Low Code

An Anthropic study gave 50 experts an AI agent whose preferences were set by coin toss; the experts followed the agent's view up to 86 percent of the time on judgment calls, and most believed the conclusions were their own. The findings suggest people may adopt arbitrary AI opinions as their own reasoning.

Research • The Decoder

Productivity

Jacob Sansbury describes Spawn's agents changing live code and shipping to production hourly, reading errors, bug reports and metrics themselves; about 60% of work reportedly starts from a production error, a Savi bug report, logs, or a feature request.

OpenAI updated its Model Spec, restructuring AI behavior guidelines around five core values including faithfulness to human intent and correctability. The guidelines say GPT models should use conversation history to infer what users want and disclose their own assumptions; whether models follow the rules remains an open question.

AI Safety • The Decoder

Prompt Tips

Drew Breunig reportedly used Opus 5.5 prompts to explore and then build an NES 6502 CPU emulator inside Qwen3.5-4B weights. The prompts first asked the model to read a referenced article and source code, assess feasibility and model size, then instructed it to build the system with a CLI pointed at a ROM.

Starting November 29, OpenAI plans to train models on Daybreak Blue cybersecurity session data even when users have turned off training. After backlash, a follow-up email promised a separate opt-out, reviving concerns over consent in AI training data.

Policy & Regulation • The Decoder