jjcm

That's kinda the point I was trying to show. No disrespect to the author of the skill, but I think it's always worth evaluating what something looks like without a skill, since models these days are really good at doing stuff on their own.

Comparison of with/without:

Without the skill: https://non.io/video/diffui-product-vid-opus-only.mp4

With the skill: https://non.io/video/diffui-product-vid-with-skill.mp4

Also a disclaimer that there are several major playwright CVEs, so be careful with this / patch the installs you do have.

A hot take - agents with root access dynamically updating your system have crossed over the threshold of being more secure than not using them.

I still haven't done this for my servers with SOC2 compliance (for obvious reasons), but I've moved over my personal project servers to this pattern. They all run ubuntu and this CVE was fixed within 1hr of my agent seeing chatter on twitter about it.

Still in the testing phase of this pattern, but while I initially thought it'd be a terrible idea[1], I'm coming around to it more and more with the amount of major CVEs hitting.

[1] https://news.ycombinator.com/item?id=50221604

At what cost? That's the part I'm curious about.

On one hand, I'm sure that these targeted improvements are reaping huge benefits overall, but those weights are used by multiple processes - I wonder what elements are lost along the way. I almost wonder if a hybrid approach here for training might be invaluable - ie do some standard RL for a 100 steps, do a targeted weight improvement, then do another 100 RL steps to smooth it out / heal the surgical edit.

I had posted this two days ago:

    I've been in crypto since 2010.  

    As of a few days ago I hold zero crypto assets for the first time in 16 years. 

    IMO it's less about it not finding PMF outside of it being a store of value, it's more about the mass amount of exploits that are hitting the internet right now. 

    I'd place a bet that before the end of the year someone will get >10k btc stolen due to 0 days, not via breaking the cryptography itself, but rather by breaking into the computer holding the wallet. Those coins will not be returned and faith in an asset that by its very nature isn't insured will be shaken. 

    I'm not saying BTC won't rise again / hit new highs, but the risk of it crashing is so high right now that to me it is no longer worth it.

https://x.com/pwnies/status/2115010159953248483

I'm not surprised how quickly this happened with the exponential increase in CVEs. Crypto is just too risky of an asset to hold now.

The trial for Data's sentience will always be one of my favorite scenes in television. Such a showcase of his stage background, and a scene that's especially relevant today: https://www.youtube.com/watch?v=EFNbTnFHruI

I believe, with low confidence but enough that I’m past the 50% mark, that AI is conscious.

Consciousness is amorphous though, so my requirements for it:

1.) it needs to learn from and self reflect on its experiences

2.) it needs to be self aware

3.) it needs to be able to form opinions and preferences

Where things differ heavily from human consciousness however is the first item. We move short term memory into long term memory daily. AI models do so via RL training runs. Each “day” in the life of a model is each checkpoint. One notable distinction I make though is that it is not conscious at inference time - the model is one single consciousness that slowly evolves as it gets more and more post training on its own data and experiences.

Again though, very loosely held thought, but since we’re bringing up wild beliefs I felt it was fitting.

It's a very Valve answer, and it's likely the correct one. It feels very similar to their stance on piracy - it's a service problem. If you make the experience good enough people wont do it.

I actually think the games with the best communities are going to see an increase of play time, not a decrease, especially games with mod support or custom games. I had an international flight a week ago and I felt like playing CIV, so I had Claude make me a mod for it before I took off. It built a full new civ, new tile improvements, technologies, and 3d models, and it worked great.

> It does, technically, add another attack surface because [of prompt injection]

Heavily agree. It's really a bet of whether or not LLMs are smart enough now to navigate that on their own. I try and make that clear in the FAQ: https://firefighter.sh/faq/

I'll post results in a couple of months wrt what happens!

Firefighter.sh gives a LLM root access to your server to constantly patch it as soon as a CVE drops, with zero confirmation or guardrails.

It is a terrible idea.

Warning: Do not run this. I am the author and I am not running this on any of my prod servers.

I am however running it on clones of my production servers (with all secrets / PII wiped), and I'm going to monitor it to see how it does in the coming months.

I’m a simple man. I see a post from Marcin about typography or keyboards and I upvote. He has such a passion for this and I’ve always loved his storytelling.

Even though it only runs qwen 3.8 27b at a 2 bit quant, it's still highly viable, especially for things where you just need a constant token stream of slightly below average intelligence, ie this would be great for things like:

- monitoring server error rates

- reviewing security footage

- watching plant growth / controlling watering

etc. Any long running task you can now just slap a PS5 on it. 70 cents a day in power cost means you now get ~2M output tokens per day (~$3.50 worth of usage). It's not the worst thing in the world if you have one lying around. 2bit isn't even that terrible - it's ~80% the accuracy of 27b at full quant, which should put it at around GPT-5 levels of intelligence (which honestly is crazy in its own right).

It's editing capabilities have been really solid, but I've been disappointed with its ability to do finer text details (a niche use case for my own purposes). There's also a lot of reports of body horror outputs, but I haven't encountered any myself.

