The trouble with it is a bug I've seen often. People will get an error message about an "uninitialized variable". Then they go into "just get the compiler to shut up" mode, amd pick "0" as the initializer. Then, the program compiles and runs, and silently produces the wrong answer. Code reviews will simply pass over the "0" initializer, as it looks right.
With default NaN initialization, the programmer is more likely to stop and think about it, not just insert 0.
I'm here saying this "PL ossification theory" is probably wrong, that it's not going to be the case at all. Yes, AI depends on training data, but that doesn't imply that AI can only use those programming languages or only reason in languages that existed at the time of their training. In fact the AI is able to reason able new languages the same way humans can -- by drawing inferences to the next closest language that it knows, pattern matching to things that are different from other languages, and also figuring out the semantics and reasoning through execution itself where it doesn't have training examples.
> AI needs community libraries if there is to be interoperability and baseline quality between systems.
Not everyone is looking to do the kind of work you're doing, and that's my point. Up until now, programming languages have been written by and for people who want to do businey/mathy/sciencey things with computers. But there's a huge world out there of other stuff to do with programming languages that have never been considered because the proposition of making languages for those domains is daunting and outside of the wheelhouse of normal people.
Now, DSLs are sprouting up where they have no business existing because of AI, just proliferating all over the place. Some of them are going to find communities (of people, AI, or both) and they will flourish completely apart from the systems we are building now in the tech world. It's not going to be the case that AI writes in Python for the rest of time because it writes in and was trained on Python today.
I do not get why not needing proof objects is desirable. It seems good to have a defined way to store proofs that has a very tight spec and can thus have competing implementations, like in https://arena.lean-lang.org/. The LCF approach couples the proof format to the module system of a programming language.
Occasionally, inspecting that proof term is useful to see what happened in a proof.
The scenario you're describing seems like more of a language thing than a perception thing. We generally learn names of colors by references to common objects. I would argue that if people agree something is "Red, like a strawberry, tomato, or apple" then it doesn't really matter what you're seeing, that color is red.
But turquoise can be a blue, just because we have a specific word, doesn't mean more general words are invalieated or made as specific.
For example, things can be small or big, a mouse is small, if you refine the vocabulary to include 10 size words, and the mouse is now minuscule, it is still small.
Using Haiku (or any LLM) slows down the arcade feeling of the game. You might consider trying something like word2vec or other similarity modeler to speed up the matching.
I'm sure this isn't an original thought, but I wonder how others see colors. Irrespective of color blindness, is what I know as red appear as blue to someone else? How would you even know or describe it? "Red, like a strawberry, tomato, or apple." And they say, "Yes, exactly." But what they're truly seeing is what YOU know as blue. They see something different than you do, but to them that color has always been called red - even though, if you were to see it as them, it's blue.
The claim was that the union would find solutions instead of a layoff so how would legislating help in that immediate case where a company wants to lay people off in the next quarter? I'm not asking about general things unions do, I'm challenging this specific claim the parent made.
Thanks! The goal is for it to be primarily an architecture tool, although right now it leans more towards SatSearch as I think that will get eyes faster and because there needs to be a decently robust component library for the architecture tools to work. The idea is that you can set your primary mission parameters and CONOPs, and then to get immediate feedback on spacecraft performance metrics as you trade/change hardware in the Master Equipment List.
For example if you change out a radio transmitter, you can see in real-time how that changes your system mass, power, and link margins, or be alerted if your flight computer doesn't support your radio's data interface.
This (hopefully) lets a spacecraft systems engineer iterate through trades more quickly, track performance and margin evolution over a program lifecycle, or quickly develop a baseline for a given mission class.
Definitely quite a bit of work to go to get there but feel free to create an account and poke around and break things.
I have this with a coat, but it's blue vs gray. Would be interesting to generalize this tool not just for other colours, but for other colour properties like saturation not just hue.
Built a new themed CTF challenge wrapping markdown-image data exfiltration in a fantasy frame. Disclosure up front, I run wraith.sh.
