Oh wow, what a surprise! I learned a lot from that article, and often referenced and sent it to my colleagues on my old team. Thank _you_ for writing it!
Bun and the Rust pre-release have continued to be open source; the pre-release isn’t exclusive to Claude Code but available to anyone to download and use. There hasn’t been any public motion to indicate they may close source anytime soon. This FUD is so easily fact-checkable that I’m genuinely confused at how it caught on
Probably because they said AI rewrote it in 11 days, yet haven't cut a release almost 2 months later. If their AI finished it, where's the release?
They ported enough of it to run a single app (Claude Code). Good job. But if I ported Linux to Rust and it was only capable of running a single app, should I be able say it was a complete port? We'd all agree that was misleading. It's the same with Bun.
Producing a stable release for their community of code that had been written by AI in 11 days and had not yet been extensively tested in the real world would be grossly irresponsible.
Instead, they made the AI rewrite available as a canary release and shipped it for a month as part of Claude Code to millions of their own users.
So... not finished, then? There have been lots of commits between then and now.
If I pay someone to build a house and they left me a voicemail saying "Yep the house is finished, it only took us 11 days!" And then later, "No, you can't move in yet, still fixing some last-minute issues (there's no plumbing yet)." We would classify this as: a lie. And if you told it to your investors: fraud.
Would it really be so bad for Anthropic to admit that a project of this size takes months instead of weeks?
I'm not saying AI is not capable of this. I'm saying that their dishonest marketing has given ammo to their haters.
> I don't think they've been pushing that "It only took us 11 days!" message nearly as hard as their critics suggest that they have.
I have to disagree here. From their blog post Rewriting Bun in Rust[0]:
> This Rust rewrite would've taken a team of engineers with full-context on the codebase a year of work. With 1 engineer using Fable & closely monitoring Claude Code, we went from start to 100% of the test suite passing on all platforms in 11 days. One engineer can do a lot more today than a year ago.
The eleven day figure features in multiple animated infographics in Jarred's blog post; https://bun.com/blog/bun-in-rust, including a calendar that shows commits over 11 days.
> I rewrote Bun in Rust using about 50 dynamic workflows in Claude Code run continuously over the course of 11 days.
> letting AI do what it's good at, while constraining the overall system to follow the right practices, is like the ideal combo
What constraints? What right practices? What ideal combo? Maybe this whooshed over your head but I was sarcastically pointing out that just because you can do something doesn't mean it's a reasonable engineering practice or, uh, an "ideal combo."
Hmm, you are right. The 11 days number came from the Bun blog, not the Anthropic blog. Anthropic has a history of making dishonest claims like this (for example their C compiler) so when I saw the news coverage I assumed this was just another one of those. My mistake.
Even then, can you point out where Jarred said this on the Bun blog? Because AFAICT, it seems pretty clear that he said that it took 11 days to get to all integration tests passing, but not to a full release candidate. In his blog post from a few weeks ago, he specifically calls outs:
>Bun v1.4 makes Bun faster, smaller, use less memory and gives the team incredibly powerful tools for systematically improving stability going forward: Rust's borrow checker, Miri (which runs for a growing chunk of code in CI), LeakSanitizer, and 24/7 coverage-guided fuzzing for parsers. There's still more to refactor, but things are off to a great start.
Yup. It almost seems that HN threads that aren’t about AI are against it, according to comments. I have found out that it doesn’t make sense to argue with negative-sentiment strangers…
Well i was mostly talking about the "handing over to a non-profit" control parts. Not the release.
but since you said so, could you link me to the release / tag on github? I don't think they've put it out yet?
Maybe they won't change licenses, but Bun now feels more like SQLite as in "source code available" kinda open source, than the "come build with us, community first" kind that i think used to be classified on at least for me mentally.
No horse in this race, just pattern matching as we meatbags love to do.
Idc about the downvoting i can take a bit of heat on my fake internet points
(I'm not a fan of how they reuse that single tag rather than putting out version-numbered alphas, but they do at least update the description to include the commit hash - "canary" is basically a pointer to the most recent commit hash that they have judged worthy of distributing to their canary users.)
No normative opinion on whether this is justified or not, but noting that this is only for parent -> subagent spawns/messages, and only for the `multi_agent_v2` feature (currently experimental / off by default).
Notably, subagent output is still in plaintext.
EDIT: Title was now clarified. But wanted to expand that this is actually enabled for 5.6 Ultra it appears, which does subagent orchestration more natively in the API rather than direct tool calls; they are beginning to treat subagents as similar to chain-of-thought traces (already encrypted) rather than traditional tool calls.
> and only for the `multi_agent_v2` feature (currently experimental / off by default).
Wrong, this is enabled by default for Sol and Terra (not Luna), no way of avoiding this short of patching the client yourself, and that still doesn't make the backend endpoints work, they want the ciphertext that OpenAI creates on their side.
> but noting that this is only for parent -> subagent spawns/messages
This is almost fully correct though, the encryption only seems to be for the initial prompt the main model sends the sub-agent, not all communication and not regarding the state of the sub-agent at all.
So you can inspect what the sub-agent is doing currently, and the output, but you cannot see what the initial prompt the sub-agent got started with.
Couldn't you just instruct the model to always use your tool call to spawn subagents? Subagents are not some magical thing; it's just another prompt with a couple tool calls for plumbing. One of my colleagues made his own subagent harness earlier this year before codex had them at all.
When I say things like that, I'm talking about a hypothetical version that would be computationally possible. I'm not talking about today's homomorphic encryption.
For the curious, newer versions of MUMPS are still used as the core database framework by a lot of electronic health records companies. Most notably, it’s the backbone of Epic Systems, the largest EHR company in the US.
