Fast Company ran a piece this week titled Welcome to the age of the underdog AI model, framing our $20M bet on training Krea 2 as evidence that disruptive AI work doesn’t have to come from the giants. Mark Wilson’s read of the moment is sharp, and we want to add to it from the inside.
The underdog thesis
OpenAI has raised around $180B. Anthropic, $72B. Krea has raised $83M, has 37 people, and just shipped its first in-house model. By the numbers, this isn’t a fair fight. By the work, it doesn’t need to be.
Our co-founder Diego put it this way to Fast Company: “Until there’s a winner — until OpenAI or someone is profitable — the Olympic Games are on.” That’s the part we want to underline. The frontier isn’t settled. A smaller, opinionated team can still ship something that the biggest labs structurally can’t.
”Trained not to fail” is the problem
Most image models today are optimized to never give you a bad result. That sounds good until you actually try to make something specific, weird, or personal — and the model quietly funnels you toward the same safe, well-lit, slightly plastic look.
Our co-founder Victor said it best in the piece:
"The models are trained not to fail and to always give you a good image. And I feel like that takes away a lot of the creative uses — breaking the barriers and letting people go off-road, letting you make 'bad' images, stuff that looks more artistic that a creative might appreciate more."
The example in the article is a cat riding a bicycle. In Nano Banana, no matter how you rephrase the prompt, you get the same coloring-book composition. In Krea 2, you get hand-drawn, grainy, illustrative, painterly, surreal — variation that actually responds to taste. Mark called it the difference between McDonald’s and a Michelin burger joint. We’d just say: one is built to please, the other is built to surprise.
A model is a sculpture, not a benchmark
Most frontier labs train for measurable metrics. Photorealism. Prompt adherence. Resolution. Those matter. But there’s a final layer of training — post-training, where a model’s actual visual voice gets shaped — that almost nobody talks about in public.
There are maybe 200 people in the world who really know how to do this well. It’s not a model architecture problem at that point. It’s a taste problem.
"Building a model is almost like crafting a sculpture."
We spent seven months building our dataset by hand, labeling it ourselves, and writing our own post-training workflows. Not because we had to reinvent the wheel — but because the default wheel rolls in a direction we didn’t want to go.
What this looks like in the product
Krea 2 isn’t just a model card. It’s wired into the way you actually work in Krea:
- Drop reference images directly into the prompt bar. The model uses them as style anchors.
- Slide their influence up or down per image. No more guessing how much “in the style of” actually means.
- Build a mood board and prompt against the whole aesthetic, not a single word salad.
- Get a personalized board of suggestions after you generate, tuned to what you’ve been making — less Pinterest, more co-pilot.
And because this is built for creatives, your inputs are yours. You can opt your own work out of training. The IP you generate is yours. We’re also experimenting with using AI itself to credit creators whose style measurably influences a generation — a sustainable royalties layer is something we want to figure out before the industry locks in the worst version of it.
Why “conservative” is a feature
Victor told Fast Company that K2 is the conservative model in our roadmap. That word matters.
The $20M GPU cluster we’re running for a year is funding K2 and two more Krea models. K2 had to work — we’d never trained a model before, and a face-plant on the first attempt would have ended the bet. Now that it works, the next two models are where we get to be reckless on purpose. New training norms. New aesthetics. The kind of risks a frontier lab can’t justify because it has to keep a billion users happy.
The bigger point
The underdog framing isn’t a marketing posture. It’s a real structural advantage. A small team can train for taste instead of averages. A small team can take a creative bet that wouldn’t survive a quarterly review at a $180B company. A small team can call its first model conservative and mean it.
Read Mark Wilson’s full piece in Fast Company — it’s the best outside read of where this is going that we’ve seen.
