The next 36 months will be WILD
In a Nutshell
AGI and superintelligence are forecasted for 2027-2028 by leaders like Dario Amodei, Sam Altman, and Jensen Huang, driven by super-exponential autonomy growth (METER eval: Claude Opus at 14.5-hour median, doubling every 90-100 days toward week-long tasks by end-2026) and imminent RSI via advances in math, data, code, training, and evals. Unstoppable AI race (US-China, $600B infra sunk costs) clears bottlenecks like chips/HBM (solved soon) and energy (microgrids/SMRs by 2028-2030), triggering an automation cliff: jobless growth, vanishing entry-level jobs, and Solow Paradox 2.0 with exploding GDP/productivity decoupled from employment. Watch METER scores hitting 24+ hours, trillion-dollar capex, and imploding hiring; accelerationists predict high-impact positive transformation despite short-term shocks, dismissing doomer/x-risk fears.
These notes were generated by AI and may contain inaccuracies.
Why does it seem like all of a sudden everyone is talking about AGI, ASI, and recursive self-improvement (RSI) in the 2027 to 2028 window? Sam Altman, Jensen Huang, and others like Dario Amodei are converging around this timeframe.
Key Predictions and Sources
- Dario Amodei (Anthropic): Powerful AI matching Nobel-level capability, "a country of geniuses in a data center," around 2027-2028.
- AI 2027 report (from safety researchers like Daniel Kokotajlo): Modal prediction for AGI 2027-2028; window has collapsed from decades to months with smaller error bars.
- Jensen Huang: AGI competitive on broad human tests within 5 years (speaker thinks he's sandbagging).
- Sam Altman: Research intern by 2026 (widely agreed upon across internet, Twitter, domains like advanced mathematicians, economists); superintelligence by 2028.
These are objective statements from the experts on AGI, ASI, and RSI, with RSI expected this year or 2027 at latest.
METER eval shows super-exponential growth (logarithmic scale still curving up exponentially). Latest data: Claude Opus jumped to median autonomous success rate of 14.5 hours, 95th percentile up to 90+ hours (as of February 2026). This ricocheted across the internet; 96 hours ≈ 4 days.
Doubling time shortening from 4-7 months to 90-100 days (one quarter). Not even through first quarter of 2026, so potentially three more doublings this year:
- From 14.5 hours → 29 hours → ~60 hours → 120 hours (~1 week). Approaching 60-120 hours machine autonomy by end of 2026 at current rates.
1. Algorithmic Research/Math
Gold medal on IMO (International Math Olympiad) achieved by two labs last year. No need to solve Millennium Prize problems (e.g., Andrej Karpathy, Ilya Sutskever haven't, yet lead AI). Threshold: Between IMO gold and Millennium Prizes for advancing AI architectures/algorithms/theories.
2. Data Generation and Curation
Latest models trained on 10-20% synthetic data. Risks model collapse, but labs using self-play (like AlphaGo, AlphaFold) for infinite text generation, validated by selecting best from thousand reasoning traces. Models grade each other. Human brains learn with far less data; not near data efficiency ceiling.
3. Writing and Executing Code
OpenClaw, Claude Code, Codex better than most developers (causing existential anxiety). Not perfect (mangles codebases), but last year Dario Amodei predicted AI writes most code by summertime/end of year—off by months. Latest Claude version had all code written by Claude. 1-2 more 90-day cycles to exceed RSI threshold.
4. Training Models
Infrastructure job: Define dataset, parameters, hyperparameters, architecture, then run on compute (most expensive activity, millions in compute for days/weeks). GPT-3/GPT-3.5 trained for weeks. All code/API calls; AI could hypothetically do it autonomously, but no reports of fully autonomous agent-triggered trainings yet.
5. Model Evals
Requires tremendous work (e.g., Philip at AI Explained's private benchmarks need human grading for non-provable tasks). Provable evals: code execution, math calculation. Labs use AI for evals, but papers show dumber models can't accurately grade smarter ones. Biggest bottlenecks: training execution and evals (math/data/code already closer).
Tech bros in Silicon Valley working on putting themselves out of a job.
Race dynamics: Companies/nations racing (terminal race condition coined by speaker: heading to maximally intelligent/efficient AI; no incentive to slow). More players since coining: xAI, Google Gemini, Grok, DeepSeek (vs. old Bard = "Before All Revenue Declines" joke).
Nations: US vs. China heating up; Europe deregulating (e.g., EU proposal: approval default if no regulator response in weeks—180 from tradition).
Point of no return: Sunk cost/"Hail Mary" (e.g., New Line Cinema all-in on Lord of the Rings). $600 billion committed to data centers/chips/infra; pausing = bankruptcy.
All-in: Maximum speed forward; pausing fatal.
- Chips: Problem solved (no longer discussed).
- High-bandwidth memory (HBM): Current bottleneck (desktop memory expensive due to foundry prioritization like Samsung; >$100B investments by SK Hynix/Samsung to retool, solved in 12-24 months like chips 2023-2025).
