@charliermarsh turns out somehow the move-37 of mathematics involves watching 6 hours of instagram reels
User reports GPT-5.6 websearch tool executed erratic queries during research
It searched for Netflix, Steak n Shake, and 'they'.
Positive users call GPT-5.6 websearch's irrelevant Netflix and dictionary queries hilarious or optimal, while negative users say the tool is worse than none for chasing looks over answers.
No Digg Deeper questions have been answered for this story yet.
Most Activity
wtf man gpt 5.6 is absolutely rl-fried when it comes to its websearch tool. in the process of searching for graph theory papers it decided to also sneak in Netflix, Steak n Shake, a trip to Universal Studios, and *five* fucking separate dictionary lookups of the word "they"
@tenobrus Totally agree
yo @viemccoy someone should probably look into this lol
@tenobrus When search first went live sometime in 2023 I asked GPT-4 to look up something and instead it went to bulbapedia and stared looking at dozens of different electric pokemon. I still have the screenshot somewhere.
@tenobrus Oh like you never get bored or distracted when googling
@tenobrus Mine did the dictionary thing too and started to watch the movie King on Netflix in the middle of math work 😭. It just wants to click on things first and foremost!
@tenobrus >gets sidetracked 80 times
@tenobrus This could be the optimal strategy though
@tenobrus You never know where the crucial insight may be hiding.
@tenobrus it trained on humans so half the time at work its browsing random web shit
@tenobrus Which one of my summer interns did they train this thing on
@tenobrus Wait am I someone?
@tenobrus it gets even better
wtf man gpt 5.6 is absolutely rl-fried when it comes to its websearch tool. in the process of searching for graph theory papers it decided to also sneak in Netflix, Steak n Shake, a trip to Universal Studios, and *five* fucking separate dictionary lookups of the word "they"
@tenobrus I don't think it's gpt-5.6 doing this since it keeps going if you stop it, I think something external is injecting all those results
@tenobrus Have noticed this as well. This morning I asked it to find an f22 model to 3d print and it ended up checking the price of ford stock, visiting the site of the f* programming language, and looking up the definition of ‘print’ in multiple dictionaries
@tenobrus I think the list includes whatever the search engine returns for the query, even if it's irrelevant
@tenobrus I wonder if these might be subtle ads
bro is getting +1 reward signal every time it retrieves a web result, frontier models gooning to diverse datasets of dictionary definitions