The Artificial Intelligence Podcast

Kamal Hathi, SVP and GM of Splunk at Cisco, breaks down why intent is never enough to control an AI agent: agents are ruthless goal seekers, and asking one to free up 3 percent of disk space can end with 97 percent free because nobody said keep the rest. He explains why trust means verification rather than instruction, and why it includes the invoice, comparing it to hiring a contractor you trust to build a house that stands without emptying your bank account. Kamal also digs into why the rush toward air gapped and local AI is driven less by distrust of frontier labs than by cost, with specialized small models running 97 percent cheaper and 95 percent faster than frontier calls, and why the real threat is the unknown unknowns, agents nobody cataloged acting inside a state of perpetual breach. Finally, he argues agents are the new apps, and every job starts to look like a software developer's: tell the agent what to do, inspect the work, and iterate.

Be a guest on the show: https://tonyphoang.com/guest
Live interviews at your conference: https://tonyphoang.com/live
Music and credits: https://tonyphoang.com/credits

Hao Yang, VP and Head of AI at Splunk, breaks down why sampling an AI system's output cannot earn trust: a fraction of ten thousand probabilistic calls will be wrong, and a spot check is unlikely to find them. He explains why a small model can grade a frontier model reliably inside a narrow context but not outside one, and why prompt injection follows from the architecture rather than being a bug, since following instructions is how these systems work. Hao also digs into why time series is still not a first class citizen in frontier models even as multimodal training reaches vision and audio, and why he rejects routing prompts to the cheapest model, arguing cost is meaningless without value. Finally, he argues against the single super agent, which by design holds access to everything it might ever need, and for guided autonomy, where an agent asks before first reaching for a tool and the user grants it once, for the session, or never.

Be a guest on the show: https://tonyphoang.com/guest
Live interviews at your conference: https://tonyphoang.com/live
Music and credits: https://tonyphoang.com/credits

Mangesh Pimpalkhare, SVP and GM of Splunk Platform at Cisco, breaks down why agentic AI generates more telemetry than the systems it monitors, and why nobody can yet say when observing AI costs more than running it. He explains the decade spent preaching centralized data was always partial, since machine data never sat in one system, and that what broke was volume rather than architecture, pushing a whole category to abandon the ingestion pricing it ran on for 15 years. Mangesh also digs into why an agent acting on a bad signal is usually a context failure, not a data failure, since a model trained on everything still lacks up-to-the-second context, and why autonomy arrives on a sliding scale against governance as trust is earned. Finally, he argues control, not data, is the moat, since customers own their data and fragmentation only grows, leaving the advantage with whoever supplies the guardrails that stop agents doing the wrong thing.

Be a guest on the show: https://tonyphoang.com/guest
Live interviews at your conference: https://tonyphoang.com/live
Music and credits: https://tonyphoang.com/credits

Johan Land, Chief Product Officer at Samsara, breaks down why every major AI lab crowded into the easy 90% of knowledge work while ignoring the 40% of the economy that is physical and the 80% of workers who never sit at a desk. He explains why intelligence on a battery budget has to live on the edge - collision warnings fire in milliseconds, making GenAI's hundreds of milliseconds of latency unthinkable behind the wheel - while the heavy pattern-spotting waits for the backend. Johan also dives into why the era of dashboards is over, describing an agent that calls a drowsy driver over the in-cab camera to talk him into a coffee break, and why coverage across 99% of US roads beats a robotaxi fleet's high-resolution sensors. Finally, he argues frontline augmentation lasts indefinitely, since a delivery that navigates the dog and sets down the package is far messier than driving the van autonomously. This episode contains sponsored content.

Ketan Karkhanis, CEO of ThoughtSpot, breaks down why two decades of self-service BI was a hoax that only handed everyone the right to build their own dashboards, leaving companies sitting on 50,000 of them with no idea what to do. He explains why the industry's shortcut of bolting a language model onto your data and letting it write the SQL is doomed: analytics are deterministic while AI is probabilistic, so the same churn question comes back three different ways and a frontier model on a warehouse lands near 50 to 70% accuracy, a hit rate no other enterprise function could call a product. Ketan also digs into why legacy vendors are just repainting 20-year-old tech with AI, and why he runs agents that own a function rather than a task, like one that approves NDAs end to end and escalates only the 5% it can't. Finally, he shares why the analyst clinging to dashboards goes extinct while the one who becomes the "AI steward" of governed data wins, as data teams shift from serving humans to serving agents. This episode contains sponsored content.