Comment by emotf | Hacker News Reader

Comment by emotf | Hacker News Reader

I have been wondering whether AI-heavy work fits the Hooked model a bit too well.

Most days I have multiple agents running. It feels a lot like pulling the handle on a slot machine. One finishes, another is already waiting. You check the result, adjust the prompt, spin up another thread, copy from one - feed the other, keep going...

Sometimes the output is bad. Sometimes it is nearly right. Sometimes it is surprisingly good. That alone seems to be enough to keep the loop going.

I'm conscious of the chaotic mental state it creates. The context switching is jarring. Code review, writing, strategy, creative, then back again. You catch yourself queuing the next prompt during a call while trying to stay present with the person in front of you.

It started out feeling like the flow-state, now I'm just tired. I have joked “time to serve the bots”, but I am not really joking anymore.

I am quite sensitive to compulsive work patterns because I have already had one serious health wake-up call from prolonged overload. Some of this feels familiar in a way I do not entirely like.

Curious whether others here feel this too.

Yes. You're dopamine hacking yourself. All probabilistic programs come with this type of risk. You need to be actively on guard against it. Yes, the providers are absolutely incentivized to keep you interacting with the system to make ROI materialize. No, you are not crazy. You are aware. Stay safe, and try not to let your skills atrophy.

It’s true about the atrophy. On these days i find it hard to get the willpower to actually engineer. I have to set that aside to the next day. New mind. New thinking.

I've got a light version of this with local models; just one coding agent, one task takes 1-5 minutes. All local on constrained hardware helps; can't really run a ton of agents in parallel at good speeds.

During each task I context switch to some other work, emails, chores etc.

Important to take breaks and assess before starting a new session.

I can do this somedays. Others I find myself staring at the reply panel in an email with my mind still remotely focused on the potential outcome of my last prompt. I think is what gets me. In theory, I should be able to both. But in reality: I’m 1.5x tasking: not entirely multi tasking when in comes to context shifts.

From articles I’ve seen, you’re not alone. Apparently it’s being called “AI brain fry”.

I have enough issues with context switching already, so I’ve just stuck to working with one instance at a time and treat it more like a colleague I’m bouncing ideas off of, or looking to for more information on concepts I’m less familiar with. This leaves me just sitting there while it’s thinking for a few minutes from time to time, which I’m fine with. The idea that we can be actively doing stuff with 0 gaps doesn’t hold.

My tools should reduce stress and load, not increase it.

I think the key phrase here is “I’m fine with”. The pace I’m working at (in an accelerator like YC) I’m not really fine with the waiting. But weighing that up.

Reply on news.ycombinator.com