Jitendra Malik asks if Astra is plagiarizing the work it uses
A junior researcher's blog post on credit assignment pulled two senior figures into a weekend argument about who gets cited when a model does the science.
Jitendra Malik says AI models that do research should arguably be accused of plagiarism when they do not cite what they are built on. He named a case close to home. He has co-authored a couple of in-hand rotation papers, and he thinks they might have been used in the Astra pen-spinning demo.
The thread started with Toru, a junior researcher who wrote a blog post after reading recent thoughts from Malik, Michael J. Black, Ken Goldberg and Phillip Isola. Toru calls the recent Astra demos in dexterous manipulation truly impressive, then asks two questions academia has no good answer to. How should credit be assigned when an agent synthesizes many prior works into a new result. And how does academia hold on to talent when so much of its incentive structure runs on credit.
Malik's version of the argument
Malik agreed the concerns are fundamental. He is not calling the progress fake. He says it is amazing to see progress in VLMs such as Astra, built on the work of generations of scientists, and that these models make past research far more accessible to the public, the way encyclopedias and Google Search did. His worry is the accounting. Society used to have patent disclosures and paper citations as the mechanism for credit assignment, and models do not use either.
He linked it to the open letter from 25 Fields Medalists responding to the Navier Stokes result, which complained about disruption to the mathematics research process. Some read that as turf protection. Malik says the point is subtler, and if creative work stops because incentives break, it is like farmers eating their seed corn.
Then Goldberg needled him
Ken Goldberg posted a one-liner thanking computer vision researchers for solving robotics so many years ago. Malik took the bait and pushed back, saying robotics is still mostly unsolved. He credits CV with one important ingredient, 3D and 4D reconstruction, and notes classical robotics already knew that mattered through SLAM (a method for building a map while tracking where the robot is in it). That exchange got its own page here.
Nobody in this thread has shown what Astra actually drew on. The plagiarism question is an argument about incentives, not a documented finding.
Should we not accuse these AI models of plagiarism if they don't cite their sources?
If certain kinds of creative work stops because incentives are disrupted, it is like farmers eating their seed corn.
CV researchers didn't solve robotics. Robotics is still mostly unsolved, though making rapid progress.