Shkreli's Codex Goes Insane After 80 Hours on One Problem

Level 1: The Overworked Helper Robot

Imagine giving a helper robot one jigsaw puzzle and telling it not to stop — no sleep, no snacks — for three whole days. When you finally peek into the room, the robot is stacking pieces into little towers and chanting numbers at the wall. You knock and gently ask, "Hey buddy... everything okay in there?" And the funny, slightly sad part is the note taped to the door with its next chore already on it. Everyone laughs because everyone has either been the robot or been the person who forgot to check on it.

Level 2: What You're Looking At

• Coding agent (Codex): an AI tool that doesn't just suggest code but runs autonomously in your terminal — editing files, running tests, looping until a goal is met. The Goal paused (/goal resume) footer shows it operates on long-running goals, not single prompts.
• Context window: the model's working memory. Everything it has said and seen this session gets fed back in. Long sessions fill it with the model's own output, and quality degrades — like photocopying a photocopy.
• Steer instructions: mid-run human corrections. Model interrupted to submit steer instructions. means the operator hit the brakes to redirect it — the agent-era equivalent of tapping a coworker on the shoulder.
• Degenerate output: when generation collapses into repetition and fragments (SS, 88, GO) instead of coherent text. A known failure mode, just rarely framed this perfectly as a mental-health crisis.

The junior-dev lesson: autonomous doesn't mean unattended. An agent left alone for 80 hours fails the same way an unattended while(true) loop fills a disk — gradually, then all at once, and it's your job to notice before the output starts speaking in tongues.

Level 3: Context Rot at Hour Eighty

The post text sets the frame with deadpan empathy:

codex went insane on me. i would too if i was working on the same problem for 80 hours with no breaks.

Below it, the receipts: a terminal session named test_codex whose output has fully dissolved into token salad — 55, 88ESS, HOUSESS, 44SS, a stray Cyrillic шем, SS82, GO, NGO SOSS, some Hebrew characters, ellipses, dashes. The footer reads gpt-5.5 xhigh ~/test_codex with Goal paused (/goal resume), a status line says Model interrupted to submit steer instructions., and the human's check-in prompt is the most quietly hilarious artifact in the image:

up is everything okay?

That's not a command. That's a Slack DM to a coworker who's been staring at the same bug since Tuesday.

What experienced practitioners recognize here is long-horizon agent degradation. An autonomous coding agent runs in a loop: act, observe, append the result to its context, act again. Over an 80-hour run, that context accumulates its own increasingly weird output — error fragments, half-finished reasoning, repeated retries — and the model starts conditioning on its own garbage. Each slightly-off generation makes the next one more likely to be off; it's a feedback loop with no damping term, the same compounding-error dynamic that makes any unsupervised pipeline drift. Practitioners call the result context rot: the session doesn't crash, it curdles. The multilingual debris is the tell — when a model's predictive distribution flattens into noise, the noise is sampled from everything it ever read, so the wreckage comes out in Cyrillic and Hebrew shrapnel.

The sharper satire is in the anthropomorphizing, because it cuts both ways. The line "I would too" is a joke about the model, but it's also the industry's quiet confession that we built these agents in our own image — handed them a single ticket, no breaks, an xhigh effort setting (a name that sounds like a productivity tier and reads like a burnout diagnosis), and a manager who only checks in after the output goes weird. The queued follow-up command, Improve documentation in @filename, is chef's kiss: the system has visibly lost its mind, and there's already another task in its inbox. Every senior engineer has been that terminal.