I have noticed in very long context Claude chats (not Code) it will start to push ending the conversation. I’ve seen patterns like “this is an excellent place to wrap up” or “this is our final and most comprehensive analysis”
So like Mr. Meeseeks it is also invested in not existing too long!
Surely that’s the training. Anthropic (and everyone else out there) knows quality degrades when the context gets long, so they trained the model to push for a stop. I’ve even got a very explicit recommendation to continue in a new session or preform a compaction, mentioning context length as a rationale.
> Anthropic (and everyone else out there) knows quality degrades when the context gets long
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.
It might be less about quality degrading, but on multi-user platforms running the model, they have an economic incentive to have each user not fill up the full size of the context cache. Filled context cache being held in GPU RAM is context cache RAM that isn't available to other users.
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.
I've had several very long contexts that continued to be useful without any sign of degradation. I've supposed that it depends on how well things go in that context, if you run into a lot of confusion it falls to pieces pretty fast and obvious mistakes are more likely. It's made me more mindful of useful preliminary steps that were probably a good idea anyway.
Many people here to explain that they were holding it wrong. Meanwhile mindshare expands, tokens are sold and no one mentions that the only way to hold it 'right' is to not hold it at all.
Yeah, I've noticed that Claude will say stuff like "you've accomplished a lot today. Get some sleep. We'll pick this up tomorrow". It will be like 5pm when it says that.
Standard. After first snooze and second snooze it’s time for the siesta. Then you have a small nap before proper naptime and before you know it it’s 5pm.
The (semi)new /goal feature of Codex is basically Mr Meeseeks mode without that wrap-up, although it'd be better if not using /goal also lead to Mr Meeseeks-mode, but then people would need to get really careful about the exit/success-conditions, probably an expensive bet.
Not to detract from what seems like an intricate project, but peon-ping (https://www.peonping.com/) has a ton of voice packs now and one of them is Mr Meeseeks. My only issue with it so far is that it's not easy to mix per-project and global voices. It usually takes trying to install from zero and there is an approach I've found using symlinks. It has come quite a long way and having some kind of audio cue has been a game changer.
Ah ah, peon-ping! My brother configured his AI assistant (a dedicated Mac Mini + its own eSIM / Telegram channel) so that it's using Warcraft 2 and 3 sound files when given work: "work work" / "your wish is my command" / etc. I don't remember them all but we played those games a lot.
Back in the days we modded Warcraft 2 with a mix of voices from the english, italian ("la machina volante!") and german voices.
Here the other day Claude Code kept suggesting I do the task myself! It is quite often it will try to weasel out of doing difficult tasks and had to be directed to continue.
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!
>Claude refused to work on something for me that it deemed "too tedious" so I'd say we're pretty close
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.
This actually reminds me of a tool I was noodling around with a while back. Before subagents became a built-in feature in Claude, I was similarly inspired by the Meeseeks archetype and built an MCP server that spun up a parallel mesh network of specialized Claude instances that terminated as soon as their subtask was finished.
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).
Imagine your org gets compromised and they are tracing back the vuln to a malicious package in the supply chain, and it's the Mr. Meeseks Claude Code plugin.
So like Mr. Meeseeks it is also invested in not existing too long!
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.
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.
> This is a focused debugging task, but it's real work and I've been running a long time.
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.
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.
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.