State of AI in 2026: LLMs, Coding, China, Agents, and AGI

The Evolving AI Landscape

• 💡 The DeepSeek moment in early 2025 marked a pivotal shift, as a Chinese company released a near state-of-the-art model with significantly less compute and cost, challenging established giants.
• 🔑 A crucial distinction emerged between open source (providing code and data) and open weight (providing only the trained model), with Chinese companies flooding the market with powerful open-weight models.
• 🌍 Geopolitically, US companies are now buying inference from cheap and effective Chinese models, demonstrating that the primary competitive moat is no longer proprietary ideas but massive compute and hardware.

Technical Breakthroughs in AI

• 🚀 The focus shifted from traditional scaling laws in pre-training to post-training advancements, particularly Reinforcement Learning from Verification/Reward (RLVR).
• 🔬 RLVR trains models using objective answer keys (e.g., in math and coding) to learn how to solve problems, rather than relying on subjective human feedback.
• 🧠 This led to the development of thinking models that pause to generate hidden chains of thought, debate internally, and self-correct, resulting in smarter answers through inference-time scaling.

The Coding Revolution

• 💻 The concept of vibe coding has emerged, where AI is directed with high-level descriptions to build software, reducing the need for manual syntax writing.
• 📈 Counterintuitively, senior developers are using more AI-generated code (over 50%) than juniors, leveraging their expertise to verify output and guide the AI effectively.
• ⚠️ While AI can skip the "desert" of boilerplate and syntax errors, there's a risk of losing the ability to build mental models if AI is overused, highlighting the need for a balance in learning.

AGI, Agents, and Infrastructure Challenges

• 🤖 The vision of Artificial General Intelligence (AGI) as a "fully autonomous remote worker" is not yet realized, with current AI agents often described as "slop" due to their unreliability in end-to-end tasks.
• 🚧 Robotics still face a significant "sim to real gap," as the messy real world presents challenges that perfect simulations cannot replicate, especially concerning safety.
• ⚡ The desperate hunger for compute infrastructure is pushing boundaries, with serious discussions about placing compute clusters in orbit (space servers) to overcome power and cooling limitations.

The Future of AI and Knowledge

• 🌱 AI is best viewed as a "bicycle for the mind," amplifying human expertise rather than replacing the effort of learning and mastery.
• 📚 The revolution lies in making the entirety of human knowledge accessible and personalized to anyone, anywhere, instantly.
• ❓ A critical question for the future is whether advanced AI insights will remain a public utility or become a luxury good, accessible only through expensive subscriptions.