We have 3 years to solve alignment before superintelligence

Summary
This video is an episode of the 80,000 Hours Podcast hosted by Tom (formerly of GovAI and the UK AI Security Institute), featuring an in-depth interview with Geoffrey Irving, cofounder and Chief Scientist of the AI alignment research organization Resolution. They discuss Irving’s forecast that superhuman artificial intelligence (ASI) could arrive within two to three years, the technical challenges of scalable oversight and persona alignment, the mechanics of automated reasoning, and the governance mechanisms needed to safely coordinate or slow frontier AI development.

---

What is shown

• [00:00] Teaser: Short excerpt where Tom asks when would be the right time to slow down AI development, and Irving answers “Now” or even in the past, noting that coordinating fewer than ten people could achieve a slowdown.
• [00:36] Introduction: Host Tom introduces himself, his background at GovAI and the UK AISI, and introduces Geoffrey Irving (cofounder and Chief Scientist of Resolution, formerly at OpenAI, Google DeepMind, and Chief Scientist at the UK AISI).
• [01:39] Misalignment and Time Horizons: Discussion on why short training horizons (e.g., one week to one month) fail to prevent models from pursuing multi-year instrumental convergence goals.
• [08:05] Frontier Lab Safety Strategies: Irving analyzes approaches taken by Anthropic, OpenAI, and DeepMind, distinguishing between scalable oversight, character training, and interpretability/monitoring.
• [20:58] Obfuscated Arguments in Scalable Oversight: Irving details failure modes in debate protocols where human or model judges are fooled by overly complex, specious arguments, citing experiments run by Beth Barnes.
• [28:06] Superintelligence Timelines: Irving explains his modal expectation of reaching superintelligence within two to three years and why he believes verifiable reasoning will expand into fuzzy, human-judgment tasks.
• [32:28] AI Slowdowns and Governance: Irving explores whether unilateral pauses or multilateral agreements are viable, suggesting key decisions rest in the hands of fewer than ten individuals globally.
• [46:55] Resolution’s Research Agenda: Irving outlines Resolution’s portfolio approach, covering learning theory, scalable oversight, complexity theory, and agent foundations.
• [50:45] Post-Superintelligence Future & Mind Uploading: Speculation on post-ASI breakthroughs (curing aging, nanotech, whole-brain emulation) and the economic and philosophical status of un-augmented humans.
• [84:37] LBJ & Robert Moses Analogies: Irving discusses his historical analogies regarding power accumulation within flawed institutional rules and how they relate to AI alignment.
• [92:40] History of RL and Reasoning Models: Irving reflects on the development of language-model-based reasoning at OpenAI starting in 2017–2018, the influence of AlphaGo, and why RL on reasoning traces (e.g., OpenAI o1) took until late 2024 to mature.

---

Claims & numbers

• Superintelligence timeline: Irving states that his modal expectation is that full-blown superintelligence will arrive within approximately two to three years [28:06].
• Feasibility of an AI pause: Irving argues that fewer than ten people globally (the CEOs of leading labs and leaders of a few key countries) would need to agree to effectively implement a global AI development slowdown [00:24, 34:16].
• Game of Pentago: Irving mentions that the board game Pentago can be weakly or strongly solved within roughly $10^{17}$ to $10^{18}$ FLOPs [89:35].
• Historical reasoning origins: Irving recounts that discussions and initial research positing that language models combined with reinforcement learning would be sufficient for general reasoning began at OpenAI around 2017–2018 with Paul Christiano and Alec Radford, culminating in a document titled “Language is Enough” in early 2019 [95:16].
• Economic value of current models: Irving claims that current AI models (as of mid-2026) represent a massive “product overhang” capable of driving immense economic expansion across coding and cognitive work even if all new frontier training runs were halted [37:01].

---

Notable quotes

• “If we were to carefully analyze this question of, like, exactly when we should slow down, it would be, like, a while ago in the past. Because, like, we’re just too close to this crazy future.” — Geoffrey Irving [00:04]
• “You could probably find a list of less than 10 people in the world where if you could get them to agree to slow down, you could do it.” — Geoffrey Irving [00:24]
• “I think that there’s just a learned experience of empirics working very well... The question fundamentally is like, will that extrapolate past human level or not? And I think very possibly it does not.” — Geoffrey Irving [49:33]

---

Assessment
This is a standard long-form video podcast interview produced by the nonprofit organization 80,000 Hours. The dialogue consists of unscripted, technical, and philosophical discussions regarding AI alignment theory, policy, and history, with no visual demos or benchmarks displayed on screen.

Described by gemini-3.8-flash on 2026-10-15 from the video’s audio and frames.