Don't Worry About the Vase Podcast AI #149: 3
20 snips
Jan 1, 2026 The discussion dives into the evolving landscape of language models and the significant impact they have on coding and job markets. Insights reveal how AI could soon write much of the code we rely on. Topics range from deepfake technology and its detection challenges to public concerns about AI's potential risks. The conversation also touches on the preference for human doctors over AI, reflecting societal fears and expectations. Finally, there's a thought-provoking exploration of regulatory needs amid rapid AI advancements.
AI Snips
Chapters
Transcript
Episode notes
Jailbreaks Reveal Model Layering
- Many jailbreaks exploit model internal layering (predictive, persona, surface) rather than magic.
- Zvi notes repeated data poisoning tests like the 'Love Pliny' string reveal training overlaps and filtering gaps.
Get Involved: Apply For NIST Internship
- Apply for agent security internships like the NIST position if you can; deadlines are time‑boxed.
- Zvi highlights this as a concrete way to get involved shaping agent security standards.
Generative AI Traffic Is Fragmenting
- Traffic share shifted in 2025: ChatGPT fell from 87% to 68% as Gemini and others gained ground.
- Zvi notes Claude remains small in consumer share despite strong coding/enterprise traction.