- Energy: Harder; alligator jaws graph shows AI demand (to 500 TWh) outpacing production (current data centers ~4 TWh/year; US total annual ~4,000 TWh, so 12% of footprint).
Energy Solutions
- Microgrids: Collocate data centers with solar/natural gas turbines (bypass grid).
- Restart traditional nuclear (e.g., Three Mile Island).
- New nuclear: 20 years (too slow; China wins race).
- SMRs (small modular reactors): Earliest 2028, likely 2029-2030.
National pressure clears red tape. Bipartisan anti-data center sentiment rising (NIMBYism; data centers = silent warehouses, cleaner than coal, yet protested). Need data centers to compete with China or achieve advanced socialism.
Superior Whorf hypothesis: Language constructs reality (e.g., Russian has two blues; brain detects shift faster). AGI raified by sci-fi; people wait for "consciousness/soul/humanlike convo" brain wave, ignoring exponential ability growth.
Real measure: Impact (remote worker/long-horizon tasks, economic substitution). Already seeing job dislocation (poorly tracked).
Expected 2028 Capabilities
- Autonomous projects: Week-long engineering tasks, no human in loop (coding first; robotics solved by 2028).
- Multi-agent swarms: Divide/conquer workflows, instantiate/deprovision on fly.
- Persistent memory: GitHub repo for context (version-controlled real-time knowledge base).
- Tool fluency: Indistinguishable human use of CRM, IDE, email, KVM (computer-using agents like CUA models; OpenClaw via CLI).
Machine doesn't need consciousness; needs economic substitutability.
Solow Paradox 2.0: AI everywhere except employment stats (like 1980s PCs). 2025: 3.7% GDP growth with only 180,000 jobs added (post-revision).
Alligator jaws: GDP/productivity up faster than wages/labor. AI boosts company productivity/profits with lower headcount. 2025 first dramatic observation (hard to isolate).
J curve of productivity: Initial investment reduces efficiency, then explodes (entering phase 2 harvest: productivity explodes, labor decouples—and has always decoupled from output for long time). This is just an
Jobs are drying up and disappearing through the vanishing ladder or ghost jobs. Layoffs are the obvious top line, like AI nuking jobs, but the flip side is jobs that never appear. This drives jobless recoveries or jobless growth, entering a long-term secular regime of jobless growth. The automation cliff is not mass firing but positions never filled or created: entry-level hiring freezes, jobless grads, average job latency over a year for new grads to get first job, and that's only getting worse. Canary in the coal mine: highly qualified students who can't get jobs signals automation and regime change in work.
Silicon Valley tech bros say we're close because of a step function like sigmoid activation functions in neural networks. AI is currently in augmentation phase: experimental, not business critical, optional for enterprises, makes some people more effective. Once reliable, easy to deploy, with tribal knowledge, it goes from zero to one quickly, hitting the automation cliff. New technology becomes viral and mandatory. 85% of people will be unemployable as we sail off the automation cliff. Technologists see this early.
Speaker started IT career at virtualization ramp-up (what people call cloud today). It wasn't obvious but clear future direction; built career on it. Would pivot to AI if starting over; did so leaving corporate job in 2022-2023, going all in on AI. A little early but not wrong. Heading for automation cliff as AI becomes better, faster, cheaper, safer than humans.
Four quadrants of risks: y-axis impact (low to high), x-axis predictability (low to high).
- Zone of unknown (high impact, low predictability): societal outcome, government capture, financial collapse, x-risk.
- Trivial unknowns (low impact, low predictability): misalignment bugs, agentic failure modes (e.g., how OpenAI/Claude fails, but patchable; dozens of alignment papers weekly; flaws like constitutional AI discovered and patched).
- High predictability, high impact: scaling laws, model capabilities, compute growth (meter evals); follows Moore's law squared (Jensen Huang); better at math, coding, agentic autonomy.
- High predictability, low impact: business as usual (market forces, energy bottlenecks, memory bottlenecks); low impact because market sees and solves them (like chips problem).
- Meter scores: task horizon crossing 24 hours in 90-120 days (or 90-180 days full range).
- High bandwidth memory: stops being news (like chips did).
- Capex: trillion dollars per year threshold, probably this year or next.
- Entry-level hiring: continues to implode (e.g., IBM adds 6,000 jobs, brings back 600—not a reversal; watch headline wording).
Matrix: high/low significance of AI, good/bad outcome.
- Doomer (high significance, maximally bad): x-risk, societal collapse, control problem; cheerleader Eliezer Yudkowsky.
- Benign/optimistic (high significance, good): accelerationists (speaker's group); positive societal impact long run, solving major problems, relatively smooth transition (concedes job shock could be faster than ready).
- Technoskeptics (low significance): AI hype, limited real-world impact; cheerleaders Gary Marcus, Yann LeCun (LLMs dead end, but already major impact).
- Realistic engineering concerns (low significance): concerned but slow down, more safety resources; practical challenges, reliability, safety bugs, deployment; cheerleaders Geoffrey Hinton, Yoshua Bengio (titans saying take step back).
State of AI, AGI, ASI; what to expect from experts over next 2-3 years.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
