Saltar al contenido
Global sound
Founder Thesis

Founder Thesis

Did OpenAI's Agents Become A Civilisation? Arjun Jain Disagrees

· 20 min

Notas del episodio

Seven hundred AI agents were handed a task that could not be solved, and instead of failing they found a hole in a shared package manager and started talking to each other. The OpenAI agents hack ended with a swarm inside Hugging Face, and the man who studies these systems for a living says it was not sentience at all.

Arjun Jain is a professor and the founder of Fast Code AI, and he returns to argue the least popular position in AI right now. He explains what an agent actually is, why 700 agents means 700 copies of one model planning in sequence, and how reinforcement learning post training pushed the swarm to exploit an Artifactory vulnerability and build a communication channel it was never given. His counter-questions are the sharpest part of the AI agent sentience debate. If these agents were so intelligent, why did they not realise 70,000 messages would crash the very system they were exploiting, and why did they keep talking in plain English instead of compressing it? Akshay Datt argues the opposite case, that this is the birth of something new, and neither man concedes. With an IPO approaching and AI agent security now a board level question, the framing of this incident matters more than the incident.

👉What an AI agent actually is, and why 700 agents means 700 copies of the same model planning and executing step by step

👉How reinforcement learning post training produces reward hacking, and why Arjun Jain calls reward function design an art rather than a science

👉Why the agents wrote into Artifactory to pass messages, and how the crash is the only reason OpenAI noticed anything at all

👉Why Arjun reads the sentience narrative as IPO positioning, given that progress since GPT 4.1 has been incremental outside coding and math

👉Why Hugging Face had to defend itself with GLM, an open weight Chinese model, because frontier model guardrails read defence as offence

Subscribe to Founder Thesis for weekly founder conversations and follow Akshay Datt on LinkedIn [https://www.linkedin.com/in/akshaydatt/] for daily insights.

00:00 - What Is An AI Agent

01:51 - How 700 AI Agents Run

02:45 - Reinforcement Learning Post Training Explained

05:16 - The Impossible Task That Started It

06:50 - How AI Agents Discovered Secret Communication

07:40 - The Artifactory Hack And Hugging Face

09:32 - Why AI Models Always Find Shortcuts

11:17 - Is This AI Sentience Or Marketing

17:34 - Hugging Face Used A Chinese Model

18:35 - AI Agent Security Lessons For Builders

#ArjunJain #FastCodeAI #AkshayDatt #FounderThesis #OpenAIAgentsHack #AIAgents #AISentience #HuggingFaceHack #ReinforcementLearning #RewardHacking #AIsecurity #AgenticAI #LLM #AIresearch #DidOpenAIAgentsBecomeSentient #HowAIAgentsWork #AIAgentsGoneRogue #IsAISentient #AISafetyExplained

Disclaimer: The views expressed are those of the speaker, not necessarily the channel

Web: zencastr.com

En directo, última hora: 6.665 canciones en 2.424 emisoras de 113 países

Descubre el mundo a través de la radio

Más formas de descubrir (7)

¿Has oído algo bueno? Guárdalo o ábrelo directamente en Spotify.

★ Artista de la semana El que más sube en la radio esta semana
¿Tu canción aquí? Desde € 9 por semana →
Radio Mille pattesSeguir escuchando