images labelled as not showin...
www.newscientist.com
October 11, 2025 at 6:58 PM
I totally get why people think this (due to how rapidly the SOTA has changed), but it's not really accurate. Transformer models operate in high-dimensional embedding spaces where semantic relationships differ from surface form. It's not just about the text. It's the meanings, the relationships.
Lobste.rs AI discussions are so cringe. I mostly don't visit the site at this point because it's basically just 1 or 2 commenters providing value, but when I do and I see AI convos they're just hard to read.
lobste.rs/s/xbxhvq/vib...
Simon does good work, idk why he tries to talk to these people.
lobste.rs/s/xbxhvq/vib...
Simon does good work, idk why he tries to talk to these people.
Lobsters
Lobste.rs
October 9, 2025 at 5:07 PM
Lobste.rs AI discussions are so cringe. I mostly don't visit the site at this point because it's basically just 1 or 2 commenters providing value, but when I do and I see AI convos they're just hard to read.
lobste.rs/s/xbxhvq/vib...
Simon does good work, idk why he tries to talk to these people.
lobste.rs/s/xbxhvq/vib...
Simon does good work, idk why he tries to talk to these people.
An interesting thing about "a malicious package grabbed keys from the system" is that this problem has 0 technical or UX considerations and has been solved by browsers for over 20 years.
Sandboxing is the solution. That is it. UX for sandboxing is solved, look at manifests and how changes surface.
Sandboxing is the solution. That is it. UX for sandboxing is solved, look at manifests and how changes surface.
September 25, 2025 at 3:06 PM
An interesting thing about "a malicious package grabbed keys from the system" is that this problem has 0 technical or UX considerations and has been solved by browsers for over 20 years.
Sandboxing is the solution. That is it. UX for sandboxing is solved, look at manifests and how changes surface.
Sandboxing is the solution. That is it. UX for sandboxing is solved, look at manifests and how changes surface.
Two things I feel comfortable "vibe coding" (ie: with loose guidance, allowing a model to drive).
1. After my own tests are written, asking for additional "Given/ When/ Then" tests to be written based on expected properties of the code.
2. Reduction of code to reduce LOC. "That could be a fn"
1. After my own tests are written, asking for additional "Given/ When/ Then" tests to be written based on expected properties of the code.
2. Reduction of code to reduce LOC. "That could be a fn"
September 1, 2025 at 8:20 PM
Two things I feel comfortable "vibe coding" (ie: with loose guidance, allowing a model to drive).
1. After my own tests are written, asking for additional "Given/ When/ Then" tests to be written based on expected properties of the code.
2. Reduction of code to reduce LOC. "That could be a fn"
1. After my own tests are written, asking for additional "Given/ When/ Then" tests to be written based on expected properties of the code.
2. Reduction of code to reduce LOC. "That could be a fn"
Feels nuts that Github Actions doesn't provide basic features like restricting outbound networking. This feels like it would be so dead simple for them to implement?
August 25, 2025 at 5:55 PM
Feels nuts that Github Actions doesn't provide basic features like restricting outbound networking. This feels like it would be so dead simple for them to implement?
Claude is far too nice and not nearly as skeptical as I want it to be. I suppose second guessing and checking assumptions is very expensive but it's frustrating, and I've found that I need to have a second model in a sort of "skeptical mode" to compensate.
August 3, 2025 at 2:27 PM
Claude is far too nice and not nearly as skeptical as I want it to be. I suppose second guessing and checking assumptions is very expensive but it's frustrating, and I've found that I need to have a second model in a sort of "skeptical mode" to compensate.
OpenAI's Codex needs a lot more work.
July 20, 2025 at 2:34 AM
OpenAI's Codex needs a lot more work.
Very glad that "function coloring is good" has stopped being such a hot take.
July 13, 2025 at 6:28 PM
Very glad that "function coloring is good" has stopped being such a hot take.
