Benaich: reasoning models and China's open-source surge define this AI moment

Key points

Key takeaways from the Peel interview with Nathan Benaich, founder and solo general partner of Air Street Capital and author of the State of AI Report (January 2026):

Reasoning and tool calling. The biggest research shift of the past year is models moving from memorising the internet to step-by-step reasoning combined with tools — web search, APIs, software — visible in ChatGPT's "thinking" traces and, says Benaich, unlocking computer use and more complex scientific and mathematical work.

Traces may not be truthful. Research questions whether chain-of-thought traces reflect what a model actually did: models are trained to reproduce reasoning that humans rate highly, so the displayed trace can be a plausible post-hoc story rather than the real process.

China's open-source surge. A year after the DeepSeek moment, Chinese systems keep reaching leaderboard headlines, particularly in world modelling and long-form video generation, and Alibaba's Qwen has taken the open-source mantle Meta's Llama held a year ago.

First Chinese model IPOs. MiniMax and Zhipu AI (Z.ai) both filed on the Hong Kong exchange — the first pure-play large-model companies to list, raising several billion dollars each and trading up sharply since; there are no ADRs yet and US-based investors cannot buy MiniMax.

Distillation question marks. Asked who created them, some Chinese models answered "OpenAI", prompting FT coverage and US concerns that Chinese labs train on American frontier outputs — though Benaich notes the answer could equally be statistical memorisation of AI-generated text online.

Open weights is not open source. In AI, training code, data and model weights can each be open or not: releasing "open weights" gives away the trained artefact but not the recipe to reproduce it — the winning F1 car without the engineering.

Bazaar over cathedral. The community wants open everything for reliability, contributions and control, so that a vendor cannot rug-pull users by closing access to tools they have built on.

Eight years of State of AI. The report started in 2018 with Ian Hogarth, is written solo from early August to September and published each October as an open-access document, wagering yearly predictions to sanity-check progress claims.

Read more: The Peel on YouTube · State of AI Report