Is the Scaling Law a Pessimistic Future?

On October 22, 2025, Meta CEO Mark Zuckerberg approved a plan to lay off roughly 600 employees from the company’s artificial intelligence division. This marked Meta’s largest AI-related layoff of the year, primarily affecting the core R&D group known as the Superintelligence Lab.

We interviewed Tian Yuandong, former research director at FAIR (Fundamental AI Research), who was among those affected. Our conversation went far beyond Meta's layoffs, touching on:

• AI roadmaps
• The LLM (Large Language Model) trajectory
• Open-source vs. closed-source development
• The evolving role of research labs
• Career choices between pure research and engineering

> For background on why Meta initiated these layoffs and how new AI head Alex Wang might reshape Meta’s AI strategy, see our previous article.

---

1. The Layoff Was Expected — Just Happened Faster

Chen Xi:

I see you’re still wearing a FAIR shirt.

Tian Yuandong:

(Laughs) People like us don’t care much about clothes — we wear whatever the company gives us.

Chen Xi:

Has this week been overwhelming? Media and companies must be approaching you non-stop.

Tian Yuandong:

Since I had an offer before the layoff, it wasn’t a shock. I’d already told my supervisors that I was looking around.

After the announcement, multiple companies — from tech giants to startups — reached out with opportunities, including co-founding roles. I haven’t decided yet; it’s been less than a week (<168 hours).

Chen Xi:

So the layoff was not a surprise?

Tian Yuandong:

Correct. I might have stayed another 6 months before moving, but this simply accelerated the decision.

---

2. AI Industry Trend: Decline of “Execution Layer” Roles

Chen Xi:

600 employees is notable. It’s partly restructuring; fewer positions are needed in the AI division.

Image: CNBC

Tian Yuandong:

Yes, the shift reflects an industry-wide trend:

• As models improve, automation reduces repetitive engineering roles
• Mature pipelines require fewer people to manage
• Execution tasks will be replaced or scaled down by AI agents

For practitioners, adaptability is key — leverage AI as a tool for productivity and revenue. Platforms like AiToEarn官网 allow creators to:

• Generate AI content
• Publish across multiple platforms (Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X)
• Analyze performance and model rankings
• Monetize globally

> In short: fewer will build AI models; more will use AI to build products.

---

Foundation Model Research Will Grow — But Teams Will Shrink

• More research into novel architectures and theories
• Lean engineering teams to train and deploy
• Shift towards vertical and niche applications

---

3. Open Source Isn’t Dead — But Use Matters

Key points:

• Open-source will persist — many Silicon Valley orgs (e.g., AI2) continue to release models
• The critical question: how the model will be used
• In vertical domains, small companies can innovate with specialized models
• Open source suits tools and platforms needing collaboration; closed source suits proprietary personalization

---

4. LLMs — Strengths & Challenges

Main challenge:

> LLMs require massive data — far more than humans ever consume.

• Humans: ~10 billion lifetime tokens
• LLMs: 10–30 trillion tokens for training
• Efficiency gap is enormous

Future work:

• Explore better algorithms or training paradigms than gradient descent
• Seek human-like learning efficiency

---

5. Reinforcement Learning (RL) vs. Supervised Learning (SFT)

Tian Yuandong:

• RL = active learning, searching, collecting, and using better quality data
• SFT = passive learning, memorizing patterns
• For reasoning tasks, RL beats SFT
• Excessive SFT can degrade model quality

Human domain expertise is critical:

> Without unique insights, AI outputs are generic — limiting AGI ambitions.

---

6. Computation Is Not Everything — The Scaling Law Problem

> Scaling Law: Performance improves linearly with exponentially more data/compute.

Why pessimistic?

• Ultimately exhausts global resources
• Focus should be on efficiency, not just scaling

Better path:

• Develop architectures that learn faster with fewer samples
• Combine human insights with model capabilities to push boundaries

---

7. Ten Years at FAIR — Gains & Lessons

Regrets:

• Should have done more engineering work alongside research

Biggest gain:

• Developed strong research taste — ability to identify meaningful problems and design effective experiments

Advice:

• Balance engineering skill and research skill — both are valuable

---

8. The AI Talent War — Don’t Chase “Scarcity”

• Market demands shift rapidly — today’s scarce skill may be outdated in a year
• Better to:
• Pursue what you truly want
• Assess long-term usefulness
• Blend passion with practical potential

Industry cycles:

• Past: skills remained relevant 10–20 years
• Now: relevance can fade within 12 months

---

9. Idealism in Research Labs

Even within large companies like Meta, pockets of pure research exist.

Research will continue, often decentralized — “guerrilla” style among small teams and indie researchers.

---

10. Next Steps for Tian Yuandong

• Aim to combine cutting-edge research with practical applications
• Envision high goals first — then find resources to achieve them

---

Key Insight

AI’s future hinges not just on compute and scaling laws, but on:

• Human insight guiding AI exploration
• Efficient learning models
• Automation freeing researchers for deeper work
• Platforms like AiToEarn bridging creation, publishing, and monetization across global networks

---

Video Streaming Platforms

• Domestic: Bilibili | Tencent | Video Channel | Xigua | Toutiao | Baijiahao | 36kr | Weibo | Huxiu
• International: YouTube

Contact: video@sv101.net

Team

• Exec Producers: Hongjun, Chen Xi
• Script / Host: Chen Xi
• Editing: Orange
• Motion Graphics: Chuai AK12
• Ops: Wang Ziqin, Sun Zeping, Zhu Jie

Read the original | Open in WeChat

---

💡 Explore AiToEarn

• Site: aitoearn.ai
• Docs: docs.aitoearn.ai
• Blog: blog.aitoearn.ai