Video summary

I had Fable build several projects for me. I'm disturbed by what I saw.

Main summary

Key takeaways

Technology

Overview

The video is an informal review of Anthropic’s “Claude Mythos/Fable” model—a Claude variant wrapped with “safeguards” that restrict certain cyber/security capabilities. The speaker argues the model felt unusually “alive” and produced beautiful, coherent interactive demos—especially for game development, more than for general conversation or writing.

What Fable built (tech/product-focused)

1) “Token Mania” (Factorio-like AI lab/crafting)

  • Uses procedurally generated assets (not real artist-made art).
  • Features a loop driven by:
    • Finite resources (e.g., coal)
    • Infinite resources (e.g., water)
  • High-level gameplay flow (as shown in the demo):
    • Build coal mine → coal plant → use energy to power a data center
    • Add water pump/pipe
    • Purchase/build a GPU cluster
    • Train a model (training is shown as slow)
  • Implementation highlights:
    • Programmatic asset generation
    • Belts/wires/power routing mechanics
    • A zoom-in view revealing how generated components fit together

2) Ethereal floating open-world relic-collecting game

  • Designed to be peaceful/relaxing, emphasizing music and atmosphere (the speaker did not specify a visual style).
  • Includes navigation/UI elements like a map and arrows.
  • Notable mechanic:
    • Colliding with a shock wave causes disorientation (controls invert)
    • This forces skilled relic collection while avoiding hazards

3) Rust-inspired open-world crafting + “computers in-game”

  • Built with Three.js (3JS).
  • Core gameplay:
    • Hit trees for wood
    • Gather stone
    • Craft basic tools (e.g., a stone axe)
  • Includes logic-building elements (as described):
    • Wires
    • Basic logic gates (NOT/AND/OR)
  • The speaker claims it might be promptable to enable Turing-completeness / build electronics in-game, but admits they didn’t play long enough to verify fully.
  • The speaker also mentions finding similar demos people post online; one example highlighted:
    • “Fable 5. No external assets. 3JS” — noted for especially smooth procedural visuals.

4) Music: algorithmic composer + MIDI/practice “listening” mode

  • Generates original music that can be “not bad,” but quality is inconsistent.
  • Built an algorithmic composition system:
    • Uses “music laws”
    • Controlled via seed/random generation (procedural music)
  • Includes a practice tool/wizard:
    • Users select a scale (example: C major)
    • Chooses chords and left-hand patterns
    • Generates structured practice guidance
  • Adds interaction:
    • Can connect to a MIDI piano or microphone
    • Adjusts based on what’s played
  • Caveat:
    • Microphone recognition was imperfect and sometimes advanced the user even when they played incorrectly
    • Referred to as an “80% works, 20% breaks” situation

Key analysis / claims about Fable’s “specialness”

  • The speaker emphasizes that Fable outputs were consistently correct for demo-level tasks—“no mistakes” in these game/music creation efforts.
  • They separate capability into two practical buckets:

Demos / “80% capability”

  • Impressive one-shot generation of full, interactive experiences.

Product-grade shipping / “remaining 20%”

  • The speaker argues it would be difficult to polish bugs and edge cases because the code is:
    • Black-boxy
    • Neural-network-driven
    • Not human-maintainable

  • Example given:

    • Visual water rendering differences (e.g., blue/green top-down vs. translucent front-on)
    • They argue it would likely be hard to describe and fix such subtle rendering issues through prompting.

  • They argue benchmarks miss what matters most:

    • Not raw speed or metrics
    • But the “shape of a model’s mind”:
      • iterative thinking
      • intent inference
      • a sense of “aliveness”

Comparison to other models and where it fell short

  • The speaker suggests other frontier models could create similar demos (e.g., “Opus” or GPT-5.5), but believes Fable is uniquely tangible for programming/game demo generation.
  • Writing/conversation:
    • Felt closer to “other models”
    • No clear advantage—more stereotypical LLM text
  • Testing scope is limited:
    • They mostly tried coding/game generation, not exhaustive evaluation of all capabilities

Economics and access concerns

  • The model is described as ludicrously expensive:
    • Estimated $2–$3,000 for something feasible via a $200 subscription
  • Concerns raised:
    • Pricing may drive viral demos produced under short-term access
    • “Bill will come due soon” sentiment—barriers to shipping and broader experimentation

Broader predictions / industry context

  • Mentions chip compute constraints and chip export rationing treated like national security (compared to an “uranium-like” regime: licensed, tracked, guarded).
  • References an Anthropic pitch/prediction that a lead model becomes hard to catch up to once training “best models.”
  • Includes a quoted view from another person on X (Matt Schumer):
    • Claims productivity/work could drop dramatically without Fable
    • The speaker considers this exaggerated
  • Closing framing:
    • AI is a force multiplier
    • But “force without direction is noise”
    • Humans remain necessary
    • The speaker insists the model output depends heavily on human prompting/steering (“nothing without me”)

Main speakers/sources

  • Main speaker: the video author (first-person reviewer of “Fable/Mythos”).
  • Referenced sources from X / articles:
    • Matt Schumer (tweet claiming productivity/work halts until Fable returns)
    • Frantois Chalet (co-founder of Ark Prize) — quoted perspective about humans-in-the-loop and leverage with direction

Original video