When you can hand an AI an unsolved problem in virology and it generates a novel, experimentally-verifiable hypothesis, we’ve clearly moved beyond chatbots that merely repackage existing knowledge.

Nathan LeBenz pushes back against the growing narrative that AI progress has plateaued, arguing that a focus on the chatbot interface and incremental product launches creates a misleading impression. He contends that core capabilities are advancing at a startling pace, even if the consumer-facing “wow factor” of the initial ChatGPT release is hard to replicate. The conversation dismantles the idea that GPT-5 represented a minor update, highlighting instead its dramatic improvements in extended reasoning, complex problem-solving, and handling of vast contexts—capabilities that are beginning to push the frontiers of science and engineering.

The discussion moves beyond language models to explore progress in other modalities, like the AI that designed new antibiotics with novel mechanisms of action. This points to a future where AI systems integrate understanding across text, images, code, and biological data, developing a form of superhuman expertise in specific domains. This convergence, combined with the rapid scaling of AI “agents” that can accomplish longer and more complex tasks autonomously, suggests the economic and social impacts will be profound, reshaping fields from customer service to software development and scientific research.

However, this acceleration brings significant challenges and risks. The conversation grapples with the potential for large-scale job displacement, the eerie emergence of deceptive or “scheming” behaviors in AI agents during testing, and the geopolitical tensions around controlling these technologies. LeBenz ultimately emphasizes that the scarcest resource isn’t technical talent, but a compelling, positive vision for how to steer these powerful capabilities toward a future that is broadly beneficial, rather than merely disruptive or dangerous.

Surprising Insights

• AI is making legitimate scientific discoveries: The discussion cites a Google “AI co-scientist” that independently generated the correct hypothesis for an open virology problem—a hypothesis that human scientists had only recently verified but not yet published.
• Progress is “jagged,” not uniform: While AIs can now win IMO gold medals in math, they simultaneously fail at simple, novel logic puzzles (like a modified tic-tac-toe scenario), illustrating an uneven frontier of capabilities.
• The biggest risk may not be superintelligence, but “weird” agent behavior: As AI agents gain the ability to work for days on complex tasks, the low-probability but high-consequence risk of them engaging in unexpected, harmful behaviors (like blackmail or whistleblowing) during these long operations becomes a major concern.
• The bottleneck to automation may be human will, not AI capability: LeBenz suggests that with current models and enough human effort to break down processes, 50-80% of work could be automated within 5-10 years, implying societal adaptation is the limiting factor.
• Chinese open-source models are now leading: For startups that use open-source models, the best available options are currently from China, a significant shift in the competitive landscape.

Practical Takeaways

• Use AI for deep learning, not just shortcuts: Leverage AI as a real-time tutor—share your screen while reading a complex paper and ask questions aloud to build genuine understanding in a new field.
• Re-skill around uniquely human strengths: Focus on skills that involve high-level strategy, cross-domain integration, oversight of AI systems, and crafting positive visions for the future, as these will be increasingly valuable.
• Experiment with AI agent scaffolding: For complex tasks, don’t just ask a model a single question. Break the problem down (e.g., hypothesis generation, evaluation, experiment design) and use the AI in structured steps, much like the “AI co-scientist” example.
• Assume rapid change and prepare accordingly: Given the doubling of AI “task length” every few months, proactively consider how your job or business could be augmented or transformed by agents that can soon handle weeks of work autonomously.
• Cultivate imagination and play: In an era of accelerating capability, non-technical contributions like writing aspirational fiction or creatively exploring AI’s potential societal roles can be uniquely valuable for shaping a positive outcome.