Vice-president, Adobe Firefly - Generative AI and Sensei ML platform
Monte Sereno, CA
Joined June 2008
- Tweets2.7K
- Following1.3K
- Followers2.6K
- Likes9K
No spoilers. But for the ones of you that saw Backrooms and are into AI, doesn't it remind you of a latent representation of all the information it has seen? @scottbelsky?
5 cups o’ coffee a day keep the doctor away ;). Feeling healthier already.
Five cups of coffee a day is fine for most adults, heart association says, as evidence grows about the beverage’s benefits on.wsj.com/3TIJLfa
One of the hottest topics these days at Adobe, too - faster builds with artifact caching, deterministic affected tests identification.
Alexandru Costin retweeted
1/ Some things I've learned recently running coding agents on large-scale projects. Most of this contradicts advice from 6 months ago!
Alexandru Costin retweeted
With Adobe Firefly, you're always in the driver's seat. Watch @kinigra steer her output to perfection. 🚗
Alexandru Costin retweeted
Your creative process is about to get supercharged .⚡
Say hello to Firefly AI Assistant, a new conversational interface powering creativity across Adobe Creative Cloud apps. By orchestrating and executing complex, multi‑step workflows, this revolutionary innovation collapses the distance between imagination and creation. Coming soon.
This is a preview of technology currently in development, coming soon to Firefly.
Learn more, plus explore what’s new in AI video and image editing: adobe.ly/4tMW5Ib
Really enjoyed the conversation I had with Michael - great questions and hopefully I did a good job sharing our agentic strategy - venturebeat.com/technology/a…
Alexandru Costin retweeted
Introducing Precision Flow and Markup in Adobe Firefly!
How many times have you struggled with AI image models not giving you precise control over what gets generated?
Today we're introducing 2 features that give you control over what, where, and how much an AI model edits your images.
📏 Precision Flow allows you to control the level of change in your image on a spectrum.
✏️ Markup lets you draw and annotate on the image on your canvas to give the AI model rich visual instructions.
Now available in Firefly and Creative Cloud Individual plans.
My InterAKT co-founder @bogdanripa recovered this old photo of t-shirt from 2005 - our Romanian startup was building code generators for Dreamweaver /PHP/ MySQL. Good foresight :)
Life's too short to code by hand!
Alexandru Costin retweeted
OpenAI acquired OpenClaw
FaceBook acquired MoltBook
Choose your company name carefully…
As an ex software engineer - I’m facing similar dilemmas. Software engineering is changing and we need to change with it.
It's a weird time. I am filled with wonder and also a profound sadness.
I spent a lot of time over the weekend writing code with Claude. And it was very clear that we will never ever write code by hand again. It doesn't make any sense to do so.
Something I was very good at is now free and abundant. I am happy...but disoriented.
At the same time, something I spent my early career building (social networks) was being created by lobster-agents. It's all a bit silly...but if you zoom out, it's kind of indistinguishable from humans on the larger internet.
So both the form and function of my early career are now produced by AI.
I am happy but also sad and confused.
If anything, this whole period is showing me what it is like to be human again.
When legends @tri_dao and @istoica05 publish a paper that beats @BossBobster 's MoMOE implementation for efficient generative models - I am still a proud dad and proud Romanian :) (a summer internship for Bobby at @tilderesearch, and some of the approaches are similar)
MoMoE - blog.tilderesearch.com/blog/…
SonicMoE - arxiv.org/abs/2512.14080
Alexandru Costin retweeted
Building world models. Using synthetic data built with Unreal Engine tune the bidirectional teachers to allow action conditioning. Then distill this teacher into an autoregressive student. Maybe the gap between plain video generation pretraining and action conditioned tuning could be the most significant problem?
Alexandru Costin retweeted
📢 We propose MoGAN, a novel post-training with few-step video diffusion models to improve motion quality. This is done during my internship at Adobe Firefly.
The idea is quite straight-forward but effective for improving motion with relatively low cost: we add a GAN loss on the estimated optical flows computed from the clean x_0 of few-step video diffusion.
Some tricks we applied: (1) use DMD-loss as regularizer (2) regularize GAN discrimiantor with R1/R2 loss. Please check our preprint papers for more details.
paper link: arxiv.org/pdf/2511.21592
project page: xavihart.github.io/mogan/
visualizaion: xavihart.github.io/mogan/src…
