That setup is unsafe - there should be no "default account". Otherwise claude-b session will occasionally access configs from that default account. This might be an accidental claude subprocess in a clean subshell (CLAUDE_CONFIG_DIR reset). Also claude has self-diagnose skills/prompts (tied to /usage), and these commands again may read from hardcoded ~/.claude dir.
I don't understand his attitude towards "anonymous maintainers". Right now ALL contributions to curl are pseudonymous, including his own. There is just no such organization as "Curl". Want to see non-anonymous contributors - go to Google/Intel/etc, they ask for IDs when they hire employees.
> A (to me) surprisingly large amount of contributions are done by people who do not state a full real name
Again, strange attitude, given that he personally had legal issues with US in the past, the reasons for which were never disclosed[1].
Typical good developers are not Rambo. When law enforcers come to them and force them at gunpoint to make them commit/add a new maintainer, they should not expect active resistance. Minor reminder: curl is not just some http library, they maintain their own CA list[2]. They don't need any intricate hidden lines for backdoor, CA list is a backdoor on its own.
> Right now ALL contributions to curl are pseudonymous, including his own.
How so? He has a rather extensive autobiography page[1] with his full name (Daniel Stenberg), plus over a dozen photos of himself[2] (and videos), and his git commits are under his full name and email address (hidden on github.com web interface just like on all repos, however). Perhaps you mistook his handle of "badger" for a pseudonym?
> he personally had legal issues with US in the past, the reasons for which were never disclosed
You thought his blog posts about it[3] were insufficiently descriptive? Or you blame him somehow for the fact that the US government never explained itself?
Of course I know name "Daniel Stenberg". But at any moment there is no guarantee that next release published under his account was actually done by person named "Daniel Stenberg". Single person outside of organization is a huge weak point.
Let me explain by comparing curl with XZ (and it was not me who started this comparison). XZ was initially released in 2009 by Lasse Collin. But how would you know that "Lasse Collin" actually exists? Is it a single person or a group? Maybe some papers? (unlikely, for papers see Igor Pavlov's dissertation in LZMA[1]). Do we know his country? No, all that we know that Lasse was in some "hotel with limited internet connection" at the time of the incident (hello to Russian and Chinese prisons). Who would benefit from having an autobiography page, which nowadays can be made up in 5 seconds in ChatGPT?
> Or you blame him somehow for the fact that the US government never explained itself?
Not for that. In his "Administrative Purgatory" post he makes some assumptions on reasons of visa denial. But the whole picture he showed there is that he is just a dutiful developer of a popular opensource library. Also he owns a domain. And that was enough for the US to equate him with criminals. Then why didn't he take a lesson from this that in order to remain free (this applies to traveling as well), it's acceptable to separate hobbies for information security, cryptocurrency, export-restricted technologies like HPC, AI and quantum computing and other legally complex stuff under a separate profile?
Pseudonymity in open-source (as in any other charity) is long overdue to be accepted as the norm. Let's stop building security theater on autobiographies from other people's personal pages.
> Who would benefit from having an autobiography page, which nowadays can be made up in 5 seconds in ChatGPT?
So now you're saying that the fact that (scare-quotes) Daniel Stenberg has a public autobiography page and photos and gives video-recorded in public makes you less convinced he's a real human individual?
OK, sure, you can never be sure that your neighbor or coworker or Linus Torvalds is not a CIA or FSB agent. But you are basing this on nothing.
The project is dishonest. Maybe AI agent was instructed to split Blender UI code, but what happened instead is AI implemented a classic immediate UI, not even reading the code.
I just tried few very standard operations that every Blender user knows:
1) Drag vertically across multiple number inputs to enter multiedit mode - failed (even though main page advertises this: "Drag down across X, Y and Z to edit all three at once")
2) Hover on field, press Ctrl+C, hover on another field - press Ctrl+V - failed
3) Press shift while dragging - it glitched
4) Scale UI - failed
5) Collapse collapsible block - failed
6) Right click - reset to default - failed
7) Jump to next field with Tab (while editing) - failed
8) Automatic alt accelerators - failed
9) Tooltips - N/A
10) AI slop: funny cursors, focus outline
So basically it is more like styled imgui (less polished than Blender), rather than Blender's UI.
I'm not saying that it is impossible to do a honest implementation, but one needs to kick agent much more to prevent cheating, and actually know how Blender UI works.
You completely missed what I wrote in my message. Web version is still neither a original decoupled Blender UI logic, nor an attempt to reimplement it in js. Yes, now when you move cursor across 3 inputs, value changes, but that's not even close to what original Blender controls do.
If you want share a demo, maybe canvas-based would work better (e. g. https://www.egui.rs/)?
I also occasionally have fun injections in my online profiles, and at some point of time it even worked. For example, more than 1 year ago there were few "Show HN" entries for "Analyze any HN profile" and "Analyze any Github profile" (using AI), and silly injections worked. Modern LLMs are a bit smarter than that, but who knows, maybe we'll see some novell small language models that are still prone to these injections.
Aaand it is down again, and I have never seen it alive.
The experiment looks like "Steal my compute/storage scaling money and uptime", and it barely matters that there is some Bitcoin behind it: a company could lose more than $2000 if agents constantly break its website.
AI verification: AI meticulously abuses all available loopholes to deliver answer expected by operator, rather than actually responding to the problem.
Simplest example: proof that `x+1 > x` (borderline impossible)
Well, you can always ask it to predict a next bit in the code. Or even a letter. Maybe combine multiple letter into one and call it a token. We'll give a name to a procedure that gets tokens - decode. To improve speed, we can predict multiple tokens speculatively. To improve quality we'll give it extra space to store Chain of Thought. Ah, wait, we just invented LLMs
reply