David Bombal

#505: GPT-5 Just Dropped… And It’s NOT What You Think

Aug 18, 2025
In this engaging discussion, Gary Marcus, a Professor Emeritus of Psychology and Neural Science at NYU and an AI expert, dives into the reality of GPT-5’s release. He argues that it’s far from a breakthrough, unraveling the truth about AI hallucinations and the cybersecurity risks posed by coding agents. The conversation also highlights the significant limitations of AI in achieving real intelligence, emphasizing that we may be years away from true AGI. Marcus shares his insights on the hype surrounding AI and the urgent need for transparency in its development.
Ask episode
AI Snips
Chapters
Books
Transcript
Episode notes
INSIGHT

GPT-5 Is Incremental, Not AGI

  • GPT-5 is an incremental improvement, not a jump to AGI.
  • Gary Marcus argues the new model is better on benchmarks but far from the promised 'humpback whale' leap.
INSIGHT

Agents Amplify Vulnerabilities

  • Agents multiply risks because each LLM step adds potential failure points.
  • Prompt injection attacks create large attack surfaces for multi-step autonomous workflows.
ADVICE

Harden Against Prompt Injection

  • Prioritize defenses against prompt injection and poisoned examples.
  • Treat external code and web-sourced examples as attack vectors when using coding agents.
Get the Snipd Podcast app to discover more snips from this episode
Get the app