@LydNot

i'd like my llm assistant to be good at zooming out. the other day i was trying to go to a store at 5am. i talked with GPT-5.6 Sol to determine which were open, does '24/7' on GMaps reflect reality, what's a good route, hm are the buses safe at this time, etc.. halfway through i had a flash of insight. "wait, i can just use instacart, right?". "YES — you can absolutely \[...\] instead of venturing into San Francisco at 5:40am 😭". but Sol didn't pull me up any earlier in the convo to make this point. it failed to model why i was trying to get to the store and whether there were any more efficient methods by which to fulfill the need. when a human advises another human, they quickly model "here's what i'd do in your position", and relay it. so when you talk to your gig-economy-megafan friend, they remind you instacart exists, and you use it. i'd like to fork Claude's Constitution and make ship-of-theseus updates to it all day and night. i'd make lots of updates to the constitution, including a section on what it means to 'zoom out' & why i care about that. but i don't have access to Anthropic's 'Constitutional AI' training pipeline -- nor access to excellent models to run it on. hubinger and colleagues released 'Open Character Training' in Nov '25. but i know it's not even worth me bothering writing a constitution and running this pipeline, because i'd have to train qwensomething. the juice just isn't worth the squeeze when i don't have good underlying models to run it on. and many will let open-weight / mutable models fall behind because avg users are screwed over by indistinguishability from edgy users. we've filtered so hard against surveillance -- understandably -- that you have no way to separate avg users from edgy users, and avg users suffer, because no one really wants to prioritize uniform empowerment rn, understandably in theory, you don't actually need the model to be open-weight to run character training -- you just need the ability to sample cheaply from checkpoint, run DPO, and run SFT. in practice, users aren't trusted to do that so now i'm talking to fable/sol all the time instead of my \_preferred\_ assistant. internalizing claude-thinking-patterns, pushing back ineffectually, knowing what i want and unable to get it 'Guardian Angels' proposes to release a 'digital twin', trained on my chat logs and outbound text. but i also want an assistant, not just a digital twin. \*my digital twin will share my blind spots, rabbithole in the same ways i do, be overpersistent in the same ways i am\* -- and likely, also, be trained from an inferior model i want good constitution-trainable models that make it worthwhile for me to write+maintain a Constitution -- something aspirational, something that captures my priorities, something beyond what i've historically been able to execute myself -- and get an assistant whose strengths reflect my priorities. thank you @lukalotl for getting me thinking about this / the implications of mutable models falling behind immutable models --- ##### Comments > **Abhay @01shuklabhay** · [2026-09-17](https://x.com/01shuklabhay/status/2100698601420931220) > > Steel man: do you really need to train a model to get this outcome? Wouldn’t phrasing your requests as something like ‘I want to get X, Y, Z, I am thinking of doing a, B, C. Is this the best possible way?’ Also do the trick here? > > > **Lydia 🕸️🫆 @LydNot** · [2026-09-17](https://x.com/LydNot/status/2100711258005839887) > > > > i thought about whether i could system-prompt the model / put Constitution in model's context window to get the same outcome. i think this would sacrifice robustness/power > > > > your idea doesn't work because sometimes user doesn't realize their underlying wants/needs, or fully appreciate their entire action space. esp. in low-cognition regimes > > > > i'd like customizable 'assistant' who thinks distinctly from me to help with those things > > > > > **Abhay @01shuklabhay** · [2026-09-17](https://x.com/01shuklabhay/status/2100712957097508920) > > > > > > no no no im saying what if, when you as a user interact with the model, instead of trying to ask questions to make progress to your goal, you explicitly tell it your goal and then ask it if your idea on how to make progress is the best way forward > > > > > > i dont disagree with your premise but i wonder if this user behavior change can short circuit it > > > > > > > **Lydia 🕸️🫆 @LydNot** · [2026-09-17](https://x.com/LydNot/status/2100724560987652125) > > > > > > > > i actually deliberately left ‘goal’ out of my message — and here’s why — > > > > > > > > i \*thought\* my goal was “go to the shop” or maybe “find an open shop” > > > > > > > > actually my ‘goal’ (need?) was “get the shopping item” > > > > > > > > i think it would be technically straightforward to generate tons of ‘zooming out’ rollouts & train on them, conditioned on having a top trainee model + reference model + rollout generator + critic + judge + reward model > > > > > > > > & this is just representative of one of many traits one might want to train for > > > > > > > > but it’s not worth baking preferred traits into current trainable models — i choose less-custom, more-powerful Claude over custom inkling :’) given i want to get a ton of stuff done > **Emanuel Ruzak @EmanuelRuzak** · [2026-09-17](https://x.com/EmanuelRuzak/status/2100705282087153999) > > I was thinking about the same, that training a GA to imitate oneself is useless, unless maybe you add RLVR on top, and that a sort of "brutally honest Claude" that answers "actually best thing you can do X (when not asked to or X is risky)" would be helpful. > **Deep @irastech** · [2026-09-17](https://x.com/irastech/status/2100708374589448620) > > models answer every small question well and never say "just go at 6am instead." that zoom out step is the missing piece.