jack-clark.net

Import AI is a newsletter about AI research written by Jack Clark and published at jack-clark.net. Each issue summarizes recent AI research and developments — often drawing on arXiv — with Clark's commentary layered on top. It is aimed at readers who want a regular digest of AI research news rather than a general technology feed. Issues are numbered and dated, and the newsletter asks readers to subscribe to support it.

Who writes it and how it is structured

The newsletter is authored by Jack Clark, and the site describes it as "a newsletter about AI research" that "runs on arXiv, cappuccinos, and feedback from readers." Issues carry a sequential number and a date in the title — for example, "Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space" — followed by a welcome note and then a series of research items.

Each item typically follows a recognizable pattern:

• A headline finding stated in plain language
• A summary of the paper or release, including who built it and what it measures
• Specific results, often with numbers and category breakdowns
• A "Why this matters" section that places the work in a broader context

That last section is where much of the newsletter's value sits: it connects a single benchmark or model release to larger questions about how AI systems are developing.

What a typical issue covers

The content is research-first but not limited to academic papers. A single issue can move from a new benchmark to open-weight model comparisons to speculative applications. In issue 476, for instance, the lead item covers "Humanity's Sixth Sense" (HSS), a benchmark built by AI startups Elorian and Scale AI that "measures how a model interprets and reasons through information a visual scene implies but doesn't explicitly show."

The item illustrates the newsletter's typical depth. It describes the benchmark's composition — 522 questions covering 288 images and 234 short video clips, totaling 17.6 hours of video — and organizes the questions into four categories:

It then reports results against a human baseline: people scored an aggregate pass@1 of 93.1 (image 94.1, video 91.9, time 97, space 94.9), while the best-scoring AI system, GPT 6 Astra, reached 53.6 (image 55.6, video 51.1, time 58.2, space 53.6). Chinese open-weight models scored lower, with GLM-5.3-flash at 26.8 and Kimi K3 at 25.5. The item also names the two most common failure modes — "missing a detail that decides the answer (21% of failures)" and "misidentifying an object, person, or role (20%)" — before moving into interpretation.

How to read it and what to expect

The newsletter is best treated as a curated briefing, not a tutorial or a primary source. It points you toward papers and releases and gives you the headline numbers and the author's read on why they matter; the underlying work is where you go for methodology and caveats.

A few practical notes for new readers:

• Expect interpretation alongside summary. The "Why this matters" framing is editorial. In the HSS item, Clark reframes the benchmark as a test of whether world models are emergent, noting that "many people have claimed that true visual reasoning will rely on developing spatially aware 2D" systems — a claim the results complicate.
• Expect numbers you can check. Results are reported with specific scores and category splits, which makes it easy to trace back to the source paper.
• Expect a mix of maturity levels. A single issue can pair a rigorous benchmark with more speculative topics, as the issue 476 title suggests with its references to creativity, recursive self-improvement, and AI in space.

Access and subscription

The newsletter is available at jack-clark.net, and the site includes a "Subscribe now" link pointing to importai.substack.com/subscribe. The page evidence shows a subscription prompt but does not state pricing or whether any content is gated, so check the subscribe page directly for current terms before assuming free access to everything.

If you want a recurring, opinionated digest of AI research with enough detail to follow up on the primary sources, Import AI fits that use case. If you need exhaustive coverage of the field or hands-on technical guidance, use it as a pointer to the papers rather than as the endpoint.

To subscribe to Import AI, use the official "Subscribe now" link on jack-clark.net, which points to the newsletter's Substack page at importai.substack.com/subscribe. Signing up there is the way to receive new issues by email and to support the newsletter; the site itself does not describe any other sign-up route, so third-party or unofficial forms should be avoided.

What you are subscribing to

Import AI is Jack Clark's newsletter about AI research. Issues combine commentary with summaries of recent work — for example, issue 476 (October 12, 2026) covers a new benchmark on intuitive visual reasoning, open-weight models, and AI in space. The newsletter describes itself as running "on arXiv, cappuccinos, and feedback from readers," and it explicitly invites readers to support it by subscribing.

Steps to subscribe

1. Go to jack-clark.net and locate the "Subscribe now" link in the newsletter text.
2. Follow it to https://importai.substack.com/subscribe — this is the official subscription page referenced on the site.
3. Complete the sign-up on that Substack page. Substack newsletters typically ask for an email address and send a confirmation message; follow the confirmation to activate delivery.
4. Check your inbox (and spam folder) for the confirmation and for the first issue after subscribing.

Expected result: new Import AI issues arrive by email, and your subscription counts as reader support for the newsletter.

Common snags

• Confirmation email missing: check spam or promotions folders before trying again; without confirming, delivery usually will not start.
• Wrong sign-up page: if a link does not lead to importai.substack.com, it is not the official route described on the site.
• Pricing questions: the site references subscribing and a Substack subscribe page but does not state terms or prices, so check what the Substack page shows at sign-up rather than assuming a free tier.

