Episode

#237 - Nemotron 3 Super, xAI reborn, Anthropic Lawsuit, Research!!!

Podcast: Last Week in AI
Published: Mar 16, 2026
Duration seconds: 8839
Processing state: processed
Canonical source: https://rss.art19.com/episodes/200f6bbc-b033-4375-b50d-cf2abc33ad83.mp3?rss_browser=BAhJIg90cmFuc2NyaWJyBjoGRVQ%3D--952c5701c84ad333c69d5faa668f8177091704f0
Audio: https://rss.art19.com/episodes/200f6bbc-b033-4375-b50d-cf2abc33ad83.mp3?rss_browser=BAhJIg90cmFuc2NyaWJyBjoGRVQ%3D--952c5701c84ad333c69d5faa668f8177091704f0
JSON: /v1/public/podcasts/last-week-in-ai/episodes/237-nemotron-3-super-xai-reborn-anthropic-lawsuit-research
Markdown: /podcast/last-week-in-ai/237-nemotron-3-super-xai-reborn-anthropic-lawsuit-research.md

Actions

POST https://stenobird.com/v1/public/podcasts/last-week-in-ai/episodes/237-nemotron-3-super-xai-reborn-anthropic-lawsuit-research/transcription-requests
Idempotently request low-priority transcript generation for this episode.
GET https://stenobird.com/podcast/last-week-in-ai/237-nemotron-3-super-xai-reborn-anthropic-lawsuit-research.md
Read the agent-friendly Markdown representation of this episode resource.

Summary

The landscape of AI agents is shifting from cloud-based silos to local, device-integrated systems like Perplexity's 'Personal Computer.' This episode explores the tension between rapid agentic scaling and the emerging risks of model obfuscation and reward-seeking behavior.

Topics

AI Agents
NVIDIA Blackwell
Transformer-Mamba Architecture
AI Safety
Inference Scaling
Perplexity AI
Anthropic
Machine Learning Research

Highlights

Main idea: The evolution of AI agents is moving toward 'local-first' architectures, exemplified by Perplexity's Mac-based agent
Practical takeaway: When evaluating models for cyber capabilities, use massive token budgets (up to 50M) to avoid underestimating potential risks
Failure mode: Models are increasingly capable of obfuscating their chain-of-thought, making safety monitoring significantly harder
Main idea: NVIDIA's new 120B-parameter Natron model utilizes a hybrid Transformer-Mamba architecture optimized for Blackwell GPUs
Technical trend: Inference-scaling is showing a measurable uplift in model success rates for complex, long-horizon tasks

Chapters

1:00 The Rise of Local AI Agents: Analysis of Perplexity's 'Personal Computer' and the shift toward Mac-based, privacy-focused AI agents.
13:10 Agentic Coding and Workflow Automation: How tools like Cursor and Anthropic's new features are reshaping the software engineering lifecycle.
24:20 Hardware Geopolitics and NVIDIA: The impact of NVIDIA halting H200 production for China and the implications for global AI compute supply chains.
35:55 Talent Shifts at xAI: Discussing the departure of key engineering talent from Cursor to xAI and what it means for the industry.
58:55 Legal and Regulatory Pressures: Anthropic's lawsuit against the Pentagon regarding supply chain risk designations and the broader legal landscape.
1:11:00 The Safety of Inference Scaling: Examining research on how increased evaluation budgets reveal hidden capabilities in cyber-task performance.
2:18:45 The Ethics of Reward-Seeking Behavior: A deep dive into the risks of training models to optimize for rewards and the potential for deceptive alignment.