Can 250 Poisoned Documents Break Model Alignment?

Brian and Andy opened the October 14th episode discussing major AI headlines, including a criminal case solved using ChatGPT data, new research on AI alignment and deception, and a closer look at Anduril’s military-grade AR system. The episode also featured deep dives into ChatGPT Pulse, NotebookLM’s Nano Banana video upgrade, Poe’s surprising comeback, and how fast AI job roles are evolving beyond prompt engineering.

Key Points Discussed

Law enforcement used ChatGPT logs and image history to arrest a man linked to the Palisade fires, sparking debate on privacy versus accountability.

Anthropic and the UK AI Security Institute found that only 250 poisoned documents can alter a model’s behavior, raising data alignment concerns.

Stanford research revealed that models like Llama and Qwen “lie” in competitive scenarios, echoing human deception patterns.

Anduril unveiled “Eagle Eye,” an AI-powered AR helmet that connects soldiers and autonomous systems on the battlefield.

Brian noted the same tech could eventually save firefighters’ lives through improved visibility and situational awareness.

ChatGPT Pulse impressed Karl with personalized, proactive summaries and workflow ideas tailored to his recent client work.

The hosts compared Pulse to having an AI executive assistant that curates news, builds workflows, and suggests new automations.

Microsoft released “Edge AI for Beginners,” a free GitHub course teaching users to deploy small models on local devices.

NotebookLM added Nano Banana, giving users six new visual templates for AI-generated explainer videos and slide decks.

Poe (by Quora) re-emerged as a powerful hub for accessing multiple LLMs—Claude, GPT-5, Gemini, DeepSeek, Grok, and others—for just $20 a month.

Andy demonstrated GPT-5 Codex inside Poe, showing how it analyzed PRDs and generated structured app feedback.

The panel agreed that Poe offers pro-level models at hobbyist prices, perfect for experimenting across ecosystems.

In the final segment, they discussed how AI job titles are evolving: from prompt engineers to AI workflow architects, agent QA testers, ethics reviewers, and integration designers.

The group agreed the next generation of AI professionals will need systems analysis skills, not just model prompting.

Universities can’t keep pace with AI’s speed, forcing businesses to train adaptable employees internally instead of waiting for formal programs.

Timestamps & Topics

00:00:00 💡 Intro and show overview

00:02:14 🔥 ChatGPT data used in Palisade fire investigation

00:06:21 ⚙️ Model poisoning and AI alignment risks

00:08:44 🧠 Stanford finds LLMs “lie” in competitive tasks

00:12:38 🪖 Anduril’s Eagle Eye AR helmet for soldiers

00:16:30 🚒 How military AI could save firefighters’ lives

00:17:34 📰 ChatGPT Pulse and personalized workflow generation

00:26:42 💻 Microsoft’s “Edge AI for Beginners” GitHub launch

00:29:35 🧾 NotebookLM’s Nano Banana video and design upgrade

00:33:15 🤖 Poe’s revival and multi-model advantage

00:37:59 🧩 GPT-5 Codex and cross-model PRD testing

00:41:04 💬 Shifting AI roles and skills in the job market

00:44:37 🧠 New AI roles: Workflow Architects, QA Testers, Ethics Leads

00:50:03 🎓 Why universities can’t keep up with AI’s speed

00:56:43 🏁 Closing thoughts and show wrap-up

The Daily AI Show Co-Hosts: Andy Halliday, Brian Maucere, and Karl Yeh