Will Artificial General Intelligence (AGI) Arrive by 2030?

AI capabilities have been doubling roughly every 12-18 months, a pace faster than the historical Moore's Law trajectory for semiconductor performance. Frontier models have saturated major benchmarks within two to three years of their introduction, and successive generations — from GPT-3 to GPT-4 to o3 — demonstrate qualitatively new reasoning behaviors. Agentic systems that autonomously browse the web, write and execute code, and coordinate multi-step tasks are already deployed in production, collapsing the gap between narrow AI and the autonomous action that AGI requires.

Whether simply scaling current transformer architectures will produce AGI is one of the most contested questions in AI research. Empirical scaling laws (Hoffmann et al., Chinchilla) show predictable gains in language tasks with more compute and data, but critics including Yoshua Bengio and Gary Marcus argue that LLMs hit a ceiling on tasks requiring robust causal reasoning, physical world models, and compositional generalization. The ARC-AGI benchmark — designed to resist memorization — still sees frontier models scoring well below human average, suggesting architectures may need fundamental changes rather than mere scale.

The global GPU buildout has dramatically increased the compute available for training frontier models. Nvidia's H100 and upcoming Blackwell chips, combined with hyperscaler investment from Microsoft, Google, and Amazon exceeding $200 billion annually, ensure that raw training compute will grow by orders of magnitude before 2030. Historical capability jumps in AI have closely tracked compute increases, and sovereign AI programs in the US, China, and EU are further accelerating infrastructure deployment.

Growing awareness of AGI risk — from existential safety researchers to mainstream policymakers — is creating genuine friction. The EU AI Act's risk-tiering framework, proposed US executive orders on frontier model oversight, and Anthropic's Constitutional AI approach all create incentive structures that may slow the most aggressive development paths. If a system meeting AGI criteria is built but deemed unsafe to deploy, public recognition of AGI arrival may be deliberately delayed or suppressed, creating an asymmetric reporting risk for prediction market participants.

“Altman has consistently moved his personal AGI timeline forward, telling staff and investors he believes AGI — defined by OpenAI as a system that outperforms the median human professional at most economically valuable cognitive tasks — could arrive within a few years. His essay 'The Intelligence Age' and subsequent public appearances place his median estimate at 2027-2028. Critics note his commercial incentives to promote AGI proximity as a fundraising narrative, but OpenAI's o3 model achieving 87.5% on ARC-AGI with high compute has given technical credibility to the accelerated timeline.”

Fonte: Sam Altman, OpenAI CEO

“Hinton, who left Google in 2023 citing concerns about AI risk, now estimates a 10-20% chance of AI causing human extinction within 30 years. His AGI timeline has compressed since 2022, moving from 'decades away' to 'possibly within a decade.' Unlike some researchers, Hinton believes current neural network architectures are on the right conceptual path to AGI and that the main remaining obstacles are engineering rather than fundamental. He strongly advocates for international safety frameworks before AGI deployment.”

Fonte: Geoffrey Hinton, Turing Award winner

“Bengio, one of deep learning's founding figures and a prominent safety advocate, argues that current architectures lack causal reasoning, physical world understanding, and the compositional generalization needed for AGI. He places the probability of AGI by 2030 below 5% and believes achieving it would require genuine scientific breakthroughs with no credible near-term path. He has testified before the Canadian Parliament on existential risk and co-signed major AI safety letters calling for coordinated international governance.”

Fonte: Yoshua Bengio, Turing Award winner