Video summary

George Hotz | Programming | Welcome to Gas Town and the future of Computer Use | Agentic AI | Part 1

Main summary

Key takeaways

Technology

Tech/Product Concepts Covered

  • Agentic AI / “computer use” shift: The speaker frames the moment as a transition away from manually controlling computers toward agents layered over the user’s computer, with new limitations to be discussed later.

  • AI agent tooling ecosystem

    • Maltbook: Described as a social network for AI agents, used to “read some great multis.”
    • Open Claw / “claw”: Presented as an agent runtime/tooling layer that actually executes actions (i.e., “the AI that does things”). The speaker notes the branding has changed multiple times.
    • Gas Town: Described as a new take on an IDE; the speaker tries it with “Gas Town with Open Claw instances” on a cloud box.
    • Beads: Required for setup; installed via quick scripts, but there were PATH/version issues initially (e.g., “Beads installed but not in path” / “Cannot verify beads version”).
    • Hyperland (tiling WM) and Tailscale: Used in the workflow; networking/VPN/Tailscale setup is mentioned as another source of friction.

  • Model/Reward/Training Analysis (high level)

    • Mentions RL with “verifiable rewards” as the long-term direction for these systems (attributed to a prior conversation with Andrej Karpathy).
    • References post-training pipeline concepts such as SFT and joint RL, and mentions vision-related training (“zero vision” / “vision stuff”).

  • Security/permissions posture

    • Mocks/condemns the modern state of permissions (“Computer security is over… yolo it”) and repeatedly uses “allow always” style permission settings while debugging.

  • Interface design claim

    • Argues that agents will handle communications across apps (WhatsApp/Telegram/Instagram/Twitter), filtering content/ads so the user doesn’t see them.
    • Then asks about alignment and ownership, concluding that the API surface is simple enough (tokens in/tokens out) that swapping models/providers will be manageable.

Product Setup + Troubleshooting (Practical Guide Elements)

Environment provisioning

  • Uses a DigitalOcean box and an OpenRouter account locked to limited credit.
  • Uses Open Claw and tries configuring providers/models via OpenRouter.

Gas Town installation path

  • Runs brew install gas town and then attempts language/toolchain setup for Go.
  • Debates “go” vs “gcc/go lang” style confusion; ultimately installs/sets up Go.

Open Claw + provider/model connectivity debugging

  • Repeated issues where the system shows no output / the assistant doesn’t respond in both TUI and web UI, suggesting websocket/streaming or provider integration issues.
  • Encounters rate limiting and “temporarily rate limited” errors for specific models/providers.
  • Finds that “Kimmy” models are problematic through OpenRouter:
    • Attempts multiple Kimmy variants.
    • Edits Open Claw JSON.
    • Restarts the gateway service.
    • Still reports intermittent breakage.
  • Works around model/provider failures by switching providers:
    • Tries routing through different providers (e.g., Fireworks, Together).
    • Finally sees a working configuration with a GLM 4.7 test after updating provider/model config and restarting the gateway.

Browser automation limitation

  • When attempting to control Twitch:
    • Notes Twitch is heavily JavaScript-driven.
    • Suggests browser control may not work until a control component (“browser control isn’t running yet”) is available.

Gas Town agent execution failure

  • After creating a “rig” and trying to start the “mayor”:
    • The process times out waiting for the mayor.
    • It reports that the Claw code runtime may be missing.
  • Additional friction: Anthropic “blocks” something in Hong Kong; the speaker starts using Tailscale as a workaround.

“Review/Analysis” Tone & Conclusions

  • Overall impression is very bullish on agentic software becoming “maximalist software,” but the live demo highlights:
    • Fragility in setup (“vibe coded shit” / things breaking unpredictably).
    • Model/provider unreliability (rate limits, intermittent provider failures).
    • Versioning/config/state problems (PATH issues, missing runtimes, gateway restart required).

Despite breakage, the speaker suggests the direction is inevitable: agents + computer-use tooling that hides complexity behind the agent layer.

Main Speakers / Sources

  • Primary speaker: George Hotz (in-stream monologue and demos).
  • Referenced source: Andrej Karpathy (mentioned via a past breakfast conversation about RL with verifiable rewards).

Original video