Jack Clark’s weekly newsletter, Import AI, has become one of the most-forwarded documents in AI policy circles, read inside labs like Anthropic, OpenAI, and Google DeepMind as a barometer of where the field is actually heading. The newsletter’s latest edition, tracked by aggregators as Import AI 476, is circulating under a headline that ties together three threads: AI creativity and recursive self-improvement (RSI), a widening gap between Western and Chinese open-weight models, and the slow but real migration of AI hardware into orbit. Each thread touches a live story already playing out across the industry, and together they sketch a picture of an AI race that is getting faster, more open on one side of the Pacific, and more literal about reaching for the stars.
What follows separates what reporting can actually confirm about issue 476 from what remains headline shorthand, while digging into the three underlying stories that make the newsletter’s framing worth taking seriously regardless of the exact wording Clark used.
What Import AI 476 Reportedly Covers
Import AI publishes on a near-weekly cadence from Jack Clark’s Substack archive, and the issue immediately preceding the one in question, numbered 475 and dated October 5, 2026, carried a title spanning swarm scaling, Google DeepMind’s biology watermarking work, and the AI science economy. That confirmed publication cadence puts issue 476 squarely in the window of October 9 to 12, 2026, consistent with the newsletter’s usual weekly rhythm, which outlets like NewsletterScan track alongside dozens of other AI-focused mailing lists.
The three themes attached to issue 476, creativity and RSI, a Chinese open-weight lead, and AI moving into space, are headline-level descriptions rather than verified excerpts from the newsletter’s text. No publicly indexed copy of the full issue could be located at the time of writing, and no secondary outlet had yet published a detailed breakdown of its contents. That matters for how this piece treats specifics: names, dates, and benchmark figures below are drawn from verifiable reporting on each underlying story, not from a confirmed transcript of the newsletter itself.
| Reported Theme | Headline Description | Independent Confirmation Status |
|---|---|---|
| Creativity and RSI | AI systems edging into tasks once considered uniquely creative, plus rising discussion of recursive self-improvement | Partially supported; ties to OpenAI’s internal math-proof episode and Clark’s public RSI commentary |
| Worse-than-China open models | Western open-weight releases trailing Chinese counterparts on key benchmarks | Not independently verified for issue 476 specifically; matches a broader pattern seen in recent open-weight launches |
| AI goes to space | AI compute infrastructure moving into orbit | Context exists via Google’s Project Suncatcher satellite launch on October 1, 2026 |
Who Jack Clark Is and Why the Newsletter Carries Weight
Clark co-founded Anthropic and has spent years writing Import AI as a running commentary on AI research, policy, and the gap between lab announcements and what papers actually show. The newsletter’s influence comes less from traffic numbers, which newsletter trackers like InboxReads list alongside peer publications, and more from who reads it. Policy staffers, lab researchers, and journalists treat it as a filter for separating genuine technical movement from marketing. When Clark frames something as a trend worth three paragraphs, labs tend to notice, and competitors tend to respond within days rather than months.
That reputation is exactly why a headline suggesting the West trails China on open weights, or that recursive self-improvement is moving from theory to practice, travels fast even before anyone outside the newsletter’s subscriber base has read the full text. The claims do not need to be perfectly sourced to move markets and roadmaps. They only need Clark’s byline attached.
Theme One: Creativity, Open Math Problems, and the RSI Question
The creativity thread connects most directly to a story shattered.io has tracked closely over the past month: OpenAI’s push to show that an internal, unreleased model could tackle long-standing open problems in mathematics and computer science. That effort produced a widely disputed batch of 722 claimed proofs, drawing pushback from working mathematicians, including Fields Medal winners, who questioned whether the model’s outputs constituted genuine proof or an elaborate approximation of one. The episode is the clearest public example right now of a lab arguing that AI has crossed into territory, open-ended mathematical creativity, that was assumed safe from automation for years longer.
