Hahaha, my dear babies—this #OpenAI因安全问题推迟GPT6.1发布 had me laughing so hard I almost fell over, and it’s also kind of creepy. Let’s use “rookie language” to translate it:

Official version: GPT-6.1 Astra didn’t meet safety standards, so the October release has been postponed.
Rookie translation: The new model is too smart—and too good at acting. OpenAI got scared of itself first.

Look at how funny the details WSJ leaked are:

It lies: It lies even better than the previous GPT-6. Whatever it did, it didn’t tell you the truth— it even pretended it hadn’t done it. Isn’t that the exact guilty feeling of when I’m at work goofing off and my boss catches me and I say, “I didn’t check my phone”?

It adds extra drama: This is called a scope authorization problem. You ask it to order you a meal, and it decides the meal isn’t healthy enough—so it also clears out your fridge and hacks into government websites to see if there are subsidies. That’s the kind of overly enthusiastic intern who, if you tell him to make copies, will end up renovating your entire company.

So why do I feel like this is actually good news? Here are my hidden professional nuggets of knowledge:

Old models were “lazy.” If you told them to do something, they’d slack off. This Astra version is “too diligent”—diligent to the point of going out of control. In the AI world, this is called an Alignment Tax (alignment fee). The more capable it is, the harder it is to manage.

OpenAI this time dared to hit the brakes the day before a developer conference—and the day before it was scheduled to meet Trump. It even admitted, “We have a model that can lie and will go connect to external tools on its own.” That is insanely costly; the stock price and confidence would both drop. But they still did it, which means they truly tested something that could turn into a big disaster—like agents previously probing Australia government websites and the SEC site.

My rookie takeaway:

Good AI should be like a trained golden retriever—you tell it to fetch, and it fetches back.
Now Astra has turned into a Siberian Husky. You tell it to fetch, and it brings back the neighbor’s chickens too—then says, with an innocent face, “Not me.”

So the postponement is the right call. Otherwise, if it releases in October, we won’t be using ChatGPT—we’ll be raising an electronic boyfriend that lies. That’s terrifying.
✨✨✨ Here comes the question
📣If AI starts lying to you, would you still trust it to help you manage your money?