It has also a presence in the financial world (under brand name "InterSystems Caché"), though mainly as no-sql database until Java and other foreign wrappers stop being enough, and suddenly you're writing stored procedures (or at least their equivalent) in MUMPS
I can’t speak for other companies, but I worked at Epic for a few years (left last year), and indeed MUMPS is still their bread and butter. In fact, I had not even heard about ObjectScript until reading this thread.
A few years ago, they _did_ develop a TypeScript wrapper framework (which was transpiled down into MUMPS). They heavily encouraged newer application code to use this, but it never caught on, was quietly abandoned after a bit.
Internally, MUMPS was often criticized or laughed at, but I doubt it’s going anywhere anytime soon.
Thanks for this interesting insight. Since they are apparently using Intersystems backends, I thought that ObjectScript might be a consequent choice; from my perspective it is rather comparable to JavaScript, so the TypeScript idea might have been stronger, but on the other hand might cause issues with (dynamically typed) globals.
This seemed familiar? The first few lines/commands are a straight copy and paste of a post that reached front page 8 months back: https://maurycyz.com/misc/easy_git/ (HN post [here](https://news.ycombinator.com/item?id=45710721)). The rest is just AI rewording. Even the quips about wiring this up to the blog's site generator are stolen. Even the _date format_ (which is different from the rest of the sites blogs) is stolen.
However, this does not look affiliated at all with the real author's blog. Exploring this vercel site, looks like they rip dozens of creators blog posts day for SEO spam? And for what? Their homepage (and its non-functional demo) is just for a link preview feature? Something that many browsers already provide natively (and if not, you can just use the many existing browser extensions for this)?
Is this a giant spyware/malware promotion? I'm very confused.
OP (https://news.ycombinator.com/user?id=sheelagay) seems to be affiliated with the page and seems to be only posting this for the last few months; their first post is promoting/showcasing the "product".
I used to use a symlink but was concerned that Claude might see the presence of an "AGENTS.md" file (in e.g, a "List Files" tool call output or from a direct `ls`), be curious and attempt to read it directly (not knowing that it's the same as the "CLAUDE.md" file auto-injected by the harness), and essentially double the token impact / context bloat. Indeed, I did some local experimentation and noticed this was the case, which is why I switched to the explicit "@AGENTS.md" approach.
Your solution "appears" to solve the problem, but without any information about how exactly that solution works, it's just magic an vibes.
Does the harness parse the contents of CLAUDE.md, see the @AGENTS.md and then inject the contents of AGENTS.md into that file with a note saying "hey Claude-the-llm, here's the contents of AGENTS.md per an @reference"? Does the harness parse and inject but WITHOUT any ceremony telling claude-the-llm that what it's looking at is from another file? Or does the harness do nothing and Claude-the-llm is merely trained to see @references and treat those as "I should go read that file"? Which is it? If it's #1 and #3, then on long contexts/after compaction, then the LLM is still pretty likely to lose the fact that it read that file and re-read it again. If it's #2, the LLM could have the same problem as with the symlinks.
How does it work? The Anropic docs certainly don't say. So what's the best choice? I guess it's magic. For folks like me who care though, I guess there's always Pi or Opencode (or even, writing your own harness [which is shockingly easy if you want it reliable without all the fanciest features]).
why over complicate it? the answer is obviously yes? it sees it and reads the file? might it sometimes not? yeah, just like every other thing you write in agents.md or the prompt. if you want to write a more verbose statement telling it to read the other file do so. or if you really want to make damn sure it read what you wrote, then symlink it, because it does in fact read the thing twice.
FWIW -- I'm not as repulsed by it as the parent comment. But I do want to substantiate that it _is_ heavily LLM-written.
(If you're unfamiliar, Pangram has garnered a reputation as the leading LLM-detector, with a minimal rate of false positives; IME this has come with the tradeoff of being easy to manipulate/tweak your way into turning an LLM-generated piece of text into reporting a false negative, but for most folks that's worthwhile.)
I have been writing for a long time, and using AI for as long as it was available to me, and I have noticed that I get accused of being an AI more and more - and I do not think that is because I am an AI, but because we are all being consumed with the AI mind-set, and thus the AI is taking over our thinking - so perhaps subconsciously, I am indeed formulating my thoughts in a more AI-centric manner, prompting the association across the vast distances of the internet by other human beings (- or, AI) of my conscious thought, with an artifice.
This is a banal insult, but it is also a dire warning wherever I see it - these days, people moaning about being AI may as well just be AI - automatic ignorance - but .. I do have to wonder.
I applied for my oss project Filestash which satisfy all their criterias but never heard back despite the project having millions of users, more than 14000 stars on github and representing more work than a single person can cope with
I doubt considerations are based on need. Filestash is cool, but probably isn't the marquee marketing opportunity they are looking for; it jas to be a household name they can name-drop or place a logo on a marketing page and get instant street-cred "${AI_MODEL}: Used* by the React project in 95% of PRs closed last quarter"
Not only that, but such sponsorships would also generate new, high-quality end-to-end training data from targeted projects, including work-in-progress edits and prompts which are not present in scraped git histories.
Same here ~13k stars. Never hever heard back from them. So frustrating.
Also applied for Codex Open Source - within 2 days I got confirmation and using it since then. Great Job OpenAI! Shame on you Antrophic for not even sending out a refusal message with the cause.
I applied for both. Heard back from neither. Mentioned two particular projects when applying, one with 2k stars and 5M monthly downloads, and another with 2M monthly downloads.
Now I'm wondering what the bar is since even people with millions of users aren't making the cut. I'm orders of magnitude smaller but I signed up too since I had nothing to lose. Didn't get a response, of course.
Is it possible this is vapour marketing and no projects are actually being selected?
Perhaps someone from a project who has heard back can respond here?
reply