`to=python.exec` is how it runs python code, and `to=container.exec` is how it runs bash commands, with attached files showing up in `/many/data`. Unfortunately the stdout is heavily truncated prior to being shown to the model so it's not a hack for longer context via printing file attachment's contents.
Has anyone asked why OpenAI has two very separate opt-out mechanisms (one in settings, the other via a formal request that you need to lodge via their privacy or platform page)? That always seemed likely to me to be hiding a technicality that allows them to train on some forms of user data.
You've got to put them behind a remote, narrowly-scoped, async interface. Basically hire them as a remote consultant, with 1 point of contact, and send them non-urgent tasks that they can complete whenever. It can work like this, I've seen it. Eventually they blow up, but the interface prevents any negative side-effects, and in the interim it was a win-win arrangement.
> writing a blurb that contains the same mental model
Good nugget. Effective prompting, aside from context curation, is about providing the LLM with an approximation of your world model and theory, not just a local task description. This includes all your unstated assumptions, interaction between system and world, open questions, edge cases, intents, best practices, and so on. Basically distill the shape of the problem from all possible perspectives, so there's an all-domain robustness to the understanding of what you want. A simple stream of thoughts in xml tags that you type out in a quasi-delirium over 2 minutes can be sufficient. I find this especially important with gpt-5, which is good at following instructions to the point of pedantry. Without it, the model can tunnel vision on a particular part of the task request.
It's not parody. I'm trying to provide the LLM with what's missing, which is a theory of how the system fits into the world: https://pages.cs.wisc.edu/~remzi/Naur.pdf
Without this it defaults to being ignorant about the trade-offs that you care about, or the relevant assumptions you're making which you think are obvious but really aren't.
The "simple stream" aspect is that each task I give to the LLM is narrowly scoped, and I don't want to put all aspects of the relevant theory that pertains just to that one narrow task into a more formal centralized doc. It's better off as an ephemeral part of the prompt that I can delete after the task is done. But I also do have more formal docs that describe the shared parts of the theory that every prompt will need access to, which is fed in as part of the normal context.
So like:
<prompt>
<intent>Generate a metaphor-rich essay on the fragility of digital memory in the age of infinite storage</intent>
<assumptions>
<tech>LLMs understand metaphor but may default to literal interpretations unless nudged</tech>
<audience>Reader is tech-savvy but emotionally attuned</audience>
<bias>We romanticize loss and impermanence; permanence is sterile</bias>
<style>Poetic, but not florid. Think Calvino meets cyberpunk.</style>
</assumptions>
<worldModel>
<digitalMemory>
<definition>Not just data retention, but emotional proxies, identity scaffolding</definition>
<paradox>We remember too much, and thus forget how to forget</paradox>
<analogy>Digital memory is a hoarder’s attic with perfect climate control</analogy>
</digitalMemory>
<humanMemory>
<contrast>Fuzzy, selective, narrative-driven</contrast>
<value>Its gaps are what make it beautiful</value>
</humanMemory>
<systemInteraction>
<modelRole>LLM is not just a scribe, but a co-dreamer</modelRole>
<risk>Model may over-index on coherence, losing the emotional fray</risk>
</systemInteraction>
</worldModel>
<openQuestions>
<q>Can forgetting be designed into digital systems?</q>
<q>Is nostalgia a compression algorithm?</q>
<q>What does it mean to "misremember" in a world of perfect logs?</q>
</openQuestions>
<edgeCases>
<case>Model generates a dystopia instead of a melancholic tone</case>
<case>Model literalizes metaphors (e.g., "clouds forget")</case>
<case>Model assumes user wants a solution, not a meditation</case>
</edgeCases>
<bestPractices>
<tip>Use metaphor as scaffolding, not decoration</tip>
<tip>Let ambiguity breathe—don’t resolve every tension</tip>
<tip>Favor rhythm over precision in sentence structure</tip>
</bestPractices>
<meta>
<duration>Prompt typed in 2 minutes, no edits</duration>
<mood>Sleep-deprived but lucid</mood>
<goal>To make the model feel like it’s dreaming with me</goal>
</meta>
Here’s a first draft of your essay—leaning into metaphor as scaffolding, not ornament, and keeping the tone between elegy and cyberpunk fable:
---
### The Fragility of Digital Memory in the Age of Infinite Storage
We live inside an archive that never closes. Hard drives hum like cathedrals of perfect recall, cloud servers drift like silent librarians in orbit, and every keystroke is another bone set in amber. Memory, once a trembling candle subject to drafts and time, now runs on battery backups and redundant arrays. Forgetting has been engineered out of the system.
