Protein folding explained

The real guest in this episode is Dario Amodei, CEO of Anthropic, not Demis as the title suggests (likely an editorial slip). Dario lays out a timeline that is both concrete and jarring: around 2027, we will see AI systems capable of independently performing PhD‑level research. He insists this is not a semantic debate about AGI definitions—once a system can reliably do what normally takes years of doctoral training, the world has already shifted into a new phase.

His benchmark is disarmingly specific: AI that can tackle the kind of work a Nobel‑laureate scientist does. He does not believe chips and compute will be the hard roadblock before 2027; those are physical bottlenecks that supply chains and scaling momentum can still absorb. The real constraints lie in data, algorithms, and organisational capability. Anthropic’s core bet is that alignment and scaling must move in lockstep—you cannot scale first and figure out safety later.

The most counter‑intuitive idea sits in the safety discussion. While most labs talk about serious alignment work only once AGI looms, Dario argues the opposite: safety research is not a cleanup step after AGI arrives, it is the very process of reaching powerful AI. You cannot build a super‑intelligent brain and then slap moral band‑aids on it; the alignment problem must be woven into training from day one. This philosophy shapes everything at Anthropic—from Claude’s RLHF to Constitutional AI—and even spills into their business model. They deliberately stay API‑first, avoiding the consumer‑app layer, so they can focus on delivering raw agentic capabilities upstream. Dario calls this “restraint,” but it is really an organisational design decision to prevent short‑term commercial pressures from hijacking the safety agenda.

This logic invites pushback. If chips and compute are not the binding constraint, why push for export controls? Dario’s answer is pragmatic: controls are not meant to halt the spread of technology but to buy time for alignment research—just enough to keep the safety curve a fraction ahead of the capability curve. He also says they are not in the race to be “first to AGI,” which sounds noble yet reads to many in the industry as a competitive narrative: while everyone stares at the shiny prize, he positions himself as the one who cares about safety. It is a powerful brand stance, but it is also a gamble that a pure‑safety focus can survive in a market driven by raw capability.

In the end, Dario offers a sober countdown, no magic tricks—just an organisation betting its entire architecture on a simple belief: the stronger something becomes, the more it needs to learn, while it grows, how to live alongside human values.