About this episode Dive deep into the daily currents of artificial intelligence with AI Unfiltered. We meticulously analyze new models, product launches, research breakthroughs, and funding rounds, dissecting the true impact from the fleeting hype. Gain a researcher's perspective on what genuinely shifts the AI landscape, equipping you with essential insights to navigate this rapidly evolving field.
0:00 Anthropic has embedded about six engineers inside the NSA.0:04 Their mission: to deploy a new AI model, named Mythos, for offensive cyber operations against networks in China and Iran.0:14 In our last episode, we focused on cutting through hype to find the signals that actually move this industry.0:23 This is not a signal.0:24 This is an air-raid siren.0:26 Frontier AI is now operationally deployed for state-level attacks.0:32 The rest of the landscape shifted hard today as well.0:36 NVIDIA just released Nemotron 3 Ultra.0:39 It’s a 550 billion parameter open reasoning model, and according to benchmarks, it’s now the most capable open model from the US.0:50 It supports a one million token context window and is built specifically for complex agentic workflows.0:58 This is a massive capability drop for the open-source world.1:03 Microsoft, not to be outdone at its Build conference, unveiled seven new in-house models.1:10 The flagship is MAI-Thinking-1, a one trillion parameter sparse model.1:15 Their entire pitch is about provenance and compliance, a direct appeal to enterprise legal teams worried about data licensing.1:26 Meanwhile, Salesforce’s cumulative investment in Anthropic just hit five billion dollars.1:33 This comes even as Salesforce stock is down nearly thirty-three percent for the year.1:40 The market is skeptical, but Salesforce is betting its entire future on integrating Claude AI directly into its core products, arguing that AI without verified customer data is unreliable.1:55 And in the robotics world, a company called ACE ROBOTICS just open-sourced Kairos-HomeWorld.2:02 It’s the first framework that can generate a fully interactive, whole-home 3D scene from a single text prompt.2:11 This is a foundational technology for training robots in realistic, simulated environments.2:18 Let’s go back to Anthropic and the NSA.2:21 The model, Mythos, is not just a research project.2:26 It’s an operational tool.2:28 According to technical reports, it autonomously reads source code, hypothesizes vulnerabilities, writes its own test cases against live software, and confirms the exploits.2:41 All without human guidance after the initial tasking.2:46 In one benchmark against a Firefox engine, it produced 181 working exploits.2:52 This is a company that built its brand on AI safety.2:56 A company that is simultaneously suing the Pentagon for barring the use of its other model, Claude.3:04 The contradiction is stark.3:06 One arm of the company is fighting for government access while another is deploying an autonomous hacking tool inside the nation’s largest intelligence agency.3:19 This isn't a future scenario.3:21 It is happening now.3:23 The debate about AI weaponization just became obsolete.3:27 The weapons are deployed.3:29 Now, look at the counter-current.3:32 On the very same day, NVIDIA releases Nemotron 3 Ultra.3:36 This is not a crippled, safety-washed model.3:40 It’s a 550 billion parameter reasoning engine with 55 billion active parameters per token.3:47 It uses a new hybrid architecture combining Mixture-of-Experts with state-space layers.3:54 NVIDIA claims it delivers up to five times faster inference on agentic tasks, at thirty percent lower cost.4:03 But here’s what truly matters: NVIDIA open-sourced everything.4:08 The weights.4:09 The training data pipeline.4:11 The recipes.4:12 It’s all available under a permissive license.4:16 They are not just giving away the car; they are giving away the entire factory.4:22 This is a deliberate, strategic move to empower an entire ecosystem of developers to build sophisticated AI agents outside the walls of the big labs.4:34 So on one side, you have a frontier model being turned into a targeted cyber weapon, deployed in secret.4:42 On the other, a model of comparable power is being given to the entire world to build with.4:50 The same underlying technology, forked into two completely different realities.4:56 This is the fundamental tension of AI in 2026.5:00 The argument over safety guardrails and ethical alignment is no longer academic.5:06 It is being stress-tested in real-time, in classified networks and public code repositories, simultaneously.5:15 The capabilities are out.5:17 The choice of how they are used is now the only variable that matters.
