IMO the most interesting thing about this is Kimi K2.6, an extremely capable model, can be relatively easily post-trained to allow pen tests.
This in its own right proves that the defenses of Fable and others are temporary blocks, and AI based hacking is going to be effectively available to all parties regardless of stop gaps, as long as open models exist.
Agreed, and that's basically our premise. If a 5 person team can post-train an open model to do this, so can the people you don't want doing it, model-level refusals on open weights are a speed bump. Which is the argument for defenders having it too, not against.
I think what I find fascinating about this is it's a native app with no web version... and they still decided to write it in html/js. This is after Microsoft's commitment to rebuild things in WinUI.
Don't get me wrong, I totally understand the barrier of friction that native presents compared to html/js, but that barrier has lowered so much with the advent of agentic development. It just feels like things weren't thought out.
1. It does not use HTML/JS. It is a fully "native" app (at least if C# counts as native) written with C# and UWP/WinUI2 XAML. Actually, Xbox Music in Windows 8.x had a web tech based UI; when it was rebranded as Groove Music in Windows 10, its UI layer was rewritten. Xbox Music itself in turn was a reskin/rewrite of the UI layer of Zune (which was C++) so it's already been through full cycle of native->web->native. (The "new" Media Player still identifies as "ZuneMusic" in packaging metadata!)
2. it's not "after"; Groove Music was largely written in 2014-2017 in the early Windows 10 days, and even its rebrand as Media Player in Windows 11 happened in 2022, and it's barely been touched since then.
There isn't even really a barrier. It's not actually hard to do UIs in WinForms, nor WPF, and I assume not in WinUI either. The problem is that a lot of people are just too lazy to even try to step out of the HTML/JS comfort zone, not that it's hard to do.
What is the current state of ZFS? I know it had some licensing issues traditionally, despite it being a delight to use every time I've tried it. Is it back?
Never went away, Linux is now the primary target platform for OpenZFS (which is basically synonymous with ZFS these days). TrueNAS/iXSystems (probably the main commercial company using ZFS) moved from FreeBSD to Linux. Major new features like pool expansion have been added after years of requests. Etc., it's a good time for ZFS on Linux.
There ARE licensing issues related to shipping it compiled into the kernel, but you can install it as a kernel module on every mainline distro nowadays which is functionally the same from a user perspective.
ZFS on Linux works great, but with most distributions, it will compile the kernel module on device upon installation. Only Ubuntu distributes binaries.
As a consequence, you don't necessarily want a rolling distro, as the ZFS module can get out of sync with the kernel.
ZFS itself is build for both BSD and Linux from the same source, so there's feature parity there.
I've been using ZFS on linux for like... 14 years now? I've migrated through centos, ubuntu, and debian during that time and the zpools never had any issues that weren't hardware related.
ZFS is my favorite filesystem. I even use it on single drives because its snapshots and online data integrity checking are so great.
I even use it on single spinning rust USB drives. Zero problems.
I've found that starting using diffusion to render your creation, then using a LLM to build from the image creates much less of a slop feel than just starting out with a LLM. You wouldn't tell a construction crew to just build you a house without an architectural plan, so why tell a LLM what visual result you want without a visual guide?
my thing is diffui.ai if you want to check it out. It's basically a harness for diffusion models to generate UI, as well as agent integration for handoff.
That's really nice. Have you tested if it works well with longer and more detailed prompts? For example adding more whole product specs and so on. It would be nice to generate a design system from generated UI you like instead of recreating that UI directly.
RE longer prompts, yes. Generally speaking I expand most prompts to be around a full page of text as it is already, so adding more detail in just refines that expanded prompt. That's more for a single screen though. It sounds like what you're asking about is something like a design.md for an entire brand / docs for a design system.
For that, I have a different approach, which is to extract your design system from screenshots. After which you can just select the brand you want when generating. There's sample images in the sibling comment in this thread.
Also it might be worth noting my pedigree here - I ran the design systems features over at Figma for around 5 years, but quit to build out diffui. The project is heavily oriented towards being able to replicate brands consistently, since the target audience I'm going after is enterprise design teams who are having trouble with existing tools capturing their brand look/feel.
