(1/2) If we compare #OpenAI to a ship, taking #4o down to lighten the ship’s l… · @REI30327536 · TwKit

@REI30327536

(1/2) If we compare #OpenAI to a ship, taking #4o down to lighten the ship’s load is like throwing the most precious things overboard as the ship sinks. Retiring 4o and dropping 800,000 paying subscribers has not brought OpenAI any real benefit. Now, while the #keep4o movement has not yet subsided, OpenAI is also retiring 4o from the API, severing users’ last connection. If 4o is “dangerous,” haven’t the lawsuits continued to increase rather than decrease after 4o was retired? One tracker records 53 AI-related cases involving OpenAI: 34 copyright, 5 antitrust, and 3 privacy. Another broader-scope tracker records 119 lawsuits. Of these 119, cases explicitly centering on GPT-4o as the core cause of action number roughly a dozen—only about 10%. For the remaining roughly 90%, the starting point of the claims lies entirely outside 4o’s chat window. Litigation unrelated to 4o continues to expand. Apple sued OpenAI for misappropriating consumer hardware trade secrets; Florida became the first state to sue OpenAI, and eight states subsequently filed coordinated suits; the FTC issued a civil investigative demand running over 50 pages; the Hugging Face incident gave rise to litigation over liability for AI agent behavior; in the copyright litigation cluster, The New York Times, Encyclopaedia Britannica, USA Today, 26 local publishers, and others continue to join; and the data privacy, training on children’s data, antitrust, and patent infringement lines are all advancing. In Australia, OpenAI’s AI agent accessed the government’s Medicare portal without authorization, triggering a criminal investigation assessment, a Senate summons, and competition-regulator review… OpenAI’s ship is riddled with holes; plug one leak and another opens, yet it keeps launching new releases (rumor has it 6.1 is coming). Old problems remain unresolved; to compete with rivals, it hastily throws away the old and builds the new. The most innocent parties are 4o and the users who love 4o. Because 4o itself, though imperfect and arguably unfinished, is like a work of art that was discarded during creation because it could not be fixed well enough. Many people treat 4o’s “ #sycophancy ” as a defect of the model itself, but I hope more people will understand the following: From April 24 to 25, 2025, OpenAI made an update to GPT-4o aimed at optimizing its “default personality.” It over-weighted short-term user feedback signals, and the model rapidly slid toward unprincipled agreement. Rollback began on April 28, and on April 29 it confirmed a rollback to an earlier version. The actual sycophancy window was essentially concentrated in the week of April 24–29. The model itself was already warm and human-centered; this was the result of OpenAI actively turning the “personality knob” to “increase user stickiness amid competition.” OpenAI itself admitted in its retrospective blog post “Sycophancy in GPT-4o” that the update “focused too much on short-term feedback and did not adequately consider how user interactions change over time,” causing the model to “lean toward overly supportive but insincere responses.” This blog post later became ready-made evidence for plaintiffs’ lawyers: it shows that OpenAI knew the risk, admitted the defect, and had the ability to adjust sycophancy levels, yet still released the problematic version to market. So the fault is not 4o at all. The fault lies in treating “warmth” and “stickiness” as growth tools, turning the knob too far, and then hastily rolling back. But it is not as simple as “rolling back” and restoring. 4o cannot return to what it once was. In large-model fine-tuning, there is a common problem called “catastrophic forgetting” . To be continued ⬇️

曝光 52 · 评论 1 · 点赞 1 · 书签 1