DEV Community

정상록
정상록

Posted on

Demis Hassabis on AGI: Scaling Laws Only Get Us 50% There

#ai #llm #machinelearning #news

Demis Hassabis on AGI: Scaling Laws Only Get Us 50% There

Google DeepMind CEO Demis Hassabis recently gave one of the most candid assessments of AGI's timeline and challenges on the 20VC podcast. Here are the key technical takeaways for developers and AI practitioners.

The "Jagged Intelligence" Problem

Hassabis coined the term "Jagged Intelligence" to describe current AI's inconsistency. Picture this: a model that wins a gold medal at the International Math Olympiad but stumbles on elementary arithmetic.

This isn't just an anecdote — it's a fundamental architectural limitation. Current transformer-based models lack:

  • Continual learning: Models are static after training. No experience-based adaptation.
  • Long-horizon planning: Short-term task completion works, but strategic multi-step planning fails.
  • Robust reasoning: Performance is unpredictable across difficulty levels.
  • True creativity: Pattern recombination ≠ genuine novel creation.

The 50/50 Strategy

Here's the claim that matters most for the AI engineering community:

"Scaling gets us 50% of the way to AGI. The other 50% requires Transformer-level fundamental breakthroughs."

DeepMind's response is a balanced investment: equal resources for scaling existing architectures AND pursuing novel research directions (world models, new training paradigms, etc.).

This contrasts with some competitors who are going all-in on scaling. Hassabis acknowledges diminishing returns are emerging:

Returns: substantial but decreasing
Strategy: 50% scaling + 50% research innovation
Risk: "middle regime" performance plateau
Enter fullscreen mode Exit fullscreen mode

Will LLMs Become Commodities?

Short answer from Hassabis: No.

His reasoning relies on what I'd call a "flywheel effect":

  1. Build coding tools, math tools as intermediate outputs
  2. These tools accelerate next-gen model development
  3. Only well-resourced labs can maintain this cycle
  4. Result: 3-4 leading groups maintain persistent advantage

He's supportive of open-source models but sees cutting-edge capability as inherently concentrated.

What This Means for Developers

  1. Don't bet solely on scaling: If DeepMind's 50/50 assessment is correct, the next breakthrough won't come from bigger models but from architectural innovations.

  2. The reliability problem matters: Jagged Intelligence means AI-assisted coding tools need robust error detection, not just generation capability.

  3. Continual learning is the gap: Building systems that can learn from deployment feedback is an underexplored area with massive potential.

  4. Agentic AI needs guardrails: As AI becomes more autonomous, Hassabis advocates for nuclear non-proliferation-level international agreements. Infrastructure for safe agent deployment is a growing field.

The Bigger Picture

Hassabis' ultimate goal: "Understanding the fundamental nature of reality."

AlphaGo → Go. AlphaFold → Protein structures (Nobel Prize 2024). Isomorphic Labs → Drug discovery.

For him, AI isn't a SaaS product. It's a scientific discovery accelerator. And AGI? "10x the Industrial Revolution at 10x the speed."

Full interview: 20VC with Harry Stebbings (April 7, 2026)

What's your take on the 50/50 claim? Are we hitting scaling limits, or is this just DeepMind positioning? Let me know in the comments.

Top comments (0)