Conference on Neural Information Processing Systems

should read.

Who's ahead in the global AI talent race? For the Carnegie Endowment for International Peace, Damian Ma and Bingyi Yang updated Macropolo's global AI talent tracker for the 2025 NeurIPS conference.

One remarkable finding.

41% of elite talent, paper authors at this top AI conference, are employed by China-based organizations, compared to just 34% based at US institutions.

This is such a stark change from the 2022 version of the data, 46% in US versus 27% in China, to the extent that I find it suspect.

I would speculate that this change might capture a change in who publishes in NeurIPS, 5,823 papers in 2025 compared to 2,671 papers in 2022, rather than such a drastic shift in the geographic distribution of elite AI talent.

Should read.

Two relevant China AI issues from the past.

We want to get out of that deadlock.

Yeah, when we met Jeff Dean and did an event with him at NeurIPS two years ago, he was quite interested in this idea of being able to put together relatively irregular neural networks like the brain and so on with all sorts of arbitrary connections.

But probably they were not quite supported by the chips and it made many of those ideas impractical.

I imagine some of those ideas will be back on the table again.

That making this comparison where the system will become equal to human intelligence and then maybe exceed human intelligence is relevant in their framework.

And Ilya gave a talk at one point at this really prominent AI research conference that happens every year called Neural Information Processing Systems.

It's a mouthful.

But he gave this keynote where he shows this chart of the size of brains and the intelligence of a species.

This is background knowledge for ML practitioners, but can be surprising to other scientists.

The venues of record are a few giant annual conferences, chiefly NeurIPS, ICML, and ICLR, not journals.

A paper accepted there is peer-reviewed, archival, citable, and is what counts for hiring and tenure.

Thus there is a single fixed annual deadline per venue, so tens of thousands of papers arrive at once, review compressed into a few intense weeks by reviewers who just submitted their own work, and accept reject decisions with only a short rebuttal and little of the revise and resubmit dialogue normal in a journal.

Significance is judged separately.

Significance judgments are made in optional, rarely awarded certifications.

Outstanding, roughly equals best paper, roughly once a year, featured, roughly equals Oral Spotlight, plus reproducibility, survey, and a journal-to-conference J2C certification that lets roughly 10% of accepted papers be featured at NeurIPS, ICML, or ICLR.

rolling submission with no deadlines.

In other news.

So NeurIPS is a conference around machine learning and AI.

They're saying we found that over 50 papers published at the NeurIPS 2025 have AI hallucinations.

Things like a totally made up paper, including several other papers, various hallucinations in the citations.

They found that John Smith and Jane Doe are getting a lot of citations.