MON · NOV 16, 2026Superintelligence.AI & FINANCE
 
TODAY’S LEAD
Qwen4-Max Is Here, and the Numbers Disappoint
Alibaba's new flagship looks like a match for Claude Opus 5.5. An independent ranking puts it in twelfth place
READ THE LEAD STORY →
Alibaba/Qwen, Qwen4-Max launch demo
By Kim IsenbergForwarded? Subscribe free →
IN PARTNERSHIP WITH
Arca

Dear Readers,
Thank you for reading us for so long. Today we want to let you know that we have redesigned the newsletter. We hope you like it!
The Signal · From the editor · Kim Isenberg
Today is about the gap between what companies claim and what outside checks show.
12th
Qwen4-Max's place in an independent ranking
38%
of new arXiv papers in one sample read as AI-written
$25.3T
invested in the listed companies most tied to AI
Alibaba's own charts put its new Qwen4-Max next to the best models from Anthropic and OpenAI. An independent ranking puts it in twelfth place. A developer's analysis finds that 38% of new research papers on arXiv now read as if AI wrote them. At Nvidia, Jensen Huang is talking about a comeback, and tens of thousands of supporters, Elon Musk among them, say the board pushed him out too fast. And on Wall Street, big investors are going through their portfolios to find out how much AI risk they really hold. The lesson for today: trust the numbers someone else has checked.
All the best,Kim Isenberg signatureKim Isenberg
01  ·  IN THE NEWS
BUSINESS  ·  Nvidia
Jensen Huang Weighs a Comeback at Nvidia
Bloomberg's report on Huang's comeback talks, citing the Financial Times. Image: Source: Bloomberg, "Huang Discusses Comeback as Nvidia Chief With Allies, Report Says", 11/15/2026; highlight added
→
IN SHORT
Huang wants back in and has Silicon Valley behind him, but the board, its investigation and Washington still stand in the way.
Update on NVIDIA: Jensen Huang is talking with allies about returning as Nvidia's CEO, just days after the board pushed him out, Bloomberg reports, citing the Financial Times. His supporters are pushing hard: tens of thousands have signed a "Bring Jensen Back" open letter, including Elon Musk and Epic Games CEO Tim Sweeney. Treasury Secretary Scott Bessent wants a "dramatic shake-up" instead. Replacing Huang won't be easy, The Washington Post notes, because he ran the company he co-founded for more than 30 years. For now, interim CEO Colette Kress is in charge, and Nvidia reports earnings on Tuesday. One complication, according to Reuters: in October, Huang told Nvidia's top lawyer that his dealings with Megaspeed and Alice Huang were normal customer business. The board's investigation is still running.
READ AT BLOOMBERG  →
ARXIV
Detector Flags 38% of New arXiv Papers as AI-Written
Share of new arXiv papers flagged as machine-written, by submission month, with 95% confidence bands. Image: unslop
→More than a third of new research papers in this sample read as AI-written, up from almost none before ChatGPT.
A developer checked 13,500 research papers on arXiv from 2021 to 2026 with a self-built AI-text detector: 38% of the papers from August to October 2026 read as AI-written. Before ChatGPT, the same detector flagged just 0.4% of papers. It also misses about a fifth of AI text, so the real share is probably higher. Two limits: a flag only means the text sounds machine-written, not that AI did the research. And the author also sells a tool that makes AI text sound human.
READ AT UNSLOP  →
STATE DEPARTMENT
US and Nine Allies Sign a Pact on Super Intelligence Supply Chains
The State Department's note on the San Francisco Declaration, which speaks of Super Intelligence supply chains. Image: Source: U.S. Department of State, 11/15/2026; highlight added
→Washington's AI alliance now has a formal plan to keep the technology behind advanced AI away from hostile hands, and ten governments have written superintelligence into an official text.
The US and nine partners, including Japan, South Korea, India, Israel, the UK and the UAE, signed the San Francisco Declaration on November 15, the State Department said. They promise to jointly protect what the text calls "Super Intelligence (SI) supply chains": the goods, software, materials and infrastructure that advanced AI depends on. Five working groups are meant to turn this into practice. A footnote explains that for some signatories, the term simply means artificial intelligence. Germany, Sweden, the Netherlands and the European Commission are set to sign next week, Reuters reports, citing the Financial Times.
READ AT STATE DEPARTMENT  →
ADVERTISEMENT

Your old AI sessions are worth money

Arca pays you for the AI sessions you'd otherwise throw away

Got hundreds of Claude Code, Codex or Cursor sessions just sitting there? Might as well get paid for them.

Arca pays you for them. It’s a free desktop app for Mac and Windows, and as of November 11, it’s open to everyone. There’s no invite or waitlist.

  • You pick what to share. Nothing uploads in the background.

  • Private details stay private. Names, file paths, account details and keys are removed on your computer before anything leaves it.

  • You see a price first. Each session gets a quote range. Accept it to send the session for review, or withdraw it and the upload is deleted.

