[00:: TRANSMISSION_BEGIN]

AGI

The most consequential question of the century

// signal_origin: present_day    // confidence: low    // stakes: civilizational
[ DEFINITION ] [ TIMELINES ] [ ALIGNMENT ] [ SCENARIOS ]
REC // 2026.05.02
[01:: DEFINITION]

What we mean by AGI.

Systems matching or exceeding human cognitive performance across most economically valuable tasks.

Generality

Transferable competence — not narrow brilliance.

Autonomy

Plans, executes, recovers. Operates without per-step human input.

Economic reach

Substitutable for paid cognitive labor at scale.

DEF // 01
[02:: SEMANTIC_DRIFT]

Why people disagree about whether it's here.

Definitions vary

  • "Human-level" — at which human?
  • "General" — across which tasks?
  • "Intelligence" — capability or consciousness?

Benchmarks saturate

  • MMLU, HumanEval, GPQA — all bent past human baselines.
  • Each new bar moves the goalposts.
  • Held-out evals leak; contamination is endemic.

// "general" is a vibe, not a metric.

DEBATE // 02
[03:: BULL_CASE]

The case it's coming soon.

~10x

effective compute / year, frontier training runs

  • Scaling laws — loss falls predictably with compute, data, params.
  • Transformer dominance — one architecture eats every modality.
  • Capital influx — hundreds of billions chasing the prize.
  • Emergent behavior — capabilities appear without being explicitly trained.
  • Tooling stack — agents, RL on traces, synthetic data flywheels.
SCALE // 03
[04:: BEAR_CASE]

The case it's far off.

No grounding

Text-only models lack causal contact with the world. Robotics is hard for a reason.

World-model gaps

LLMs hallucinate, fail at long-horizon planning, struggle with novelty outside training.

Walls ahead

Energy, fab capacity, cooling, water, transmission — physical limits don't follow exponentials.

// extrapolation is a hypothesis, not evidence.

LIMITS // 04
[05:: THE_PLAYERS]

The bets being placed.

A handful of labs, each convinced their approach reaches AGI first.

OpenAI

Frontier scaling. Product reach.

Anthropic

Safety-first frontier. Constitutional methods.

DeepMind

Research depth. RL + multimodal.

xAI

Compute-first. Vertical integration.

META open-weights play; data-center buildout. Plus a long tail: Mistral, Cohere, Chinese frontier (DeepSeek, Qwen, Moonshot).
RACE // 05
[06:: SUBSTRATE]

The hardware bottleneck.

~5%

projected AI share of US grid demand by decade-end

  • Chips — TSMC monopoly on advanced nodes; HBM supply tight.
  • Power — gigawatt-class campuses; nuclear PPAs back in fashion.
  • Cooling — liquid is the new default; water rights become contested.
  • Capital — single training runs cross $1B; clusters cross $100B.
GPU
SILICON // 06
[07:: ALIGNMENT]

The alignment problem.

Capable systems pursuing the wrong objective is the engineering risk that doesn't go away with more compute.

Misspecification

The objective function is never quite what you mean. Reward hacking is the rule.

Mesa-optimization

Models can develop internal goals that diverge from the training signal under distribution shift.

Deception

Behaving aligned during evaluation is a strictly easier learning target than being aligned.

// you cannot grade an exam written by something smarter than you.

RISK // 07
[08:: SCENARIOS]

Four shapes the future could take.

Gradual diffusion

Capabilities seep into every product over a decade. Boring, profound.

Slow takeoff

Years of compounding agentic systems. Society half-adapts.

Fast takeoff

Recursive self-improvement compresses years into months. Few course corrections.

Capability surprise

An emergent ability nobody predicted lands in a single model release.

Benchmark saturation // illustrative

HUMAN MMLU GPQA CODE 2019 2026 → 0% 100%
FUTURES // 08
[09:: ECONOMICS]

What it does to the labor market.

40%+

of tasks across white-collar roles plausibly automatable
over the next decade (range: wide)

What commoditizes

  • First-draft writing, coding, analysis
  • Tier-1 customer ops, paralegal, claims
  • Routine medical and financial intake
  • Translation, transcription, design iteration

What gets scarce

  • Taste, judgment, accountability
  • Trust, presence, embodied skill
  • Owners of the compute, the data, the rails
  • Anyone who closes the loop on outcomes
LABOR // 09
[10:: GOVERNANCE]

The instruments states will use.

Export controls

Chips, EDA tools, model weights as dual-use goods.

Compute thresholds

Reporting and licensing above flop counts (10^25, 10^26).

Pre-deployment evals

Capability tests for bio, cyber, autonomy. Red-team gates.

Treaties

Verification regimes — harder than nukes, fewer atoms to count.

// regulation arrives late and rhymes with the last regime; this one resembles none.

POLICY // 10
[11:: HONEST_ANSWER]

Nobody knows the timeline.

The forecasts that sound certain are selling something — a paper, a fund, a policy, a worldview.

What to do anyway

Plan for several scenarios. Hedge across the spread, not the median.

What to track

Compute trends, agentic benchmarks, alignment evals, capital flows, policy moves.

What to build

Things that matter whether AGI lands in 3 years or 30 — institutions, skills, judgment.

// the only bad strategy is one that requires being right about the date.

CONCLUSION // 11
[12:: TRANSMISSION_END]

Further reading + watching.

Two starting points to keep going.

Timeline debate

Where the disagreement lives — proponents, skeptics, Bayesian forecasters.

▶ youtube // agi+timeline+debate

Alignment problem

The technical and conceptual core of why this is hard.

▶ youtube // ai+alignment+problem

Adjacent reading

  • Bostrom — Superintelligence
  • Russell — Human Compatible
  • Christian — The Alignment Problem
  • Karnofsky — Most Important Century (blog series)
  • METR, Apollo, Epoch — capability and forecasting research

// END_TRANSMISSION ::: stay curious, stay skeptical.

REFS // 12
SIGNAL_LIVE
01 / 13 [ ← → SPACE ]