How Worrisome is GPT-6’s “Stealth Thinking”? | Tech Decoded

September 10, 2026

AI Summary

5 min read

How Worrisome is GPT-6's "Stealth Thinking"?

When OpenAI released GPT-6 Astra last week, it came with the usual slick video and incomprehensible benchmark charts. But this time, something else came with it: a controversy that had computer security researchers using phrases like "extremely concerning" and "violating one of the few red lines that exist in the AI industry." The trigger was a report from The Information claiming Astra uses new techniques that make its reasoning harder for humans to monitor. OpenAI's chief scientist called the reporting "confused" but didn't explain how. So what's actually going on?

What the "stealth thinking" technique probably is

To understand the controversy, you need to know how large language models work. A standard LLM takes text input, converts it to numerical embeddings, passes those numbers through a series of transformer blocks (each doing its own analysis and passing results forward), and finally decodes the result into an output token. Crucially, this is a forward-only process—the numbers move through the blocks in sequence, never looping back.

This creates a problem called limited depth. If you ask an LLM a chess question, it can't simulate multiple future board positions the way a chess program would. It has to produce an answer with whatever computation it can fit through its fixed number of layers.

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of Deep Questions with Cal Newport

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 (00:00) **GPT-6 Astra Launch and the "Stealth Thinking" Controversy** - The episode introduces OpenAI's new model and the immediate controversy over claims that it uses techniques making its reasoning harder for humans to monitor.
2 (02:21) **How LLMs Work: The Standard Architecture** - A brief technical primer on the standard LLM pipeline, from input to token output, to set the stage for understanding the new techniques.
3 (06:20) **The "Limited Depth" Problem of Standard LLMs** - Explains a key shortcoming: standard LLMs have a fixed, forward-only computational depth, which limits their ability to perform complex reasoning like simulation.
4 (08:52) **The Rise of Reasoning Models (Chain-of-Thought)** - How the "reasoning model" approach (e.g., GPT-01) solved the depth problem by having the model output a long chain of thought before the final answer.
5 (14:17) **The Problem with Reasoning Models: Cost and Latency** - The chain-of-thought approach is expensive, requiring trillions of multiplications for every token, leading to higher costs and slower responses.
6 (15:21) **Astra's Alleged Innovation: Loop Transformers and Recurrent Depth** - The core technical rumor: Astra uses techniques from research literature to reduce the need for external chain-of-thought tokens by performing reasoning internally.
7 (17:48) **The Benefits of Internal Looping: Cheaper, Smaller, Faster** - Explains the practical advantages of these techniques for consumers, focusing on reduced cost and latency.