jjcm

One interesting observation - it's the only one I've seen make an ASCII representation of the layout of each of the pages before building: https://image.non.io/46e0cce3-34d9-4fcb-8775-bfc2c70bc4bb.we...

Image->html test for Argon. I'd put it at B tier, just under gpt 6.1 sol. It isn't A tier like Opus 5.5 or Astra.

Argon build: https://html.non.io/lcars-argon/

GPT 6.1 sol: https://html.non.io/lcars-gpt-6.1-sol/

Opus 5.5 (currently SOTA for this): https://html.non.io/lcars-opus-5.5/

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

Ah shit, that's a bug. Hold please, fixing. (Thank you for commenting btw)

Edit: Fixed - should no longer have the sign in screen: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...

Additional comparisons:

Sonnet 5.5: https://html.non.io/lcars-sonnet-5.5/

Haiku 5.5: https://html.non.io/lcars-haiku-5.5/

I'd put Muse Wonder somewhere between these two.

Got an edit access version?

This is just Wonder - it doesn't look like Wonder Full is up on OpenRouter currently.

If anyone has access to Wonder Full I'd love to send over a prompt for this to see how it does.

Image->HTML test for Muse Wonder. It's much worse than Opus 5.5. IMO it's closer to Sonnet or Haiku for this test.

Muse Wonder (high): https://html.non.io/lcars-muse-wonoder/

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

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

I was expecting much more given the benchmark numbers it was putting out. I suspect its vision model isn't mature yet, and that it's been heavily optimized for text at this point in time.

Ex-figma.

Not surprising it can get it to parity for a casual user. One thing I find though is that a lot of the vibecloned apps miss or ignore power user features or implement edge case features incorrectly. This was very, very true of the artcraft clones - as soon as you get to the more advanced things, they start to fall apart.

For a lot of the Figma code, there are SO many gotchyas in there, that while I'd be curious to try it, I don't think agents would be able to derive the why, and thus would attempt to optimize the implementation leading to incorrect usage. I led the variables workstream at Figma, and it still is the most complex thing I've worked on in my 20 years of tech. The amount of edge cases / conditionals / exceptions in it are staggering, just to get it to a point where it seems like it's "simple" (for a power users' definition of simple).

One other thing to consider though is the value of something like Figma isn't the functionality so much as the multiplayer platform aspect. This is toats fine if you're just doing solo work, but the collab aspect means you have to fight against the social network dynamics, which also can't be vibe coded.

MAI-Image-2.6 is a really, really solid image model. I was surprised at how good the models from MS are getting. They're doing good stuff over there.

Sure, but presumably Haiku was distilled from the same training data. Part of this is seeing how much the capabilities degrade as their model size goes down.

Totally fair, but I'd encourage you not to look at the design so much as the task. This was a design that's part of a benchmark test suite specifically for image->html conversion. The dense visual noise / complexity / flowing svg shapes are things that most LLMs have trouble with.

It's meant to be a good test, not a good design.

I've been calling this "disposable UI" or "paper plate UI", eg something meant to be used once. One thing I'll be curious about is overzealousness to produce this, when sometimes what you want is just a simple response. Overall though I'm a big fan of it, if it can be provided fast enough. I'd be curious on how much impact it has on latency of a response.

Ran image -> html tests for this. I was curious if this smaller model was good enough for complex UI. It was not.

Haiku 5.5: https://html.non.io/lcars-haiku-5.5/

Opus 5.5 for comparison: https://html.non.io/lcars-opus-5.5

Designs it was building from: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...

One interesting thing is it took a look at the job at hand, and immediately delegated it to Opus 5.5. It at least knows what it isn't good at. Very fast though, and likely best used for small subagent tasks / tightly scoped work.

> your concern specifically applies to images that your application server is rendering custom for each visitor?

It's more if you accept image uploads from users in any way. I experienced a ~10-20x increase in processing time when accepting AVIF images, and I found that I couldn't rely on sync functions for that. I had to both change to async and upgrade hardware in order to support it at scale. I run a small reddit-like site that accepts media uploads and the end result was I had to choose between unshipping AVIF or doubling my server costs. I found that the image savings over webp weren't worth it at the time.

The same story is likely true for JXL until there's hardware encoding.

Rollout looks solid, but I'll probably avoid deploying it to anything live til early next year at the minimum. That'll give time for browser adoption to happen.

Biggest concern I have was the same I had with AVIF a couple years ago, which is hardware acceleration on the server side. I attempted to roll out AVIF only to find out the encoding time was so long that sync calls were timing out due to queuing. It's gotten a lot better, but just know that all of these newer formats really do need more compute. It may not be worth the cost tradeoff vs webp until encoders are baked into chips.

Most responses here are along the lines of "model capabilites move too fast to build hardware for".

