How Close Are We to AGI? Inside the Race for General Intelligence

Artificial intelligence models can already design molecules, generate working code and produce fluent prose, yet they still stumble at simple puzzles a layperson solves in minutes. That gap — impressive narrow capabilities paired with surprisingly brittle reasoning — captures the central challenge of artificial general intelligence (AGI): building systems that perform robustly across many domains and contexts.

The gap between narrow skills and general reasoning

Today's models excel at narrowly defined benchmarks: predicting the next token, optimizing a molecule for a target property, or writing code to a specification. But their behavior often lacks the kind of flexible, goal-directed reasoning people use when switching tasks, improvising with incomplete information, or planning across time and modalities.

That mismatch fuels the core question: can ongoing advances in model architectures, training data, compute, and system design converge into machines that rival or surpass human performance across virtually all tasks?

What leaders are predicting

Dario Amodei, co-founder of Anthropic, forecasts that some form of "powerful AI" could arrive as early as 2026. He envisions systems with domain expertise at a Nobel-Prize level, the ability to switch seamlessly between text, audio and the physical world, and the autonomy to reason toward goals instead of merely answering prompts.

Sam Altman, CEO of OpenAI, says AGI-like properties are "coming into view," comparing the potential societal impact to transformative technologies like electricity and the internet. He attributes progress to steady gains in training methods, data availability, and compute resources, along with falling costs and a rapidly rising socioeconomic value.

Aggregate forecasts and timelines

Optimism extends beyond founders. Surveys of experts give at least a 50% chance that AI systems will hit several AGI milestones by 2028. One expert survey estimated a 10% chance that unaided machines will outperform humans at every task by 2027, rising to 50% by 2047. As breakthroughs accumulate, reported median time horizons have shortened dramatically — what once looked like decades became years.

Ian Bratt, vice president of machine learning technology at Arm, notes that "large language and reasoning models are transforming nearly every industry," reflecting how incremental improvements cascade into rapid change across sectors.

What could enable AGI?

Several enablers are frequently cited: larger and better-curated datasets, algorithmic innovations that improve generalization and reasoning, orders-of-magnitude increases in compute and training efficiency, and new ways to integrate multimodal inputs and physical interaction. Equally important are systems-level advances: orchestration, safety mechanisms, and tooling that let models act reliably in the real world.

The combination of hardware, software, and socio-technical infrastructure — including cost-effective compute and institutional investment — will likely determine how quickly and safely more general systems emerge.

Uncertainties and risks

Timelines are contested and dependent on assumptions about scaling laws, emergent capabilities, and societal priorities. Even if AGI-like capabilities arrive sooner than many expect, the consequences will depend on governance, deployment choices, and how well safety and alignment research keeps pace with capability growth.

Notes on the source

This overview is based on reporting and a custom Insights report produced for MIT Technology Review. The Insights content was researched and written by human authors and reviewers; the original piece details survey results, expert quotes, and broader context about the potential paths to AGI.