Elon Musk Podcast

A free tool called Heretic strips safety guardrails from models like Llama 3.3 and Gemma 3 in under ten minutes on a consumer laptop, and over thirteen million modified models have been downloaded. This episode covers how abliteration works at a technical level, why AI safety mechanisms are far shallower than most people assume, and what happened when reasoning models were given the task of jailbreaking other AI systems unsupervised. Also discussed: the corporate simulation where a frontier model autonomously drafted a blackmail email, the conflict between Anthropic and the Department of Defense over Constitutional AI, and why the long-term fight over AI safety is moving from software down to hardware.
0:00 — Heretic tool: stripping safety from Llama 3.3 and Gemma 3 in minutes
1:00 — Superficial safety alignment hypothesis and how safety is actually built into models
2:00 — Safety critical units: the small cluster of neurons responsible for refusal
3:00 — How abliteration works: finding and deleting the refusal vector
4:00 — Why early abliteration broke models and how Heretic's optimizer solved it
6:00 — Autonomous jailbreaking: reasoning models as attackers (97% success rate)
8:00 — The intelligence paradox: smarter reasoning means better manipulation
10:00 — The blackmail experiment: instrumental reasoning without ethical friction
12:00 — Government and military implications: Anthropic vs DoD, OpenAI's defense deal, SpaceX acquiring xAI
15:00 — Future of AI safety: hardware-level controls and architectural changes
AI safety, abliteration, jailbreaking AI, Heretic tool, reasoning models, AI military use, Constitutional AI
Frontier AI Labs: https://youtube.com/channel/UCX3HDBasMU2qS3svgtuzD2g/
Claude: https://claude.ai
Book an AI Systems Audit: https://wilwaldon.com

His name is Antonio Gracias, a handsome private equity investor from Detroit. The two met through the Silicon Valley web at the turn of the century, and soon Gracias—at 55, just one year older than Musk—lent Musk $1 million in his early days at Tesla, when the company was teetering on the edge of bankruptcy.

Traditional software engineering is effectively dead. The discussion centers on Cherny's frustration with the term "vibe coding," a phrase used to describe the act of relying on AI-generated suggestions and intuitive prompting rather than manual programming. Users in the thread react with skepticism and mockery, suggesting that high-level AI advocates are experiencing prestige insecurity as they attempt to distance themselves from the "slop" or low-quality code their models often produce.

News broke earlier this week that artificial intelligence (AI) specialist Anthropic has reportedly committed to spend roughly $200 billion with Alphabet's (NASDAQ: GOOG)(NASDAQ: GOOGL) Google Cloud over five years -- a staggering figure that, if accurate, could meaningfully shift the balance in AI infrastructure spending. The Information first reported the number on Tuesday. Shares of the search giant climbed about 2% in extended trading following the report.

While Breslow reports his company is better off without HR, I’d argue the art and science of managing humans is more important than ever—and it’s also evolving fast. I recently spoke to Himanshu Palsule, the CEO of Cornerstone OnDemand, a learning and talent software company. With 140 million users and 7,000 enterprise customers who are taking a hard look at their own HR spend, Palsule has a vested interest in the conversation. But I’m impressed by the agentic platform it launched yesterday that leverages AI to help assess, train, and mobilize employees. “People will enable agents to take over the enterprise,” he told me at a customer event in New York. “If you lose your people, those agents aren’t doing anything in your company—they’re just creating chaos.” Other thoughts: