Skip to content

README.md

Latest commit

 

History

History
60 lines (45 loc) · 2.34 KB

README.md

File metadata and controls

60 lines (45 loc) · 2.34 KB
Raw

AI Engineer World's Fair 2026 talk demo

This directory contains the maintained live probe from the talk “RLM: Recursive Language Models for Large Codebases”, delivered by Shashi Jagtap at AI Engineer World's Fair 2026.

The original presentation repository remains available for the complete React slide deck and the exact event-day snapshot:

The probe here is intentionally maintained against the current RLM Code APIs. It replaces the talk repository's copied source snapshot and hard-coded evidence builders with RepositoryContextBuilder, while retaining the demonstration's observable contract:

  1. repository evidence is placed in a Python context variable;
  2. the root model receives metadata rather than source contents;
  3. model-written code executes in a Docker sandbox with networking disabled;
  4. the code makes exactly one focused llm_query subcall; and
  5. the same code block completes through FINAL.

Local Ollama

Start Ollama and choose a model you have installed:

RLM_TALK_PROVIDER=ollama \
RLM_TALK_MODEL=qwen3.6:35b-mlx \
uv run --frozen python examples/aie_world_fair_2026/rlm_probe.py

For another local model, change RLM_TALK_MODEL accordingly.

Gemini

export GEMINI_API_KEY="your-key"
RLM_TALK_PROVIDER=gemini \
RLM_TALK_MODEL=gemini-2.5-flash \
uv run --frozen python examples/aie_world_fair_2026/rlm_probe.py

Options

Variable Default Purpose
RLM_TALK_REPO Current checkout Repository to investigate
RLM_TALK_CONTEXT evidence mini, evidence, or full context profile
RLM_TALK_TIMEOUT 120 REPL execution timeout in seconds
RLM_TALK_BASE_URL Provider default Optional OpenAI-compatible endpoint
RLM_TALK_TASK Mechanics validation Override the root task
RLM_TALK_NO_DOCKER unset Set to 1 only for trusted local smoke tests

The in-process option executes model-written Python in the host process. Use it only with a trusted model and repository; Docker is the recommended path.

For the API-key-free harness generalization proof, see ../harness_generalization.