Yes, several organizations, researchers, and AI safety labs evaluating new AI models have developed extrapolations suggesting AI could outperform humanity in most or all cognitive tasks within this decade, with some estimates pointing toward 2026–2028.
These projections are often based on the trend of rapid improvements in AI reasoning, coding, and agentic capabilities, which indicate AI systems could soon automate complex, long-horizon tasks.
Key Organizations and Extrapolations
METR (Model Evaluation and Threat Research): METR, which specializes in evaluating frontier AI agents, has highlighted a rapid, exponential trend in the "time horizon" of AI capabilities (how long a task takes a human expert to complete, which AI can now manage reliably). Extrapolations from their RE-Bench benchmark suggest that by late 2028–2031, AI could automate many tasks that currently take humans a month.
Anthropic: In early 2023, researchers at Anthropic suggested that if current scaling trends (larger models, more compute) hold, simple extrapolations show AI systems could match or exceed human-level performance at most intellectual tasks within the next decade.
AI Futures Project (and similar forecasts): Researchers associated with organizations like the AI Futures Project have authored scenarios ("AI 2027") predicting that AI systems will become fully autonomous agents better than humans at everything by late 2027 or 2028, largely by automating AI research itself.
Lab Leadership Estimates: Leaders of major AI labs have suggested shortened timelines. Anthropic CEO Dario Amodei previously mentioned a "human-level" AI could be developed within 2-3 years (from 2024), while OpenAI CEO Sam Altman has indicated AGI could be reached within 4-5 years. They get money for producing the hype so their projections should be considere with a pound of salt.
Stanford AI Index (2025): The 2025 report indicates AI is rapidly surpassing human performance across critical benchmarks, with steep gains in 2022-2024, closing the gap in high-level math, visual reasoning, and PhD-level science.
Key Milestones in the Projections
2026: Widespread integration and models that match human expert performance on many tasks.
2027–2028: Potential for AI to reach "superintelligence" or become better than humans at all, or most, tasks, driven by agentic AI that can perform AI research.
50% Probability: A large survey of machine learning researchers estimated a 50% chance of AI outperforming humans in all tasks by roughly 45 years (from a 2017 survey), but recent surveys (2023–2024) have seen these timelines compressed significantly toward the 2030s–2040s.
Uncertainties and Skepticism
While these forecasts are widely discussed, they are not universally accepted.
The "Messy Task" Limitation: While AI excels at well-defined, short-term tasks, it still struggles with "messier," high-context, long-horizon tasks.
Validation Gaps: Independent testers, such as for the "Humanity's Last Exam" (HLE) benchmark, have found that reported AI capabilities can be inflated, with models often performing lower than companies claim.
Skepticism of Scaling: Critics, such as researchers at the Allen Institute for Artificial Intelligence, argue that current forecasts often rely on "fictional scenarios" rather than grounded scientific evidence and that scaling may encounter bottlenecks.
Komentarų nėra:
Rašyti komentarą