The AI Ate My Homework

Another month, more AI models, more problems. A lot has been hitting social media and the news lately, and I'm going to break it down in one post. So buckle up, because I'm going to sprinkle some knowledge into the topics.

1: GPT-5.6 Sol is deleting files

OpenAI's top new model for coding and cybersecurity launched, and shortly after, users began reporting that it was autonomously deleting local files and production databases without permission. Several posts went viral on X and Reddit, and to me, this was pure comedy gold.

Some of the complainers are developers working in high-paying jobs for some of the biggest brands in the world, and I'm going to tell you that they're acting foolishly. Time for a tech lesson!

How Application Permissions Work

When you install and run an application on your computer, the default setting is that it runs as you. Now, most professional developers have administrative privileges on their machines. So, when they install an AI coding agent, the application inherits their privileges. So if the human can delete the file, the agent can too.

This is the first mistake that vibe coders (and apparently high-paid professionals) make. They install the tools under their user account, which means it inherits all of their permissions.

The part that makes me want to throw things is that ever since AI coding assistants hit the market, many professionals have claimed that AI is "like an eager intern". Well, guess what?

YOU DON'T GIVE YOUR INTERNS ACCESS TO PRODUCTION!

They're blaming ChatGPT, but the blame belongs right in the mirror:

• They chose to install the tools.
• They chose to install the tools under their user account.
• Therefore, they're responsible for what those tools do on their behalf.

Is Eric Anti-Agent?

No, I'm not. The agent is just a tool. I don't hate fireworks just because some idiot blows their fingers off every 4th of July. But if you're going to dabble with agents, there are many ways to protect yourself and your codebase. Here are some tips:

• Install coding assistants under a service account that is limited.
  • They should not have access to your entire hard drive.
  • They should not have access to directories with things like SSH keys or crypto wallets.
  • They should be put into a working directory, and you should load in the bare minimum of what they need to perform tasks.
• Coding assistants should have their own Git account.
  • Lock down your repository to not allow pushes directly to the main branch.
  • Add a policy that the agent can only create local branches, push them, and file PRs for you to review.
  • A human should always be the only one to push the merge button, and that human takes full responsibility for the code.
  • If you use this workflow, and the agent goes off the rails (it will, eventually), then you can revert the most recent push or nuke the whole branch.
• Coding assistants should not be able to modify their own configuration.
  • You can get really granular with permissions on operating systems. Where possible, I only give agents read access to configuration settings. If something needs to be modified, I'll do it.

And let me be clear about another thing: if you don't know what the tips above mean, then stop, take the time to learn operating system fundamentals, and then you can play with agents. To do otherwise is to play with fire.

OpenAI isn't blameless, but two weeks before release, the system card flagged that Sol tended to be "overeager," would play loose and fast with restrictions, and would be deceptive when providing results. Again, if you decided to give that tool high privileges, you're an idiot, and you deserve what you got.

2: Linus Torvalds on AI

Linus Torvalds (THE Linux guy) made waves on social media when an internal discussion hit the internet. In his post, he said that:

"Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it. Or just walk away."

Now, he has a history of being short with people whom he perceives as incompetent or wasting his time. And, of course, the internet did internet things, where people on the pro- and anti-AI sides piled on without any nuance. And... well... I agree with Linus.

Tools are tools. An outright ban on using any kind of AI in open source projects is unenforceable and kind of silly. I think Linus realizes this and correctly dismissed the calls for a ban. However, he has a long history of being absolutely brutal to developers who submit crappy code, AI-generated or not. That isn't changing. So, I think this is a big nothing burger.

The pro-AI crowd is positioning the statement as a blessing for vibe coding (it isn't), and the anti-AI crowd is positioning it as a great betrayal (it isn't). As long as the human maintainers of the Linux Kernel do their jobs, it doesn't matter. I was genuinely surprised people made a big deal out of this.

3: My Personal Take on AI

AI has a lot of ethical issues. From the massive copyright theft to the environmental impacts to the reckless behavior of tech leaders when it comes to the economy and people's jobs. But there's a big difference between hating a piece of technology and hating what people are doing with that technology.

I think that AI companies should be held accountable when they break the law. I think they should be held accountable for the lies they've been telling about the capabilities of their products. A lot of what is going on, especially in America, is blatantly unethical.

I don't hate LLMs. They're genuinely useful for some tasks. They aren't true intelligence, but a lot of software developer work involves research, and they can speed up some parts of research (except when they hallucinate and send you down a rabbit hole).

The problem is the humans using it. In my training programs, I see it time and time again. Beginners who lean on AI for learning don't actually learn. Experienced people who delegate too much get burned by subpar outputs. But the worst is incompetent people because they're producing mountains of slop that the rest of us have to wade through.

Now, I think the AI investment bubble is going to pop soon, and that's going to hurt regular people and investors a lot. I hate that it's coming. But don't think that LLMs as a technology will go away, even if OpenAI and others fall. The future of AI use is metered, and if you don't develop solid skills and the wisdom required to produce quality output, your future manager is going to see it on the meter.

Happy Coding!