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

Core Thesis on AI Progress and Economic Growth

• Tame Bessaroglu and Ege Erdog argue that explosive economic growth (e.g., 30%+ annual rates) is plausible due to the simultaneous scaling of labor, capital, and data, rather than a single "intelligence explosion" driven solely by reasoning.
• They characterize the "intelligence explosion" as a misleading concept, analogous to calling the Industrial Revolution a "horsepower explosion," which ignores complementary innovations in law, finance, urbanization, and supply chains.
• The speakers predict that the automation of remote work (tasks involving digital labor) will not be complete until around 2045, a timeline significantly longer than the 2027–2030 estimates held by many San Francisco-based AI proponents.
• Ege Erdog adopts a more bullish stance than Tame, citing rapid progress from ChatGPT (2 years ago) to current coding and reasoning models to argue that full automation of remote work could be achieved in roughly 20–25 years.
• The speakers contend that AI reasoning capabilities, while impressive, do not yet produce novel mathematical concepts or scientific discoveries; they excel at solving known problems but lack the creative innovation of human researchers.
• Tame Bessaroglu identifies a "more of X" paradox: tasks humans find difficult (abstract reasoning, playing Go, complex math) are where AI progresses fastest, while tasks humans find easy (general agency, long-term planning in unstructured environments) remain AI's biggest hurdle.
• Evidence for AI progress is tracked via "task length," with the speakers noting a doubling of the time horizon for tasks AI can complete every seven months, suggesting a trajectory toward long-term coherent agency by 2030.

The Limits of "Software-Only" Singularities and Alignment

• The speakers reject the "software-only singularity" argument, positing that AI progress is not driven solely by cognitive effort but is fundamentally constrained by the need for experimental compute, data collection, and hardware scaling.
• They estimate that 9–10 orders of magnitude of compute have been used since 2012 to unlock capabilities, and only 3–4 orders of magnitude remain before the cost of scaling (energy, GPU production, infrastructure) reaches a non-trivial fraction of global GDP.
• Regarding alignment and safety, the speakers argue that pausing development is counterproductive because progress on alignment requires the same scaling of compute and data that enables capability growth; 2016-era compute was insufficient to learn the principles now in use.
• The cost of delay is quantified as potentially tens of trillions of dollars per year in lost utility for the current population, far outweighing the marginal reduction in risk from pausing.
• They argue that "alignment taxes" are likely overstated because the primary bottleneck for AGI is not just intelligence but the integration of AI into the physical economy and supply chains.
• On the risk of AI takeover, the speakers express skepticism that superintelligent AI would seek to dominate humans, noting that in many historical conflicts (e.g., East India Company), integration and trade were more profitable than conquest due to the high value of the existing human economy.
• They suggest that as AI systems become more capable, their incentives will shift toward negotiation and mutual benefit rather than conflict, as the cost of war (destroying the very economy that powers them) becomes too high.

Macro-Economic Mechanisms and "Learning by Doing"

• Technological progress is viewed as a result of "learning by doing" and complementary inputs rather than isolated genius; for example, the light bulb required not just an idea but the development of power grids, filament materials, and manufacturing infrastructure.
• Explosive growth requires the simultaneous scaling of multiple factors: capital accumulation, labor force expansion (human and AI), and the development of supply chains; scaling just R&D or just capital is insufficient.
• The speakers predict that regulatory heterogeneity will drive differential growth rates, with some jurisdictions (potentially those with more permissive norms like the UAE) adopting AI faster than others, similar to the initial spread of the Industrial Revolution.
• They refute the "Baumol cost disease" objection to explosive growth, arguing that even if some sectors (e.g., healthcare) grow slowly, the reallocation of labor from automated sectors to non-automated ones will drive massive overall productivity gains.
• The "O-ring" theory (where one failed component collapses the whole system) is deemed optimistic for growth, as AI's flexibility allows for substitution and scaling even if specific bottlenecks exist, provided the bottleneck is not total.
• The speakers anticipate that the primary obstacle to explosive growth is not technical capability but the political and regulatory coordination required to scale deployment globally; however, national security and economic competition create strong incentives to bypass such constraints.

Organizational Evolution and the Future of Firms

• AI firms will possess a distinct evolutionary advantage over human firms due to high-fidelity replication, allowing for the copying of tacit knowledge and "hyper-agents" (e.g., copying the collective knowledge of an entire engineering team or a CEO like Jensen Huang).
• This replication eliminates the "principal-agent problem" and cultural dilution, enabling organizations to maintain a single coherent vision and scale indefinitely without the friction of human turnover.
• Central planning is argued to be more viable in an AI-dominated economy due to superior communication bandwidth, the ability to disaggregate sensing and processing (e.g., Tesla's data flywheel), and the elimination of misaligned human incentives.
• The speakers suggest that AI firms will not operate in isolation but will integrate deeply into existing global supply chains, leveraging the "data stock" of the internet and the existing physical infrastructure of the human economy.
• They predict that the transition to an AI economy will not be a "software singularity" confined to a desert, but a broad integration where AI drives growth across all sectors, potentially leading to 10x or 100x increases in global GDP over decades.

Career Advice and Epistemic Strategy

• The speakers advise aspiring researchers and thinkers to avoid deliberate, linear career planning, instead favoring curiosity-driven exploration and high-bandwidth collaboration with existing experts.
• They emphasize the importance of reading "key literature" (e.g., Romer, scaling law papers) and prioritizing high-quality information sources (like specific podcasts or social networks) over general reading to maximize information density.
• A critical strategy for productivity is "reaching out" aggressively to build a network of collaborators, as the marginal utility of individual insight is low without access to the collective knowledge and feedback of a community.
• They recommend treating current AI uncertainty with "classical liberal" principles (decentralization, flexibility) rather than rigid, high-volatility central planning, as specific policy plans will likely become obsolete given the pace of change.
• The speakers suggest that the most valuable contribution one can make is to participate in a community of thinkers working on these problems, rather than attempting to solve them in isolation.