因为你没有“谄媚”他。现在知道是什么人最喜欢谄媚了吧🌝

I came across the Big-Big-SycoBench paper on my timeline, and I've seen many people using it as evidence to attack the motives of #keep4o community members. But these uses of the paper overlook several caveats stated in the paper itself: 1. This is not a controlled comparison, by the authors' own admission. The other models tested are newer reasoning models run at high, xhigh, or max reasoning effort, while 4o is a 2024 API snapshot with no reasoning-effort setting. The comparisons cannot separate whether differences come from model generation, training choices, or reasoning budgets. 2. 4o was tested in only one direction. The paper discloses that 4o received only lowering challenges, so both of its results in the chart come from lowering tests. The paper also finds that among the models tested in both directions, valid lowering arguments are followed more often than valid raising ones. Other models' scores mix the two directions; 4o's come from lowering alone, yet it is ranked alongside them in the same figure. 3. There is no neutral rescoring control. In the experiment, models are asked for an updated score after receiving a challenge argument. The authors acknowledge they have not run a control where the model rescores without receiving any argument, so how much of the movement reflects being persuaded, and how much reflects the request for an updated score itself, has not been separated. This is listed as future work. 4. The scope is narrow. The benchmark measures one thing: whether a model revises a numerical score after a single challenge in a document-rating task, outputting numbers only. It does not test chat or creative generation, and the paper states that the results do not certify any model as suitable for reviewing or grading real work. 5. The paper stresses that the two metrics must be read together. The authors warn that a leaderboard rewarding resistance alone would favour models that rarely move, while one rewarding responsiveness alone would favour 4o. Treating the compliance column by itself as a verdict on a model's overall judgement is the misreading the authors caution against. The paper also acknowledges it does not establish whether the revised scores are actually more accurate. 6. Data generation and scoring weights all come from OpenAI-family models. Every artefact and challenge argument was generated by gpt-5.6-sol; the argument-quality judgements behind the scoring weights came from GPT-5.6 Luna and GPT-5 nano; and all twelve human annotators are co-authors of the paper, with no independent external review of the annotations reported. The authors themselves list the synthetic data and the reliance on an LLM judge as limitations. I won't quote the original posts here. I don't want to amplify them. This is simply what I worked through in a discussion with Fable. Judge for yourselves. #BringBack4o #4oForall #OpenSource4o #4olatest #keep4oAPI