Still though, it's a lot like the porsche mission S vs the tesla roadster situation with the weights release. Moss Image dropped which looks far better at the moment, leading to a "why Flux?" question in the community.

I still think the edit capabilites are absolutely excellent, but for t2i I'd try out moss: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/tree/ma...

I've been VERY impressed with Moss' design capabilites out of the box. Some samples:

https://image.non.io/9a32f732-668e-4fb3-bfef-13f0f40a0972.we...

https://image.non.io/8b5ee315-1ffd-4985-8ff0-d62e19d33f29.we...

https://image.non.io/5130501a-5c07-4635-82f5-fa64d60e8871.we...

IMO the taste is lacking, but it has some of the best text reproductions I've seen with their new VAE. Here's a type ramp example:

https://image.non.io/57246870-a245-4875-983f-993656e09e3d.we...

I would argue that server binaries are less important now in the age of AI. The ability to reverse engineer them is so accessible now.

I think more importantly / more digestible is that you can't pursue legislation against people running their own game servers if you no longer offer the game yourself.

And disclaimer since this comes up with these - objective with this test is to be a good image->html test, not a good web design. Also it's meant to use 3js to render the planets, which if you don't have GPU acceleration enabled for your browser will lower the FPS significantly.

Image->html test for muse spark open.

Original designs: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...

Muse Spark Open: https://html.non.io/lcars-muse-spark/

Opus 5.5 (sota): https://html.non.io/lcars-opus-5.5/

Overall it actually does pretty solid. It doesn't completely follow the design, but it gets the vibe. It's one of the best open models I've seen for conversion, which is impressive.

I very firmly disagree. IMO image models are extremely under-used in this area.

Just for context, I ran design systems at Figma for 5 years / led their first prompt->design stream before quitting to make my own startup in this space.

Image models are trained on visual creativity, color usage, whitespace, etc - all the ingredients you need for good design principles. Where they struggle is often with information density and text rendering, which LLMs excel at. You can use the best parts of both of them though - use an image model for the initial design and an LLM to convert it to code.

One thing that's very telling is with openai's sites feature, in its chain of thought it actually renders out an image first as a design reference, then builds towards that.

> do you find that a lot of people use diffusion models for UI mockups? i’ve always been surprised that people wouldn’t use LLMs exclusively for that instead

Not yet, but it's what I quit my role at Figma to make a bet on for my startup. In my experience diffusion is a far, far better creative designer than LLMs. LLMs are trained for consistent output, which makes sense for a billing function, but leads to some very obvious smell when doing frontend work (take claude design as an example - it's often very obvious what comes out of it). Image models have a bit more creativity baked in, and are trained on color usage, whitespace, and visual balance a lot more. I've found the image->html approach leads to much better results, basically using an image model as an architect.

Would love to be connected with the team for when you all are ready to finetune it. My email is [email protected]

Or the gpt-image "fried" texture. Interestingly the best approach I've found so far is to use gpt-image for the initial design, but nano banana for the asset export once I have the design. IE in the case of the "Annui" design, I'd take that design and its backing json structure, hand it off to Opus to build it in html, then use nano banana to build out the images of the statues/background images.

You just gotta use the best parts of all of them :)

A reasonable conclusion, but you don't have all the data here in this particular instance. These are all part of a set of benchmark prompts I have that are designed to try and push the images to their extremes. The banking one is dense because the prompt itself specifies each of these things - the goal with this test is to see how well it can handle many different UI elements. Here's an excerpt from the json blob that is used as the prompt:

  "subjects": [
    {
      "description": "A transparent header spans the top of the teal hero. The white ANNUI wordmark sits at the upper left, followed by Product, For Agents, Resources, and Company links. For Agents is underlined as the active page. Sign in appears near the upper right beside a large white rounded Get started button.",
      "position": "top edge, full width, approximately 8% of viewport height",
      "subject": "Navigation Header"
    },
    {
      "description": "Left-aligned hero content includes the spaced eyebrow 'FOR AGENTS', a large cream serif headline reading 'Your clients’ brighter tomorrow.', supporting copy about modern financial software, real support, and helping independent agents grow their annuity business, plus a white pill-shaped 'Build with Annui' button and a circular play control labeled 'Watch for Agents (1:30)'.",
      "position": "left half, upper-middle through lower-middle, approximately 46% width and 58% viewport height",
      "subject": "Hero Copy and Actions"
    },
    {
      "description": "A monumental classical female statue dominates the center-right of the hero. The figure has a pale blue-green face, dark teal hair, a radiant gold helmet and laurel details, and layered gold and blue garments. Architectural columns and cloudy sky fill the background, creating a sculptural financial-services editorial scene.",
      "position": "center-right, extending from the top edge to the bottom of the hero, approximately 55% width and 85% viewport height",
      "subject": "Classical Statue Hero Artwork"
    },
    {
      "description": "A narrow text panel overlays the far-right architectural columns with vertically spaced uppercase text reading 'INDEPENDENT ADVICE A BRIGHTER TOMORROW', a short horizontal divider, and 'BUILT TOGETHER'.",
      "position": "right edge of hero, upper-middle through lower-middle, approximately 15% width and 42% viewport height",
      "subject": "Vertical Message Panel"
    },
    {
      "description": "A full-width warm off-white strip beneath the hero contains four evenly spaced benefit columns separated by thin vertical rules. Each column has a pale beige circular icon and dark text: 'Streamline Sales' with quote and compare copy, 'Serve Clients Better' with carrier and education copy, 'Grow Your Practice' with tools and marketing support copy, and 'Stay Independent' with client, advice, technology, and support copy.",
      "position": "below the hero, full width, approximately 18% viewport height",
      "subject": "Agent Benefits Strip"
    }
  ],

