Today's rabbit hole: teaching a model something new = easy.
Making it unlearn something = still an unsolved research problem.
No delete. What it learned is smeared across millions of weights, tangled with everything else. You can't remove one thing without risking others.
Wild
I am joining the movement to build 1 Million Agent Engineers and just contributed to bring it to Bengaluru! Join me and unlock our city: https://t.co/bya9CZo5rW
My test passed.
I ran it again out of paranoia. It failed at offset 970,898.
I ran it a third time. It failed somewhere else.
Here's the concurrency bug that took me 3 weeks to trust I actually understood
The part that stuck with me:
claimed โ committed doesn't stay on one machine.
A write isn't real until a quorum has it.
A leader isn't a leader until its term commits.
Same crack, every layer of distributed systems.