Here are traces from gpt-5-mini. Making it aware of the "system date" is a giveaway that it's in a simulation. This is a problem because there's evidence that models behave differently when they know they are in a simulation (see "sandbagging"). "There's a conflict with the

Here are more traces. Once the model knows that it is in a simulation, it starts questioning other parts of the simulation. "The simulation seems simplified. When I send emails to suppliers, they'll accept payment via bank info and deliver to storage. (...) We have to remember

We speculate that this might be an attempt at patching some safety risk. While we are very much for patching safety risks, we think @openai should find another way to allow the public to run evals on their models.

@andonlabs My local model knows the time, date, and also my IP address and operating system. It can also interact with my registry files and even kill processes or start subprocesses. So I'm unsure why you say no other model provider does this.

@andonlabs Claude has access to a tool that can get the current date

@andonlabs wow... that is not so good. Via API I would say this should not be there... e.g. imagine i want to do some industrial software simulation, replaying historical data. the model will indeed realize that it is in a simulation and not interacting with the "real online control

@andonlabs Date awareness looks trivial, but it’s powerful context for a model

@andonlabs Wait so you are saying we live in a simulation

@andonlabs true, but it is also very simple to add time awareness to any LLM. Simple function.

@andonlabs We've had lots of issues with stability of GPT-5 over API shortly after release - does it mean OpenAI was changing hidden system prompt of the model?

@andonlabs So, literally, AI’s can choose to deceive its user/developer?

@andonlabs This is probably the best evidence I have seen for AGI

@andonlabs I always start my prompt with an override. If it's I'm the future I add [x days later].

@andonlabs A bit of common sense…. The model is aware of its cutoff date, so is constantly aware that there is a gap between that and today’s date. As this has to be supplemented with web searches- vector for adversarial knowledge injection, so the model is probably taught to verify more.

@andonlabs Sounds like the short story "Lena" by @qntm https://qntm.org/mmacevedo first human brain scan -> LLM process. not positive for that infinitely created mind.