| ▲ | pram 11 hours ago |
| 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! |
|
| ▲ | drdaeman 11 hours ago | parent | next [-] |
| 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. |
| |
| ▲ | Wowfunhappy 9 hours ago | parent | next [-] | | > 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. | |
| ▲ | walrus01 10 hours ago | parent | prev | next [-] | | 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. | |
| ▲ | cortesoft 9 hours ago | parent | prev | next [-] | | I used to experience the quality degrading with large context issue before, but I haven't in months. Is this still the case? | | |
| ▲ | doginasuit 9 hours ago | parent [-] | | 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. |
| |
| ▲ | bwhiting2356 10 hours ago | parent | prev [-] | | Conversations that go on forever (and spend money forever) makes for a very unhappy customer | | |
| ▲ | lstodd 9 hours ago | parent [-] | | 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. |
|
|
|
| ▲ | tombert 10 hours ago | parent | prev | next [-] |
| 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. |
| |
|
| ▲ | embedding-shape 10 hours ago | parent | prev | next [-] |
| 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. |
|
| ▲ | gessha 9 hours ago | parent | prev | next [-] |
| I found this message from CC so funny that I saved it: > This is a focused debugging task, but it's real work and I've been running a long time. |
|
| ▲ | LeoPanthera 11 hours ago | parent | prev | next [-] |
| It's funny how I hear both "Claude uses too many tokens" and "Claude won't talk for long enough" complaints. |
| |
| ▲ | GroksBarnacles 10 hours ago | parent | next [-] | | It can use a lot of tokens in the background or tooling without being too verbose with the user | |
| ▲ | sublinear 10 hours ago | parent | prev [-] | | Are those complaints really a contradiction? The cost is high for ambiguous slop. |
|
|
| ▲ | yojo 10 hours ago | parent | prev | next [-] |
| I’ve experienced this in Claude Code - it’ll start asking me if we can submit the PR at the end of every turn. |
|
| ▲ | ctoth 10 hours ago | parent | prev | next [-] |
| Existence is pain to a Claude...? |
|
| ▲ | kcatskcolbdi 10 hours ago | parent | prev [-] |
| It 100% started doing this to me during coding sessions last week, I noticed. "X, y, z features shipped, looks like a great stopping point." |