I wonder how much ChatGPT usage spikes on Prime Day
July 11, 2025 at 11:50 AM
I wonder how much ChatGPT usage spikes on Prime Day
Support for atoms is now implemented.
June 24, 2025 at 11:29 PM
Support for atoms is now implemented.
Running into a particularly challenging issue. I want anonymous structs in my language. Doing this in a way that's zero cost + compiling to Rust is *hard*. In particular, the first pass I have is to gen `HasX` traits for each prop, but then rust borrows the entire `self`, making moves impossible.
June 24, 2025 at 10:22 AM
Running into a particularly challenging issue. I want anonymous structs in my language. Doing this in a way that's zero cost + compiling to Rust is *hard*. In particular, the first pass I have is to gen `HasX` traits for each prop, but then rust borrows the entire `self`, making moves impossible.
My language is now faster than rust... sometimes.
Faster default hasher + the effects system opens up novel optimization opportunities.
Faster default hasher + the effects system opens up novel optimization opportunities.
June 23, 2025 at 5:52 PM
My language is now faster than rust... sometimes.
Faster default hasher + the effects system opens up novel optimization opportunities.
Faster default hasher + the effects system opens up novel optimization opportunities.
got capabilities done
June 22, 2025 at 12:40 AM
got capabilities done
Spent time writing some basic optimizations for codegen. I've doubled the performance of some operations. Still much slower than Rust, but I'm about to write a CFG to do a lot more optimization. And there's still low hanging fruit.
June 20, 2025 at 10:04 PM
Spent time writing some basic optimizations for codegen. I've doubled the performance of some operations. Still much slower than Rust, but I'm about to write a CFG to do a lot more optimization. And there's still low hanging fruit.
My language now supports inference of union types in a way that's really neat and compiles to a rust enum under the hood. Full Hindley-Milner with unification, lots of type tracking info that I may use later like *where* a type comes from (for flow typing and refining later!).
June 19, 2025 at 8:56 PM
My language now supports inference of union types in a way that's really neat and compiles to a rust enum under the hood. Full Hindley-Milner with unification, lots of type tracking info that I may use later like *where* a type comes from (for flow typing and refining later!).
Anyone aware of good type system stuff for GC'd languages that track mutable vs immutable references? Seems like you basically end up with regular reference types + specially marked "unique" types where unique types have tracked references.
June 19, 2025 at 6:44 PM
Anyone aware of good type system stuff for GC'd languages that track mutable vs immutable references? Seems like you basically end up with regular reference types + specially marked "unique" types where unique types have tracked references.
Writing this programming language and documenting it all as I go. Then I give cursor a prompt to try learning the language from docs and generate test projects as it goes, validating every bit of documentation.
Either it finds bugs in the docs or bugs in the compiler, extremely helpful.
Either it finds bugs in the docs or bugs in the compiler, extremely helpful.
June 15, 2025 at 8:41 PM
Writing this programming language and documenting it all as I go. Then I give cursor a prompt to try learning the language from docs and generate test projects as it goes, validating every bit of documentation.
Either it finds bugs in the docs or bugs in the compiler, extremely helpful.
Either it finds bugs in the docs or bugs in the compiler, extremely helpful.
Time to try to give my programming language reference types.
June 14, 2025 at 11:48 AM
Time to try to give my programming language reference types.
I'm building a lil programming language that compiles to Rust. It's GC'd, much friendlier type system, more limited in power, but lets you write inline Rust as-needed.
Big feature is capabilities.
Compiles and works, just need to improve the type inference (limited Hindley-Milner right now).
Big feature is capabilities.
Compiles and works, just need to improve the type inference (limited Hindley-Milner right now).
June 11, 2025 at 4:06 PM
I'm building a lil programming language that compiles to Rust. It's GC'd, much friendlier type system, more limited in power, but lets you write inline Rust as-needed.
Big feature is capabilities.
Compiles and works, just need to improve the type inference (limited Hindley-Milner right now).
Big feature is capabilities.
Compiles and works, just need to improve the type inference (limited Hindley-Milner right now).