Other
The cases for and against AGI by 2030 (article by Benjamin Todd)
- Experts including OpenAI's Sam Altman, Anthropic's Dario Amodei, Google DeepMind's Demis Hassabis, and Benjamin Todd project AGI is a plausible outcome by 2030, with Amodei expressing high confidence in reaching powerful capabilities within two to three years and Hassabis estimating a window of three to five years.
- Current progress is attributed to four key drivers that are expected to continue accelerating until at least 2028 and potentially until 2032, with Todd predicting that a lack of acceleration by 2030 will likely cause significant progress to slow or stall.
- Reinforcement learning has emerged as a critical paradigm shift, enabling models like O1 and DeepSeek R1 to achieve 70% accuracy on PhD-level questions and surpass human experts in scientific reasoning within a year, creating a data flywheel for future iterations.
- Computational demands are escalating rapidly, with projections suggesting GPT-6 training could cost $10 billion and total training runs could reach $10 trillion by 2028, constrained by compute availability, electricity usage (projected at 4% of US power by 2028), and the finite AI talent pool.
- By 2026, benchmarks are expected to saturate, and models are projected to possess expert-level knowledge across domains, handle day-long tasks autonomously, and potentially perform multi-week projects, while by 2028 systems may achieve "beyond-human" reasoning abilities.
- Agent-based architectures are becoming the top priority for leading labs, with expectations that digital workers will automate significant portions of the economy, potentially allowing startups to generate billions in revenue with minimal human staff.
- Risks to the timeline include a potential slowdown in compute growth post-2028, diminishing returns on pre-training, the high cost of future training runs ($10 billion+), bottlenecks in handling ill-defined high-context tasks, and external disruptions such as geopolitical conflicts or regulatory crackdowns.
- The probability of AGI is described as a 50-50 split with a range of 30% to 80%, contingent on whether AI can trigger its own acceleration before 2030; failure to do so may reduce the annual probability of discovery significantly by the mid-2030s.