What Import AI is
Import AI is a weekly newsletter written by Jack Clark, one of the co-founders of Anthropic. Each issue is a curated digest — typically three or four distinct items — that synthesizes recent developments across AI research, safety, industry, and geopolitics. It is not a news wire; it is closer to an annotated reading list with editorial commentary, written by someone who has worked at the frontier of AI for years.
Why it's worth reading
Most AI coverage either goes very deep on one technical paper or very shallow on a product announcement. Import AI occupies a different niche: it connects dots across domains. A single issue might cover a new Chinese AI benchmark, a reinforcement-learning result in robotics, and a philosophical question about what it means for AI to participate in its own development — and it treats all three as part of the same unfolding story.
That editorial sensibility reflects Clark's background. As an Anthropic co-founder, he has a practitioner's eye for which technical signals actually matter and a safety researcher's habit of asking what could go wrong.
What a typical issue looks like
Each issue is numbered sequentially. Recent issues (439–464, spanning roughly January through July 2026) follow a consistent structure: a handful of short summaries of papers, products, or events, each with Clark's framing of why it matters, sometimes followed by a speculative "framing question" or a short piece of narrative writing.
Topics covered in this window include:
- AI self-improvement and agents: Issues 455 and 456 covered AI systems beginning to automate AI research itself and the economic implications of recursive self-improvement. Issue 441 asked whether AI agents are reliably working yet — and covered adversarial "poison fountain" attacks against them.
- Cybersecurity and offensive AI: Issues 450, 452, and 457 covered scaling laws for cyberattacks, AI-enabled Stuxnet-style scenarios, and the industrialization of cyber espionage.
- Geopolitics and compute: Issues 446, 449, 463, and 448 covered Chinese AI benchmarks, a 72B-parameter distributed training run, a reported 10,000-GPU Chinese compute cluster, and ByteDance's CUDA-writing agent.
- Benchmarks and evaluation: Issues 445, 444, and 461 introduced or discussed new ML research benchmarks, ChipBench (for chip design), and FrontierCode (for coding).
- Big-picture framing: Issues 458, 462, and 463 included essays on paths to artificial superintelligence, the quasi-religious nature of singularity beliefs, and reflections on the transition away from a human-dominated era.
The speculative layer
One thing that distinguishes Import AI from purely technical digests is its willingness to sit with hard questions. Issues regularly pose "framing questions" — when will financial markets price in transformative AI? Is superintelligence a discrete phase change or a gradual shift? How might a timeless, persistent AI agent relate to time? These aren't answered; they're offered as lenses.
This makes the newsletter useful not just as a technical update but as a map of what serious people at the frontier are worried about and excited by.
Where to situate it
Import AI is a secondary source — it synthesizes and contextualizes rather than breaking primary research. Its value is curatorial: Clark reads widely and flags what matters, saving readers the work of monitoring dozens of preprint servers, lab blogs, and conference proceedings. For anyone trying to stay oriented in a fast-moving field without drowning in it, that curation is the product.
