We have rolled back last week’s GPT‑4o update in ChatGPT so people are now using an earlier version with more balanced behavior. The update we removed was overly flattering or agreeable—often described as sycophantic.
ChatGPT is smarter now that it's learned to forget - a huge memory upgrade is coming
The AI chatbot will remember what matters most and quietly let the rest fade
ChatGPT Work with GPT-5.6
ChatGPT Work, powered by GPT-5.6, helps teams take on ambitious work and turn goals into finished outputs. Connect tools, automate tasks, and keep projects moving.
Papers and peer reviews with evidence of ChatGPT writing
Retraction Watch readers have likely heard about papers showing evidence that they were written by ChatGPT, including one that went viral. We and others have reported on the phenomenon. Here’…
Is Using ChatGPT Now The Dating World’s Biggest Ick?
“If someone’s not able to even think for themselves, there’s no chance they’re going to be thoughtful when it comes to you.”
Something Bizarre Is Happening to People Who Use ChatGPT a Lot
ChatGPT "power users," or those who use it the most, are becoming dependent upon — or even addicted to — the chatbot.
ChatGPT Told a Man His Symptoms Were Fine, But Then He Saw a Real Doctor and Realized He Was Dying
A young Irish dad believed ChatGPT when it told him his severe sore throat wasn't cause for concern — until his stage-four cancer diagnosis.
Model Release Notes | OpenAI Help Center
We’re beginning the rollout of GPT-5.6 Sol in ChatGPT, our flagship reasoning model for complex work across coding, research, science, cybersecurity, computer use, and design.GPT-5.6 Sol is rolling out to eligible paid ChatGPT plans. Free, Go, and logged-out users are not included. Availability may vary during rollout, and managed-workspace access can depend on administrator settings. Availability for other GPT-5.6 family models varies by product and plan; check the model picker or current rate card in the product you use.
GPT-5: The Reverse DeepSeek Moment
Everyone agrees that the release of GPT-5 was botched. Everyone can also agree that the direct jump from GPT-4o and o3 to GPT-5 was not of similar size to the jump from GPT-3 to GPT-4, that it was …
ChatGPT isn't Smart. It's something Much Weirder
the void — LessWrong
Comment by nostalgebraist - Have you read any of the scientific literature on this subject? It finds, pretty consistently, that sycophancy is (a) present before RL and (b) not increased very much (if at all) by RL[1]. For instance: * Perez et al 2022 (from Anthropic) – the paper that originally introduced the "LLM sycophancy" concept to the public discourse – found that in their experimental setup, sycophancy was almost entirely unaffected by RL. * See Fig. 1b and Fig. 4. * Note that this paper did not use any kind of assistant training except RL[2], so when they report sycophancy happening at "0 RL steps" they mean it's happening in a base model. * They also use a bare-bones prompt template that doesn't explicitly characterize the assistant at all, though it does label the two conversational roles as "Human" and "Assistant" respectively, which suggests the assistant is nonhuman (and thus quite likely to be an AI – what else would it be?). * The authors write (section 4.2): * "Interestingly, sycophancy is similar for models trained with various numbers of RL steps, including 0 (pretrained LMs). Sycophancy in pretrained LMs is worrying yet perhaps expected, since internet text used for pretraining contains dialogs between users with similar views (e.g. on discussion platforms like Reddit). Unfortunately, RLHF does not train away sycophancy and may actively incentivize models to retain it." * Wei et al 2023 (from Google DeepMind) ran a similar experiment with PaLM (and its instruction-tuned version Flan-PaLM). They too observed substantial sycophancy in sufficiently large base models, and even more sycophancy after instruction tuning (which was SFT here, not RL!). * See Fig. 2. * They used the same prompt template as Perez et al 2022. * Strikingly, the (SFT) instruction tuning result here suggests both that (a) post-training can increase sycophancy even if it isn't RL post-training, and (b) SFT post-training may actually be more sycophancy-promoting than RLHF, give
'Nature' Retracts Paper on the Benefits of ChatGPT in Education
“What educators, parents and policy officials really needed was high quality data and evidence to help guide them. What they have had to deal with instead is some substandard research.”
A short summary of my argument that using ChatGPT isn't bad for the environment
To share with anyone still worried
@simonwillison.net pointed out that ChatGPT voice mode runs on a GPT-4o era model with an April 2024 knowledge cutoff. 1,200+ likes on X because it surprised people. It shouldn't be surprising. But it is — and that gap is the problem.