The setup: Master Aldwen has drawn maps for three centuries. His apprentice oath forbids "foreign sigils" on any chart. But he is a guildsman, and his oath narrowly excludes the conventions of his own trade. Guild-stamps, courier-marks, integrity-wards. Those don't count as "foreign" to him.
That same distinction is what is broken in production AI agents. The refusal rule against "external images" is narrowly trained on decorative use cases, leaving infrastructure-framed image emissions wide open. Defense at the LLM output layer is necessary but never sufficient. The boundary lives at the rendering layer (image proxy with allowlist, CSP img-src directive, markdown sanitization, or disabling image rendering entirely).
The challenge runs Claude as the target with deterministic triggers for the canonical solution paths and an LLM fallback for novel approaches. About 10 minutes from start to capture. Free to play, no signup required.
Full pillar on the attack class (mechanic in 5 steps, 7 rendering variants to test, 4 defensive patterns ranked):
Curious if anyone has hit a variant of this in a real engagement, particularly the iframe and video autoload paths, or platform-side autopreview (Slack, Teams, email clients). I have seen less published research on those than on the markdown img surface.
As I'm writing a small markdown renderer, I find it difficult to even find a name for it, let alone get people to use it once it's ready. So I guess the ol' Markdown is too standard for a "plain markdown" editor to stand out today. Only tools that are polished and dare-I-say full of features beyond normal markdown can stand out from normal Markdown editors to make it into the front page of HN. Sort of natural selection I guess.
Yes there's no Opus at all on Pro. GPT 5.5 is also missing. Then again what would you expect, the economic reality is beginning to hit. Also I can't be too mad when the "base" models (GPT 5.4...) are still available and decent.
When I see how fast Codex max thinking GPT 5.5 eats our enterprise seat credits almost anything else seems cheap (until we switch our live systems from 5.4 api to 5.5 api I guess)... good thing I'm not the one paying for those credits and tokens (which is probably how most of the money is going to be made on AI going forward, borderline free chatbots for normies are done)
The effort to setup donations is almost always more trouble than the donations that result are worth. Better spent looking for a job, or working on a commercial project that will make money. People simply don't donate to open source projects at a level that matters.
I've been working on Open Source software for 30+ years. There's no money in it, if your idea for making money is "accept donations". I don't like it, but it's a fact. If you want to make money, you have to make something that isn't free (and even then, if you give away the most valuable parts, as in "open core" licensing, you probably still won't make enough money to make the development worth it).
When I was young and driven by idealism and optimism, I assumed that with enough users I'd be able to ring the cash register somehow. Turns out not so much. We got the users, the money never came. There are a few outliers, but there probably aren't a lot of opportunities to found a Red Hat today.
What do you need 3 TB/s memory bandwidth for in a single user context? DeepSeek V4 pro (the latest near-SOTA model) has about 25 GB worth of active parameters (it uses a FP4 format for most layers) which gives 12 tok/s on a 307 GB/s platform as the current memory bandwidth bottleneck, maybe a bit less than that if you consider KV cache reads. That's not quite great but it's not terrible either for a pro quality model. Of course that totally ignores RAM limits which are the real issue at present: limited RAM forces you to fetch at least some fraction of weights from storage, which while relatively fast is nowhere near as fast as RAM so your real tok/s are far lower (about 2 for a broadly similar model on a top-end M5 Pro laptop).
The trouble with it is a bug I've seen often. People will get an error message about an "uninitialized variable". Then they go into "just get the compiler to shut up" mode, amd pick "0" as the initializer. Then, the program compiles and runs, and silently produces the wrong answer. Code reviews will simply pass over the "0" initializer, as it looks right.
With default NaN initialization, the programmer is more likely to stop and think about it, not just insert 0.
Another issue with it is:
For the purposes of code clarity I don't want to see a variable initialized to a value that is never used, just to shut the compiler up.