AGI is still 30 years away — Ege Erdil & Tamay Besiroglu

Tamay argues that calling AI progress an “intelligence explosion” is like describing the Industrial Revolution as a “horsepower explosion”: it singles out one measurable capability while overlooking the complementary changes that transformed production. The Industrial Revolution involved agriculture, transport, finance, law and urbanization as well as physical power. Likewise, highly capable AI may be one important component of a transition, but its economic effects depend on other sectors and systems changing alongside it.

The disagreement becomes concrete in the timeline for a fully automated remote worker. Ege estimates around 2045 if “drop-in replacement” means doing literally everything; Tamay is more optimistic but says the forecast depends on what counts as full replacement. Dwarkesh points to rapid recent advances in reasoning and coding, while Ege argues that extrapolating visible progress does not establish when the whole economy will be automated.

The guests offer a capability-based way to reason about timelines. Tamay describes roughly nine to ten orders of magnitude of training-compute growth since AlexNet, accompanied by a handful of major advances: complex games, language, abstract reasoning, coding and perhaps mathematics. If broad human-level performance requires further abilities such as long-horizon coherence, autonomy and multimodal understanding, the pace of past breakthroughs offers one reference point. But the earlier period also benefited from rapid compute scaling, which is getting harder as energy, GPU supply and data-center infrastructure become constraints.

Dwarkesh notes that the current AI build-out is still a small share of GDP and that even leading-edge chip production is not mostly devoted to AI. The speakers treat this as evidence that substantially more investment may be needed to unlock additional capabilities, and ask whether the economy can sustain that level of scaling. Ege also challenges the claim that current models are nearly complete remote workers: he says computer-use systems still struggle with ordinary tasks such as booking a flight, and argues that an isolated task does not represent a whole job.

The booking example highlights a distinction between automating a task and automating a role. A travel agent may book flights, but that may occupy only a small fraction of the job; other tasks must be integrated as well. Ege sees this as a major difference between his view and more optimistic forecasts: many jobs require a broader set of competencies than their simplest visible task suggests.

Dwarkesh presents the “unhobbling” hypothesis: current models may already contain substantial general intelligence, but limited training data, short context, lack of inference-time deliberation and missing workplace context prevent them from being productive. He points to ChatGPT’s post-training and reasoning models as cases where relatively small additions appear to unlock striking capabilities. Tamay responds that impressive systems such as AlphaZero did not become general simply by adding text and context; new capabilities may require changes to how models are trained.

Tamay cautions against judging a capability by how little compute its final training stage seems to require. Reasoning models depend on years of prior work: large-scale hardware, internet-scale training data, faster inference and distillation. What looks like a small final step is built on a deep technological stack. He allows that agency could later appear similarly simple in retrospect, while arguing that the complementary innovations needed to make models competent agents may take years.

Ege offers AlphaGo and AlphaZero as a warning against assuming that an impressive narrow capability will transfer broadly. Their success at games did not make them competent mathematical agents. He therefore doubts that current language models will become generally capable at agency through the same kind of incremental additions that improved reasoning. The discussion leaves open whether future methods will unlock that capability; this excerpt establishes the disagreement rather than resolving it.

The speakers distinguish a system that can complete one bounded computer task from an agent that can sustain a long sequence of actions. Dwarkesh cites METR evaluations, where the length of tasks models can complete appears to double every seven months, and asks whether extrapolating that trend could yield month- or year-long tasks by 2030. He also argues that reasoning may have seemed more complicated than it was: chain-of-thought methods can produce large gains with relatively little additional compute.

Ege accepts that AI has solved difficult reasoning tasks, but stresses that games such as chess, Go and StarCraft take place in narrow, fixed environments. A system that reacts well to a specific game does not thereby acquire general sensorimotor skills or the ability to learn an unfamiliar game. The Pokémon example sharpens this distinction: Claude may know what a player should do in Mount Moon, yet still get stuck there for 48 hours when it has to act. Explicit knowledge does not automatically translate into competent behavior.

Tamay names several developments that would make him more optimistic: robust long-context and multimodal reasoning, integration across systems, and sustained agency that can complete long, unfamiliar tasks. One concrete test would be finishing a newly released Steam game with little or no training material. He also discusses OpenAI revenue as a possible signal, but distinguishes a $100 billion milestone, which he considers relatively likely, from $500 billion, which would change his view much more. Ege and Tamay caution that willingness to pay for a product does not by itself show that the product will transform the economy.