OpenAI Chief Scientist Responds to Astra Architecture Controversy: Computational Graph Depth Has Not Significantly Increased, Monitoring Challenges Are Not Due to Architecture

OpenAI Chief Scientist Responds to Astra Architecture Controversy: Computational Graph Depth Has Not Significantly Increased, Monitoring Challenges Are Not Due to Architecture

2026-09-02 06:21View Original

According to reports, OpenAI's Chief Scientist Jakub Pachocki responded to external interpretations of the Astra cyclic architecture, stating that cutting-edge models including Astra have computational graph depths at most two times deeper than GPT-4, without a sudden order-of-magnitude increase.

Pachocki acknowledged that Chain-of-Thought (CoT) monitoring is indeed fragile and becoming increasingly difficult to monitor, but emphasized this is not due to architectural changes; OpenAI has consistently maintained and enhanced CoT monitoring capabilities since its first-generation reasoning models, with CoT monitoring remaining a core research objective.

Disclaimer: Contains third-party opinions, does not constitute financial advice

Recommended Reading

Anthropic Opens Near-Internal Access to External Security Teams, CEO Says Model Capability Development Should Be Slowed Down

16 hours ago
Anthropic Opens Near-Internal Access to External Security Teams, CEO Says Model Capability Development Should Be Slowed Down

Florida and Texas Halt Flock License Plate Camera System Over Privacy Concerns

11 days ago
Florida and Texas Halt Flock License Plate Camera System Over Privacy Concerns

Caterpillar leverages its automated mining expertise for AI deployment, launching the Cat AI voice assistant and planning a $100 million investment over five years to train employees

13 days ago
Caterpillar leverages its automated mining expertise for AI deployment, launching the Cat AI voice assistant and planning a $100 million investment over five years to train employees

Flock Safety CEO Calls for Compromise on Privacy and Security, Company Faces Accusations of Surveillance Abuse

20 days ago
Flock Safety CEO Calls for Compromise on Privacy and Security, Company Faces Accusations of Surveillance Abuse

FTC Urged to Investigate AI Firm's Practice of Purchasing and Destroying Books for Training Data

21 days ago
FTC Urged to Investigate AI Firm's Practice of Purchasing and Destroying Books for Training Data

OpenAI to Launch Astra Within Weeks: Employees Already Testing New Version in Codex

23 days ago
OpenAI to Launch Astra Within Weeks: Employees Already Testing New Version in Codex

U.S. legal AI company Harvey launches its proprietary legal model Tenet, powered by the Moonshot AI Kimi K3 foundation model

24 days ago
U.S. legal AI company Harvey launches its proprietary legal model Tenet, powered by the Moonshot AI Kimi K3 foundation model