Jason Corso

My take: White-collar jobs aren't being replaced by GenAI; they are being redefined by high-stakes curation.

Achieving a maintainable, professional result requires bridging the "Pareto Gap" with domain expertise. Your moat is knowing when to push back.

Read the essay: medium.com/@jasoncorso/...

See you there!

or a donkey.

Elegant code doesn't pay the bills; solved pain points do 🛠️

In the lab, you are the customer. In a startup, it's a trap.

TAF Series Week 7 covers finding your ICP (Ideal Customer Profile) & avoiding the "Mirror Fallacy" pulling from @voxel51.bsky.social

Read more: medium.com/@jasoncorso/...

This whole conversation reminds me of my favorite book I read last year. Quit by Annie Duke.

Be careful with that word courage.

Job postings links:

Postdoc (Research Fellow) careers.umich.edu/job_detail/2...

Technical Project Manager careers.umich.edu/job_detail/2...

Senior Software Engineer careers.umich.edu/job_detail/2...

Hourly (Student) forms.gle/3b1NLcRKD5Dn...

- architect, build, and maintain stationary and mobile physical infrastructure
- planning and conducting human factors user studies, possibly in the field
- preparing papers, presentations, documentation, and videos for the work
The job postings below have more details about requirements and such.

The work involves
- developing, training models and learning approaches for the AI agent
- architecting, implementing, and maintaining a services oriented codebase
- curate datasets to support model development and capabilities assessment

MULTIPLE RESEARCH & ENGINEER POSITIONS OPEN

We are opening up multiple new roles to support multiple funded projects at the University of Michigan. These projects focus on human-AI teams, physically-grounded CV, continually learning, RL, and AI agents.

Girl dad engineer geek moment of joy --- Barbie and sisters are Math wizards in our Barbie-land.

Super super grateful for all of the hard work the thousands of reviewers, hundreds of ACs, dozens of senior ACs, and the PCs did to make this review cycle possible.
Thank you thank you.

Yes I told it I was driving at the beginning of the chat to avoid it spewing code when I could not read it.

On a long drive last night, I was geeking out with Claude on how to evolve my orgmode configuration. Really dig the voice-to-voice dialogue.

Two-plus hours into it, Claude seemed to get bored and basically told me "we're done."

Is Anthropic business-logic or is a reward to get to the point...

I’m afraid… Maybe 75% or 80% of papers at recent conferences however violate this idea. Perhaps this is what the recent edition of the findings workshop is designed to help with.

Total FOMO man.

And how many prompt iterations?

GitHub - Comfy-Org/ComfyUI: The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. - Comfy-Org/ComfyUI

Really impressed by @comfyui.bsky.social ComfyUI codebase and UX experience.

Just stumbled into it for some video diffusion testing and was delighted: as a builder, it's always nice to see good work...

Repo: github.com/Comfy-Org/Co...

Did you see this post from me on LinkedIN today about my C64? Here's the post: www.linkedin.com/posts/jason-...
And the video is here in this reply! (That is me, sort of)

Stop Building. Start Listening....

See my essay here about the greatest pitfall technical founders make in the earlier days of their startup: medium.com/@jasoncorso/...

In cv/ml, there’s a big gap b/w papers and production

We're measuring it. We're asking computer vision people, MLEs, data dogs-folks who build visual AI systems everyday-to share how you’re actually building, deploying, and scaling for real

Take it link.voxel51.com/visual-ai-su...

Thank you

Sometimes Loss Overwhelms People

Weekend Reminder as your write your @cvprconference.bsky.social
rebuttals : your reviewers are human too. They are busy.They make mistakes. They are give their time with no compensation. Be grateful be respectful.

(If you don't think they're human, vote, like, repost this survey lnkd.in/e6-9e2Cw)

#cvpr2026 | Jason Corso With the #CVPR2026 reviews just released yesterday, there is so much complaining going on about how reviews are from LLMs. I've seen things like "I am so frustrated with my reviews, they are LLM ge...

You think your #CVPR2026 reviews were generated by LLMs. I don't buy it.
Let's get some data.
Here's a poll: www.linkedin.com/posts/jason-...

I think far fewer reviews are LLM generated than people think. The reviews may be terse and not what you want but that doesn’t mean they are not human generated. They may be verbose and not what you want but that doesn’t mean they are human generated.

Wait, are you sitting in the front seat? (or just leaning forward)

Eh, certainly not...
But it is interesting that certain "failures" seem to be in the eye of the beholder. EG it's not clear to me that not being end-to-end is necessarily a failure.

I’m teaching a computer vision course this term and building a fun 1-hour “wins vs. fails” kickoff lecture.

What are the biggest successes & failures in CV in last 3 years? Please share specific examples (paper/product/deployment/news) so I can cite them. Photo contains my hot-take on wins...

Wow. Good luck. No words

Thanks @escorciav.bsky.social really appreciate the nod here. Our meetups are open to all and great fun. Also check out our open source tool via pip install fiftyone. Think of it like vscode for computer vision work!

If those plans are changeable, have him reach out to me :)