Open Source, Faster Jev is HERE

Technological concept: “System 1” classification models (ultra-fast decisioning)

• The video introduces Leia as a local, open(-ish) “system one” model designed for classification/structured decisions rather than free-form text like typical LLM chat.
• Key performance claims:
  • ~30 decisions per second in a local demo (Snake)
  • Sub-33ms decision latency (e.g., “less than 33 milliseconds” / 39 milliseconds in another timing)
  • After idling, the Snake run length grew past 140 while still not failing

Core product idea: structured outputs for “types” and probabilities

Unlike LLM chat, the model is used by providing:

• States: variables describing the input context
• Questions: formatted as JSON
• Criteria: custom rules/definitions for what counts as true/false

The expected output is structured, typically including:

• Yes/No (or a choice selection)
• Confidence scores / probabilities
• Additional metadata such as request IDs and evaluation time (as shown in the demo examples)

Comparisons and ecosystem impact (Jeff vs Leia)

• Jeff is presented as a fast, hosted, closed-source model from Tape Safe AI (also described as system-one classification).
• Leia is presented as an open-source competitor that can run inside the user’s own computer, enabling:
  • more local/free usage
  • potentially faster or similarly fast interaction than hosted Jeff (with possible accuracy tradeoffs)
• The presenter claims Leia may be less precise/accurate than Jeff, but is “a great start,” especially because it appeared very soon after Jeff’s announcement.

Community use cases showcased (real-time + browser automation)

1. Games controlled by classification decisions

• Demo: AI plays Snake locally using Leia
• Mentioned: people made Jeff play Super Mario Bros. in real time, with rapid action decisions (e.g., “jump/left/move/pause”)

2. Lead scoring + personalization workflows

• Example: with ~700 leads, Jeff can classify:
  • which leads are most likely to respond
  • which message template is best per person
  • how to outreach effectively

3. Browser automation / agentic workflows

• Example: booking flights using Browser-use + Jeff
  • Claim: flight booking took about 7 seconds and cost roughly $0.0039 (hosted API pricing)
• A Leia-based browser agent demo:
  • Uses Leia to control actions in a Chrome session (via browser automation)
  • Demonstrates:
    • manual mode (“choose next” etc.)
    • run automatically mode (very fast automated clicking/navigation)
  • Example constraints:
    • route (Zurich → London)
    • date (Oct 20, 2026)
    • economy, 1 adult
    • stop when matching flight options are visible
    • do not book/select automatically

4. Another demo: “finite choices” for browser agents

• Uses a fixture “reading room” style site
• The model infers the page’s meaning and follows instructions about finite choices to control browser agents

Open-weight / open model alternatives

• Mentions an open competitor “Kev”:
  • a small family of decision models built on Qwen 3.5
  • examples of model sizes like 9B / 4B / 0.8B
  • uses open weights and can be downloaded and run locally
  • outputs structured JSON with probabilities/scores

Tutorial / reproducibility element: integrating an open model into an existing project

• The presenter describes using Cloud code to take an existing open-source browser-automation project (from Gregor / browser-use demo, described as open source) and integrate Leia.
• They provide an example prompt to a coding assistant to:
  • import the source
  • integrate Leia
  • allow switching between Jeff (API) and Leia (local)
• Result: a working demo “completely independent” of Jeff’s API and without paying (per the presenter).

Key API / workflow details (Jeff / Tape Safe AI example)

Jeff request structure (described)

• Inputs: states + questions
• questions are JSON objects with:
  • a type (e.g., true/false classification, choice selection, etc.)
  • an instruction like “Is sandwich?”
  • a criterion defining the true/false logic

Example classifications

• “Is food a sandwich?”
  • state food = hamburger → ~94% true
  • state food = avocado toast → ~94% false

Timing comparison

• Jeff example mentioned ~93ms evaluation time
• Leia example mentioned ~39ms, described as about 3× faster

Speaker / source list (main sources)

• Tape Safe AI (Jeff model; newly available; also framed as “system one”)
• “Gregor” / browser-use demo author (source of the demo project adapted for Leia)
• Presenter (primary narrator demonstrating Snake + browser demos; unnamed in subtitles)
• Browser-use (automation framework referenced)
• “cloud code” (tool used to integrate Leia)
• Daniel Kahneman (referenced conceptually via Thinking, Fast and Slow for the system-one framing)