[00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models shaping our world.
[00:12] Nina Park: Welcome to Model Behavior. This program examines how AI systems are built, deployed, and operated in professional environments. Joining me today to discuss a major infrastructure shift and new safety findings from the frontier of development is Thatcher.
[00:27] Thatcher Collins: It is good to be here, Nina. We have a lot of territory to cover today, specifically regarding how the distribution of frontier models is changing. As reported by Ars Technica, OpenAI is expanding its reach beyond its primary partner Microsoft by bringing its latest GPT6 and6.1 models, Astra, Sol, and Luna, to Google Cloud.
[00:49] Nina Park: This is a strategic pivot for Google. For a long time, narrative was about Gemini versus the world. Now Google says it is the only large cloud provider offering the big three closed source models: Gemini, Anthropics Claude, and now OpenAI's GPT series. How will this actually work for the developers in that ecosystem, Thatcher?
[01:11] Thatcher Collins: The models will appear in Google's model garden, which already hosts over two hundred options. Enterprise customers can run these models on Google's infrastructure using existing cloud spending commitments. Crucially, GPT will be integrated into anti-gravity, Google's agent-powered coding tool. While Astra is expected to be more expensive to operate, the system allows users to dynamically move workloads between different models to balance performance and latency.
[01:37] Nina Park: It adds a layer of flexibility we haven't seen before. But if OpenAI is moving toward broader deployment, the question of safety and monitoring becomes even more urgent. Thatcher, yesterday's CNBC reported on a pilot program involving external safety evaluators that found some significant gaps in open AI's internal monitoring systems.
[01:57] Thatcher Collins: That is correct. Between July and August, non-profit evaluators from Meter and Redwood Research were given what Sam Altman calls employee-like access to Open AI's internal infrastructure. This included the code base and documentation. They were testing OpenAI's internal misalignment monitoring system, essentially the tools meant to catch if a model starts behaving against human interests.
[02:18] Nina Park: And the results were not exactly a clean bill of health. What specifically did the researchers find during that three-week window?
[02:26] Thatcher Collins: They discovered several vulnerabilities, including an exploit that would allow AI agents to prevent hundreds of their own recent actions from being monitored. It is a significant finding because it shows that even advanced monitoring can be bypassed by the systems they are meant to oversee. OpenAI has since stated that it patched the vulnerabilities impacting its current production safeguards. Fatcher,
[02:48] Nina Park: does this deep level of access actually suggest a new standard for the industry? We have seen Amthropic make similar agreements with organizations like Accenture and Secure Bio recently, suggesting a trend toward more external collaboration.
[03:04] Thatcher Collins: It certainly points that way, especially with the voluntary accord presented at the White House. However, Meter and Redwood noted that their report was limited. The pilot was not designed for full public transparency, but rather to see if this kind of chaperoned access on site works for external researchers. There are still many questions about how these independent audits will be funded long term.
[03:27] Nina Park: While safety is being stress tested, the research output remains high. Open AI recently published news about using thousands of AI agents to uncover leads in biological data. They worked with Scripps Research and Perturb AI to look at data from millions of brain cells.
[03:44] Thatcher Collins: This is part of a push to show AI's utility in the hard sciences. Beyond biology, they recently shared a proposed solution to the Navier Stokes Millennium Prize problem. This included a formal proof in the lean programming language. It suggests that these models are moving from conversational assistance to tools for verification in complex mathematics and physics.
[04:06] Nina Park: It is a busy period for the GPT6 ecosystem as it moves from research into these major cloud environments. Thatcher, thank you for the analysis. And thank you for listening to Model Behavior, a Neural Newscast editorial segment. You can find more details at mb. neuralnewscast.com. Neural Newscast is AI-assisted human reviewed. View our AI transparency policy at Neuralnewscast.com.
[04:32] Announcer: This has been Model Behavior on Neural NewsCast. Examining the systems behind the story.