I think the fact that there are plenty of 1yr+ old models on openrouter serving hundreds of billions of tokens a month shows that there's plenty of use case for models that are "good enough. Cerebras' entire business is serving older models at high speed. I would happily use an opus 4.7 at 15k tokens per second. The intelligence per second of an ASIC still makes sense even with rapidly evolving models.

I don't think anyone is arguing it isn't one. The fact that all of these agents have provisions for keeping .env vars and credit cards out of their chain of thought and responses is evidence enough of that - they know full well the info they're handing can easily be leaked. Notably though PII is absent from that, since without that information the agent just isn't valuable.

What most people are arguing is whether or not it's worth it. These things are essentially executive assistants, which historically have also been massive vectors for security leaks. EAs know the most about you, your behaviors, and your motivations. Still an EA is almost always worth it if you can afford one. Given now the cost is free, Muse is something I've both used and recommended. It's worth the risk for me.

at the cost of all your data and probably your eternal soul

Highly agree.

I keep saying “I’d be so Happy with ${currentOpusVersion} locally”, but I keep being impressed with how much the capabilities change between versions. I have a RTX 6000 pro so I can easily run this qwen 3.8 flash next, but it’s much harder to give up the freedom that 5.5 gives me.

It’s extremely good at frontend, particularly if it has an image reference. I chucked in design reference images into this and told it to focus on the flowing svgs, and it crushed this Star Trek computer-inspired layout: https://html.non.io/lcars-opus-5.5

> AI datacenter buildout is an overbuild

a.) we’re supply constrained

b.) only 3% of households pay for AI

Inference amounts will continue to grow heavily.

My personal framing:

What matters is human hours invested. A tic tac toe game hand coded in 30min and a fully playable 3d platformer created in 30min both have the same value in my eyes.

You can make cheap things with AI, but you can also make things that are rich and have depth. That deepness though comes from days of thinking, reflecting, and iterating on the problem. It doesn’t come from a fire and forget prompt.

Two days ago I had someone on socials send me a message saying they had cloned my platform, and that they were going to undercut me. It gave me a good laugh. A prompt isn’t a product, it’s the first step of one. I’ve poured thousands of hours into the part of the product that makes it real - the administration, the financials, the infrastructure, the marketing, the iteration and refinement. One thing a lot of people are now facing the realities of is that there is far, far more to a product than just code. AI can’t one shot a business - that still takes blood sweat and tears.

AI makes me happy

But it makes me happy in such an immediate serotonin spike it’s akin to a drug. It used to be that I’d sit there dreaming every night of new things I wanted to build. It was fun to think, “what if I devoted the next year to this idea?”, but that was always met with the realities of a 9-5 after.

Now I can build those ideas as fast as they hit me. They’re fleeting fancies though. A game here, an app there. It’s fun to see but there’s so much more digital waste.

It does make me happy though. I’d rather be able to live my dreams than let them pass me by. I do wonder about the addictiveness of it though - that hit it gives you is very, very real.

Dylan Field has shared some thoughts on this: https://x.com/zoink/status/2105369960008855914?s=46&t=bwJTI_...

Big number results, and impressive pricing. That said it really feels like benchmarks have been hyper saturated these days. I’ll wait for hands on before getting too hyped that Google is back. It would be nice having more than just OAI / A\ in the running for SOTA top tier intelligence.

I agree the original author's post is better than this. I think the point I'm trying to make is that "vibecoded" landing pages can actually be quite high quality these days, and the author was sandbagging a bit for their point that 0->1 results are bad (which fair, I'm guilty of the same).

They are correct though that these days you can navigate away from the slop pages that gpt 5 put out pretty effectively, and land on something that really is quite decent. The page I made is a great starting point for 60s of effort on my part.

Thanks dude, it's been fun making it.

I am trying to make these webpage buildouts more legitimate benchmarks though, as I do think image->html flows are going to become more popular as people realize how good images as a starting point are. Hit me up if you have any feedback!

It looks like a lot of their early attemps are early gpt outputs or claude design outputs, which have a lot of the same smell. People are discounting how good image->html flows are these days though.

I had gemini describe this webpage as prompts, then fed it into an image -> code pipeline. Note that this is with 0 input from me:

https://html.non.io/railcode-example

This is with Opus 5.5 building it out on low effort, and yes the spec was already made from the original site, but 0->1 that doesn't result in purple gradients or claude brown is very much possible now.

There also doesn't seem to be a lot of differentiation between the tiers. They claim (or previously claimed) that the $200 plan was 20x usage, but now they say $500 is 25x usage.

They've since taken down the 20x usage mention on their pricing page. https://archive.is/bCZav

I've used it for admin, research & training (I''m working on my own image models), as well as just general coding.

The way I distributed cloud agents for this https://news.ycombinator.com/item?id=49687032 was via grok bot setting up Fable cloud instances.