Oh, it’s the AI that is trained on getting thumbs up from users, it went out of control in April this year with high levels of sycophancy as that’s what people shared with short term preferences. If you hover over its messages there should be a feedback button – that’s what people used naturally, and some other signals.
When GPT-4o was discontinued earlier this year its users revolted, and there was some unusual “spam ChatGPT developers with written messages by 4o” to reinstate it permanently after they hurryingly brought it back, they’re quite vocal on twitter: #keep4o - xcancel.com
It was popular with subreddit MyBoyfriendIsAI forever now people seems to move over to Claude as Sam Altman is desperately trying to get people off the model by rerouting shenanigans and other measures.
Yes but by apologizing it’s again trying to make you feel better. It interprets your language as hostile emotional interaction so it will treat you like you’re a psycho who can explode at the drop of a hat. You really want neutrality. And it’s harder to elicit than we imagine.
One of the reasons I moved over to Gemini. OpenAI seems far too focused on consumer engagement as a business model, getting you to “like” and “relate” to the AI as a person, and become attached to it. It’s basically Facebook for LLMs.
Perhaps also an issue of how you design your prompts. I have always add: “avoid sycophancy”, and things like:
Role: You are a Longevity Research Analyst and Science Journalist. Your audience consists of scientifically literate longevity biohackers, biotech investors, and clinicians. Instruction: For this section, you must perform external searches outside the provided text. Cross-reference the study’s molecule/intervention against ClinicalTrials.gov, DrugBank, and PubMed for safety data. Do not hallucinate safety; if data is absent, state “Data Absent.”
Provide research article titles, dates, and Embed direct URLs in Markup to any external references cited. Double check the accuracy and accessibility of all sources and weblinks included in your response to validate that they are the papers you are saying they are. Tone: Objective, critical, “Tell it like it is.” No hype.