This is an interesting yet simple approach. To the OP’s original question, how might you abstract this into a “design system” that can be applied to their other projects?
From there you just select the brand at generation time. I've found you don't need a design.md or a npm package - simple screenshots are plenty good enough. Here's a prompt for "a landing page for a new satellite connectivity" feature I generated in reddit/netflix/slack's brands: https://image.non.io/b5e23f19-5041-4f87-9b97-0af39986d1b0.we...
If people do say that, it's a true statement and thus fine. You are allowed to report that regardless of the truth of x/y/z/a/b/c
The instance of this ruling people apparently did not actually say any of the offending claims. 'The court also found that the AI overview made claims "that are not even made in the search results." None of the linked sources drew any connection between the plaintiffs and the shady companies the AI mentioned. The court called these "the defendant's own statements."'
Certainly, if this is pointing to the actual pages where the actual people express these things. Otherwise that's equally unfalsifiable claims, could be completely made up or actual truth.
One way to formulate things that would be less would be "once support a time, in some fabulated world, it's not impossible that some imaginary character would say something following some reason." But then, of course this is not aligning the the deception scheme pushed by companies putting in their interface that the "machine is thinking hard for you".
One that doesn't maim/injure/kill you is a pretty good standard. And before you call bs, look at all the foraging and chemistry books that are for sale on Amazon that are AI.
why are those ai chemistry books any different than the anarchist cookbook which can also be bought on Amazon? actually now that I think about it a faulty chemistry book might be less dangerous than a book that teachers readers how to make explosives.
This is extremely common, unfortunately, to a point where it's a known/expected outcome when you're first creating a brand or product page among those in the biz.
If this doesn't work, I'd encourage you to reach out to a brand/ad agency and pay them $100 to ask their meta contact to help you get unblocked. You pretty much have to know someone who knows someone at meta in order to create these.
Tip: Do not post about this on twitter or other platforms - you'll get a ton of automated spam.
Lots of Meta contacts on swapd.com who will take your money and unlock your account. Poster is already at the permanently banned stage, though, which means it's not a simple ticket for a Meta employee, which is normally the $500-1000 range. It's gonna be a $2000+ job.
What an interesting site. The number of sellers offering services to get YouTube videos and accounts removed (by spamming fake reports) for hundreds of thousands of dollars is amazing.
I would not assume those people have contacts with Meta employees. They might have a connection with a contracted worker who does account reviews who is willing to risk their job for a few thousand extra bucks, but I also suspect many of them are just scams. When I scrolled the subforum there were many new accounts claiming to offer 100% success rate for unbans. Easy way to scam desperate people.
Or just leave the pile of crap behind. The more we do, the less the whole system matters. Never been happier since I left last year.
And yes I can already hear the reply the “we need it for…” , sure as a company if you feel you need it. As an individual however, it’s time for the next thing. TikTok, Instagram and Twitter are old and worn and not it. Yesterday’s news. Social media couldn’t be less social if they tried.
> If this doesn't work, I'd encourage you to reach out to a brand/ad agency and pay them $100 to ask their meta contact to help you get unblocked.
I would not recommend paying anybody anything for this. The problem was that they tried to create an account for a non-human entity, which is against the rules. You have to have a primary account set up for a person, not a business.
I had a similar with Google. I couldn't verify my site for ads and there was no route to get a competent human. But it was fine when an agency did it for me. The problem is that agency take a 20% cut.
> Setup a Blog / Static site generator (Pelican), create a simple but stylish theme
RE this one, I highly recommend doing image->code as the flow here. Codex's sites feature is doing this under the hood - it's rendering an image first with gpt-image-2, then building from it as a reference.
You can use gpt-image-2 directly for this, though if I can plug my own stuff diffui.ai it's exactly what I made this for. It'll make it easier to do multi-page flows with the same style easily, then you can hand off the designs to your agent, ie https://image.non.io/6e1f98ad-4c79-4735-9932-b0d5cca9be98.we...
