AGI timelines: expert disagreements
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
From model scaling and inference levers to societal preparedness, the 80,000 Hours podcast's compilation of expert disagreements reveals that the only consensus on AGI timelines is that there is no consensus—yet even the most cautious experts now believe transformative AI is within a decade.
A Year of Whiplash
In early 2025, the release of OpenAI's o1 and o3 reasoning models sent shockwaves through the AI community. Sam Altman declared that "we are now confident we know how to build AGI," and Demis Hassabis estimated AGI was "probably three to five years away". The "AI 2027" scenario—in which AI research and development is fully automated, triggering a recursive self-improvement loop and intelligence explosion—dominated public discourse.
But by late 2025, sentiment had swung dramatically in the opposite direction. Forecasts for transformative AI blew out further than they had been even before reasoning models arrived. Host Rob Wiblin described this as "the timelines madness of 2025"—a period of extreme whiplash driven by deeper technical revelations about what reasoning models could and could not achieve.
The Technical Reality Behind the Swing
The initial optimism rested on two pillars: models appeared to be genuinely smarter, and they were given much longer to think through problems. But as 2025 wore on, the shine wore off.
The key revelation was that more than two-thirds of the improved performance from reasoning models came from inference scaling—giving models more time to think—rather than fundamental improvements in architecture or learning. This approach is fundamentally unsustainable: it requires exponentially more compute per query, hitting hard physical and economic limits. "There actually aren't enough computer chips in the world to go on and give models 10 minutes or 100 minutes to think," Wiblin explained.
The second pillar—reinforcement learning on checkable domains like maths and coding—also failed to generalise as expected. AI leaders had hoped that teaching models to reason in easily verifiable domains would transfer to messier real-world tasks. It did not. A senior staff member at an AI company told Wiblin that this experience "actually updated them towards longer timelines to artificial general intelligence".
A Spectrum of Expert Views
The 80,000 Hours compilation features 15 opinionated highlights from years of interviews, revealing the full spectrum of disagreement .
The Extremes: Ian Morris, a historian, argued that "the one thing we can bet the farm on is that we go extinct" and dismissed the "business as usual" scenario as "staggeringly unlikely".
The Optimists: Tom Davidson made the case for "crazy-sounding explosive growth stories," while Michael Webb explained why he does not buy them.
The Pragmatists: Rohin Shah questioned whether humanity's work ends at the point it creates AGI or whether "it's just the start of a new set of difficult and annoying choices".
The Sceptics: Hugo Mercier argued that even superhuman AGI might not be that persuasive, while Rob Long presented the case for and against digital sentience.
Rob Wiblin himself admitted that his views had evolved: "I wouldn't say all of these things now and the focus of the conversation would be different. I was more relieved listening back to this than anything else".
The RAND Analysis: Definitional Ambiguity Isn't Everything
A comprehensive RAND Corporation analysis published in March 2026 examined AGI forecasting methodologies and identified a crucial finding: much apparent disagreement reflects different definitions of AGI, but "substantial disagreement remains even when definitions and information are held constant." People with similar training, working in the same organisations and looking at the same data, often reach very different conclusions.
The analysis identified three key limitations in the forecasting infrastructure: the lack of resolved forecasts for calibration, benchmarks resistant to saturation, and independent validation of influential models. The authors concluded that "decisionmakers are making decisions based on methodologies that are in nascent stages of development".
The Unexpected Consensus
Despite the volatility of 2025, the new consensus among experts who once dismissed the idea is stark: even the most skeptical voices now project transformative AI within a decade. This shift is alarming because societal institutions are ill-equipped to handle the scale of disruption—from labour markets and scientific norms to democratic processes and epistemic trust—that AGI would cause.
Benjamin Todd, 80,000 Hours co-founder, argued that even a 10-year horizon is "dangerously short, demanding urgent action despite the uncertainty of exact timing". He laid out three scenarios—from AGI by 2029 to a decades-long plateau—and explained that the optimal career strategy depends on where you are in your career and what leverage you have.
The GFN Context
For Global Future Nexus, the AGI timeline debate underscores the urgency of anticipatory governance. Whether AGI arrives in three years or three decades, the governance, identity, and trust frameworks cannot wait for certainty. As the RAND analysis concluded, the policy question is not "when will AGI arrive?" but "how should we prepare for a range of possible AI futures?" GFN's Borderless AGI Vision—with its focus on Artificial Personhood, cross-species trust, and adaptive governance—is precisely the infrastructure required for a world where the only certainty is transformation.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
