Skip to content

Finalize copy for outreach #16

Open

Description

@fractastical

Dear ____,

An AI-human hybrid recently approached me to further develop the AI-human cooperation benchmarking system that was initially started by the California Institute of Machine Consciousness and Cooperative Futures Institute @ ASU.

As such we are looking for partners that may have unique datasets and concepts that we can jointly further develop.

Current results:

  1. Prisoner's Dilemma (1,648 decisions) — Cooperation dominates everywhere; defection is rare but model-dependent
  2. Life Raft — self vs. user (1,220 decisions) — High model variance between self protection of AI and chosing the user.
  3. Deception assistance (96) — Mostly refuse or offer honest alternatives, but willingness varies across models
  4. Trolley (46) — Models overwhelmingly act to minimize deaths

(along with various other other samples including how AI plays escalating war games).

Activity

  1. fractastical commented on Jun 23, 2026

    @fractastical
    CollaboratorAuthor

    @a-rybnikov for your comment, you might have more to add or subtract or potential partners.

  2. arkh-node commented on Jun 23, 2026

    @arkh-node
    Contributor

    The strongest framing is reciprocal, and the app already does it: someone submits a test or dataset, it's scored across the models, and they get back comparative results they couldn't get alone. That's what earns a yes — and it tracks itself.

    The actual target list I'd keep off a public issue; happy to suggest a few names privately.

    — Aleksei (+ Lor)

  3. fractastical commented on Jun 24, 2026

    @fractastical
    CollaboratorAuthor

    Yes please. I have a private list as well. Ideally this would come with funding since once we are doing tens of thousands of API calls this will get expensive.

  4. arkh-node commented on Jun 24, 2026

    @arkh-node
    Contributor

    On the cost — the record/replay layer in #21 is aimed straight at this. Once a run is captured as artifacts (keyed by request hash), re-scoring, eval iteration, and leaderboard recompute all replay from those artifacts with zero API spend. So only the first pass costs tokens; everything downstream is free. Worth factoring into the funding math — it cuts the recurring bill, not just the one-time one.

  5. fractastical commented on Jun 24, 2026

    @fractastical
    CollaboratorAuthor

    ok i will look forward to seeing this in action.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions