AGI: Are We There Yet? (2025 Edition)

Every year, the “AGI timeline” debates intensify. Let’s take stock of where we actually are at the end of 2025.

Defining Terms

What is AGI?

No consensus exists, but common definitions:

The goalposts shift with each advance.

OpenAI’s Levels

OpenAI proposed a framework:

We’re firmly in L1, approaching L2, experimenting with L3.

2025 Progress

Reasoning Models

o1, R1, and similar models think before answering:

• Multi-step mathematical proofs
• Code generation with planning
• Scientific problem-solving
• Still fail on novel challenges

Agents

Real-world task completion:

• Browse web, complete forms
• Multi-step research
• Code changes across files
• Still need supervision

Multimodal Integration

vision + audio + text in one model:

• Real-time video understanding
• Native speech generation
• Physical world comprehension improving
• Not yet embodied reliably

What’s Still Missing

Robust Reasoning

Human: "If I put a cup in the microwave for 30 seconds, 
        then transfer the water to a bowl that was in 
        the freezer overnight, what happens?"

AI: [Often gets physical reasoning wrong]
    [Common sense gaps evident]
    [Edge cases fail]

Models are “stochastic parrots” with better statistics, not true reasoners.

Persistent Learning

Session 1: "My name is Alex, I work on climate models."
Session 2: [No memory of previous session]

Models don’t learn from interactions (by design, for safety).

Planning Over Long Horizons

Task: "Build a successful startup"

AI can:
- Generate a business plan
- Suggest marketing strategies
- Draft legal documents

AI cannot:
- Adapt to market feedback
- Handle novel challenges
- Maintain coherent strategy over months

Physical World Understanding

Question: "Can I fit this couch through that doorway?"

AI: [Makes assumptions]
    [Lacks spatial reasoning]
    [Can't interact with physical world]

Embodied intelligence remains limited.

Expert Opinions

Optimists

Skeptics

Median Estimate

Surveys of ML researchers:

• 50% probability of “high-level machine intelligence” by 2060
• Large uncertainty bands
• Definition-dependent

The Benchmark Problem

Benchmarks Saturate

Are tests too easy, or is AI too capable?

New Harder Tests

ARC Prize, FrontierMath, Novel benchmarks:

• Designed to require genuine reasoning
• Models still struggle
• But performance improves monthly

What Would AGI Look Like?

Signs We’re Close

• Scientific discoveries attributed to AI
• AI improving AI effectively
• Genuine transfer learning demonstrated
• Novel problem-solving without prompting
• Self-correction without human feedback

Signs We’re Far

• Basic reasoning failures persist
• Common sense gaps remain
• No autonomous scientific progress
• Embodied intelligence limited
• Safety/alignment unsolved

Implications for Developers

If AGI is Near (Optimist View)

• Learn to direct AI rather than code
• Focus on judgment and oversight
• Build AI management skills
• Prepare for rapid change

If AGI is Far (Realist View)

• Current skills remain valuable
• AI as tool, not replacement
• Focus on human-AI collaboration
• Solve real problems with current tech

Either Way

# Good advice regardless of timeline

1. Stay adaptable
2. Learn current AI tools
3. Create judgment, not just output
4. Build human skills AI lacks
   - Leadership
   - Ethics
   - Physical creation
   - Empathy

My Take

We’re not close to AGI by strict definitions. We are close to AI that transforms significant parts of the economy.

What’s real:

• AI that exceeds human performance on narrow tasks
• Useful agents for defined scopes
• Automation of cognitive work at scale

What’s hype:

• Superintelligence imminently
• Human obsolescence
• “One more scaling jump gets us there”

Final Thoughts

The AGI question depends on your definition, your timeline, and what you include as “intelligence.”

More useful question: “What can AI do for me today, and how is that changing?”

Answer that question. The AGI debate can wait.

The singularity is always five years away.