Sam Altman Shares Warning on Chain-of-Thought Monitoring and AI Model Depth

Key Info

Sam Altman shared a post warning against a "race into unmonitorability" driven by confused reporting, noting that the computation-graph depth of current frontier models, including Astra, is within a factor of two of GPT-4. He reiterated OpenAI's commitment to chain-of-thought monitoring since its earliest reasoning models, while acknowledging the technique remains fragile.

Highlights

• Frontier models are not arbitrarily deeper than GPT-4 — depth is within 2x, per the shared post.
• OpenAI has used chain-of-thought monitoring since its first reasoning models and sees it as a window into how alignment generalizes.
• Altman called the technique fragile, emphasizing the need to avoid misleading narratives that push AI into unmonitorable territory.