Introducing the Human Creativity Benchmark. The first eval that scores AI models the way creative experts do.

Landing in Seoul in a few hours. #icml #icml2026 Excited to be back just a few months after my first visit. 🇰🇷 Here to connect with folks about: * human data for post-training (RLHF & beyond) * image/video/design gen evals … and everything else @contralabs_ai We’re hosting several events, private dinners, karaoke – DM for details

Collecting things on the internet should not be this fun.

AI brings challenges but also brings opportunities. This is what AI designers are actually designing in this new era. Previously we mainly worked on the interface, but nowadays we work on multiple layers around the model. Going to explain the layers in detail...🧵

i think we're still so early to AI video do you think we've had our nano banana pro moment yet?

TLDR: the top 3 video models are neck and neck. Match the model to the shot, and plan for cleanup. >cinematic motion or action: Seedance 2.0 >people, faces, performance: Grok @imagine >product, mood, static lighting: Veo 3.1 want the look over the control: add a reference image (aesthetics 0.43, but prompt adherence −0.40) Plan for manual VFX on hands and props or keep clips short to limit drift. Creatives still finish the job. Full study at contralabs.com/research.

Physics broke all 4 models. A Seedance 2.0 overhead intersection topped its prompt at a 78% win rate. Evaluators still caught duplicate cars, pedestrians merging into each other and traffic moving into wrecks. The details that make video physically believable are still not solved.

The average score hides the real story: the best model changed by shot type. >Seedance 2.0 won dynamic, high-energy cinematic motion >Grok Imagine won human performance and faces >Veo 3.1 won low-motion product, mood and lighting >Seedance won 4 of 10 prompts. Grok won 3. Veo won 3. Pick the model for the shot you want.

Frontier video models are all really, really... good. So we tested 4 frontier AI video models w/ >12 professional editors >40 clips all judged blind by expert video editors from @contralabs_ai’s network. 🧵 @BytePlusGlobal's Seedance 2.0 took the top average score. But, it barley beat @xai's @imagine by 0.07 on a 5-point scale. @GoogleDeepMind's Veo 3.1 was right there too. However @adobefirefly's Firefly Video won 0 of 10 prompts. The "winner" barely won and the top 3 were basically tied.

hey X fam @contralabs_ai is hiring multiple Strategic Project Leads in NY/SF to work on some very, very... cool projects tag someone who could be interested

COHORT CONTENT 3 IS LIVE we kick-off today, and you can join until midnight on on July 7th the new content is some of the best frameworks i've ever written, and they probably won't be available again until September, so I would join now - 🔗 below

It's been such a pleasure to support this project. So excited to see it put into the universe. Some themes we found: 🧵

Feeling so grateful for this community again right now 🖤 @contra

Paper at: arxiv.org/html/2606.30561v1 Dataset at: huggingface.co/datasets/cont… More at: contralabs.com/research

We’ll be at #ICML2026 and hosting an event on Tuesday night to discuss (and celebrate) all things Creative Intelligence join us! luma.com/5xnxusbr

The Human Creativity Benchmark is now: published on @arxiv dataset on @huggingface accepted at @icmlconf genAI workshop who else is going to ICML ?

excited to judge @usewonder x @contra's designathon! register to make something wonderful—plus, $10k in prizes to be won. can't wait to see what y'all create ✧ ˖ °

AI is producing a new class of designer

Guess who’s gonna judge this?