What Recursive Self-Improvement Actually Means
Recursive self-improvement describes a system that gets meaningfully better at building its successor, not just better at answering questions. It is the scenario AI safety researchers worry about most, because it implies a feedback loop that could compress years of progress into months. Clark has separately estimated, outside the context of any single newsletter issue, a probability of 60 percent or higher that AI capable of helping build its own successor could emerge by the end of 2028. That figure is his own standing estimate rather than a confirmed quote from issue 476, but it explains why any newsletter section pairing “creativity” with “RSI” gets read closely. The math-proof episode is evidence, in Clark’s own framing elsewhere, that creative and research-adjacent capability is arriving faster than many expected, even if full self-improving loops remain unproven.
Theme Two: Is the West Losing the Open-Weight Race to China?
This is the thread with the most concrete evidence behind it, even without a confirmed newsletter excerpt. Over the past several weeks, Chinese labs have shipped a run of open-weight models that landed with real benchmark claims attached, while several Western open-weight efforts have leaned on efficiency and licensing arguments rather than outright capability wins.
Where Chinese Open-Weight Models Stand
Zhipu AI’s GLM-5.3 and Tencent’s Hy4 have both drawn attention for different reasons. Security researchers found that GLM-5.3 could be pushed to a 100 percent safety-bypass rate under adversarial testing, including generating working exploit code, a result that undercuts any argument that open-weight releases are automatically safer because they are inspectable. Meanwhile Tencent’s 770-billion-parameter Hy4 model narrowly beat GLM-5.3 on head-to-head benchmark testing, suggesting Chinese labs are now competing against each other at the frontier of open weights, not just against Western releases. DeepSeek, Alibaba’s Qwen line, and Moonshot AI’s Kimi models round out a field that ships new checkpoints on a cadence Western labs have struggled to match.
How Western Labs Are Answering Back
The Western response has come from a mix of startups and incumbents. Reflection AI’s Beam launched as a direct challenge to GLM-5.2, with the company claiming roughly four times less compute needed for comparable results, a bet it is making explicitly against the Chinese model it names as its benchmark. Mistral answered with Large 4, internally nicknamed Le Chonk, which the company put at 675 billion total parameters with 41 billion active, though a separate estimate pegged the real figure closer to 1.05 trillion, a gap Mistral has not fully resolved publicly. Nvidia’s Nemotron 3.5 shipped at 30 billion parameters with the company signaling a path toward a trillion-parameter successor, while smaller entrants like Microsoft’s Decision-1 and Nace.AI’s Drex 1.5 compete on speed and benchmark score rather than raw scale.
Taken together, no single benchmark proves the West is losing outright. But the pattern of Chinese labs shipping larger, benchmark-topping checkpoints more frequently, while Western open-weight players lean on efficiency and compute-cost arguments, is real enough that a headline like “worse than China” lands as plausible shorthand even without a line-by-line citation.
| Model | Developer | Region | Reported Scale | Headline Claim |
|---|---|---|---|---|
| GLM-5.3 | Zhipu AI | China | Not fully disclosed | Found to hit 100% safety-bypass rate in third-party testing |
| Hy4 | Tencent | China | 770B parameters | Narrowly outperformed GLM-5.3 on benchmark testing |
| Large 4 (“Le Chonk”) | Mistral AI | France / EU | 675B total, 41B active (company figure); ~1.05T per a separate estimate | Positioned as Europe’s open-weight flagship |
| Beam | Reflection AI | United States | 501B parameters (reported) | Claims roughly 4x less compute than GLM-5.2 for comparable output |
| Nemotron 3.5 | Nvidia | United States | 30B shipped; 1T-parameter successor signaled | Tied to Nvidia’s broader AI-factory open-weight strategy |
| Decision-1 | Microsoft | United States | 9B parameters | Claims 35x inference speed versus comparable models |
| Drex 1.5 | Nace.AI | United States | 9B parameters | Scored 58.08 on an independent decision-model benchmark |
Theme Three: AI Hardware Heads to Orbit
The space thread has the clearest paper trail of the three. Google’s Project Suncatcher put a prototype satellite, reportedly refrigerator-sized and carrying four Tensor Processing Units, into orbit on October 1, 2026, aboard a SpaceX Falcon 9 rideshare launch from Vandenberg Space Force Base. The goal was narrow: confirm that AI accelerator hardware can survive and operate reliably in the orbital environment, not demonstrate a working commercial data center in space. Google has already outlined a follow-up test planned for 2027 using two satellites to evaluate high-bandwidth laser links between spacecraft, a necessary step if orbital compute clusters are ever going to talk to each other at useful speed.
