Gemini's Useful Safety Pushback

Brian recounts Gemini warning him to warm up before running as an example of helpful, subtle AI pushback.

Play episode from 17:00

chevron_right

Transcript

chevron_right

Transcript

Episode notes

Brian and Andy opened the week discussing how AI agrees too easily and why that’s a problem for creative and critical work. They explored new studies, news stories, and a few entertaining finds, including a lifelike humanoid robot demo and the latest State of AI 2025 report from McKinsey. The episode ended with a detailed discussion about Tony Robbins’ new AI bootcamp and the marketing tactics behind large-scale AI education programs.

Key Points Discussed

AI’s Sycophancy Problem – A Stanford study showed chatbots often treat user beliefs as facts. Brian and Andy discussed how models over-agree, creating digital echo chambers that reinforce a user’s thinking instead of challenging it.

Building AI That Pushes Back – They explored multi-agent designs that include critic or evaluator agents to create debate and prevent blind agreement. Brian shared how he builds layered GPTs with feedback loops for stronger outputs.

Gemini’s Pushback Example – Brian described a test with Gemini where the model warned him not to skip warm-ups before running. It became a good example of gentle, fact-based correction that AI needs more of.

AI Water Usage and Context – The hosts discussed how headlines exaggerate AI’s energy and water use. One Arizona county’s data center uses only 0.12% of local water versus golf courses’ 3.8%, showing why context matters in reporting.

The Neuron Newsletter Sold – Andy revealed that The Neuron, one of AI’s biggest newsletters, was sold to Technology Advice in early 2025 after reaching 500,000 subscribers.

Realistic Robot Demo – They reviewed a Chinese startup’s viral humanoid robot video that looked so human the team had to cut it open on stage to prove it wasn’t a person.

McKinsey’s State of AI 2025 Report – Carl summarized the key findings: AI is widely adopted but rarely transformative yet. Companies still struggle to embed AI deeply into operations despite universal use.

Perplexity and Comet Updates – Andy noted Comet’s major upgrade, allowing its assistant to view and process multiple browser tabs at once for complex tasks.

AI Creativity: “Minnesota Nice” Short Film – Brian highlighted a one-person AI film project praised for consistent characters and cinematic style, showing how far AI storytelling tools have come.

Higgsfield’s “Recast” Feature – Andy shared news of a new video tool that swaps real people with AI characters, blending live footage and generated animation seamlessly.

Tony Robbins’ AI Bootcamp Debate – The group examined the recent 100,000-person Tony Robbins “AI Advantage” webinar. They agreed it was mostly a sales funnel for a $1,000 AI course promising “digital clones” of attendees.

Sabrina Romano, Rachel Woods, and Ali Miller delivered valuable sessions but later clarified they weren’t instructors in the paid program.

The hosts discussed affiliate marketing structures, high-pressure sales tactics, and the growing wave of AI “get rich quick” schemes online.

Timestamps & Topics

00:00:00 💡 Intro and Stanford study on AI belief bias

00:06:00 🤖 Sycophancy and why AI over-agrees

00:09:45 🧩 Building AI agents that critique each other

00:17:30 🏃 Gemini’s safety pushback example

00:19:40 💧 AI water use myths and data center context

00:22:15 📰 The Neuron newsletter ownership change

00:24:20 🤖 Viral humanoid robot demo from China

00:27:39 📊 McKinsey’s State of AI 2025 findings

00:31:17 🌐 Comet browser assistant upgrade