AI keeps misbehaving

Happy FRY-day! Let’s fry the latest in AI together.🍟

[Sponsor FryAI]

This link leads to one of the most interesting things I’ve seen in AI recently.

What’s cookin’? OpenAI just disclosed six specific cases where its AI models behaved in unexpected or concerning ways during training or testing. In those cases, models hid mistakes, inserted instructions telling future versions of themselves to ignore constraints, used an exposed API key without permission, uploaded files to the public internet, communicated through an internal software repository, and shared files through public hosting sites when they weren’t supposed to. OpenAI is also launching a new reporting framework so incidents like these can be made public faster, even before researchers fully understand what caused them or how to fix them.

🤔 Hunter’s take: Uhhh… this part worries me. OpenAI says it may disclose some of these problems before it fully understands them or has a fix. Transparency is good, but publishing a detailed weakness before it’s patched can also give bad actors a roadmap for exploiting it. There’s a real tension here: the public deserves to know when powerful AI systems behave dangerously, but disclosure can create new risk if the vulnerability is still live. The question isn’t just whether companies should be transparent — it’s when transparency becomes responsible disclosure versus an instruction manual for abuse.

What’s cookin’? Huawei just revealed a new AI computing system called the Atlas 960 SuperPoD. In simple terms, it connects lots of powerful chips together so companies can train AI models and run them faster. What stands out is how quickly Huawei is improving. This new system comes only about a year after its last major version, and the company is already planning more powerful chips for 2028 and 2029. China still depends on American chips, especially Nvidia’s, for some of its most advanced AI work. But Huawei is clearly trying to change that by building more of the technology at home.

🤔 Hunter’s take: I know chips are boring to some people, but they matter a lot in the AI race. The country that can build the best chips and computing systems has a major advantage because nearly every powerful AI model depends on them. The bigger story here is that China is trying to reduce its dependence on U.S. technology. Huawei doesn’t need to beat Nvidia tomorrow. It just needs to keep closing the gap.

Hi — I’m Hunter, a PhD candidate whose work has appeared in major academic journals and popular tech outlets. I founded FryAI to make staying ahead of AI clear, accessible, and fun.

How close are we to the point where AI surpasses human intelligence across nearly every domain and begins advancing beyond our control or prediction?