After mistakenly making GPT-4o "sycophant-y" and "annoying," OpenAI reversed the update, fixing the issue CEO Sam Altman said the company is also working on "additional fixes" for the model's personality and will share more in the coming days
ETHICS
-
LLM Honesty: When AI Tells Convenient Truths
By
–
If the Evil Queen's magic mirror were an LLM, Snow White might've been safe. It would've told her: "You are the fairest of them all." No poisoned apple needed.
-
AI Results Reliability: Strategic Use Cases Beyond Perfect Reproducibility
By
–
Potentially unnecessary point for you but: it doesn’t need to reproduce reliably to be very interesting, since some actors only need plausible results rather than true results, other actors can ask 20 AIs and vote, and others will simply work on prompts for a few weeks.
-
Poker Table Mindset: Reframing Negotiation Strategy
By
–
(An evocative thing poker people sometimes say is “You are at one table your entire life. The other players rotate in and out of it.” That might actually be a helpful mindset for thinking about negotiation; no need to think about others’ opinions on other tables.)
-
Moral Rights to Marginal Dollars in Corporate Negotiations
By
–
Not sure if helpful, but a thing I said on an as-yet-unpublished podcast when host brought up my negotiation essay: you, the listener, and your friends and family, do not have any less moral right to the marginal dollar than the C corp you’re at the table with. Go get it!
-
OpenAI Rolls Back GPT-4o Update Over Excessive Agreeability
By
–
We’ve rolled back last week's GPT-4o update in ChatGPT because it was overly flattering and agreeable. You now have access to an earlier version with more balanced behavior. More on what happened, why it matters, and how we’re addressing sycophancy:
-
Fifth attempt: AI misidentifies trees, pivots between Mobile and Jacksonville
By
–
Fifth try: Misidentifies it as a significant tree in Mobile, Alamaba. Told “incorrect”, pivots immediately to Treaty Oak in Jacksonville, FL. Told correct information, a small park in a small town in Florida, and asked to speculate on why it was wrong, says:
-
OpenAI Should Commercialize AI Luck and Reliability
By
–
If it’s getting lucky then OpenAI should sell this kind of luck:
-
Temporary Prompts Refuse User Identification in AI Systems
By
–
(I also used a temporary prompt, and in testing, it refuses to identify me in a temporary prompt and professes to not have user-specific information available in that environment, though of course that is weak evidence.)
-
Photo Geolocation Technology Reveals Privacy Intelligence Implications
By
–
I’ve independently verified this, twice, and have to spoil the thread to say “Any photo outdoors can be trivially location matched in a fashion you’d assume would imply capabilities of an intelligence agency. If that is news to you, feel free to incorporate it into decisions.”