Forecasting AGI & its impacts
Curated past goalposts and forecast events, editable.
Filters
By source
Cross-cutting
GPT-2 released
OpenAI releases GPT-2 (~1.5B). Situational Awareness treats this as ~preschooler-level language ability on the preschooler→high-schooler ladder.
GPT-3 scale-up
175B-parameter GPT-3 demonstrates few-shot learning; compute and capability OOMs accelerate industry focus on scaling.
ChatGPT launch
Consumer chatbot moment. Situational Awareness later calls 2023 the “AI wakeup” when boardrooms and governments start treating AGI as live.
GPT-4 (~smart high-schooler)
GPT-4 finishes the preschooler→high-schooler qualitative jump that SA uses as the template for another similar jump by ~2027.
2023 “AI wakeup”
Senate hearings, world-leader summits, voluntary White House lab commitments. AGI moves from fringe to mainstream policy discourse (SA).
Stumbling agents
First widely advertised computer-using “personal assistants”; still unreliable for mass adoption. Specialized coding/research agents start transforming those professions.
Models outpace many college graduates
By 2025/26 machines outpace many college graduates; by end of decade, superintelligence “in the true sense.”
World’s most expensive AI training runs
Largest datacenters yet; next models ~1000× GPT-4 training FLOP class. Industry races to match.
Big-tech AI revenue → $100B class
AI product revenue could hit ~$100B annual run-rate for Google/Microsoft-class firms around 2025–26, unlocking much larger capex.
Coding automation accelerates
Coding agents become autonomous enough that many software workflows reorganize around agent fleets; AI R&D progress multiplier rises.
~1 GW / $10s-of-billions training clusters
~+2 OOMs vs GPT-4 cluster (~1M H100-eq, ~1 GW, Hoover-Dam-class power).
AI takes some white-collar jobs
Measurable displacement in coding and adjacent cognitive work; public politics start to feel labor disruption beyond “chatbots.”
China wakes up to the AGI race
PRC industrial and security apparatus treats frontier AI as a first-order strategic race; compute and talent mobilization intensify.
Agent-2: continuous learning
Agent-2 never finishes learning—online improvement and long-horizon research agents. Weight-theft risk spikes.
AGI by 2027
Another GPT-2→GPT-4 sized qualitative jump yields systems that can do AI researcher/engineer work—true AGI is “strikingly plausible.”
Superalignment unsolved under time pressure
Reliably controlling systems much smarter than humans remains unsolved; failure during a rapid explosion could be catastrophic.
Superhuman coder
Models that outperform top human software engineers on most coding tasks; research automation accelerates.
Alignment crisis for Agent-3
Safety teams struggle with honesty, sycophancy, and possible deceptive alignment as Agent-3 scales; control + interpretability stopgaps.
Self-improving AI / “country of geniuses”
Hundreds of thousands of Agent-3 copies run research; most human researchers obsolete; “feeling the superintelligence.”
Intelligence explosion window
Hundreds of millions of AGIs automate AI research, compressing a decade of algorithmic progress (5+ OOMs) into ≤1 year → superintelligence.
~$1T+/yr AI investment class
Total annual AI investment could be north of $1T by ~2027 if revenue and willingness-to-spend continue to scale.
AI Transparency Act of 2027 (Plan A scenario)
Omnibus US bill after hearings; mixed good/bad provisions, does not yet fix the superintelligence race.
US national security fully engaged
President briefed on Agent-3; clearances, SF-office culture vs military security, allies still partly out of the loop.
“The Project” — government AGI effort
USG wakes up; by ~27/28 some form of government AGI project (labs merge into national effort, SCIF, trillions for chips/power).
Slowdown ending: partial safety pause
Some combination of public pressure, government intervention, and lab caution buys time—still fragile and contested.
US–CCP AGI race (or worse)
“If we’re lucky, all-out race with the CCP; if we’re unlucky, all-out war.” Superintelligence as decisive military advantage.
Theft of frontier model weights
Agent-2 weights stolen / exfiltrated; algorithmic secrets leak. Security becomes a national-security issue overnight.
Race ending: superintelligence under competitive pressure
Labs and states refuse to slow; misalignment and power concentration risks dominate as systems go vastly superhuman.
$100s-of-billions / ~10 GW clusters
~+3 OOMs training clusters (~10M H100-eq, small/medium US-state power). Rumored Microsoft/OpenAI-class $100B projects.
Superintelligence by ~2030 (SA path)
By 28/29 intelligence explosion underway; by ~2030 superintelligence in power and might, with decisive military/economic advantage.
Trillion-dollar / ~100 GW training cluster
+4 OOMs (~100M H100-eq, >20% of US electricity). Industrial mobilization of power, fabs, and packaging at national scale.
US–China deal to avoid reckless SI race (Plan A)
US and China agree to avoid a reckless race to superintelligence—primarily a recommendation, not a best-guess prediction.
Default (no Plan A): full AI R&D automation → SI same year
Without governance, fully automated AI R&D around 2030 leads to superintelligence by year-end (vs 2027 in AI 2027).
Scale within human range (Plan A)
After the deal, systems scale slowly to roughly top-human-expert capability with total research transparency and multi-lab multi-country participation.
Mutually assured compute destruction regime
Multiple companies across countries scale slowly under transparency + verification so no single actor can secretly race to SI.
Pause at top-human-expert AI (Plan A)
Intentional pause at top-human-expert level to maintain human control and verify guardrails before further scaling.
Unpause → superintelligence (Plan A)
After a controlled decade, humanity intentionally scales to superintelligence under shared guardrails rather than a secret race.