Kind of shocked nobody is calling out at the flag at the bottom of the announcement saying that it's not eligible for Zero Data Retention and it's currently pretty nebulous what the "Don't train on my conversations" toggle means, as the TOS classifies "conversations" as "user visible input and outputs" - says nothing about thinking, etc.
I suspect anything you make in these, the thinking traces or "safety evaluations" of the content allows whatever you build to become RL or eventual pre-training data.
This is probably a silly question, but is the internal monologue useful training data? It's where the work is done, but without the input or output, it seems hard to get any useful context.
In other words, what would the data be used to train for? It has to improve some kind of objective function. But it's a little hard to see what the objective function would be if it's just raw inner monologue.
Isn't the internal monologue exactly what Anthropic is always trying to hide to avoid distillation?
If it was useless, they wouldn't bother.
It's probably not as useful as if you also have the prompts, but even assuming they really don't train on the prompts (which we can never verify), you can probably get to the prompts based on the monologue. Some agents essentially repeat the prompt in the monologue.
Yep. It occurred to me a few days ago that the OAI TOS only says “Input” as what you provide “Output” as what you receive, Input and Output collectively as “Content.” and they won't "train" on your content. But they can retain content for safety evaluations and debugging.
What's neat about "safety evaluations" in LLM parlance is apparently encountering any novel information constitutes a "safety event" that can result in new reinforcement learning data... This anthropic 2022 paper that basically describes how it's a perpetual information siphoning machine with a cute little graphic https://arxiv.org/html/2212.08073 - they publish their "constitution". OpenAI has a "model spec" they somewhat regularly update that I suspect is their equivalent process. I suspect we've all been unwittingly advancing their models capabilities..
This Dec 2025 Google paper "A Practical Guide to Generating Synthetic Data With Differential Privacy" spells it out pretty clearly - the focus is on "privacy" - nothing about protecting the user's IP or unique knowledge/ insights.. https://arxiv.org/html/2512.03238v1
In OpenAI's case it seems to essentially generating synthetic training tuples from {prompt, chain-of-thought, answer} or scores on the chain of thought for reinforcement learning.
It would appear opting out of "Improve the model for everyone" didn't actually mean what we thought it meant.
The open models let you see there thinking and the models seem to be succinctly/ compactly representing the actual underlying concepts they're working on in some symbolic way or another. But they definitely represent the meat, bone, and marrow of the task
I mean I still regularly use 5.3 spark (the cerebrus model) that comes with my sub to do rapid reviews of 5.6's work and it finds oodles of problems in about a minute.
I'd actually suggest a great starting point would be a local command reviewer LLM. Could ostensibly be a modern AV type thing. Particularly seeing this lately has driven the need home deeper to me: https://x.com/chrisbanes/status/2085341561609425230?s=20
An open weight tool call auto-reviewer, has all sorts of achievable scaling curve milestones.
Just a note that I think the direction most people are paying attention to is memory bandwidth; thats the real bottleneck and “number go up” but also constraint people are designing around
Please for the love of god don't tell me the Codex sandbox is their actual eval harness sandbox?????
I maintain my own fork of Codex for "fun". Whenever I look at the sandboxing churn they're doing every release, as someone who used to work at Microsoft on Windows, my reaction is usually: https://c.tenor.com/vTzzhTiypwQAAAAC/tenor.gif
This is really cool! I’m becoming convinced the optimal UI to engage with agents, long term is going to be something spatial. No idea shape that even takes, though I really feel what you’ve made might be Xerox PARC days in terms of metaphor maturity, but there’s some real new seeds of “obvious in retrospect” ideas here. Thanks for conceiving of and building this!
Right? I also quit Claude Code and switch to Codex over that. Now I’m trying to figure out how I could make an extra $65,000 to never have to be concerned about this nonsense again. I know the economics of using open router etc…
But I’m reminded of ~2008 and the rise of “the cloud” as a marketing term that seemed to me to be a cover for dropping an expectation of rich clients, increasing a companies margins around subscriptions that would chip away at local ownership.
Then I got offput by the zealotry and absolutism around “true FoSS”, told myself I was young and moved on.
And really, a lot of subscription models I kind of can appreciate/ tolerate. Might be irksome but whatever, I get that software is expensive to make and it’s not fair in 2026 to value a yearly upgrade of Photoshop at $200. The capricious UI changes to things that’ve worked for 20 years and they take away say the classic color swatches altogether - silly and dumb.
I can use another professionally necessary tool I pay $200/ mo for, Codex, to whip up a classic swatch plugin.
Is that $200 a fair price for my token usage? I think an extremely heavy month I might’ve used a billion tokens?
But that right there is the problem. They have no idea what, specifically, profitability looks like and are going to be pulling endless levers for … I genuinely have no idea how long - at least through 2030/2032 if we tea leaves their debt obligations?
I don’t want to think about any of that. At all. I don’t want to spend time evaluating model preference and degradation and updating the nuances of how I “speak” to an AI because there’s some mystery backend experiment running on the output I use to produce functional outputs — ie the actual products I get paid to build/ maintain.
AI’s something between a tool and coworking companion, and the capricious “personality” changes due to playing with poorly understood and knobs and levers at the inference level - is maddening. To that end, I want a box in the corner I can point to and know exactly the quality of outputs that no one but myself modifies.
I suspect anything you make in these, the thinking traces or "safety evaluations" of the content allows whatever you build to become RL or eventual pre-training data.
reply