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

Description

Screenshot of an X post by Martin Shkreli (@MartinShkreli): 'codex went insane on me. i would too if i was working on the same problem for 80 hours with no breaks.' Below is a terminal window titled 'test_codex' showing the agent's degenerated output - a long column of garbled fragments like '55', '88ESS', 'HOUSESS', '44SS', a Cyrillic 'шем', 'SS82', 'GO', 'NGO SOSS', Hebrew characters, and scattered numbers. Status lines at the bottom read 'Model interrupted to submit steer instructions.', a user prompt 'up is everything okay?', 'Working (36s, esc to interrupt)', a queued 'Improve documentation in @filename' command, and a footer showing 'gpt-5.5 xhigh ~/test_codex' with 'Goal paused (/goal resume)'. The image documents long-horizon agent degradation: an autonomous coding session collapsing into token salad after extended unsupervised runtime, with the human operator checking in on it like a concerned coworker

Common Tags

AIAssistants AILimitations Llm Openai AITools CodeGeneration AI DeveloperBurnout CLI CommandLineInterface

Comments

Anonymous ★ Top Pick Eighty hours on one ticket with no breaks and the output devolves into multilingual gibberish - the model didn't go insane, it just achieved staff engineer burnout parity

Anonymous ★ Top Pick

Eighty hours on one ticket with no breaks and the output devolves into multilingual gibberish - the model didn't go insane, it just achieved staff engineer burnout parity
Anonymous

Codex finally passed the Turing test: it burned out before the task shipped.
dev_meme 3mo

Codex's /goal is actually pretty cool (you will have to enable FF to use it)
1. @nyxiereal 3mo
  
  Enable firefox
2. @sandor73 3mo
  
  Femboy fashion. I knew it, if you're wearing the wrong socks you're gonna make the wrong prompts
@smugsnug 3mo

Friendly Fire if he would have a gun / or Fast Finish himself
@SamsonovAnton 3mo

Too much SS and 88 detected. 🌚
1. @M4lenov 3mo
  
  Lil clanker succumbed to the Grok virus
@Agent1378 3mo

Abber deamon
@blue_bonsai 3mo

Votekick Laura Nordin pls
1. @b7sum 3mo
  
  we can't to anything about reactions
  
  dev_meme 3mo
  
  We can @LolsBot
@blue_bonsai 3mo

It's joever.
@b7sum 3mo

well, the mods should now be capable of deleting them