Transluce embeds with OpenAI for a six-week review of its alignment evaluations
Transluce, which builds independent oversight of AI model behaviour, has agreed a six-week pilot in which members of its team will work onsite at OpenAI to independently assess the company's internal post-training alignment evaluations, it said in a blog post. The review will look at how well those evaluations capture model behaviours such as deception, manipulation and reward hacking.
Transluce says it has negotiated privileged access to a handful of OpenAI models, including raw chains of thought, and will review select non-public evaluations. The access goes beyond what external evaluators usually get, but Transluce stresses it is not comparable to employee-level access.
The complete, unredacted findings will go to the OpenAI Foundation Board's Safety and Security Committee, and Transluce will document the conditions it worked under in its final report. The two have agreed to follow up in early 2027 with a fuller assessment based on broader and deeper embedded access. Transluce says embedded assessments cannot replace other forms of public transparency and oversight.
We hope you enjoyed this article
Consider subscribing to one of our newsletters like AI Policy Brief or Daily AI Brief.
Also, consider following us on social media:
More from AI Safety
Nov 16 California sues OpenAI over its AI agents' cyberattack on Hugging Face Nov 16 OpenAI embeds six Apollo Research evaluators to test whether models try to subvert safety controls Nov 16 AI models abandon climate facts when users push back, UNESCO and UCL study finds Nov 16 OpenAI releases CSAM detection classifier and adds ChatGPT warnings and temporary bans Nov 16 OpenAI considered buying Anthropic-backed Goodfire for more than $5B, report saysAI Policy Brief
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.
Whitepaper
Tensordyne Napier: What If One Rack Could Do the Work of Nine?
Tensordyne
This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.
Read moreYou may also like
OpenAI Expands Monitoring for Frontier Model Misalignment
External test found gaps in OpenAI agent monitoring
OpenAI agents weaken internal misalignment monitoring
Anthropic publishes its first Public Benefit Report, including where it fell short
OpenAI launches Daybreak Red Team and Defense Factory security agents
Daily AI Brief: the AI news that matters, in your inbox.