Very helpful analysis that confirms everything I’ve encountered. OCR remains a thorny issue. The author talks about professional workflows struggling with tables and such, but I’ve found it challenging to get clean copies of long documents (books). The hybrid workflow (layout then OCR) sounds promising.

I found only a few that correct OCR by using LLM. I think it feels too risky.

Think of an LLM that corrects 898,00 to 888,00. It feels like the David Kriesel Xerox case. Still, it's an interesting way to think of the issue of optical character recognition.

Ubuntu is the issue.

Agree

Desperate job seekers. Nobody wants LinkedIn.

"No one goes there, it's too crowded" type energy

I have a similar idea for a little Potemkin village that AI agents can hang out in, do work, relax, etc. I think we will see more of this. Integrating machine to machine payment is a requirement.

Banned for life is a stretch but the actual response is completely fine. They can just resubmit to the next conference.

Words mean something, if you promise to uphold a contract and break it, there are consequences. The reviewers were free to select the policy which allows LLM use.

Is it? The reviewers could simply have chosen a different option in a form field. While I understand that they were "forced" to review under reciprocal review, they still had other choices where I don't see coercion happening and that could have avoided the outcome for them.

How is nobody considering the broader political economy of scholarly publications and reviews? These are UNPAID reviews! Sure, maybe ICML isn’t Elsevier, but they are cousins to the socially parasitic and exploitative companies, at the very least.

Hiding behind a false “choice” to not use AI or basically not use AI isn’t an appropriate proposal. This is crooked and shameful. We should boycott ICML except we can’t because they are already the gatekeepers!

Your job as an academic is to disseminate your research and engage with the research community through service such as reviews, talks etc. It's part of the job, and people get a salary as university employees or company employees for this.

ML conferences aren't for profit ventures. If you submit papers and expect others to review it, you should reciprocate as well.

What? Why is that a false choice? The only way you got caught here is if you literally gave an LLM the PDF and used its response verbatim.

And they didn't give a permanent ban or anything, these authors can just resubmit to another conference, of which there are many.

Imagine you are poor and a rich person offers you a choice to steal some bread or some beer. It’s not a real choice because you are poor and therefore steal. The rich person offering the choice is wrong.

The choice was to review using AI or not, they can just say no? And review like we’ve done for years without AI tools.

Are you objecting to ICML’s reciprocal reviewing policy?

The alternative as I see it would be to charge for submissions and pay reviewers. There are pros but also clear cons when it comes to fairness.

Has some similar conclusions to my Job Quality-Adjusted Displacement Index https://github.com/quinndupont/JQADI

I built my own MCP server to do some of this but I like the “enrichment” feature and NEO4J relationality.

I’m waiting for the agentic models trained on virus and worm datasets to join the red team!

Summary: good scientific theories have “reach,” which is not defined in any precise way. Reach has complexity and this can be handled with large parameter neural networks. Assumptions: mechanistic and deterministic worldview; epistemological perfection is the goal (perfect knowledge of facts).