2025 AGI Predictions Scorecard: Who Got AGI Timelines Right?

UPDATED MAY 2026

Executive Summary

No major AI figure’s 1–3 year AGI prediction has held up, and several have quietly pushed their dates forward without admitting the miss. The most credible forecasters — Kurzweil, Hinton, Karpathy — share three traits: decade-scale thinking, rigorous definitions, and willingness to self-correct. Claims tied to product launches or funding (Huang’s “AGI already achieved,” Anthropic’s Mythos hype) deserve scrutiny, especially after independent testing undercut Anthropic’s own claims. Timelines have compressed from 50–100 years to 5–15 — but that’s not the same as consensus.

Introduction

Hype or reality? Based on extensive research, I review which AI predictions came true and which have failed, including predictions by the biggest names in AI including Musk, Altman, Amodei, Hinton, Kurzweil, Karpathy, and Marcus. See what they predicted, what actually happened, and what they’re saying now.

Key Takeaways for 2026

What the Best Predictions Have in Common

1. Long-term thinking over short-term promises. Kurzweil’s 25-year forecast has outperformed every 1–3 year claim on record.
2. Evidence-based updating. Hinton’s credibility comes from correcting himself publicly on the topic of radiologists. The willingness to say “I was wrong” is a signal of reliability, not weakness.
3. Rigorous definitions. Suleyman says “including physical robotics,” whereas Huang says “passing tests.” Those are different bars, so the best predictions are the ones that have rigorous definitions and clear standards to measure AGI success.
4. Conflicts of interest named, not ignored. Huang sells chips, Altman raises billions, and Musk competes through xAI. These incentives don’t make predictions wrong but they make them less than neutral.

What Consistently Fails

1. Aggressive short-term specificity without accountability. No 1–3 year AGI prediction from any major figure has been verified. Furthermore, several have been quietly moved forward without acknowledgment.
2. Domain overconfidence. Radiologists, truck drivers, and software engineers — specific industry replacement predictions have consistently underestimated human adaptability and AI’s tendency to augment rather than substitute on the first wave.
3. Unchanged positions despite new evidence. Rigidity in a fast-moving field is a credibility problem.

At a Glance: Every Expert, One Line

What Is AGI? A Working Definition

The definition problem matters before you score anyone. When Musk says AGI is here and LeCun says it’s a decade out, they are often describing different thresholds — not disagreeing about the same thing. The most credible predictors define their terms first.

2025 Predictions: Hit or Miss?

The following predictions had 2025 or earlier as their explicit target. With that year behind us, these can be scored against what actually happened.

Mustafa Suleyman

Target: 2029–2031

Suleyman calls 2-year AGI timelines “ungrounded” and “over the top.” He is the only major figure whose definition explicitly requires physical robotics capability, not just digital intelligence. Jensen Huang can declare AGI by redefining the threshold; Suleyman’s definition doesn’t move.

Sam Altman

Target: 2025–2028

Altman has the strongest delivery track record of anyone on this list — he ships things. The problem is the gap between what he ships and what he claims. His January 2025 blog post declared OpenAI was “confident we know how to build AGI as we have traditionally understood it” and predicted 2025 arrival. Yet, it didn’t arrive. He has also told Bloomberg he thinks AGI “will probably get developed during Trump’s term” while simultaneously calling AGI “a very sloppy term.”

Geoffrey Hinton

Target: 2028–2035

In 2016 Hinton declared “People should stop training radiologists now,” predicting AI replacement within 5 years. On the contrary, radiologist employment increased and AI became an assistive tool. He has publicly acknowledged this miss. In late 2025 he told CNN he is “probably more worried” about AI than when he left Google in 2023 — “it’s progressed even faster than I thought.”

Dario Amodei

Target: 2026–2027

Amodei’s definition is the most specific and evaluable on this list — when 2027 arrives, it can actually be scored. His two essays are the most substantial written record of any AI CEO’s worldview: Machines of Loving Grace laid out the upside; The Adolescence of Technology confronted the risks, arguing we are “considerably closer to real danger in 2026 than we were in 2023.”

