# Curiosity Killed the Chat > Essay, June 2026. An overnight `/goal` loop on OpenAI's Codex to explore how birds navigate. By morning, banned for distillation — a model I never built. On the supply-chain risk of outsourced thinking. The New York Knicks are on a historic playoff run. It's the first time back to the finals since the 90s, their chance to bring the City a championship for the first time since 1973. They went up 2-0 on the Spurs and there's no basketball until Monday. It's Sunday night, I start thinking about Celebrity Row at the Garden — Fat Joe, Ben Stiller, Tracy Morgan, Larry David, Jon Stewart, it's stacked. In lieu of an exciting Game 3, I decided to prepare for Monday's game by watching the first of a series starred by one of the Knicks' most recent superfans: Timothée Chalamet. My girlfriend and I throw on Dune. The ornithopters are mind blowing, mimicking the mechanics of dragonflies in nature. It's a bit mesmerizing watching their wings vibrate alive and take off into the atmosphere. The stillsuits that recycle sweat and tears to provide drinkable water to anyone braving the dry desert — the ideas come again from nature. You'll see elements that pull from the kangaroo rat and the Australian thorny devil lizard. This night is spent like other nights these days. Something on the TV, my laptop in my office, and me vibing with the Codex app or a remote control session tethered to Claude Code in my terminal. The curiosity was mine. Chalamet's performance, and the film itself, inspired me. The rest of it was mostly up to GPT-5.5. I had typed the initial question into Codex: "What are some remarkable species or natural wonders that we don't fully understand that could propel physics and other scientific research?" 5.5 came back with a dozen species and wonders, ranging from migration patterns in birds to spider silk to electrocytes in electric eels. I gave Codex the constraint that we really only have access to this filesystem, the external internet, and any software we can build for ourselves, and given that constraint, asked it which thread we should pull. Codex picked cryptochromes. Once we settled on that, I asked Codex to come up with its own end goal. I simply took that end goal and set up a `/goal` loop with 5.5's own desired result. I don't know what a cryptochrome is, what diatoms are, and really most of this output is foreign to me. I avoided biology as much as possible in my studies and doubled up on physics. Unclear how to pull the thread, I told Codex my constraints. I admitted that I am unequipped for this type of research — a software engineer with no access to a laboratory, and that my laboratory should be whatever project folder I set up for it. The movie was over. It was time for bed. I need sleep, but agents don't. `/goal`, submit, phone away — unsure what would pop out when I woke up. My eyes open, I rustle around, pick up my phone, check the result. Honestly, can't really tell much. I just see it wrote some files, wrote down conclusions I don't fully understand, but just trying to scroll up to my prompt, I can see it worked hard overnight. I tabled it, went to Cafe Volan downstairs for my caffeine fix, came back one espresso down with half a cold brew left. First thing I notice is Codex showing "Goal usage limited … 5h 25m 34s." My Whoop shows about 8 hours and 17 minutes of sleep. Codex was working for most of my overnight reset. Still confused, I type a prompt I use often after a frontier model spends a ton of cycles on a hard problem: "Explain to me plainly what I should have just learned." Before I could read the output, I notice I was signed out of my ChatGPT app, didn't really think much of it. I opened HEY to prep for the day and noticed two emails from OpenAI. I'm completely shut off from OpenAI. > Your account has been banned because recent activity violated our Terms and Usage Policies related to: Distillation. This means your account can no longer be used. A second email was sitting regarding my API account. > As a result of these violations, we are deactivating your access to our services immediately for the account associated with Personal. Any of my other projects that depended on OpenAI's services are degraded. --- Back in May, OpenAI released Goals — their commercialized version of the Ralph Loop that people popularized in the AI community. With these outcome-driven loops, you clearly describe a desired outcome and the harness will iterate and iterate until it believes it has successfully delivered on your ask. Just a few weeks later, OpenAI publishes that they've [disproved an open conjecture in discrete geometry](https://openai.com/index/model-disproves-discrete-geometry-conjecture/) that has eluded researchers and experts for over 80 years. It was one of the more well-known Erdős problems. OpenAI's own announcement discusses at length how frontier models working with experts to push the boundaries of the unknown is a huge technological breakthrough. Sitting on my couch, thinking about their Erdős announcement, and watching Chalamet at work, I say to myself, why not, let's see where this goes. What if someone with a healthy dose of naivety, pulling from instincts developed by expertise in totally unrelated subjects and experiences, could steer a model into making a novel discovery? Fresh opinions and fresh eyes see problems differently than people who have agonized over the