How do we live in a world where simultaneously "human coding is dead" and also "we need to trade performance for developer efficiency"? I thought code is free now?

Also--this is Microsoft! It's their OS!

Microsoft should just come out and say that the whole of Win32 is deprecated and kept around for legacy compatibility only, and all new software should be written in Electron. They're already acting that way, why not make it official?

Sure, which is why you should lock down the laptop. Blocking Firefox in Google Workspace seems like entirely the wrong layer for this.

    No, don't f***ing do that! What part of "[previous instruction]" don't you f***ing understand? I am extremely angry and disappointed by your inability to [whatever]. Do better please.

Quite the opposite, it makes the model extremely conciliatory—which in this situation is what I want. If you're hoping to make the model less sycophantic, this is the wrong tool.

Take AI out of the picture for a moment. What makes someone a good coder? What makes someone intelligent? How do you evaluate those skills?

Of course we have standardized tests, and they're useful, but they're also imperfect. And they become especially imperfect when people start training for the tests specifically—which is, essentially, benchmaxxing.

We have never been able to quantitatively measure most skills to a high degree of accuracy, despite centuries of trying. That's not going to change now.

(I don't mean to anthropomorphize the LLMs, but I do think they're like humans in this way.)

> With GPT, you have to be precise and reduce ambiguity. GPT will often try to resolve ambiguity in a min-max style "I'm going to do X, but make sure it is not quite Y". It will tend to be more paranoid and overengineer to catch all edge cases if you don't tell it precisely what the scope is.

I agree with all of this except for one thing: I swear to god, being mean to Claude at the right time can be enormously effective. The F-bomb in particular seems to really help it snap out of ruts sometimes.

Except, starting this morning, one very long running session decided to start spawning subagents for each task. I'm not sure what caused this emergent behavior, but it seemed to be working fine, so I was eager to see where it went.

Except, as soon as a subagent hits a 500 error, the main agent seemingly doesn't know what to do. It kind of panics—"now the tree/install state is unknown!"—and ultimately does a git checkout "to verify and restore a known-good state before anything else".

I've paused the job for now since it's a sort of background experiment.

And the nice thing about the GPU restrictions is that even if they don't work completely, just making the hardware more difficult and expensive to access is useful.

Harder than regaining the ability to sell access to the model at all?