Version 1

A converging discourse among AI's most prominent figures has crystallized around a shared but strategically calibrated message: today's AI systems are genuinely limited, even as their builders remain bullish on what comes next. Andrej Karpathy coined the term 'jagged intelligence' in August 2024 to describe AI's uneven capability profile—enormous performance spikes in coding and mathematics coexisting with failures on tasks humans find trivial[1][2]. Rather than a smooth capability curve that rises with model scale, Karpathy frames the landscape as an irregular profile where 'enormous spikes' sit beside deep troughs[2]. The concept has since propagated widely: by April 2026, the New York Times was using it to 'reframe the AI debate'[3], and a year-end retrospective had already declared jagged intelligence 'a fault line' for 2025[4]. Karpathy has extended this framing into a broader paradigm he calls 'Software 3.0,' in which LLMs become the core compute interpreter in an agent-native future—a vision now circulating alongside the jagged intelligence thesis[5][6].

Sam Altman has offered his own version of this candor. In a recent podcast widely circulated across social platforms, he stated plainly that 'models are still quite dumb relative to what they will be' and acknowledged a further limitation: current AI has only shallow awareness of individual users' lives, forcing users to 'massage them, cajole them, and try to get the thing that you want'[7]. Altman has also claimed he does not expect to be personally smarter than GPT-5[8], a framing that simultaneously humanizes current AI's inadequacy and implies imminent surpassing of human cognitive benchmarks. The message is consistent across formats—short social video clips, podcasts, and written posts[9][10][11][12]—suggesting a deliberate communications posture rather than an isolated admission.

Demis Hassabis has staked out a distinct but complementary position. While predicting human-level AI within five to ten years as of early 2025[13][14], he has also reframed AGI's purpose: in recent public remarks, he characterized AGI not as a project to replace humans but as 'a science question about what counts as truly general computation'[15][16]. This scientific-curiosity framing is paired with acknowledgment that current AI still falls short of human reasoning in key respects[17]. Hassabis's dual message—transformative timeline, non-threatening purpose—mirrors the broader pattern among AI leaders of acknowledging limits while sustaining long-term ambitions.

The discourse has attracted notable skepticism. Critics have questioned whether these public acknowledgments of AI's 'dumbness' function more as expectation management than genuine epistemic humility, with outlets like Jacobin and independent commentators explicitly scrutinizing the gap between leaders' rhetoric and their organizations' commercial ambitions[18][19]. The 'jagged intelligence' framing itself has also divided commentators: some view it as a clarifying conceptual tool[20][6], while Forbes and others have deployed it to highlight what they frame as the 'illusion of reasoning' in modern LLMs[21]. As of early May 2026, social media amplification of all three figures' positions is accelerating, with accounts summarizing Karpathy, Altman, and Hassabis in the same breath[2][7][15]—signaling that the public is increasingly treating these as a unified, if uneasy, leadership consensus.