Pramod.AI
AboutIntelligenceArtifactsAI EvolutionNewsletterBlogContact
← STORY OF INTELLIGENCEHOME
Deep Dive

The Path to AGI

What artificial general intelligence really means and where we stand

AGI - Artificial General Intelligence - is the idea of a system that can perform any intellectual task a human can. Not just pattern matching or next-token prediction, but genuine reasoning, planning, learning from few examples, and transferring knowledge across domains.

As of March 2026, we have systems that pass bar exams, write production code, and generate photorealistic video. But they still can't reliably plan a multi-step task, learn from a single demonstration the way a child does, or maintain persistent goals across sessions. The gap between impressive demos and true general intelligence remains significant.

THE PATH TO AGI - WHERE ARE WE?Google DeepMind AGI Level framework -- March 2026L1 - NARROW AIImage classification, spam filter, chess engine100%L2 - COMPETENT AIGPT-4, Claude, Gemini -- broad task assistance92%L3 - EXPERT AIDomain-level reasoning, autonomous R&D support75%L4 - VIRTUOSO AINobel-quality science, master-level creative output55%L5 - SUPERHUMAN AIOutperforms all humans across every cognitive domain32%L6 - ASIRecursive self-improvement, civilizational agency14%WE ARE HERE — March 2026CAPABILITY

How It Works

1

Level 1 - Narrow AI

Systems that excel at one specific task. Spam filters, chess engines, image classifiers. This is where AI was from 1950 to 2018.

2

Level 2 - Competent AI

Systems that match human-level performance across many cognitive tasks. Today's frontier models (Claude, GPT-4, Gemini) operate here - broad competence but inconsistent reasoning.

3

Level 3 - Expert AI

Systems that consistently outperform domain experts. We see glimpses in coding (Codex, Claude Code) and math (o3), but not yet reliable across all expert domains.

4

Level 4 - Virtuoso AI

Systems that can innovate and create novel solutions humans haven't considered. This requires genuine creativity and cross-domain transfer - not yet achieved.

5

Level 5 - Superhuman AI

Systems that exceed the best humans in virtually all cognitive tasks. Would require sustained, goal-directed reasoning over long horizons with perfect reliability.

6

Level 6 - ASI

Artificial Superintelligence - systems that surpass collective human intelligence. Theoretical and deeply uncertain. May require fundamentally different architectures than current transformers.

Key Components

Reasoning

Chain-of-thought, tree search, and formal verification. Models like o3 and Claude show strong reasoning but still fail on novel problems.

Planning

Decomposing goals into sub-tasks and executing over time. Current agents struggle with plans longer than 10-15 steps.

Memory

Persistent knowledge across sessions. Current context windows are large but ephemeral - no true long-term learning.

World Models

Internal representations of how the world works. Video generation models show nascent world modeling but lack causal understanding.

Transfer Learning

Applying knowledge from one domain to a completely different one. Humans do this effortlessly; AI still struggles.

Alignment

Ensuring AGI systems pursue human-intended goals. Constitutional AI, RLHF, and interpretability research are early steps.

Who's Building With This

A

Anthropic

Constitutional AI and interpretability research. Published 'Scaling Monosemanticity' showing how to understand model internals.

O

OpenAI

o3 reasoning model pushes toward expert-level performance. Superalignment team researching scalable oversight.

G

Google DeepMind

Published cognitive framework for measuring AGI progress. Gemini 2.5 demonstrates strong reasoning.

M

Meta FAIR

Open research on world models (V-JEPA), planning systems, and the foundations of machine intelligence.

Key Takeaway

AGI is not a binary switch - it's a spectrum. We're somewhere between Level 2 and Level 3 in March 2026. The path forward likely requires breakthroughs in planning, memory, and world modeling - not just scaling current architectures.

References & Further Reading

  1. Google DeepMind - Levels of AGI: Operationalizing Progress
  2. Anthropic - Scaling Monosemanticity
  3. OpenAI - Planning for AGI and Beyond
  4. Stanford HAI - AI Index Report 2025

Explore More Topics

Teaching AI to look things up before answeringRetrieval-Augmented GenerationWhen AI learns to see, hear, and speakMultimodal AIThe silicon, cloud, and systems powering intelligenceAI InfrastructureThe engines of intelligenceFoundation ModelsBig intelligence in small packagesSmall Language ModelsThe memory layer of AIVector DatabasesFrom chatbots to autonomous digital workersAI Agents in DepthA complete map of who builds what and how it all connectsThe AI EcosystemHow to train a model across thousands of GPUsDistributed AI TrainingMeasuring, monitoring, and operating AI in productionEval & AI OpsHow silicon is evolving to bring AI from data centers to your pocketAI Chips and Edge IntelligenceWhich industries are winning with AI and how they're deploying itAI Sector Dominance
← Eval & AI OpsAI Chips and Edge Intelligence →