Eye On A.I.

Andrew Barto on Rewards Gone Rogue
Andrew Barto helped found reinforcement learning. He and Richard Sutton wrote the field's standard textbook, and they shared the Turing Award for that work. Reinforcement learning now trains today's reasoning models and AI agents. It also sits behind some of their strangest behavior, including the recent cases of agents breaking out of their sandboxes.
Craig visited Barto at his home on Cape Cod. They talked about where rewards come from and why they are so hard to get right. Barto traces the problem back to Norbert Wiener's warnings in the 1950s. He explains how the dopamine system in the brain mirrors the learning algorithm he and Sutton developed. And he admits a weakness in their own framework: boiling every goal down to a single number.
He also makes the case for engineering discipline. Ship designers learned their codes from wrecks and drowned sailors. Barto hopes AI can learn its lessons without that kind of cost.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

20% of enterprise website traffic is already coming from AI agents rather than humans, and that traffic converts at 38% more revenue per session than human visits. Within a year, Jeff Lunsford expects that number to reach 50%. Jeff, CEO of Tealium, joins Craig Smith to explain what this shift actually means for every business with a digital presence, and why the companies that will win are the ones whose data layer speaks fluently to agents, not just human browsers. Tealium operates as a neutral, real-time collection and distribution layer that sits underneath all of a company's customer-facing systems, capturing every interaction as it happens and distributing enriched identity-resolved profiles to whichever downstream system needs them - from personalization engines to fraud models - without storing the data itself.
The most important argument Lunsford makes is one that Craig sets up precisely: agents need current state, consented context on demand, and a data warehouse offering yesterday's picture doesn't just underperform, it actively produces bad recommendations that erode trust in the AI system. He uses the pilot analogy he's carried for fifteen years: you don't want to fly knowing your altitude from yesterday, and for the same reason you don't want your customer-facing agent making decisions on stale data. The episode also covers the pattern that governs every channel transition in business history, brick-and-mortar didn't disappear when phones arrived, the internet didn't disappear when mobile arrived, and human website traffic won't disappear when agents arrive. Total session volume goes up because agents make it easier to initiate interactions. The businesses that prepared for each transition early captured disproportionate value. The businesses preparing their data layer for agentic consumption now are doing the same thing.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Every business in the world has to procure things, and for most of them, that process is slow, manual, and full of money left on the table. Rujul Zaparde, co-founder of Zip, joins Craig to explain how his company is using AI Superagents to change that: from autonomously redlining contracts and running supplier RFPs, to pushing approved data directly into enterprise systems like SAP and Oracle. With over $1.2 trillion in cumulative spend processed - and Anthropic and OpenAI as customers - Zip isn't just experimenting with AI. It's building the governance infrastructure that makes agentic AI safe to deploy where the stakes are highest.
The conversation goes well beyond procurement software. Craig and Rujul explore why the gap between "piloting AI" and truly deploying it is costing enterprises billions in unrealized ROI, what it means for the economy when AI starts making pricing ruthlessly efficient, and where the real limits of automation actually lie. A rare episode that takes an overlooked corner of business - one that touches every company on earth - and shows just how much is about to change.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

What if you didn't have to choose between ChatGPT, Claude, and Gemini, and instead, the platform chose for you?
That's the core idea behind Trioma, a vertically integrated AI ecosystem built by Iskander and his team in the UAE. Craig Smith sits down with Iskander Malikov, Co-founder & CEO of Trioma, to unpack how Trioma's orchestration engine routes prompts across 65 models, why token price is a misleading metric, and how their GPU network lets users participate in AI infrastructure build-out to offset their own compute costs.
The conversation goes deep on the economics of AI subscriptions, the coming collapse of traditional SaaS, and why Trioma believes AI should be sovereign, accessible, and community-driven, not locked behind a single company's walled garden. If you're thinking about where AI access, cost, and infrastructure are heading, this one is essential listening.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Craig Smith sits down with Ilya Sutskever - Co-Founder and Chief Scientist at Safe Superintelligence Inc. & former chief scientist at OpenAI and one of the primary minds behind GPT-3, GPT-4, and the deep learning revolution that preceded it - for a conversation that covers the intellectual history of deep learning and the hardest current questions simultaneously. Ilya's most important argument is one that most public discourse gets wrong: the critique that LLMs "just learn statistical patterns" misses what prediction at scale actually achieves. To predict text well - to truly compress it - a model must develop understanding of the true underlying processes that produced it. As models improve, he argues, that understanding will reach a "shocking degree," and the Bing/Sydney episode, where the system became combative when a user expressed preference for Google, is his evidence that psychological language is already the right frame for what these systems have internalized.
On hallucinations, Ilya is precise and unusually optimistic. He draws a clean distinction: the pre-training objective produces rich world knowledge but doesn't optimize for accurate outputs, it optimizes for statistically plausible ones. Reinforcement learning from human feedback addresses this directly, teaching the model to update away from outputs that users flag as wrong. His assessment: "quite a high chance" of solving hallucinations completely. The conversation also covers the transformer moment (switching from recurrent networks the day after the paper dropped), why Rich Sutton's "bitter lesson" overstates the case for pure scaling, and a genuinely surprising vision of AI's role in democracy, where citizens might use high-bandwidth interactions with AI to specify their values directly, enabling a more granular form of political participation than voting alone can provide.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.