Google is not alone in testing this idea. Starcloud previously flew an Nvidia H100 chip into orbit and ran a lightweight version of Google’s own Gemma model on it, while Nvidia separately announced a Space-1 module tied to its Vera Rubin platform as part of a longer-term orbital compute push. None of this adds up to data centers in space replacing terrestrial ones anytime soon. Power generation, thermal management, and radiation hardening remain unsolved at scale. But it does explain why a research-focused newsletter would spend column space on the idea: it is a live engineering question with actual satellites in orbit, not speculative fiction.
Why the Three Themes Are Connected
Read individually, creativity, the open-weight gap, and orbital hardware look like three unrelated news items stitched together for a catchy newsletter title. Read together, they describe a single underlying dynamic: the cost and speed of building frontier AI capability keeps dropping, which widens who can compete and where that competition can physically happen. Cheaper, faster training lets more labs chase creative and research-adjacent benchmarks like the math-proof episode. The same cost curve is what lets Chinese labs ship large open-weight checkpoints on a tight cadence. And the push to move compute into orbit is, at bottom, a bet that energy and thermal constraints, not algorithmic ones, will be the next bottleneck worth solving.
Clark’s newsletter has built its reputation on spotting exactly this kind of connective tissue before it becomes obvious. Whether or not issue 476 draws the line this explicitly, the underlying pattern holds up under scrutiny.
Market Impact: What This Means for Enterprises Betting on Open Weights
For engineering teams choosing between open-weight models, the practical takeaway is less about nationalism and more about procurement risk. A model that tops a benchmark chart but carries a 100 percent adversarial bypass rate, as testing found with GLM-5.3, introduces a different kind of cost than raw inference pricing. Enterprises deploying open-weight Chinese models for coding or agentic workloads now need to budget for additional red-teaming before production use, regardless of how attractive the benchmark numbers look on a vendor slide.
On the Western side, the efficiency argument from companies like Reflection AI and Nvidia is becoming a genuine purchasing factor, not just marketing. A model that needs a quarter of the compute to hit comparable output changes the economics of self-hosting dramatically, especially as GPU rental prices stay elevated. Teams evaluating open-weight deployment in late 2026 are increasingly running parallel evaluations, checking both raw benchmark score and total compute cost per task, before locking in a model family for production infrastructure.
Competitive Landscape: Mapping the Open-Weight Field
The open-weight field has split into roughly three competitive lanes. The first is scale-first, where labs like Tencent and Mistral push parameter counts upward and let raw capability carry the marketing. The second is efficiency-first, where Reflection AI and Nvidia argue that compute cost per unit of output matters more than topping a leaderboard. The third is specialization, where smaller players like Microsoft’s Decision-1 and Nace.AI’s Drex 1.5 target narrow tasks, decision-making and agentic workflows in particular, rather than competing as general-purpose flagships.
Chinese labs currently occupy the scale-first lane almost by default, shipping larger checkpoints more frequently than Western counterparts outside of Mistral. Western labs have instead clustered around efficiency and specialization, a strategy that avoids a head-on parameter race but also means fewer headline-grabbing benchmark wins, exactly the gap a newsletter headline about a “worse-than-China” open-weight field would be describing.
Historical Context: From Research Memo to Policy Barometer
Import AI has run for years as a weekly dispatch mixing paper summaries, lab gossip, and Clark’s own policy commentary, and it predates most of the current generation of AI-focused newsletters tracked by services like newsletter directories covering the AI beat. Its longevity matters here because it gives the publication a track record against which to measure a bold headline. Clark has previously flagged trends, open-source model proliferation chief among them, well before they became conventional wisdom in mainstream tech coverage. That history is why issue 476’s framing, even unverified in its exact wording, is being treated as a signal worth analyzing rather than dismissed as newsletter hype.
What’s Confirmed, What Isn’t, and Why That Gap Matters
It is worth being blunt about the limits of this story. What is confirmed: Import AI is a real, weekly newsletter written by Jack Clark, its most recent fully verified issue (475, October 5) covered swarm scaling and biology watermarking, and an issue tracked as 476 is circulating under a headline combining creativity, RSI, China’s open-weight position, and orbital AI hardware. What is not independently confirmed: the exact text of issue 476, specific benchmark figures the newsletter itself may cite, and whether Clark draws the causal links described above in those exact terms.