Alternatives if you only want to browse

You can read issues on jack-clark.net without subscribing; subscribing is presented as the way to get updates delivered and to support the work. The newsletter also welcomes feedback from readers, which the site lists alongside subscriptions as part of how it keeps running.

Import AI issue 476, dated October 12, 2026 and written by Jack Clark, leads with a new benchmark on intuitive visual reasoning and also touches on creativity and recursive self-improvement (RSI), open weight models, and AI in space. The issue is aimed at readers who follow AI research and want a concise digest of recent papers and results; the newsletter describes itself as running "on arXiv, cappuccinos, and feedback from readers" and invites readers to subscribe.

The main item: Humanity's Sixth Sense benchmark

The centerpiece is Humanity's Sixth Sense (HSS), a benchmark built by AI startups Elorian and Scale AI. It "measures how a model interprets and reasons through information a visual scene implies but doesn't explicitly show" — that is, intuitive visual reasoning rather than explicit recognition.

What's in it: 522 questions covering 288 images and 234 short video clips (17.6 hours of video total). Questions fall into four categories:

Example tasks include looking at a bookcase and judging whether it can fit more books, or a photo of horses pulling a plough and judging whether the chains bear the plough's drag weight.

How humans and models score

The human baseline is high but not perfect: people score an aggregate pass@1 of 93.1, broken down as image 94.1, video 91.9, time 97, and space 94.9.

The best AI system, GPT 6 Astra, reaches pass@1 of 53.6 (image 55.6, video 51.1, time 58.2, space 53.6). Chinese open weight models do notably worse: GLM-5.3-flash at 26.8 and Kimi K3 at 25.5.

The two most common single mistakes, per the benchmark authors, were missing a detail that decides the answer (21% of failures) and misidentifying an object, person, or role (20%).

Why the issue says it matters

Clark frames HSS less as a list of deficiencies and more as evidence of how adept frontier LLMs already are at tasks many assumed would need different, specialized AI systems. The open question the benchmark speaks to: whether world models are emergent in today's frontier models, given that true visual reasoning was long expected to require spatially aware 2D systems.

Other items in the issue

The issue's title also flags creativity and RSI, worse-than-China open weight models, and AI goes to space. The available excerpt details only the HSS benchmark, so treat the remaining items as listed topics rather than summarized findings.

Who should read it

Issue 476 suits readers tracking benchmark design, multimodal reasoning, and the open weight model landscape. If you want the full text and future issues, the newsletter points to its Substack subscription page; the excerpt does not state pricing or access terms, so check the subscribe page for those details.

Humanity's Sixth Sense (HSS) is a benchmark that tests whether AI models can reason about what a visual scene implies rather than what it explicitly shows. In the results reported in Import AI 476, humans scored a pass@1 of 93.1 overall while the best AI system, GPT 6 Astra, reached 53.6 — a gap of roughly 40 points. The benchmark was built and released by AI startups Elorian and Scale AI.

What the benchmark measures

HSS targets "intuitive" visual reasoning: given an image or short video, can a model infer facts the scene suggests but never states? The benchmark's own framing is that it "measures how a model interprets and reasons through information a visual scene implies but doesn't explicitly show."

Two example task types from the newsletter:

• A picture of a bookcase, with the question of whether it can fit more books.
• A photo of horses pulling a plough, with the question of whether the chains bear the plough's drag weight.

Dataset composition

The four question categories

1. Temporal & causal dynamics — "What happened before the scene, and what happens next."
2. Physical & spatial logic — "What fits, what's in reach, and what someone else can see from where they are standing."
3. Social understanding — "Intent, attention, belonging."
4. Abstract & contextual inference — "The patterns and rules a scene implies."

Results: humans vs. the best AI system

The human baseline is described as "not perfect, but very high." The breakdown by category:

Chinese open-weight models performed notably worse on this benchmark: GLM-5.3-flash scored 26.8 and Kimi K3 scored 25.5.

How the AI systems failed

The benchmark authors report two dominant error patterns:

• Missing a decisive detail — 21% of failures. The model overlooks the one element in the scene that determines the answer.
• Misidentifying an object, person, or role — 20% of failures.

Together these account for about 41% of failures, meaning most AI errors are not exotic reasoning breakdowns but basic perception-and-attention misses on details that carry the answer.

Why the result matters

Import AI frames HSS as a test of whether "world models" are emergent in current systems. The benchmark can be read two ways:

• As a deficiency report: frontier models still fail at inference tasks humans handle almost effortlessly, especially where a small visual detail decides the answer.
• As a capability surprise: many people assumed true visual reasoning would require purpose-built, spatially aware systems. The fact that general-purpose frontier LLMs reach the mid-50s on tasks of this kind is itself informative about how much implicit scene understanding emerges from scale.

The practical takeaway for anyone evaluating models on visual reasoning: aggregate scores hide category-level differences, and the human–AI gap is largest on time-based inference (97 vs. 58.2) and smallest on video (91.9 vs. 51.1) — though even the smallest gap is about 40 points. If your task depends on inferring what a scene implies rather than describing what it contains, expect current systems to miss the decisive detail roughly one time in five.