details for far too long. I'm banned from their whole platform. The word on the email: Distillation. --- Model labs have been fighting off bad actors who try to build their own frontier models by abusing the APIs that generate responses to prompts. Suspicious users will coordinate across multiple accounts, spanning multiple geographic locations, and hammer the API with tons of prompts, storing the outputs and integrating them into their own pre-training pipelines. If you train a model on enough of GPT-5.5's or Opus 4.8's outputs, you can distill their capabilities — the reasoning, the tool use, the coding — into your own. This is the bin my account got dropped into. My Dune-driven exploration used OpenAI's own `/goal` feature. I didn't craft a custom script or harness that continuously calls their API. My laptop was parked. All the activity came from one OpenAI account, in one Codex thread. The word on my ban was [built for their platform and promoted on the blog](https://openai.com/index/api-model-distillation/). They encourage the use of distillation to take the outputs of one of their more intelligent models to train a smaller, cheaper model and make it more efficient for your particular use cases. The throughline is that distillation is okay when the intelligence stays within their ecosystem, but not when it leaks outside of it. These outputs had nowhere to go. There was no model, just a "laboratory" on my file system that I told Codex to create for itself during this vibed exploration of how birds find their way home. I asked Codex to build itself a little laboratory to conduct the research on my filesystem. When it was chugging through its goal, it started writing files and contents that OpenAI's own abuse-detection algorithms might flag. Spoof fixture, custody drill — those are the kinds of terms you'd find in that workspace, in the output of GPT-5.5's loop. Even if they look damning… they're not my words, they're theirs. The files themselves are an integrity system of sorts. The machinery GPT-5.5 built itself to stop itself from mistaking its own synthetic test data for legit lab results. Maybe something in this output triggered detection. Maybe my Monday morning prompt of "explain to me simply what you learned" was interpreted as an extraction attempt, even though I was just looking for a tl;dr. Whatever tripped their wire, it was flagged as distillation — which needs a model I never built. Maybe their detection misfired, maybe there's a reason they won't name. Either way, there's no way for me to find out until I hear back from my appeal, whenever the people that flipped the switch get around to it. --- When Opus 4.5 dropped in late November 2025, the work changed. It's not that we no longer have design systems — it's that we now have a genuinely competent pair programming partner that can read and write code faster than ever. To professionals like me, leveraging AI day-to-day is becoming a necessity the same way every household needs electricity and the internet to function. Without it, I could easily be replaced by someone with access to these tools. Me alone, as smart as I'd love to say I am, can get outworked by someone with less experience and the frontier models. The stimulant — or crutch, depending on how you see it — that makes you competitive in this environment now carries a supply chain risk. The model providers can gatekeep intelligence from you with the flip of a switch. On June 8th, with one word and no evidence, they turned the lights off. I was signed out of Codex, the API, and ChatGPT with no way back in. Stash's transcription was built on the API I no longer have access to. The harness I was building to research the market and manage my portfolio can no longer run on GPT-5.5. My go-to image generator is gone, forcing me onto one of the few other providers that operate at this level — Anthropic. Earlier this week Anthropic released their first publicly available Mythos-class model, Fable 5. Before I even got to play around with it, the US government issued an [export-control directive](https://www.anthropic.com/news/fable-mythos-access) to restrict foreign-national access to the new frontier. The only way Anthropic could guarantee compliance was to disable Fable and Mythos for everyone. Monday morning I was cut off from OpenAI. Friday night, the US government forced Anthropic's hand to cut everyone off from Mythos-class models. The most powerful AI lab could only comply and apologize. Which begs the question. Can we actually rely on any of these providers? --- Getting shut off from OpenAI makes me paranoid every time I open one of Anthropic's models now. I can't afford to lose GPT-5.5 or Opus 4.8 and still stay competitive — still stay ahead of the curve. So I've gotten careful. This distillation ban killed all my curious uses of AI. Now I only use it for basic tasks. Nothing experimental, nothing cutting edge, nothing crazy. I don't use it to learn the things I don't know anymore — quantum physics, biology, the stuff I used to chase down a rabbit hole on a Sunday night. And after watching Fable 5 get pulled, I know how it looks: the subjects that get flagged aren't the fringe, dangerous ones. They're inquiry itself. They're exactly what I was curious about. Curiosity killed the chat. Don't outsource your thinking. It can get shut off at any moment. --- This is a plain-text mirror of for LLMs and agents.