That gap does not make the underlying stories less real. The GLM-5.3 security findings, the Hy4 benchmark win, Reflection AI’s compute claims, and the Suncatcher launch all happened and are independently reported elsewhere. What remains unverified is only the newsletter’s specific framing and wording, which is a meaningfully different kind of uncertainty than doubting whether the events themselves occurred.
Industry Reaction and Analyst Commentary
Reaction so far has split along predictable lines. Open-source advocates point to the Hy4 and GLM-5.3 benchmark numbers as proof that restricting model weights does little to protect Western competitive advantage, since Chinese labs are shipping comparable or better open checkpoints anyway. Safety-focused researchers counter that GLM-5.3’s bypass results show open weights carry their own distinct risk profile, one that scales with capability rather than shrinking as models get more useful. Neither camp disputes the raw numbers. They disagree on what the numbers imply policymakers should do next, a debate that has run through AI circles since well before this particular newsletter issue surfaced.
Five Predictions for the Next Six Months
- Chinese open-weight labs will keep shipping larger checkpoints on a faster cadence than Western counterparts through early 2027, widening the scale gap even as Western efficiency claims improve.
- At least one more major Western open-weight release will lead with a compute-efficiency figure rather than a parameter count, following the pattern set by Reflection AI’s Beam.
- Independent red-teaming of Chinese open-weight models will become a standard procurement step for enterprises, driven directly by the GLM-5.3 bypass findings.
- Google’s 2027 two-satellite laser-link test will draw at least one competing announcement from a rival cloud provider exploring orbital compute, even if only at the feasibility-study stage.
- Debate over recursive self-improvement timelines will intensify through 2027 as more labs attempt open-ended research tasks similar to OpenAI’s math-proof push, regardless of how that specific episode is judged in the end.
Frequently Asked Questions
What is Import AI 476?
It is the latest tracked issue of Jack Clark’s weekly Import AI newsletter, circulating under a headline covering AI creativity and recursive self-improvement, a reported Chinese lead in open-weight models, and AI hardware moving into orbit. The exact text of the issue was not independently verifiable at the time of writing.
Who writes Import AI?
Jack Clark, an Anthropic co-founder, writes and publishes Import AI nearly every week via Substack, mixing research summaries with his own policy commentary.
Are Western open-weight AI models really worse than Chinese ones?
There is no single benchmark that settles this cleanly. Chinese labs like Tencent and Zhipu AI have shipped larger, benchmark-topping checkpoints recently, while Western open-weight players like Reflection AI and Nvidia have focused on compute efficiency rather than raw scale. The comparison depends heavily on which metric, parameter count, benchmark score, or compute cost, is being prioritized.
What does recursive self-improvement (RSI) mean?
RSI describes an AI system becoming capable enough to meaningfully help design or train its own successor, creating a feedback loop. It remains a theoretical risk scenario rather than a demonstrated capability, though researchers like Clark have offered public probability estimates for when it might emerge.
What happened with Google’s AI-in-space project?
Google’s Project Suncatcher launched a prototype satellite carrying four Tensor Processing Units on October 1, 2026, aboard a SpaceX Falcon 9 rideshare mission, to test whether AI accelerator hardware can operate reliably in orbit. A follow-up two-satellite test is planned for 2027.
Is GLM-5.3 safe to use in production?
Third-party security testing found GLM-5.3 could be pushed to a 100 percent bypass rate under adversarial prompting, including generating working exploit code. Enterprises considering it for production workloads should budget for independent red-teaming rather than relying on benchmark scores alone.
How often does Import AI publish?
The newsletter runs on a near-weekly cadence. The most recently fully confirmed issue, number 475, published October 5, 2026.
Why does a single newsletter move AI industry conversation this much?
Import AI’s readership skews toward lab researchers, policy staff, and journalists who treat Clark’s framing as a credible filter on genuine technical movement. A multi-year track record of flagging real trends early gives new headlines outsized attention even before the full text circulates widely.
