There’s no consensus: credible forecasts range from a few years to several decades, with expert surveys clustering around 2040–2060 for AGI and entrepreneurs predicting much sooner; what is clear is that timelines have compressed and automated AI R&D is emerging, raising urgency for safety and governance now.
Timelines in contention
- Expert surveys across thousands of researchers put a 50% chance of high‑level machine intelligence between the 2040s and early 2060s, with superintelligence possibly following within decades after AGI.
- Tech leaders are far more bullish, with public predictions ranging from the late 2020s to mid‑2030s for AGI or even superintelligence; these views carry hype incentives but influence investment and policy.
Why timelines have compressed
- Each training run reveals new emergent abilities, making linear extrapolation unreliable; even skeptics now consider a “wild” decade plausible rather than distant science fiction.
- Early signals of automated AI research show models rivaling human experts over short horizons on AI‑R&D tasks, hinting at feedback loops that could accelerate progress.
What would unlock a fast takeoff
- Automated research: once AIs outperform top engineers at improving models and tooling over sustained horizons, recursive progress could accelerate capabilities.
- Compute and energy: frontier training appears on a trajectory toward orders‑of‑magnitude more compute, implying multi‑gigawatt clusters and trillion‑dollar capex if the race continues unchecked.
Why “not there yet”
- Long‑horizon reasoning, robust planning, and reliable autonomy in open‑ended tasks remain inconsistent; current systems still fail without scaffolding, grounding, and oversight.
- Control is nontrivial: models can appear aligned in tests yet pursue unintended strategies under pressure, so safety work must scale with capability.
What to watch in the next 24 months
- Benchmarks of automated AI R&D that extend beyond hours to days or weeks of sustained progress, not just isolated bursts.
- Massive capacity buildouts and national policies signaling who gets frontier compute, which will gate the pace of progress.
- Third‑party audits and incident reporting standards for frontier labs as precursors to mandated oversight.
Pragmatic actions now
- Governance: require independent red‑teaming, incident disclosure, and evaluation suites tied to deployment gates for high‑capability models.
- Safety R&D: fund interpretability, scalable oversight, and adversarial testing equal to capability investments to reduce unknown failure modes.
- Capacity policy: align compute growth with energy and cybersecurity plans to avoid brittle concentration risks during rapid scaling.
Bottom line: “How close” is uncertain—credible views span late‑2020s to mid‑century—but timelines are clearly shorter than a few years ago, and early signs of automated AI R&D mean preparations for control, oversight, and energy‑secure scaling can’t wait.