Sam Altman on AGI, GPT-5, and what’s next — the OpenAI Podcast Ep. 1

Sam Altman pegged the next major leap in OpenAI’s model line—GPT-5—at “probably sometime this summer,” while arguing that the real milestone isn’t a fixed definition of AGI but the emergence of systems that can accelerate scientific discovery. In his framing, “AGI” keeps sliding as capabilities improve, so the more meaningful question becomes what would count as something like “superintelligence”: autonomous discovery of new science or a tool that dramatically boosts researchers’ ability to find new science. That distinction matters because it ties the pace of model progress directly to measurable human outcomes—faster breakthroughs in areas like medicine and fundamental research—rather than a single benchmark that gets outdated.

Altman also described how current AI tools are already shifting workflows in ways that feel “very AGI like,” especially when models can act on computers and gather information rather than only summarize it. He pointed to internal progress from o1 to o3 as a pattern of rapid iteration, and he cited “operator” on o3 as a personal turning point for seeing agentic behavior that doesn’t collapse when tasks get tricky. He contrasted that with earlier brittleness in agentic systems—promises of autonomy that fail the moment a problem falls outside the model’s comfort zone. Beyond internal benchmarks, he highlighted real-world adoption signals: ChatGPT’s classroom use can improve learning when paired with strong teaching and curriculum, while relying on it as a solitary homework crutch can encourage shallow copying.

A major theme was infrastructure and compute. Altman said Project Stargate is designed to finance and build an unprecedented amount of compute—he characterized it as a gap between what AI can do today and what it could do with 10x, and someday 100x more compute. He described Stargate as a global effort to coordinate capital, technology, and operations, with large-scale sites already under construction, including a first site in Abilene representing about 10% of initial commitments. The scale is so large—he referenced a figure around $500 billion—that it’s less about “having the money in the bank” and more about deploying it over the next few years, assuming the hardware can be built.

Altman connected compute constraints to the broader product roadmap: memory features are becoming central because they let models hold context about a user’s life, and he called it one of his favorite recent ChatGPT capabilities. He also addressed privacy and monetization pressures. On privacy, he said OpenAI would fight a New York Times request to preserve consumer ChatGPT records beyond a 30-day window, calling it an overreach and arguing privacy must be a core principle for AI systems handling sensitive conversations. On ads, he said OpenAI hasn’t launched an advertising product yet and warned that altering the LLM output in exchange for revenue would be trust-destroying; any future approach would need to keep the response stream uncompromised.

Finally, Altman discussed hardware and interaction design. He said OpenAI is exploring new devices because today’s computers were built for a world without AI, and he expects the interface to become more context-aware and less dependent on typing. He suggested a future where an AI can sit through meetings, understand what can be shared, and then handle follow-ups—while noting that a consumer-ready device is “gonna be a while.” He closed with practical advice for adults: learn to use AI tools in your role, and build resilience, adaptability, and creativity—skills that remain valuable even as AI capabilities expand.