New Lex Fridman Insight: State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex
Sent May 31, 2026
Key Insights
- DeepSeek's R1 model set a new benchmark in 2025 by achieving state-of-the-art performance at reduced compute costs.
- Chinese AI firms like Z.ai are gaining ground with open-weight models, challenging DeepSeek's dominance.
- GPT-5.2's context score improvement from 30% to 70% highlights significant algorithmic advances.
- Reinforcement Learning with Verifiable Rewards (RLVR) can dramatically improve model accuracy in just 50 steps.
- The AI ecosystem in China is rapidly advancing, with multiple open-weight models emerging in 2026.
How the conversation moved
The episode begins with Lex Fridman framing the discussion around the current state and future trajectory of AI, focusing on the competitive landscape between US and Chinese AI firms. Sebastian Raschka and Nathan Lambert dive into the dynamics of AI models and companies, highlighting the impact of DeepSeek's R1 model, which set a new benchmark in 2025. They discuss the evolving dynamics between US and Chinese firms, noting that Chinese companies like Z.ai are gaining ground with competitive open-weight models, challenging DeepSeek's dominance.
Raschka argues that no single company will have exclusive access to technology due to frequent job changes among researchers. Lambert adds that the hype around Anthropic's Claude Opus 4.5 has overshadowed Google's Gemini 3, despite Gemini being a strong model. They explore the advancements in AI models, with GPT-5.2's context score improvement from 30% to 70% indicating significant algorithmic advancements. The conversation also touches on the role of open-weight models, with Chinese firms releasing models that are gaining popularity due to their unrestricted licenses.
Despite the comprehensive exploration of AI advancements, there is little pushback from Lex on the guests' claims. The conversation lacks explicit challenges to the optimistic view of AI's trajectory, such as the potential risks of open-weight models or the sustainability of current AI development practices. However, Raschka does question the notion that AI could replace the joy of debugging, arguing that the struggle is part of the learning process, which could be diminished by over-reliance on AI tools.
The discussion concludes with an examination of the broader implications of AI advancements, including the economic impact and the potential for AI to drive significant GDP growth over time. Lambert and Raschka emphasize the importance of open models in the US to compete with China's rapidly advancing AI ecosystem. They also touch on the historical impact of GPUs on deep learning and future technological advancements, suggesting that the future of technology may not solely be defined by AI but by the broader concept of computing and its evolution.
Surprising moments
In-depth
AI Competition
- DeepSeek's R1 model set a new industry standard in 2025.
- Chinese firms like Z.ai are challenging DeepSeek with open models.
- US models outperform Chinese models in outputs, affecting market dynamics.
Model Performance and Usage
- GPT-5.2's context score improved from 30% to 70%.
- Open models are known for weights, not usability.
- Claude Code offers an engaging interface for projects.
Reinforcement Learning and AI Training
- RLVR can improve model accuracy significantly in 50 steps.
- Scaling laws predict relationships between compute and accuracy.
- RLHF scaling does not show linear performance improvements.
Chinese AI Ecosystem
- China's AI advancements challenge US dominance.
- Multiple Chinese open-weight models emerged in 2026.
- Chinese models use larger Mixture of Experts architectures.
Notable Quotes
I don’t think nowadays, in 2026, that there will be any company having access to a technology that no other company has access to.
Still open
- Raschka questioned how future developers will learn if they rely too heavily on AI, suggesting it could hinder their development into experts.
- Lambert expressed skepticism about the feasibility of fully automating programming in the near future.