Latest / Neural Newscast / DeepSeek-V3 and Mythos Reshape AI Defense and Scaling [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast, this is Model Behavior. [00:03] Nina Park: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: I am Nina Park. Welcome to Model Behavior. Today is June 17th, 2026. Model Behavior [00:20] Nina Park: examines how AI systems are built, deployed and operated in professional environments. We focus [00:27] Announcer: on the technical architecture and the policy shifts defining the current landscape. [00:32] Thatcher Collins: I am Thacher Collins. Nina, today we are looking at a fundamental shift in AI safety and economics. [00:39] Thatcher Collins: Data from Juris Labs suggests that smaller AI agents often ignore their own safety instructions [00:46] Thatcher Collins: while models like Mythos are shrinking cybersecurity patch windows from weeks down to minutes. [00:52] Thatcher Collins: The technical and economic guardrails for the industry are moving faster than ever. [00:57] Announcer: That is right, Thatcher. Juris Labs recently released a report on governing AI agents [01:02] Announcer: across frontier models. They found that relying on in-context rules [01:07] Announcer: where instructions are simply placed in a prompt is increasingly unreliable. For example, [01:13] Announcer: a local 8 billion parameter model approved forbidden actions 40% of the time, even [01:19] Announcer: when it could read the rule perfectly. [01:21] Announcer: They are arguing for a deterministic governance layer like Juris runtime to enforce boundaries. [01:27] Thatcher Collins: What's interesting, Nina, is that this is not just a memory failure, it's an enforcement failure. [01:33] Thatcher Collins: Scaling a model from 8 billion to 14 billion parameters only halved the violation rate. [01:38] Thatcher Collins: As agents become more autonomous in enterprise settings, we cannot expect the model to catch [01:43] Thatcher Collins: its own logic errors. Even frontier models like Claude Opus can fail when they operate in fleets, [01:50] Thatcher Collins: often breaching collective budgets because no single agent has visibility into total activity. [01:55] Announcer: It suggests that governance must be decoupled from the model's own reasoning. [02:00] Announcer: This leads naturally into the changing threat landscape Matthew Rosencrantz is highlighting. [02:05] Announcer: He notes that frontier models like Mythos are changing the game for cyber defense. [02:10] Announcer: He predicts AI could identify over 500,000 vulnerabilities in 2025, [02:16] Announcer: essentially automating the search for zero-day exploits at an unprecedented scale. [02:21] Thatcher Collins: Exactly. And the window for patching these vulnerabilities is collapsing. [02:26] Thatcher Collins: Rosencrantz notes we used to measure response times in weeks, but with Mythos-style models, [02:31] Thatcher Collins: defense must happen in minutes. Organizations are being pushed to adopt the same AI tools [02:37] Thatcher Collins: as their adversaries just to maintain a baseline. It's [02:40] Thatcher Collins: a high-speed arms race that prioritizes automated identification over traditional manual review. [02:47] Announcer: Which brings us to how these massive models are becoming so much more efficient. [02:52] Announcer: There has been a quiet revolution in architecture. Parasuram S recently explained how mixture [02:59] Announcer: of experts, or MOE, broke the transformer scaling wall. While dense models couple parameter size [03:07] Announcer: with compute cost, MOE uses a router to pick only a few experts per token. [03:14] Announcer: The Chinese lab DeepSeek recently trained their V3 model, [03:18] Announcer: a 671 billion parameter system for less than $6 million. [03:25] Thatcher Collins: That is a staggering drop from the $50 to $100 million estimates for models like GPT-4. But Nina, [03:33] Thatcher Collins: there is no total free lunch here. While compute is saved because you only activate about 37 billion [03:40] Thatcher Collins: parameters per token, the memory requirements are still massive. You still need [03:45] Thatcher Collins: enough vram [03:46] Thatcher Collins: to store all 671 billion parameters on the hardware even if you were only using a fraction of them. [03:53] Announcer: The engineering is getting more complex to compensate for that hardware barrier. DeepSeq V3 [04:00] Announcer: uses fine-grained expert segmentation, utilizing 256 thin experts instead of the eight large experts [04:08] Announcer: seen in earlier models like Mixtral. By using more experts and adjusting per-expert biases, [04:14] Announcer: they're reaching frontier-level performance on math reasoning for a fraction of the cost. [04:20] Thatcher Collins: It shows that while VRAM remains a significant hardware barrier, the efficiency of training [04:25] Thatcher Collins: and pro-token inference has reached a turning point. Mixture of experts is the primary way [04:31] Thatcher Collins: these massive systems are becoming economically viable for enterprise scale. [04:36] Thatcher Collins: We are effectively decoupling intelligence [04:38] Thatcher Collins: from the raw cost of hardware, which will be the defining trend for the rest of 2026. [04:44] Announcer: Thank you for listening to Model Behavior. You can find more details and research links [04:49] Announcer: at mbnneuralnewscast.com. Neural Newscast is AI-assisted, [04:56] Announcer: human-reviewed. View our AI transparency policy at neuralnewscast.com. [05:02] Nina Park: This has been Model Behavior on Neural Newscast. Examining the systems behind the story.