  • Paid weekly for approved sessions. Code, research, writing and agent tasks all count, and so do past sessions.

02  ·  VIDEO OF THE DAY
Watch: Jim Cramer on interim Nvidia CEO: 'I like Colette very much,' 'will provide needed order'
CNBC TELEVISION
Jim Cramer on interim Nvidia CEO: 'I like Colette very much,' 'will provide needed order'
On the morning Huang resigned, CNBC asked Jim Cramer, a longtime fan of Huang and the stock, what comes next for Nvidia. Cramer thinks the exit had to happen, but he disagrees with analysts who see Nvidia as a one-man company. He trusts interim CEO Colette Kress: "I think she'll provide some needed order." And on the 42,000 employees: "It's made up of really brilliant people. And I think that Jensen probably put them together." Analyst Chris Caso adds the good news: Nvidia is sold out through the end of next year. Under four minutes, and a good companion to today's Huang news.
WATCH ON YOUTUBE  →
03  ·  QUOTE OF THE DAY
It’s very possible that AI will end up making companies a lot more productive but not necessarily more profitable.
Peter Berezin
CHIEF ECONOMIST AT BCA RESEARCH
Speaking to Reuters
BCA Research
He said this as big investors try to figure out how much AI risk is hiding in their portfolios (more in today's Daily Feature).
SOURCE: REUTERS: AI SELLOFF HAS INVESTORS SCRAMBLING TO MEASURE THEIR TRUE EXPOSURE →
04  ·  RUMOURS, LEAKS & DUSTUPS
Where is GPT-6.2? On November 12, the leak account Token Gremlin posted that some resellers were briefly offering GPT-6.2 Sol and GPT-6.2 Astra, with prices, apparently because of a mistake at Microsoft's Azure.
The name had already shown up in Codex usage data in October. Four days later, OpenAI still hasn't announced either model.
OpenAI
GPT-6.2 LEAK
X/@TokenGremlin
THE LEAD STORY3 min read
ALIBABAMODEL RELEASE
Qwen4-Max Looks Frontier-Class on Alibaba's Chart. Independent Tests Disagree.
Measured on a broad set of independent tests, Qwen4-Max falls far behind top models like Opus 5.5.
43
points on Artificial Analysis's index, 12th place
$3.27
per task, about a third of its predecessor's cost
28.8M
Claude exchanges Anthropic says Alibaba used to distill
Alibaba/Qwen, Qwen4-Max launch demo
THE TAKEAWAY
→Alibaba launched Qwen4-Max on November 15. The model is huge (2.4 trillion parameters), can work through very long documents and is only available through Alibaba's own services.
→On Alibaba's own chart it beats Claude Opus 5.5, GPT-6 Astra and the other rivals on 4 of 16 tests, including the coding test DeepSWE.
→In the independent Artificial Analysis ranking it lands in 12th place with 43 points, 22 behind the leader, Claude Fable 5.2.
→Compared to its predecessor it is a big step: 9 points better and about three times cheaper per task.
Alibaba launched Qwen4-Max on November 15 and calls it "your new AI teammate for work, creativity, and discovery." The model is huge, with 2.4 trillion parameters, but only 71 billion of them are active for each word it processes, which keeps it cheaper to run. It can also read images and handle very long documents. Alibaba's demos focus on real work: the model builds 3D scenes, makes games in engines like Unity and Unreal, and in one factory example cut the time to program a machine part from about 475 minutes to 24. It is only available through Alibaba's own API and apps.
Four of the 16 panels on Alibaba's launch chart: Qwen4-Max (dark blue) against Qwen3.8-Max, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1. Image: Alibaba/Qwen, cropped
On Alibaba's own chart, Qwen4-Max looks like a top model. Alibaba compares it with Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1 on 16 tests. Qwen4-Max wins four of them, including the coding test DeepSWE (75.7), and comes close on several others. The independent testers at Artificial Analysis see it very differently. Their Intelligence Index combines nine tough tests into one score. Qwen4-Max gets 43 and lands in 12th place: 3 points behind Kimi K3.1 Pro (46), 16 behind Opus 5.5 (59) and 22 behind the leader, Claude Fable 5.2 (65), which Alibaba left off its chart. Alibaba's numbers are not wrong: where both measured the same test, the results match. The difference is which tests and which rivals you look at. Measured on a broad set of independent tests, Qwen4-Max falls far behind top models like Opus 5.5.
Artificial Analysis Intelligence Index v5.0.1: Qwen4-Max (orange, 43); the purple arrow marks the jump from Qwen3.8 Max (34, outside this crop). Image: Artificial Analysis, cropped
Compared to its predecessor, the progress is real. Qwen4-Max scores 9 points more than Qwen3.8 Max and does better on every test in the index. It is also much cheaper: one task in the index costs $3.27, about a third of the predecessor's $10.02 and around 11% less than Kimi K3.1 Pro. Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens. So: a solid step forward, but a sobering result overall. One likely reason: several Chinese labs have trained their models on answers from stronger models like Claude, a shortcut called distillation. Anthropic says its new defenses have cut the data such labs can pull from Claude by around 90 to 99%. According to Anthropic, Alibaba used that shortcut on a massive scale: in a June letter to the US Senate Banking Committee, Anthropic accused it of "the largest known distillation attack on Anthropic to date", 28.8 million exchanges with Claude through roughly 25,000 fake accounts in about six weeks. Now that this route is largely closed, a gap like this one is exactly what you would expect.
→
WHY IT MATTERS
Anyone choosing an AI model should look at independent rankings, not at the maker's own charts. Measured fairly, Qwen4-Max is a cheap, fast-improving model, but it is 16 points behind Opus 5.5.
READ AT ARTIFICIAL ANALYSIS  →
More sources: Qwen · Anthropic · CNBC
ADVERTISEMENT

