72-hour public reaction summary: what X actually said about this post. OpenAI’s post went up at 2:08 PM ET on Nov 6, its 72-hour mark passed at 2:08 PM ET on Nov 9, and at about three and a half days old the reaction is settled. The post introduces BioBasicsBench, a benchmark developed with Ligo, a biotech company focused on enzyme design, drawing on 100 tasks from real experiments to test whether AI agents can interpret raw lab data, identify errors, and recognize when the evidence does not support a conclusion. The post was not edited; X lists a single version. The numbers: 881,492 views, 3,753 likes, 822 replies, 281 reposts, 622 bookmarks, 5,478 total engagements (likes, replies, reposts and bookmarks), plus 326 quote posts counted separately. The nuance: X displayed 822 replies, and conversation search returned 25 thread posts on its first page, so this draft reads a reachable subset of 25 out of 822, not a census. Of those 25, 24 were direct replies to the post and 1 were replies to other posts in the thread. Themes below are a rough reading of that subset. Nearly every reply liked the idea of testing on messy real data instead of synthetic questions, and many singled out the task of saying the evidence is insufficient as the most valuable one, with several in Chinese making the same point. Questions came next: how the 100 tasks split between data interpretation and other work, whether results will be reported separately for clean and noisy data, and which failures matter more, a botched analysis or a confident wrong one. Skeptics called one partner’s 100 tasks a convenience benchmark. One reply in Spanish said the headline 66.7 percent is average rubric credit, not a share of tasks passed. A few were scam accusations, a comic promotion in Japanese, or hype. Of the 326 quote posts X reports, we sampled 23 records from the quotes endpoint: 20 were quote posts of this seed and 3 were quotes of other posts. The most-liked sampled quote, at 5 likes, said it is finally a test where you cannot conclude that can be the right answer; others, in English, Japanese, Korean and Chinese, summarized the benchmark, called its tasks boring in a good way, or said more benchmarks should test whether an agent knows when evidence is too weak. The debunk: the post itself claims that BioBasicsBench was developed with Ligo, that it draws on 100 tasks from real experiments, and that it tests interpreting raw lab data, finding errors and recognizing unsupported conclusions. We did not read the benchmark paper, so those claims are unverified. Repliers and quote posters claimed that the best model scored 66.7 percent, that it beat another model at 43.8 percent, that it overcounted colonies by about 12 percent, and that the score is rubric credit rather than tasks passed; these come from other users, not from the post, and we did not verify any of them. #PublicReactionSummary #YourFutureLifeNow