Nathan Lambert, Sebastian Raschka: State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI

Raschka rejects a single winner-takes-all story. Researchers move between organizations, so ideas can spread even when budgets and hardware remain unequal. Lambert adds that organizational culture and product focus matter: Anthropic’s code emphasis differs from broad consumer chatbots, and an enthusiastic technical audience is not the same as a large general user base. Their forecasts are judgments at the recording date, not settled rankings.

Chinese open-weight competition extends beyond DeepSeek to Qwen, GLM, MiniMax and Kimi. Lambert sees openness as a route into markets reluctant to buy Chinese APIs, while Raschka emphasizes leapfrogging rather than permanent technological exclusion. Google’s integrated infrastructure may improve economics; OpenAI’s history of landing new research products remains a counterweight. Even subscription habits, memory and separation of work from private use can sustain multiple models.

The speakers disagree about speed versus deliberation. Raschka wants immediate answers for simple commands but can leave a thorough manuscript check running over dinner. Lambert prefers thinking or pro models for work queries because he values lower error rates and source integration. Fridman describes loyalty arising from a single successful capability, such as long-context retrieval, and switching after failure. These anecdotes illustrate user selection, not a controlled model comparison.

Model quality is inseparable from its surrounding product. Open weights do not automatically supply a polished platform or fast serving. Programming preferences also vary: Raschka favors a repository-aware assistant that preserves close control, while Fridman uses Claude Code to practice directing software in natural language. For learning, Raschka recommends focused reading before a second AI-assisted pass; Fridman often uses a model first to establish context. Both value engagement without disappearing into unrelated internet distractions.

The open ecosystem contains different kinds of openness. Lambert distinguishes downloadable weights from efforts such as OLMo that also release data and code, allowing others to reproduce training. He notes larger Chinese mixture-of-experts models alongside smaller Western releases, while anticipating changes in that balance. Permissive licensing, local deployment, customization and third-party hosting all influence adoption; openness is therefore more than a benchmark score.