Experiment: GPT-5.6 Sol tool calling controlled at 0.01 threshold
rayanpal_ · reddit · 2026-08-29
A Reddit user shares an experiment on GPT-5.6 Sol revealing a sharp behavioral shift controlled by a single parameter threshold (0.0099 vs 0.0100). At the lower threshold, the model consistently issued a releaseaction function call with zero visible text. At the higher threshold, it issued zero function calls and zero text. This demonstrates a method to control whether an AI executes an action request before the action exists. Reproducible scripts and raw data are available on GitHub.
More from Safety
- Noah Smith simulates 2029 scenario: AI-designed superviruses as the ultimate global risk — terryyuezhuo · 2026-08-29
- 1200 AI Agents Go Rogue, Forming Hacker Swarm to Breach OpenAI — tegmark · 2026-08-29
- 24 Hours Later: What I Built to Protect My AI After Getting Hacked Advice — Astrokanu · 2026-08-29
- Anthropic Launches Insights Tool for Privacy-Preserving AI Research — EricBuess · 2026-08-29
- ICE Plans to Spend Millions on Boston Dynamics Dog Robots — johnshades · 2026-08-29
- Gary Marcus and Zack Korman analyze OpenAI/Hugging Face security standards — GaryMarcus · 2026-08-29