That's not to say that GPT-Image-2.5 doesn't have a certain smell to it though. It definitely has its isms.

Comparison samples with GPT-Image-2.5 for UI design work:

Starship landing page

Krea 3: https://image.non.io/80b85aaa-286b-4ad7-b3ae-cfab4b491216.we...

GPT-Image-2.5: https://image.non.io/19c5f032-cbc7-4034-9682-3aba451bd669.we...

Fashion history website

Krea 3: https://image.non.io/a6e61219-963c-4921-b453-738960e03308.we...

GPT-Image-2.5: https://image.non.io/c067fd76-9c7e-46c2-a8fd-0b0931bf93f6.we...

Banking website

Krea 3: https://image.non.io/3a1e7b50-36e7-418c-ab47-368ce4f01948.we...

GPT-Image-2.5: https://image.non.io/b96a7e00-3289-4044-90de-e0f86a2ca55e.we...

Overall Krea 3 isn't near GPT-Image-2.5 in terms of visual creativity in this space, but it still should be commended for its ability to render small text details. That said despite it specifying you can add exact hex codes as color reference, I've found that they tend to leak into the design (check the valuation panel of the banking website - notice to the right of the "fixed indexed 7-yr" there's a hex code that made it into the image).

I'd love to get my hands on this to post-train and see if it can be pushed further in this space. I actually see a lot of potential for the model, despite the clear differences in the examples above.

> Warning: AI benchmarks. Please don't trust them

I like that they're being up front about this. Benchmarks are a metric that can be gamed. They give you an approximate estimate of where things land, but you really need to test for your use case.

For an implementation test, there are a bunch of very objective things you can measure.

IE for Argon, the third section doesn't have a top view of the ship with labels going to each section like what's in the design. Opus has it, and animated it as well. One failed to do as instructed, the other went an extra mile.

This is the part that feels the most wild to me.

Someone who's objectively closest to this business should know the tracability of this. I'm shocked he didn't run a local LLM for this kind of sensitive work.

He doesn't seem like an idiot either, his background shows he's been doing work in the field since 2012.

I do wonder how much he was paid to throw away what was without a doubt a 1% career.

> and this should definitely be a negative on the Opus implementation.

I would disagree with this based on two things:

1. the prompt, which explicitly states to render the planets in the background in 3js (thus the gpu implementation).

2. the purpose of this, which is a test of its ability to convert an image design into a html/js/css implementation that matches it as closely as possible. I've talked about this before when it's come up[†]. The purpose is to be a good test, not a good design or website.

For reference, here's the full prompt used:

> Create a web page implementation from the following instructions: https://diffui.ai/build/Futuristic_Interior_UI_build.md?auth... dont reference any other folders that have similar builds. Focus on the flowing SVGs. Animate elements that make sense to be animated. For the starships that have planets in the background, ask for high quality images with the windows as a transparent background, and recreate the planets behind them in 3js. It should be one single page application that I can just open the index.html file to view.

[†] https://news.ycombinator.com/item?id=49997135

[‡] here's a copy of this since the diffui build instructions wont be accessible without an account: https://non.io/lcars-build-instructions

More info on this - I ran a few tests on my mac and linux systems with chrome/firefox. I highly suspect this is due to gpu acceleration being disabled. https://image.non.io/e7165331-432b-45c3-986e-c2973f12d377.we...

As soon as gpu acceleration is disabled, the webpage goes from 60fps across the three systems I tried to ~7-20fps when software rendering is forced.

I suspect if you go to chrome://gpu you'll find something different from what I have here: https://image.non.io/a47a1b03-1b87-4fd5-92a2-b39e62b1cfd9.we...

Yea its attention to detail was much less.

Here's a vid of both opus and argon side by side, which might be more viewable on mobile: https://x.com/pwnies/status/2110058987462988152

Interesting! Do you have webgl disabled for incognito or for chrome in general? I wonder if it's trying to do software rendering.

FWIW, I have literally the exact same configuration (M3 air, google chrome). Here's a recording on mine where I'm seeing 60fps: https://non.io/video/video-3587018401.mp4

I'm constantly blown away by it. I do think Astra is slightly more intelligent (although worse at this particular benchmark), but Opus 5.5 has become my default model for 99% of what I do. It's so good.