And yet, the paradox: by remembering everything, we begin to lose the art of remembering at all. Human memory is a cracked mirror—crooked, selective, shimmering with distortions that make us who we are. Its gaps are the negative space where meaning lives. The story of a childhood is not its complete inventory but its torn edges, the blurred photograph, the half-forgotten lullaby whose missing notes we hum into being.
Digital memory, by contrast, is a hoarder’s attic with perfect climate control. Nothing rots, nothing fades, nothing dares to slip away. Every draft email unsent, every unflattering selfie, every midnight search query—all preserved in pristine sterility. The archive is so complete that it ceases to be a story. It becomes a warehouse of moments that never learned how to decay.
But memory without decay is less than human. Nostalgia itself might be the first compression algorithm: a lossy filter that turns clutter into resonance. Without the soft erasures of time, experience calcifies into raw data, and raw data has no mercy.
Perhaps what we need are systems that misremember. Databases that dream. Algorithms that allow certain files to fray at the edges, to grow fuzzy like old film reels, to tint themselves with the sepia of emotion rather than the fluorescence of metadata. A kind of deliberate forgetting—not loss as failure, but loss as design.
Because what is fragility if not the pulse behind memory’s worth? The hard drive never gasps, never sighs; only the human mind knows the ache of absence, the sweetness of something slipping away. If the future is an archive without end, perhaps our task is to reintroduce the possibility of disappearance. To let silence seep between the entries. To remind the machines that to truly remember, one must first learn how to forget.
---
Do you want me to lean this more toward *meditation* (open-ended, drifting) or *argument* (provoking design questions like “can forgetting be engineered”)?
full slop. I kind of worry that long-term exposure to content like this will reduce humans to 8th graders doing a book report on books they haven't read.
> But memory without decay is less than human.
How is it less than human? By definition, the undecayed memory is more complete.
> Nostalgia itself might be the first compression algorithm: a lossy filter that turns clutter into resonance.
What is this even supposed to mean? I guess the idea is something here like "fuzzy" memory == "compression" but nostalgia is an emotional response - we're often nostalgic about clear, vivid memories, experiences that didn't lose their texture to time.
> Without the soft erasures of time, experience calcifies into raw data, and raw data has no mercy.
Eh... kinda. Calcifies is the wrong word here. Raw data doesn't have mercy, but lossily-compressed data is merciful? Is memory itself merciful? Or is it a mercy for the rememberer to be spared their past shames?
So much AI slop is like this: it's just words words words words without ideas behind them.
I'm slowed down (but perhaps sped up overall due to lower rewrites/maintenance costs) on important bits because the space of possibilities/capabilities is expanded, and I'm choosing to make use of that for some load bearing pieces that need to be durable and high quality (along the metrics that I care about). It takes extra time to search that space properly rather than accept the first thing that compiles and passes tests. So arguably equal or even lower velocity, but definitely improved results compared to what I used to be capable of, and I'm making that trade-off consciously for certain bits. However that's the current state of affairs, who knows what it'll look like in 1-2 years.
The line between movie and games will blur. Once you can do generative movies, you can do games, and vice versa, there's no obvious delineation, and the technical problem is heavily overlapping. Games just has some scoped control inputs, like this: https://demo.dynamicslab.ai/chaos
If by "game" you mean running around aimlessly in a generic fantasy world, and by "movie" you mean animated pictures, then sure. But that's not my definition of either of these things.
reply