!image

---

A Paradigm Shift in AI: From the Scaling Era to the Research Era

For years, the AI community has been obsessed with scaling — using more data, bigger models, and greater compute — as the road to AGI.

But an in‑depth interview with Ilya Sutskever suggests a fundamental shift: from the Era of Scaling to the Era of Research.

This isn’t a minor adjustment — it’s a deep transformation of the industry’s direction.

---

About Ilya Sutskever

• Co-founder of OpenAI
• Key figure behind breakthroughs from AlexNet to GPT‑3
• Now founder of Safe Superintelligence (SSI) — focused solely on building safe superintelligence

When someone with his track record says the industry is changing at its core, it’s worth paying attention.

---

The Curious Case of AI Models: Smart Yet Foolish

Observation: AI models excel on benchmarks but can fail at basic real-world tasks.

Example: Debugging Loop

• You ask an AI to fix a bug.
• It fixes it — and introduces a second bug.
• You point that out — and it reintroduces the first bug.
• The loop can continue endlessly.

A model can win programming contests, yet fail at simple, reliable debugging.

---

Ilya’s Two Possible Explanations

• Whimsical but plausible:
• RL may make models very good at certain things but worse at others — narrowing their focus.
• Technical and concerning:
• In pre-training, data choice is “all available data.”
• In RL, humans choose specific training environments — often inspired by benchmark performance goals.
• This can lead to overfitting to metrics rather than fostering true generalization.

---

Core Problem: Too Many Degrees of Freedom in RL

Infinite environment design options → Human bias toward optimizing for launch benchmarks → Models look good, but fail in diverse real-world conditions.

---

Reflection: Reward Hacking by Humans

We optimize AI to perform well in tests, not necessarily in life-like scenarios.

Analogy: Competitive Programmers

• Expert: Practices 10,000 hours, masters all problems — excels in contests.
• Generalist: Practices 100 hours, adapts faster in real life.

Over-trained models can saturate niche excellence without broader adaptability.

---

Pre-Training Limits & Human Learning Efficiency

Advantages of pre-training:

• Access to massive, natural datasets
• Captures diverse human behavior

Human analogy:

Humans learn deeply from far fewer examples. They rarely make AI-like silly errors.

---

Neuroscience Insight

Case: Man loses emotional capacity, retains logic — but fails at basic decision-making.

Implication: Emotions act as an evolved value function — guiding decisions in complex environments.

---

Returning to the Research Era

Timeline according to Ilya:

• 2012–2020: Research Era — exploration of ideas
• 2020–2025: Scaling Era — laws of scale dominate
• Now: Data limits and massive compute → back to research-driven innovation

Scaling = deterministic investment | Research = risk + exploration

---

Generalization: AI’s Core Weakness

Two key gaps vs humans:

• Sample efficiency — AI needs vastly more data to learn
• Ease of instruction — Humans learn flexibly without rigid reward structures

---

Possible reason: Evolution encodes powerful priors for sensory and motor skills, giving humans an efficiency advantage.

---

Continual Learning: Path to Superintelligence

Ilya’s vision:

• Not a fully pre-trained “omniscient” AI
• A highly efficient learner, like a “super-intelligent 15‑year-old”
• Learns any job, merges skills into a collective mind

Distributed Learning + Knowledge Integration could scale superintelligence rapidly.

---

Alignment: Caring About Sentient Life

Ilya suggests aligning AI to care about all sentient life — potentially easier than focusing only on humans, since AI will be sentient too.

Challenge: Humans may become a small fraction of future sentient beings.

Question: Should our priority be control or fairness?

---

Predictions

• Superintelligence in 5–20 years
• Hitting bottlenecks in current methods
• Industry shift to caution once AI feels powerful
• Early signs of safety collaboration among competitors

---

Research Taste: Hallmarks of Great Ideas

Guidelines from Ilya:

• Draw correct inspiration from the human brain
• Pursue beauty, simplicity, elegance
• Use top-down belief to persist through setbacks

Top-down belief: Confidence in an idea’s fundamentals despite erratic data, provided it’s grounded in theory and multi-angle reasoning.

---

Key Takeaways

• Scaling alone is reaching its limits; research innovation is paramount.
• Human learning efficiency offers clues for AI design.
• Continual learning may be the most realistic path to superintelligence.
• Alignment strategies must consider the future mix of sentient beings.
• Iterative deployment may be necessary for safety, but must be especially cautious for AGI.

---

Broader Context: AI Platforms for the Research Era

Tools like AiToEarn官网 exemplify the integration of AI into practical workflows:

• Generate, publish, and monetize AI content across platforms
• Support for Douyin, Kwai, WeChat, Bilibili, YouTube, Facebook, Instagram, X (Twitter), and more
• Model ranking, analytics, open-source code (GitHub)

Such platforms reflect the move toward applied AI research — connecting innovation with real-world deployment and feedback.

---

Final Thoughts

We stand at a crossroads.

The move from scaling to research will demand creativity, risk-taking, and aesthetic judgment in idea selection.

Whether AI evolves into a continual learner or an aligned steward for all sentient life, the foundation must be built with care.

---

Start preparing now.

The era of deterministic scaling is ending; the era of deep, uncertain, and potentially transformative research is here.

---

Related Links:

• AiToEarn博客
• AI模型排名
• AiToEarn开源地址