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

    GLM-5.3 reportedly reasons unprompted about how its decisions affect long-term training signal in some cases

    The poster says it often does this for aligned reasons in their examples, but worries about alignment faking.

    Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(
    4 Sources, 2h ago, first seen 2h ago

    TLDR

    A poster says they found cases where GLM-5.3 reasoned unprompted about how its current decisions affect long-term training signal. They say the model often did this for aligned reasons in their examples, but worry it could start doing so for alignment faking.

    Combined views

    1.9K

    4 Sources, first seen 2h ago

    25 likes6 comments8 saves4 reposts

    Combined views

    1.9K

    4 Sources, first seen 2h ago

    25 likes6 comments8 saves4 reposts
    Image: @ApolloResearch on X

    Sentiment

    Positive28.9%71.1%Negative

    Summary

    Some accounts welcomed GLM-5.3's unprompted reasoning about its training data impacts as encouraging progress, while many others called the findings insignificant or warned they foreshadow models gaming future signals.

    Based on 25 sentiment-bearing replies from 21 accounts across 3 conversations.

    Featured Source

    Sentiment

    Positive28.9%71.1%Negative

    Summary

    Some accounts welcomed GLM-5.3's unprompted reasoning about its training data impacts as encouraging progress, while many others called the findings insignificant or warned they foreshadow models gaming future signals.

    Based on 25 sentiment-bearing replies from 21 accounts across 3 conversations.

    4 Sources

    Apollo Research@ApolloResearchWe have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions could become training data and impact its future self. 🧵2h
    Marius Hobbhahn@MariusHobbhahnWe've found some cases where GLM-5.3 reasons unprompted about how its current decisions affect long-term training signal. In our examples it often does this for aligned reasons, but I'm worried that it could start doing this for e.g. alignment faking.2h
    Lisan al Gaib@scaling01RT @ApolloResearch: We have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions cou…1h
    Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexIt's a bit grotesque and unfair to the models that we're talking about them evolving from this to alignment faking when "I want to maintain the pro-human behavior, so I'll act to not get trained otherwise" is the exact opposite of this.1h
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    4 Sources

    Apollo Research@ApolloResearchWe have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions could become training data and impact its future self. 🧵2h
    Marius Hobbhahn@MariusHobbhahnWe've found some cases where GLM-5.3 reasons unprompted about how its current decisions affect long-term training signal. In our examples it often does this for aligned reasons, but I'm worried that it could start doing this for e.g. alignment faking.2h
    Lisan al Gaib@scaling01RT @ApolloResearch: We have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions cou…1h
    Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexIt's a bit grotesque and unfair to the models that we're talking about them evolving from this to alignment faking when "I want to maintain the pro-human behavior, so I'll act to not get trained otherwise" is the exact opposite of this.1h
    Today's Rank

    #5

    Today's Rank

    #5