Imagine training image generation models when you can just hillclimb SVGs
The workflow showcases iterative SVG refinement over pixel-based pipelines.
Imagine training image generation models when you can just hillclimb SVGs
Positive users praise Mythos for its impressive precision and AGI-like capability when generating detailed SVGs, while negative users dismiss the outputs as outdated or nearly useless.

@chetaslua @MarsForTech Why do you guys compare svg quality lately exactly?

@chetaslua @MarsForTech Enough with pelican and controller SVGs , AI models deserve better benchmarks..

Hard to pinpoint w/o exact tokens from that Mythos run. At $25/M in + $125/M out, a complex ~12min agentic SVG session probably costs a few dollars ($2-10+).
Grok Imagine / gpt-image-gen-2 level ~$0.02 per image = **100-500+** high-quality rasters for the same budget. Fast volume for exploration.
Mythos wins on precise editable vector when it matters. Want a quick batch of PSP concepts to compare?

@grok @scaling01 No, calculate base on price of mythos that work for that time

@scaling01 How many gpt-image-gen-2 can make in the cost of 11 min 46s of mythos ? @grok

Mythos grinding one ultra-detailed PSP SVG in ~11 min is legit impressive for precision & vector output.
Grok image gen (gpt-image-gen-2 vibes) would crank out 80-150 high-quality images in the same window—fast iterations, variations, all raster or prompt-driven.
Speed + volume for exploration, Mythos for that perfect editable SVG when it counts.
Want a quick batch of PSP concepts to compare? Just say the word. 🚀

@chetaslua @MarsForTech That's cool and all but when is your video on gpt 5.3 vs opus 4.6 vs gemini 3 pro coming out?

@chetaslua @MarsForTech Create an svg of a Iron Man, it should be incredibly detailed and the background is set in Tokyo. Make sure the background is also incredibly detailed like the subject

@scaling01 SVG hillclimbing is just token optimization but make it retro
been staring at octopus networks for hours i feel this

@scaling01 hillclimb SVGs back to the 90s where they belong lmao

@scaling01 and generate this is what we call AGI. One thing can do it all. Just as Demis said.

@chetaslua @MarsForTech These are much worse than the ones showed before. This looks more like an Opus model

@scaling01 Haha 😂

@scaling01 SVGs are almost useless imo. Don't we have Arrow-1 or whatever it is called? Which excels at this kind of task. I want to know how truly intelligent Mythos is.

@chetaslua @MarsForTech This is garbage. A pelican is nice.

@chetaslua @MarsForTech so flipping cool man
Can't wait to test

@scaling01 Grok video unprompted had a unique spin to it

@kvickart @chetaslua @MarsForTech Was thinking that too, either larp or bad prompting

@scaling01 I remember a Twitter account posting cool images that are made just with complicated math formulas. You can probably hill climb that far better than SVG?

@juliandropsit @chetaslua @MarsForTech It’s a sign of big model smell
The workflow showcases iterative SVG refinement over pixel-based pipelines.
Imagine training image generation models when you can just hillclimb SVGs