State of AI 2026: No Winner Takes All

Why No Single Company Will Dominate AI in 2026

Lex Fridman's 490th episode brings together two of the most respected voices in the ML community: Sebastian Raschka, author of Build a Large Language Model from Scratch and Build a Reasoning Model from Scratch, and Nathan Lambert, post-training lead at the Allen Institute for AI (AI2) and author of the definitive book on RLHF. Recorded in January 2026, it's a sprawling three-hour technical tour of where AI actually stands — behind the hype cycles.

On the AI race between the US and China: "I don't think nowadays in 2026 that there will be any company who has access to a technology that no other company has access to. Researchers are frequently changing jobs, changing labs, they rotate." Sebastian Raschka argues the real differentiator isn't ideas — it's budget and hardware. Nathan Lambert adds that Chinese companies like Zhipu AI, MiniMax, and Kimi Moonshot have surpassed DeepSeek in some areas, using DeepSeek's own published techniques. The open-weight movement from China is a strategic play for global influence, not charity.

On the Opus 4.5 moment: "The hype over Anthropic's Claude Opus 4.5 model has been absolutely insane... culturally Anthropic is known for betting very hard on code, which is the Claude Code thing that's working out for them right now." Nathan sees Anthropic's focus on code as a cultural advantage, noting the company presents as "the least chaotic" of the major labs. Meanwhile, Google's Gemini 3 was technically impressive but lost the narrative — a recurring theme where organic community enthusiasm matters more than marketing.

On scaling laws and pre-training: Nathan pushes back on the "pre-training is dead" narrative. "I still think most of the compute is going in at pre-training because you can still make a model better... in a few years that'll saturate and the RL compute will just go longer." The discussion reveals a nuanced view: pre-training improvements haven't plateaued, but reinforcement learning is where the next wave of capability gains will come from. They expect $2,000 subscriptions to arrive in 2026 as models offer more cutting-edge capability.

On how AI researchers actually use AI: The conversation gets refreshingly honest about personal AI usage. Sebastian uses ChatGPT's fast mode for quick lookups and Pro mode for thorough document reviews. Nathan exclusively uses thinking models, running five Pro queries simultaneously for paper searches. For coding, both use Claude — Sebastian through the Codex plugin in VS Code, Lex through Claude Code for "programming with English." Nathan uses Gemini for quick explanations and Grok for finding AI Twitter posts. Everyone uses multiple models.

On agents and the interface problem: "The problem is for arbitrary tasks, you still have to specify what you want your LLM to do. What is the environment? How do you specify? You can say what the end goal is, but even getting it to that point — how do you as a user guide the model?" Sebastian raises the fundamental challenge for autonomous agents: specification is hard. The current agent paradigm works for code because the environment is well-defined, but general-purpose agents face an interface problem that may take years to solve.

On open-weight models and policy: Nathan describes leading the TÜLU post-training effort at AI2, one of the few fully open training pipelines in the US. He notes a critical gap: while Chinese labs are flooding the market with open models, US open-weight efforts face consolidation. The 2025 White House AI Action Plan's embrace of open source is encouraging, but translating policy into sustained funding remains the challenge.

7 Key Takeaways on the State of AI in 2026

• No winner-take-all scenario — Ideas flow freely between labs; the differentiating factor is compute budget and hardware, not proprietary techniques
• China's open-weight strategy is working — DeepSeek kicked off a movement, but Zhipu AI, MiniMax, and Kimi Moonshot are now releasing equally competitive models as a deliberate strategy for global influence
• Pre-training isn't dead — Most compute still goes into pre-training, but reinforcement learning post-training is where the next big capability jumps will come
• Multi-model usage is the norm — Even top researchers use 3-4 different models for different tasks, suggesting no single model dominates all use cases
• Code is the killer app for agents — Anthropic's bet on Claude Code is paying off because coding provides a well-defined environment; general-purpose agents still face fundamental specification challenges
• $2,000 AI subscriptions are coming — The jump from $200 to $2,000 tiers will happen in 2026, driven by models offering measurably more capability for professional use
• Open-source AI needs sustainable funding — The US has strong policy support for open AI but lacks the government-backed incentive structures that make Chinese open-weight releases possible

What This Means for Organizations Building with AI

The picture that emerges from this conversation is one of rapid commoditization at the model layer. If no company can maintain a technology moat, the value shifts to how you use AI — the applications, workflows, and integrations. Organizations shouldn't bet on a single model provider. The researchers who build these systems themselves use multiple models daily, switching based on task and quality. That multi-model future isn't a temporary phase; it's the new normal. The real competitive advantage lies in building agent-powered workflows that leverage whichever model is best for each specific task.