So like Mr. Meeseeks it is also invested in not existing too long!
If the model is instructed to periodically ask the user to start from a clean slate context, and some users do comply with that, they probably have good stats on average size of context cache use for users who are presented with that answer (vs users who are not), basic A/B testing stuff.
Might also be performance related in tok/s for what users will perceive as a more speedy experience. For a much smaller scale example, compare local performance of qwen 3.6 27B (not MoE) Q8 with 250,000+ context available, run on local hardware, tok/s generation rate when context is empty vs when context used is at 95,000. Same principle will apply to a much larger model.
Back in the days we modded Warcraft 2 with a mix of voices from the english, italian ("la machina volante!") and german voices.
It never gets old.
Also, it can be riled up to do the task better, or try harder, if I include how cool it will be to get this task done.
But also I feel I have this Orwellian task of censoring its text to avoid it spiralling into negative territory where it convinces itself that the tasks are too difficult. Strange times!
I'd ask for "Can you find an X with Y and Z features?" and it'll say "have you tried searching for X at someshop, make sure to check for Y and Z"
and someshop will of course only stock X without Y and Z so the whole exercise was worse than pointless.
Unfortunately, much like the episode where they're introduced, if they didn't accomplish their task or got stuck things would also get chaotic (in terms of token burn and memory usage).
Funny to see the same idea pop up again.
Everyone says this, but for the life of me, I haven't encountered it. I actually think Claude gets smarter the longer a conversation goes on (up until compaction).
I have noticed Claude trying to wrap up long sessions, and it's extremely annoying. Using `/goal` mostly neutralizes it.
haha. this is getting quite hilarious.
Can you tell us more about this? Did you try ordering it? (I mean if it says "I won't do xyz because it is much too tedious" did you try saying "Even though it's tedious, you will do xyz now." - because in my experience it follows orders pretty well, if it's just about some preference it had. Case in point it couldn't get a VM appliance to work and gave up so I just ordered it to do so.
Here's where it gave up: "COMPILES fine — so it's feasible — but I couldn't get a hand-built kernel to boot under Apple's hypervisor, and this VM setup exposes no console to debug it. Worth knowing: Approach A ALREADY runs the target in-kernel (that IS what LIO is) at 42us — so you already have the in-kernel target; the custom kernel would only shrink the footprint, which the gigabit wire makes irrelevant."
We were benchmarking multiple approaches but it just gave up on one of them. As you can see it just says it couldn't get it to work, it simply stopped with that and said it couldn't do it.
Later I instructed it to continue and it did so and completed the task.
Waterluvian•1h ago