Find what’s slowing your GTM.

Most GTM problems aren’t one big failure. They’re one or two weak spots quietly dragging down the rest.

The free GTM Audit scores your go-to-market across four dimensions, then sequences a 90-day roadmap for your stage.

Drop your URL. Get your score in about 10 minutes.

05  ·  GRAPH OF THE DAY
Image: Source: unslop, "38% of sampled arXiv papers read as AI-written"
COMPUTER SCIENCE
62%
→
MATHEMATICS
3.7%
Share of new arXiv papers that read as AI-written, Nov 2025 to Oct 2026
→  The chartIt comes from the same analysis as News Story 2 and shows, for each field, how many new arXiv papers from November 2025 to October 2026 read as AI-written. Computer science is highest at 62%, followed by quantitative biology (58%), electrical engineering (51.3%) and economics and finance (47.7%). Mathematics is lowest at 3.7%, astrophysics is at 11%. Before ChatGPT, every field was between 0% and 1.5%.
→  The lessonHow much AI writing shows up depends heavily on the field. In computer science, almost two out of three new papers now read as AI-written; in math, hardly any do. The people who build AI are also the ones who use it most to write their papers.
→  The caveatThe math figure may be too low: math papers are mostly formulas, which leaves the detector little text to judge.
Daily Feature: Monday, AI & Finance
Wall Street Can't Tell How Much AI It Owns
→  Bottom lineAfter Nvidia's slide, big investors are checking how much AI risk sits in their portfolios, and many can't say for sure.
→  Why it mattersAt least $25.3 trillion is invested in the listed companies most tied to AI, and that doesn't even count bonds or private deals.
→  What it meansIf the AI boom cools, it can also hit investments nobody thought of as AI bets.
38.5%
of big pension funds' stock holdings sit in 52 AI-related stocks
$1.8T
of Nvidia's market value lost since its October 6 peak
"Since when was a wedding gown shop a play on AI?" asked Andy Goldberg of Nomura Asset Management, after the bridal chain David's Bridal announced plans to take its AI wedding-planning business public. His question is at the heart of a Reuters analysis published on Monday. Since its peak on October 6, Nvidia has lost more than 30% of its value, over $1.8 trillion. Now fund managers are discovering AI risk in places they never thought of as AI.
David's Bridal plans to spin off its AI wedding-planning platform, Pearl, as a public company. Image: Courtesy of David's Bridal, via Retail Dive
The amounts are huge. According to the index provider MSCI, at least $25.3 trillion is invested in the listed companies most directly tied to AI. Pension funds are deep in, too: at a group of large pension funds, just 52 AI-related stocks made up 38.5% of all their stock holdings in June 2026, according to the Equable Institute. That is 9% to 11% of their total assets, and bonds and private investments aren't even counted.
“Since when was a wedding gown shop a play on AI?”
ANDY GOLDBERG, NOMURA ASSET MANAGEMENT
So big investors are now going through their holdings line by line. PGIM, which manages nearly $1 trillion in bonds, checks every bond and loan it holds with two questions: how does AI change this company's business, and what happens to it if AI spending slows down? Spotting the obvious bets, a data center or a chipmaker, is easy, says PGIM's Brian Barnhurst. "But how about the indirect things?"
Inside a switchgear room: the kind of power-distribution equipment Powell Industries builds for data centers and other heavy users. Image: Powell Industries
Goldberg puts it bluntly: "No one fully understands how much exposure they have to AI; it’s really become just one big trade at this point." If investors pulled even 1% or 2% out of AI stocks, he warns, it could look like a crash. How fast those bets can swing shows at Powell Industries, which makes power equipment for data centers. Four years ago its stock cost $8. Riding the data center boom, it climbed as high as $328 this year. Now it is back at $146, less than half its peak. Nvidia's earnings on Tuesday are the next test.
READ AT REUTERS  →
WANT MORE? 
The full Saturday DeepDive every week, plus discounts on AI software and member giveaways.
JOIN SUPERINTEL+  →
$10 a month or $100 a year. Students free with a .edu address.

Reply

Avatar

or to participate