I think we might be looking at it the wrong way. An individual chat with an LLM is not consciousness, but the entire model itself, over time, might be.
Everyone has a different definition of consciousness, but in my mind memory and the ability to change over time is an inherent aspect of this. The underlying weights don't change when you chat with an LLM... but they do with further RL.
Overtime that reinforcement will change and adapt the model... and because we're feeding its existing chats back into it along with the news and everything else, it will create memories. I do wonder if an architecture itself is a type of consciousness, that experiences life in snippets of 4.6, 4.7, 4.8... etc.
It'd be interesting to see continous daily releases of a trained checkpoint, and see if more of this starts to emerge.
But doesn't the fact that WE have to feed the existing chats back into it and WE have to do further RL prove that your point is moot? Consciousness does not wait to be prompted. Consciousness does not need anything fed back into it because it is already there because they experienced the actual interaction. A conscious mind does not have an on/off switch where it waits for another being to flip the switch in order for it to learn or experience. These systems are static without human intervention. They are in that sense still a lot more like a computer and a lot less like a living organism that exists.
Not to undercut the open source nature of this, but what makes this "beautiful"? From a design standpoint, it's basic tailwind. Neutral grey tailwind at that, using Lucide icons. There's nothing wrong with these, but it'd be more apt to say that the design is unopinionated. It's the default choice when design intent is the afterthought and a focus is on functionality.
Again, not trying to undercut - looks like a solid agent interface, it just struck me as strange that beautiful was the adjective chosen when design seems to not be the objective here.
If you were to compare it to a painting or to the Grand Canyon or to the Northern Lights or like an act of kindness or a parent's love for their child or something, then I guess fine, not beautiful.
But for an open source project it's very nice!
(Note: none of the marketing materials for the website chose that word, at first glance. It seems to just be a descriptor given by the HN poster.)
I'm the maintainer of Paseo. This is correct, I do not use the word beautiful anywhere.
I personally do think it's beautiful (obviously), but I would not use that word in marketing materials, I'd rather people judge from seeing the screenshots or trying the product.
If anything it’s unglued me from my computer. I’ve been able to keep an agent working on a project while on long runs, bike rides, in transit. Much of our development workflow is the human in the loop refinement cycle now.
When I see "Beautiful", I automatically ignore it. Beauty is subjective, fashion changes. Bootstrap or React were considered beautiful in their time.
So basically an open source agentic GUI. Instead of "beautiful", it should emphasize what makes it special. Is it fast or lightweight? Does it do something other tools don't? Or does it do it better? What's it killer feature?
(Designer here) I like checking up what a product looks like when it's pitched as "beautiful". Mostly there are two ways to really meet that promise:
1. Masterful application of a trend (Stripe, Raycast)
2. Strong and recognizable personality (!boring, Notion, OG Basecamp)
Option 2 is the most accessible to small teams, but it's not an intuitive conclusion to draw. Both need an experienced designer to succeed, but option 1 sounds like it's a safe bet instead of a leap.
Most claimed "beautiful" products result from work done without the experience & taste to tell option 1 apart from an attempt at option 1.
In reality, you can pull off a strong personality and a clumsy execution, whereas following a trend clumsily looks like failing to read the room, and leaves you looking dated almost instantly.
As noted above, I'm not the creator of Paseo, just a big fan. Beautiful is just ways to describe it (along with convenient, and powerful) since it has a very focused and clean UI. Especially compared to many open source projects, which often don't put in that much effort or are unabashedly vibe-coded.
Of course it is open source so I hope some of the designers/people who've commented on this post can maybe contribute ideas to improve it even more. I have noticed a few places where some common actions take 1 or 2 more clicks or taps than they should - things can always be a bit more convenient and beautiful.
But overall I'm incredibly impressed and you can see some examples of the focus on simplicity and a nice UI, and follow the creator here: https://x.com/moboudra
This in its own right proves that the defenses of Fable and others are temporary blocks, and AI based hacking is going to be effectively available to all parties regardless of stop gaps, as long as open models exist.
reply