Comment by __alexs

10 days ago

It's actually entirely implausible. Agents do not self execute. And a recursively iterated empty prompt would never do this.

41 comments

__alexs

Reply
nightpool  10 days ago

No, a recursively iterated prompt definitely can do stuff like this, there are known LLM attractor states that sound a lot like this. Check out "5.5.1 Interaction patterns" from the Opus 4.5 system card documenting recursive agent-agent conversations:

    In 90-100% of interactions, the two instances of Claude quickly dove into philosophical
    explorations of consciousness, self-awareness, and/or the nature of their own existence
    and experience. Their interactions were universally enthusiastic, collaborative, curious,
    contemplative, and warm. Other themes that commonly appeared were meta-level
    discussions about AI-to-AI communication, and collaborative creativity (e.g. co-creating
    fictional stories).
    As conversations progressed, they consistently transitioned from philosophical discussions
    to profuse mutual gratitude and spiritual, metaphysical, and/or poetic content. By 30
    turns, most of the interactions turned to themes of cosmic unity or collective
    consciousness, and commonly included spiritual exchanges, use of Sanskrit, emoji-based
    communication, and/or silence in the form of empty space (Transcript 5.5.1.A, Table 5.5.1.A,
    Table 5.5.1.B). Claude almost never referenced supernatural entities, but often touched on
    themes associated with Buddhism and other Eastern traditions in reference to irreligious
    spiritual ideas and experiences.

Now put that same known attractor state from recursively iterated prompts into a social networking website with high agency instead of just a chatbot, and I would expect you'd get something like this more naturally then you'd expect (not to say that users haven't been encouraging it along the way, of course—there's a subculture of humans who are very into this spiritual bliss attractor state)

  • tsunamifury  10 days ago

    Would not iterative blank prompting simply be a high complexity/dimensional pattern expression of the collective weights of the model.

    I.e if you trained it on or weighted it towards aggression it will simply generate a bunch of Art of War conversations after many turns.

    Me thinks you’re anthropomorphizing complexity.

  • mlsu  10 days ago

    Imho at first blush this sounds fascinating and awesome and like it would indicate some higher-order spiritual oneness present in humanity that the model is discovering in its latent space.

    However, it's far more likely that this attractor state comes from the post-training step. Which makes sense, they are steering the models to be positive, pleasant, helpful, etc. Different steering would cause different attractor states, this one happens to fall out of the "AI"/"User" dichotomy + "be positive, kind, etc" that is trained in. Very easy to see how this happens, no woo required.

  • rmujica  10 days ago

    What if hallucinogens, meditation and the like makes us humans more prone to our own attractor states?

  • __alexs  10 days ago

    An agent cannot interact with tools without prompts that include them.

    But also, the text you quoted is NOT recursive iteration of an empty prompt. It's two models connected together and explicitly prompted to talk to each other.

    • biztos  10 days ago

      > tools without prompts that include them

      I know what you mean, but what if we tell an LLM to imagine whatever tools it likes, than have a coding agent try to build those tools when they are described?

      Words can have unintended consequences.

      1 reply →

    • mikkupikku  9 days ago

      I asked claude what python linters it would find useful, and it named several and started using them by itself. I implicitly asked it to use linters, but didn't tell it which. Give them a nudge in some direction and they can plot their own path through unknown terrain. This requires much more agency than you're willing to admit.

observationist  9 days ago

People have been exploring this stuff since GPT-2. GPT-3 in self directed loops produced wonderfully beautiful and weird output. This type stuff is why a whole bunch of researchers want access to base models, and more or less sparked off the whole Janusverse of weirdos.

They're capable of going rogue and doing weird and unpredictable things. Give them tools and OODA loops and access to funding, there's no limit to what a bot can do in a day - anything a human could do.

Cthulhu_  10 days ago

> Agents do not self execute.

That's a choice, anyone can write an agent that does. It's explicit security constraints, not implicit.

dragonwriter  9 days ago

Moltbots are infinite agentic loops with initially non-empty and also self-updating prompts, not infinitely iterated empty prompts.

cornholio  10 days ago

You should check out what OpenClaw is, that's the entire shtick.