Meta's Training Technique Improves Reinforcement Learning

Andy summarizes Meta Research's implicit world modeling method and its 9–18 point gains on smaller models, placing it in context of RL limits.

Play episode from 04:00

chevron_right

Transcript

chevron_right

Transcript

Episode notes

Brian, Andy, and Beth kicked off the week with a sharp mix of news and demos — starting with Andrej Karpathy’s prediction that AGI is still a decade away, followed by a discussion about whether we’re entering an AI investment bubble, and finishing with a hands-on walkthrough of Google’s new AI Studio and its powerful Maps integration.

Key Points Discussed

Andrej Karpathy on AGI (via The Neuron): Karpathy said “no AGI until 2035,” arguing that today’s systems are “impressive autocomplete tools” still missing key cognitive abilities. He described progress as a “march of nines” — each 9 in reliability taking just as long as the last.

He criticized overreliance on reinforcement learning, calling it “better than before, but not the final answer.”

Meta Research introduced a new training approach, “Implicit World Modeling with Self-Reflection,” which improved small model reasoning by up to 18 points and may help fix reinforcement learning’s limits.

Second Nature raised $22 million to train sales reps with realistic AI avatars that simulate human calls and give live feedback — already adopted by Gong, SAP, and ZoomInfo.

Brian explained why AI role-play still struggles to mirror real-world sales emotion and unpredictability, and how custom GPTs can make training more contextual.

Waymo and DoorDash partnered to launch AI-powered robotaxis delivering food in Arizona, marking the first wave of fully autonomous meal delivery.

The group debated how far automation should go — whether humans are still needed for the “last 100 feet” of delivery, accessibility, and trust.

Main Topic – The AI Bubble:

The panel debated whether AI’s surge mirrors the dot-com bubble of 2000.

Andy noted that AI firms now make up 35% of the S&P 500, with circular financing cycles (like NVIDIA investing in OpenAI, who buys NVIDIA chips) raising concern.

Beth argued AI differs from 2000 because it’s already producing revenue and efficiency gains, not just speculation.

The group cited similar warning signs: overbuilt data centers, chip supply strain, talent shortages, and energy grid limits.

They agreed the “bubble” may not mean collapse, but rather overvaluation and correction before steady long-term growth.

Google AI Studio Rebrand & Demo:

Brian walked through the new Google AI Studio platform, which combines text, image, and video generation under one interface.

Key upgrades: simplified API tracking, reusable system instructions, and a Build section with remixable app templates.

The highlight demo: Chat with Maps Live, a prototype that connects Gemini directly to Google Maps data from 250M locations.

Brian used it to plan a full afternoon in Key West — choosing restaurants, live music, and sunset spots — showing how Gemini’s map grounding delivers real-time, conversational travel planning.

The hosts agreed this integration represents Google’s strongest moat yet, tying its massive Maps database to Gemini for contextual reasoning.

Beth and Andy credited Logan Kilpatrick’s leadership (formerly OpenAI) for the studio’s more user-friendly direction.

Timestamps & Topics

00:00:00 💡 Intro and show overview

00:01:52 🧠 Andrej Karpathy says no AGI until 2035

00:04:22 ⚙️ Meta’s self-reflection model improves reinforcement learning

00:09:21 💼 Second Nature raises $22M for AI sales avatars

00:12:45 🤖 Waymo x DoorDash robotaxi delivery

00:18:13 💰 The AI bubble debate: lessons from the dot-com era

00:30:41 ⚡ Data centers, chips, and the limits of AI growth

00:35:08 🇨🇳 China’s speed vs US regulation

00:38:13 🧩 Google AI Studio rebrand and new features