Mythos: In March 2026 a data leak revealed Anthropic had built Claude Mythos — described in a leaked draft as “currently far ahead of any other AI model in cyber capabilities,” able to autonomously find thousands of critical vulnerabilities and collapse the window between discovery and weaponization to hours. Anthropic restricted access via Project Glasswing and warned government officials. By May 28, 2026, Anthropic announced plans to widely release Mythos-level models within weeks — from too dangerous to shipping in approximately two months, weeks before a reported IPO.

The vendor’s claims deserve the same scrutiny applied to everyone else here. Daniel Stenberg, creator of curl, received a Mythos scan of curl’s 178,000-line codebase in May 2026:

• Mythos reported 5 “confirmed” vulnerabilities
• The curl security team confirmed 1: a low-severity bug planned as a CVE in curl 8.21.0
• The other 4 were 3 false positives and 1 deemed “just a bug”
• curl had already been scanned by AISLE, Zeropath, and OpenAI Codex Security before Mythos ran — those tools triggered 200–300 bugfixes in the prior 8–10 months
• Stenberg’s conclusion: “the big hype around this model so far was primarily marketing. I see no evidence that this setup finds issues to any particular higher or more advanced degree than the other tools have done before Mythos”

Amodei was also VP of Research at OpenAI during the 2019 “GPT-2 is too dangerous to release” episode — the same arc, now widely read as a marketing moment.

Andrej Karpathy

Target: ~2035

Karpathy co-founded OpenAI and ran Tesla Autopilot for five years. He is not a structural AI pessimist — he described a hard shift in December 2025 where he essentially stopped writing code by hand. He calls the current moment real and consequential, but won’t call it AGI. His Waymo example is the load-bearing one — perfect self-driving demos in 2014 but widespread deployment still not economical or fully reliable in 2025. That gap is the march of nines.

Hassabis, Kurzweil, Huang, Marcus, and the Skeptics

Demis Hassabis

Target: 2031–2036

Ray Kurzweil

Target: 2029

Jensen Huang

Target: Claims already achieved

Gary Marcus

Target: ~2035+ (different architecture required)

Marcus’s position breaks into three claims:

1. Current LLMs are structurally incapable of AGI — “deeply flawed imitators preying on the Eliza effect.”
2. AGI is not impossible — “Anyone who thinks AGI is impossible: wrong. Anyone who thinks AGI is imminent: just as wrong.” The current path cannot get there; the technology that achieves it probably hasn’t been invented yet.
3. AGI and ASI are different thresholds, and we may reach something resembling ASI through compounding narrow capabilities without ever clearly crossing an AGI threshold. ASI — recursively self-improving superintelligence — is “not very likely in the next decade and not super likely in the decade after, but certainly possible.” This is a more unsettling scenario than either the optimist or standard skeptic positions describe.

Yann LeCun and Andrew Ng

LeCun (Meta) argues AGI requires missing architectural components — persistent memory, world models, common-sense reasoning — and estimates roughly 10 years, a dramatic compression from his prior “decades” framing. A 2022 prediction that GPT-class models could never acquire basic physical intuition was falsified by GPT-4 within a year. Andrew Ng maintains that near-term AI utility matters more than the AGI milestone debate and holds a decades-long timeline for full generality.

Frequently Asked Questions

Conclusion

The field moves faster than skeptics expect and slower than optimists promise. The most reliable forecasters think in decades, define their terms precisely, and correct themselves when the evidence demands it.

By those criteria, Kurzweil’s 2029 target — confirmed as recently as April 2026 — remains the most defensible AGI prediction in the historical record. Amodei’s 2026–2027 window is the most specific live test on the board. Musk’s end-of-2026 prediction is on the clock with the same definitional ambiguity that has shielded every prior miss.

The macro shift worth keeping: in 2010, mainstream researchers estimated AGI was 50–100 years away. By 2026, the expert median sits at 5–15 years. Even the skeptics have moved. That compression is the most significant data point in this entire analysis — not consensus on when AGI arrives, but consensus that arrival is no longer a question measured in generations.

Related reading:

• • The Ultimate Guide to 15 Types of Artificial Intelligence
  • Karpathy Joins Anthropic: Why This Is Important for AI Visibility
  • What Is Answer Engine Optimization? The Complete Guide