Elon Musk (xAI)
2026“AGI 'smarter than the smartest human' will arrive as early as 2026, driven by Grok 5.”
Chapter 06 · AGI & ASI
What happens when we build a mind smarter than ours?
AGI (Artificial General Intelligence) is the point where a software system performs any economically valuable cognitive task at or above human level. ASI (Artificial Superintelligence) represents what follows: a system that recursively improves, outthinking the entirety of humanity.
AGI · Artificial General Intelligence
A coworker for every job.
A system capable of learning new skills at human speed, reasoning across disciplines, utilizing tools, planning, and executing action loops without supervision. By 2026, the remaining gaps consist of long-horizon planning, robust memory retention across context sessions, and embodied common sense in physical reality.
- · Frontier benchmarks (GPQA, SWE-bench Pro, AIME) are nearing saturation.
- · Test-time reasoning models (OpenAI o1/o3, DeepSeek-R1) trade latency for accuracy by emitting 20K-60K thinking tokens.
- · Chinese token consumption has surged to 140 trillion tokens daily in Q1 2026, indicating that active inference is replacing training as the dominant compute cost.
ASI · Artificial Superintelligence
A scientist thinking at machine-velocity.
An ASI is a system that vastly exceeds the best human minds in every cognitive, creative, and technical domain. Operating millions of parallel threads in a data center, it could execute decades of scientific progress in months, introducing recursive self-improvement loops that demand safety paradigms we have yet to verify.
- · Direct physical control: Automated laboratories executing chemistry, genomics, and hardware designs at machine speed.
- · Decisional Safety: Safety frameworks like Anthropic's RSP define strict triggers (ASL-3/ASL-4) for autonomous cyber-offense capabilities.
- · The Alignment Problem: Steering systems smarter than their creators, preventing sycophancy, reward hacking, and deceptive scheming.
When? · Predictions from the builders
The timeline, in their own words.
Dario Amodei (Anthropic)
~2027“AGI 'likely within a few years' (estimated around 2027). Reaffirmed at WEF Davos 2026.”
Shane Legg (DeepMind)
2028“50% probability of achieving 'minimal AGI' by 2028 (reaffirmed in January 2026).”
Demis Hassabis (DeepMind)
~2030“50% chance of achieving AGI by the end of the decade (~2030). Reaffirmed at WEF Davos 2026.”
Eric Schmidt (ex-Google)
2028-2030“AGI is likely 3 to 5 years away (estimate from April 2025).”
Jensen Huang (NVIDIA)
2029“AI will be able to pass any test/benchmark designed by humans by 2029 (estimate from March 2024).”
Ray Kurzweil (Futurist)
2029“Maintains his decades-long prediction of AGI by 2029, and Singularity by 2045.”
Sam Altman (OpenAI)
2035“AGI will arrive within a 'few thousand days' (essay from late 2024).”
Ajeya Cotra (Open Phil)
2040“50% chance of AGI by 2040, based on bio-anchors framework (2024 update).”
Metaculus (Forecasters)
2033“Aggregate superforecaster median points to 25% probability by 2029, and 50% by 2033 (as of Feb 2026).”
Samotsvety Forecasting
2041“Aggregated consensus: 10% chance of AGI by 2026, and 50% by 2041 (Jan 2026 update).”
Existential Risk: Cyber-Physical Convergence
Machine-velocity conflict.
In 2026, the convergence of IT and OT (Operational Technology) has created a strategic convergence of cyber-physical threats. Advanced Persistent Threat (APT) groups deploy AI-orchestrated attacks targeting critical national infrastructure: water treatment plants, regional power grids, and automated industrial control systems.
Because AI systems can compile, scan, and deploy zero-day exploits autonomously at scale, the window for human intervention has shrunk from hours to milliseconds. Safe AGI development demands hardening physical infrastructure against autonomous cyber-penetration vectors.
Benchmark reality check · 2024 → 2026
Two years. Almost every frontier benchmark fell.
ARC-AGI-1 (abstraction)
GPT-4o (2024): 5.0%
o3 (2025): 87.5%
ARC-AGI-2 (human baseline)
Human Expert: ~72.0%
Gemini 3.1 Pro: 77.1%
SWE-bench Pro (hard code)
Gemini 3.1 Pro: 54.2%
Claude Opus 4.8: 69.2%
SWE-bench Verified (curated code)
Sonnet 4.6: 79.6%
Claude Opus 4.8: 88.6%
GPQA Diamond (doctoral science)
Human PhD baseline: ~65.0%
Frontier 2026: 85.0 - 94.0%
AIME (test-time math)
DeepSeek-R1 (base): 15.6%
DeepSeek-R1 (scaled): 86.7%