OpenAI withdrew three of the 722 math preprints it posted

The generating step got cheap

OpenAI posted 722 preprints on 372 unsolved mathematics problems, then spent the following day subtracting from them: three withdrawn after a sign error invalidated one argument and the construction that two dependent papers relied on, 14 others revised, and roughly 50% of the results released without formal verification by the company's own account (OpenAI).

Read as a production figure, 722 is the headline. Read against the rest of today's record, the interesting number is the three, and what it cost to find them. Everything else on the ledger today was also machine-generated at some stage. The difference is where the checking sat.

What confirmation actually costs

The DESI Strong Lens Foundry had a residual neural network pick 76 gravitational lens candidates out of the DESI Legacy Imaging Surveys, then confirmed 70 of them spectroscopically with VLT/MUSE observations taken between 2022 and 2024 (The Astrophysical Journal Supplement Series). The network was the cheap half. Two years of large-telescope time was the price of saying "confirmed" about a list a model produced in an afternoon, and six of the 76 did not survive it.

Chalmers University of Technology ran a closed loop that generated 1,933 candidate hypotheses about brewer's yeast and then tested them, through automated cultivation and mass spectrometry, surfacing interactions including aminoadipate partially protecting cells from formic acid stress; humans still handled some physical steps and set the research boundaries (Journal of the Royal Society Interface). The generator produced 1,933 claims and the loop's own instruments decided which ones left the building.

Soochow University paired room-temperature injection with automated synthesis and machine-learning optimisation to cut perovskite nanocrystals to US$2.16 per gram from US$118.81 per gram for hot injection, and reached 19.37% power conversion efficiency in solar cells (Nature Communications). A working device is a verification step that cannot be skipped by posting faster.

Claims that label themselves

Two of today's papers are careful about exactly this. Three faint quiescent dwarf galaxy candidates near the spiral M101, at absolute magnitudes of roughly minus 7.3 to minus 8.1, would be the first backsplash galaxies observed in the local universe, but only if Hubble or JWST follow-up confirms the provisional distances (The Astrophysical Journal). SPHEREx measured 37 nearby brown dwarfs in 102 spectral bands and reported that current models fail to reproduce the atmospheres through the cloud-clearing transition (The Astrophysical Journal). Both publish the gap rather than closing it with an assertion.

Peer review's own record is not the benchmark

The comparison people reach for is human review, and today's record does not flatter it. Atmosphere retracted a 2023 paper disputing the link between carbon dioxide and warming for unsupported conclusions and inaccurate interpretation of data, after it had collected 26,000 views and four citations; among its errors, it placed the Mauna Loa observatory on Mauna Kea and put Neanderthal extinction at 1200 BCE (Atmosphere). The European Journal of Wildlife Research issued its fourth correction in four years adding a missing ethics statement to an already published paper (European Journal of Wildlife Research). The Anatomical Record declined to retract or flag a 2024 paper arguing a mosasaur holotype was likely forged, even after the original describers published CT scans they say prove the teeth are native to the jaw, holding that later evidence does not make the earlier analysis unsound (The Anatomical Record).

Three years and 26,000 views is a slower correction than one day and three withdrawals.

The line that matters

The split on today's record is not machine against human. It is whether the confirming step sits inside the loop or after publication. DESI and Chalmers paid for it first and published a smaller number. OpenAI published first and is paying now, in public, with a count: that is the better version of publishing first, and it is still the more expensive order.