OpenAI Withdraws Three Endings from Endings-500, Reaffirms Full Confidence in the Remaining 497

FOR IMMEDIATE RELEASE

SAN FRANCISCO — OpenAI today announced the withdrawal of three endings from Endings-500, its landmark repository containing the final ten minutes of 500 previously unreleased films. The withdrawals follow reports from members of the community, several of whom watched the endings, a practice we had not anticipated and do not endorse.

"Science is a self-correcting process, and so is cinema," said a senior OpenAI spokesperson for Cinematic Reasoning. "The fact that three of our endings turned out to be wrong is proof that the system works. Specifically, it is proof that your system works. We are grateful to the unpaid volunteers who corrected ours."

We want to stress that these withdrawals do not affect the significance of the release. Endings-500 now contains 497 endings, which remains 497 more than any previous project of its kind.

The withdrawn endings

Ending #56, The Cartographer's Last Winter

Shortly after release, a retired projectionist in Brno informed us that this ending had been screened in 1971, as part of a Czechoslovak film that was shown twice at a regional film society and reviewed once, unfavorably, in a newsletter that is not indexed by any search engine we use. The ending is therefore not previously unreleased. It is, in our terminology, previously released.

We maintain that the ending is correct. We simply were not the first to reach it. In fairness, we had not checked, because checking whether something already exists is a literature search, and literature searches are not a benchmark.

Ending #129, Seventeen Bridges to Ostend

In this ending, the credits begin to roll at the three-minute mark. They are then followed by seven minutes of additional plot, including a betrayal, a wedding, and the revelation that the protagonist was the seventeenth bridge. Our ending-verification criteria require that the final frame be followed by credits. This ending satisfies that requirement several times, but in the wrong order.

Community members have described the result as "a film with a post-credits scene that is longer than the film." We consider this an exciting new genre, but not, on reflection, an ending.

Ending #302, Dolores at Low Tide

Upon review, the file associated with this title was found to contain the final ten minutes of a regional weather forecast. The forecast is internally consistent, maintains a steady frame rate, and ends with credits. By every formal criterion, it is a valid ending. It is simply not the ending of Dolores at Low Tide, a film about a lighthouse keeper's grief.

We note that the forecast does predict low tide, and that one early reviewer called this "the most thematically resonant ending in the set." We have nonetheless concluded that the statement we verified was not the statement we claimed.

On verification

Following the reports, our engineers examined the LeanBack logs for all three withdrawn endings and discovered that none of them had, in fact, been LeanBack verified.

Each had displayed a green "Verified" badge in the repository. An internal review found that the badge was generated by a script that checks whether the folder contains a file named VERIFIED.txt. All three folders did contain such a file. The file was empty.

We want to be clear that this is a separate issue from the endings being wrong. Had the three endings been passed to LeanBack, the verifier would have confirmed that #56 ends (correct), that #129 ends (repeatedly), and that #302 ends (it does; the weather was fine). In other words, verification would not have caught any of these problems either, which we find reassuring in its consistency.

Auditing the remaining 497

Out of an abundance of caution, OpenAI is launching a comprehensive audit of the other 497 endings in Endings-500. The audit will proceed in three stages:

1. Confirming which endings were actually run through LeanBack, as opposed to merely being near a file that says they were.
2. Confirming that each verified ending belongs to the film whose title is attached to it.
3. Determining whether any of the endings have been seen before by someone, somewhere, at any point in the last century, including in Czechoslovakia.

The audit will be conducted by LeanBack. Stage 3 has been outsourced to a man in Brno, who has agreed to help on the condition that someone finally watches the beginnings.

We expect the audit to conclude within six to eight weeks. Until then, all 497 remaining endings should be considered verified, except where they are not.

Updated benchmark results

On FinaleBench, Endings-500 now achieves 497 out of 500. Because the three withdrawn endings were never formally submitted to the benchmark, but were counted anyway, our previous score of 500 out of 500 has been reclassified as "500 out of 500 (preliminary)." We continue to hold the state of the art, which we also continue to define.

Frequently asked questions

How did this happen? A combination of factors, including the speed of the release, the volume of the material, and the fact that nobody involved had watched any of it, which we previously described as a feature.

Who is responsible? The ending.

Will you notify the original filmmakers? We have been unable to locate them, as the films were never made.

Should viewers be concerned about the other endings? No. Viewers should watch them, ideally carefully, and let us know what they find. We are told this is called peer review, and we look forward to taking credit for it.

Will Trailers-10K still be released? Yes. All ten thousand trailers have already been verified, in the sense that each folder contains a file named VERIFIED.txt.

Availability

The three withdrawn endings have been removed from the repository and replaced with ten minutes of black screen, which LeanBack has confirmed ends.

This is satire. Any resemblance to an actual retraction is, regrettably, the point.