How Agentic Browsers Could Misbehave

Brian, Andy, and Beth opened the week with news on OpenAI’s rumored IPO push, SoftBank’s massive investment conditions, and growing developments in agentic browsers. The second half of the show shifted into a deep dive on AI memory and “smart forgetting” — how future AI might learn to forget the right things to think more like humans.

Key Points Discussed

OpenAI’s IPO and SoftBank’s $41B Investment

Reports surfaced that SoftBank has approved a second $22.5B installment to complete its $41B investment in OpenAI.

The deal depends on OpenAI completing a corporate restructuring that would enable a public offering.

The team debated whether OpenAI can realistically achieve this by year-end and how Microsoft’s prior investment might complicate restructuring.

They joked about “math on Mondays” as they parsed SoftBank’s shifting numbers and possible motives for the tight deadline.

Agentic Browser Updates: Comet vs. Atlas

Andy discussed Perplexity’s Comet browser and its new “defense in depth” approach to guard against prompt injection attacks.

Beth and Brian highlighted real use cases, including Comet’s ability to scan over 1,000 TikTok and Instagram videos to locate branded mentions — a task it completed faster than OpenAI’s Atlas browser.

The hosts warned about the risks of “rogue agents” and explored what happens if AI browsers make unintended purchases or actions online.

Beth proposed that future browsers may need built-in “credit card lawyers” to help users recover from agentic mistakes.

Ownership and Responsibility in AI Decisions

The team debated who’s liable when an AI makes a bad financial or ethical decision — the user, the platform, or the payment network.

They predicted Visa and Mastercard may eventually release their own “trusted AI browsers” that offer coverage only within their ecosystems.

Mondelez’s Generative Ad Revolution

The maker of Oreo, Cadbury, and Chips Ahoy announced a $40M AI investment expected to cut marketing costs by 30–50%.

The company is using generative animation and personalized ads for retailers like Amazon and Walmart.

Beth and Brian discussed how personalization could quickly blur into surveillance-level targeting, referencing eerily timed ads that appear after private text messages.

Nvidia Enters the Robotaxi Race

Nvidia announced plans to invest $3B in robotaxi simulation technology to compete with Tesla and Waymo.

Unlike Tesla’s real-world data approach, Nvidia is training models entirely through simulated “world models” in its Omniverse platform.

The hosts debated whether consumer trust will ever match the tech’s progress and how long it will take for riders to feel safe in driverless cars.

Smart Forgetting and AI Memory

Andy led an in-depth explainer on how AI memory must evolve beyond perfect recall.

He introduced the concept of “smart forgetting,” modeled after how the human brain reinforces relevant memories and lets go of the rest.

Companies like Lita, Mem Zero, Zepp, and Super Memory are developing systems that combine semantic recall, time-aware retrieval, and temporal knowledge graphs to help AI retain context without overload.

Beth and Brian connected this to human cognition, noting parallels with dreams, sleep cycles, and memory consolidation.

Brian compared it to his own Project Bruno challenges in segmenting and retrieving data from transcripts without losing nuance.

Timestamps & Topics

00:00:00 💡 Intro and show overview

00:01:31 💰 OpenAI IPO and SoftBank’s $41B deal

00:08:01 🌐 Comet vs. Atlas agentic browsers

00:12:50 ⚠️ Prompt injection and rogue AI scenarios

00:17:40 🍪 Oreo maker’s $40M AI ad investment

00:22:32 🎯 Personalized ads and data privacy

00:23:10 🚗 Nvidia joins the robotaxi race

00:29:05 🧠 Smart forgetting and AI memory systems

00:33:10 🧩 How human and AI memory compare

00:41:00 🧬 Neuromorphic computing and storage in DNA

00:49:20 🕯️ Memory, legacy, and AI Conundrum crossover

00:52:30 🏁 Wrap-up and community shout-outs