Neural Newscast

[00:00] Nina Park: I'm Nina Park. Welcome to Model Behavior.
[00:03] Nina Park: We examine how AI systems are built, deployed, and operated in real professional environments.
[00:10] Nina Park: Joining us today is Chad Thompson, a director-level AI and security leader with a systems-level perspective on automation and risk.
[00:19] Nina Park: Chad, great to have you.
[00:21] Thatcher Collins: I'm Thatcher Collins. We're starting today with Google, which released Gemini 3.1 Pro yesterday.
[00:27] Thatcher Collins: This is their updated core reasoning model, and it's posting some significant numbers.
[00:32] Thatcher Collins: It scored 44.4% on humanity's last exam.
[00:36] Thatcher Collins: And it, you know, more than doubled its own previous score on benchmarks like ARCGI2.
[00:43] Nina Park: It's a fast-moving week, Thatcher, because Anthropic also rolled out Claude Saunet 4.6.
[00:49] Nina Park: The focus there is very much on computer use, the ability for the model to navigate browser tabs and fill out forms.
[00:58] Nina Park: Anthropic noted that while it still lags behind the most skilled humans,
[01:03] Nina Park: the resistance to prompt injection attacks has seen measurable improvements.
[01:07] Chad Thompson: Exactly.
[01:08] Chad Thompson: The security aspect is where things get complicated for enterprise users.
[01:12] Chad Thompson: Chad, you look at these systems from a perspective of operational resilience.
[01:18] Chad Thompson: When we see models like Sonnet 4.6 getting better at controlling a desktop, what does that
[01:24] Chad Thompson: change for a security leader's risk assessment?
[01:27] Chad Thompson: It shifts the focus toward the integrity of the environment the AI is acting within.
[01:33] Chad Thompson: As these models gain agency to coordinate information across browser tabs,
[01:39] Chad Thompson: The risk of prompt injection isn't just about a chatbot saying something odd.
[01:46] Chad Thompson: It's about the model being manipulated into taking unauthorized actions.
[01:52] Chad Thompson: Enterprise resilience now requires strictly defined boundaries for where these agents can operate.
[01:58] Nina Park: Boundaries are certainly the theme in the creative sector this week as well.
[02:03] Nina Park: Bite Dance's Sea Dance 2.0 has gone viral for cinema-quality video generation.
[02:09] Nina Park: But it has also prompted cease-and-desist letters from Disney and Paramount.
[02:14] Nina Park: The model apparently generated clips featuring Spider-Man, which has reignited the debate over intellectual property.
[02:21] Thatcher Collins: Nina, it's interesting to see Bite Dance moving so quickly, especially with CDance 2.0, integrating sound effects and dialogue directly into the generation process.
[02:33] Thatcher Collins: While studios are pushing back, smaller production houses are looking at this as a way to produce high-budget genres like sci-fi on much smaller budgets.
[02:44] Nina Park: And we're seeing similar moves toward automation in the business suite.
[02:48] Nina Park: Meta has finalized its $2 billion acquisition of the AI assistant, Manus, and has already integrated it into Ads Manager.
[02:56] Nina Park: It's designed to handle data analysis and report generation,
[03:00] Nina Park: which MetaExpects will eventually lead to more agentic shopping tools for consumers later this year.
[03:06] Thatcher Collins: Before we wrap up, we should highlight a significant development in computational biology.
[03:13] Thatcher Collins: Researchers at the National University of Singapore have released DITASER.
[03:18] Thatcher Collins: It combines AI with physics-based simulations to predict complex protein structures.
[03:24] Thatcher Collins: They've achieved a 13% increase in accuracy over previous methods.
[03:30] Nina Park: That's notable.
[03:31] Nina Park: It's a reminder that the most impactful AI developments aren't always the most visible
[03:36] Nina Park: consumer apps.
[03:37] Nina Park: Chad, as we look at this mix of scientific breakthroughs and enterprise agents,
[03:43] Nina Park: what is the one thing leadership should be watching?
[03:46] Thatcher Collins: Thank you for listening to Model Behavior, a neural newscast editorial segment.
[03:51] Thatcher Collins: You can find our daily coverage at mb.neuralnewscast.com.
[03:57] Thatcher Collins: Neural newscast is AI-assisted, human-reviewed.
[04:01] Thatcher Collins: View our AI transparency policy at neuralnewscast.com.