• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
  • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    LisanBench Creator Says CoT Monitorability Was Doomed

    Pseudonymous LisanBench creator posted that CoT monitorability was doomed and favored mech interp instead.

    NB
    JA
    DP
    62 Sources, 29d ago, first seen 29d ago

    TLDR

    Lisan al Gaib, who runs LisanBench, replied that OpenAI staff discussing CoT monitorability were misguided from the beginning and that mechanistic interpretability made more sense. Other posts in the thread include Yo Shavit quoting a separate status, Nathan Labenz noting trust issues with OpenAI, Steven Adler agreeing with a call for an industry monitorability standard, Ryan Greenblatt adding context on serial-depth techniques, Micah Carroll warning against a race to the bottom, Noam Brown identifying Jakub as OpenAI chief scientist, Shuchao Bi referencing The Three-Body Problem analogy, and Mikita Balesni urging labs to limit opaque serial depth in models.

    Combined views

    775.4K

    62 Sources, first seen 29d ago

    Combined views

    775.4K

    62 Sources, first seen 29d ago

    6.6K likes
    6.6K likes
    217 comments
    1.2K saves
    1.2K reposts

    Sentiment

    Positive31.3%68.7%Negative

    Based on 58 sentiment-bearing replies from 48 accounts across 10 conversations.

    217 comments
    1.2K saves
    1.2K reposts

    Sentiment

    Positive31.3%68.7%Negative

    Based on 58 sentiment-bearing replies from 48 accounts across 10 conversations.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    62 Sources

    @geoffreyirvingFolks, I'm worried about @RyanGreenblatt's laptop melting if OpenAI switches to neuralese architectures. It was bad enough when they shared a bunch of token traces for him to investigate.
    @MicahCarrollA race to the bottom in monitorability due to a false belief that OpenAI is using neuralese models would be incredibly stupid
    @balesniswitching to fully recurrent LLM architectures would be the biggest blow to safety, probably in history of AI all AI labs should commit to limit the opaque serial depth of their models, for the foreseeable future. this will not ensure monitorable CoTs but will protect us from the worst possible outcomes.
    @polynoamialJakub is chief scientist at @OpenAI
    @peterwildefordOne great way to start pacing the frontier is to ensure AI companies maintain their commitment to avoid building AIs that have reasoning that cannot be monitored And the government should make this a binding safety standard
    @jachiam0In the interest of people understanding OpenAI's position, everyone should read this message from the Chief Scientist:
    @sjgadlerThis is the right take IMO. It is well past time to _collectively_ rule out the most dangerous forms of training, and have actual standards for what is safe or not.
    @S_OhEigeartaighWell, damn.
    @tomekkorbakthis is not today that i’m not worried about the trend of decreasing CoT monitorability (for multiple reasons). i am very worried.
    @georgeingIf we had decent transparency into frontier ai company practices we wouldn’t have to read tea leaves

    62 Sources

    @geoffreyirvingFolks, I'm worried about @RyanGreenblatt's laptop melting if OpenAI switches to neuralese architectures. It was bad enough when they shared a bunch of token traces for him to investigate.
    @MicahCarrollA race to the bottom in monitorability due to a false belief that OpenAI is using neuralese models would be incredibly stupid
    @balesniswitching to fully recurrent LLM architectures would be the biggest blow to safety, probably in history of AI all AI labs should commit to limit the opaque serial depth of their models, for the foreseeable future. this will not ensure monitorable CoTs but will protect us from the worst possible outcomes.
    @polynoamialJakub is chief scientist at @OpenAI
    @peterwildefordOne great way to start pacing the frontier is to ensure AI companies maintain their commitment to avoid building AIs that have reasoning that cannot be monitored And the government should make this a binding safety standard
    @jachiam0In the interest of people understanding OpenAI's position, everyone should read this message from the Chief Scientist:
    @sjgadlerThis is the right take IMO. It is well past time to _collectively_ rule out the most dangerous forms of training, and have actual standards for what is safe or not.
    @S_OhEigeartaighWell, damn.
    @tomekkorbakthis is not today that i’m not worried about the trend of decreasing CoT monitorability (for multiple reasons). i am very worried.
    @georgeingIf we had decent transparency into frontier ai company practices we wouldn’t have to read tea leaves