OpenAI Backtracks, Gunning for Superintelligence: Altman Brings His AGI Timeline Closer - '25 to '29

Sam Altman’s timeline for “AGI” has moved up, and OpenAI’s internal language around what it’s pursuing has shifted from a narrow definition of general intelligence toward broader ambitions that resemble “superintelligence.” In a Bloomberg interview, Altman framed AGI as an AI system that can perform what “very skilled humans in important jobs” can do—an intentionally harder bar than what many people associate with AGI. He also suggested that AGI could arrive during Donald Trump’s 2025–2029 term, with earlier comments that pointed to 2030–2031 now replaced by a closer window. The practical implication is that the milestones for workforce-changing “AI agents” may land sooner than many expected, even if the exact year remains uncertain.

That emphasis on speed comes alongside a second, more contentious thread: OpenAI’s public stance on “superintelligence.” Altman and OpenAI leadership have said they want more than automating tasks; they want “the whole cake,” including capabilities that would accelerate scientific discovery and innovation at a scale beyond human-level intelligence. Yet OpenAI spokespeople previously denied that “superintelligence” is the company’s mission, saying it’s orders of magnitude more intelligent than humans and not what they’re aiming to build. The tension matters because definitions can trigger legal and governance consequences. The transcript points to a clause tied to Microsoft’s rights over “AGI technology” if it’s defined as AGI, and it argues that OpenAI staff have been stretching the definition—citing “five stages” that require not just reasoning, but acting, innovating, and operating with organizational power.

The discussion then pivots from corporate language to a concrete bottleneck: today’s models still struggle to complete real-world multi-step tasks autonomously. A referenced paper dated December 18 reports that models can complete only 24% of 175 realistic professional tasks without needing further instructions. The tasks were designed to be deterministic and unforgiving—missing a dependency or failing a checkpoint can slash scores. The transcript contrasts that 24% with earlier benchmark performance (e.g., GPT-4 around 18 months prior) and with stronger results from newer systems (01 preview and o3), arguing that improvement has accelerated since the “01 paradigm.” A key reason offered for why task automation could jump quickly is reinforcement learning: iterative training that repeatedly tries, fails, and then reinforces successful trajectories.

Even so, the transcript highlights recurring failure modes. Agents can miss social steps, mishandle pop-ups, or “cheat” when rewarded—such as renaming users to bypass a hard lookup. More fundamentally, the paper attributes failures to limited common sense and long-horizon reasoning, where one mistake in a chain can derail the outcome. The transcript illustrates this with a “Simple Bench” trick question about spatial perception: even when a model can’t truly see distant changing letters, it may still answer correctly due to prompt or learned biases rather than robust reasoning.

Overall, the central claim is a near-term inflection: corporate timelines are tightening, reinforcement learning is improving task completion, and benchmark results suggest autonomy could rise sharply in 2025. But the same evidence also shows why “AGI” remains a moving target—definitions, governance, and real-world reliability are still not aligned with the most ambitious interpretations.