2026-11-16 · ~9 min read · Past edition

9 picked from 35 candidates · ordered by significance

Today's Insight

Congress's chip-export push matters more to Nvidia than Huang's comeback

Allies of former Nvidia CEO Jensen Huang say they are organizing a Silicon Valley executive 'war room' to lobby the board and the White House. The same day, Treasury Secretary Scott Bessent said Nvidia should bring in an outside CEO. The board has not finished its independent investigation into the relationships that led to the resignation, and CFO Colette Kress is interim CEO ahead of earnings on the 17th.

Continue reading (2) Show less

Congress is moving toward tighter limits on exports to China. Democrats and Republicans have demanded testimony from Huang and Nvidia leaders, and the Remote Access Security Act, introduced in the House in April 2025 and stalled in the Senate, is back in discussion. House Speaker Mike Johnson said there is growing consensus in Congress on limiting chip exports to China. The comeback campaign is about who runs Nvidia; the export measures would decide whom Nvidia can sell chips to.

Meituan, a Chinese delivery platform, released the weights of LongCat-2.5, a 1.6-trillion-parameter model it says was trained without Nvidia GPUs, under the MIT license. It scores 88.4 on the v2.1 coding test, close to Claude Opus 5's 89.1, but 38.5 against 53.9 on the harder v4.0. Every score is Meituan's own, so independent verification has to come before anyone uses it to judge export controls.

Signal to watch Watch whether Nvidia's board and outside counsel answer Bessent's 10-day demand on export-control compliance, and how the interim CEO explains China revenue on the 17th.

Congress wants to tighten export rules on AI chips after Nvidia spy scandal

Summary

Congress wants to tighten export rules on AI chips after the scandal over former Nvidia CEO Jensen Huang's relationship with a suspected Chinese spy. Democrats and Republicans have backed various bills and measures and demanded that Huang and other Nvidia leaders testify. Senate Banking Committee Chairman Tim Scott sent letters to Huang and Commerce Secretary Howard Lutnick asking both to testify publicly. The Remote Access Security Act, introduced in the House in April 2025 but stalled in the Senate, is drawing renewed interest. House Speaker Mike Johnson said there is "growing consensus in Congress on limiting chip exports to China," according to Reuters.

Why it matters

Why It Matters

No single bill is certain to pass, but Democrats and Republicans are now pushing in the same direction, and the scandal has revived export bills that had stalled. The Remote Access Security Act stalled in the Senate once, so the Senate is where to watch first.

Story so far Jensen Huang's exit and the question of Nvidia chips in China

Jensen Huang's allies set up a 'war room' to bring him back as Nvidia CEO

젠슨 황, 엔비디아 CEO 복귀 논의…측근들 백악관 설득 추진

Summary

Within days of resigning, former Nvidia CEO Jensen Huang has begun discussing a return with allies, the Financial Times reported on the 15th. One ally said he is organizing a Silicon Valley executive 'war room' to lobby Nvidia's board and win White House backing, and that Huang supports the effort. The same day, Treasury Secretary Scott Bessent told Fox News that Nvidia should hire an outside CEO, and that the board and its outside counsel must explain within 10 days how the company complied with chip export controls. The board is running an independent investigation into the relationships that led to the resignation, and CFO Colette Kress is interim CEO ahead of earnings on the 17th.

Why it matters

Why It Matters

The comeback camp says it will seek White House backing, but the same day the Treasury Secretary called for an outside CEO instead. The board's independent investigation is still open, and neither Huang's return nor a successor has been decided.

Story so far Jensen Huang's exit and the question of Nvidia chips in China

Where was Nvidia’s security team?

Summary

Nvidia's chief security officer, David Reber, stepped down in August for health reasons, and CEO Jensen Huang resigned about two and a half months later, but nothing public links the two departures. Reber had been CSO for four years and at Nvidia for seven. His replacement is Saša Zdjelar, who spent 17 years at ExxonMobil. The questions about Huang's relationships began with an executive at Megaspeed, a Singapore data center company, and the New York Times cited five employees describing a relationship with a vice president who reported to him, while Reuters cited two people on a second relationship. Retired senior intelligence officer Nicholas Eftimiades says security teams are often underfunded and CEOs sometimes brush off warnings with "not a nickel more than compliance." The board's independent investigation and federal inquiries are ongoing, so what Nvidia's security team knew remains unclear.

Why it matters

Why It Matters

The two departures sit close together on the timeline, but nothing public links them. The question the article leaves open is whether a security team could have caught an executive's personal relationship turning into a national-security risk, and Eftimiades's complaint that firms spend only what compliance requires is not about Nvidia alone.

Story so far Jensen Huang's exit and the question of Nvidia chips in China

Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

Summary

China's Meituan has released the weights of LongCat-2.5, a 1.6-trillion-parameter model it says was trained on Chinese hardware without Nvidia GPUs. As a mixture-of-experts () model, it uses about 48 billion , roughly 3% of the total, for each token. Its sample self-hosting guide runs on 194 Huawei Ascend accelerators and has no Nvidia GPU instructions. On the v2.1 coding test it scores 88.4, close to Claude Opus 5 (89.1) and GPT-5.6 Sol (88.8), but on the harder v4.0 it scores 38.5 against Opus 5's 53.9. It is released under the MIT license, and all scores were measured by Meituan itself.

Why it matters

Why It Matters

On a day when lawmakers are debating tighter limits on Nvidia chip exports, a Chinese company claims a near-frontier model built without those chips. But every score is Meituan's own, and on the harder test the gap is more than 15 points. For judging export controls, these numbers fit 'still behind on harder tasks' better than 'caught up.'

What to do now

Meituan says the model's API supports OpenAI and Anthropic formats, so it can be plugged into agent tools such as Claude Code or Codex. The price of about 0.27 euros per million input tokens is a limited-time promotion.

AI DJs Completely Kill My Groove

Summary

Wired software reporter Reece Rogers tried the AI DJs on YouTube Music and Spotify and decided to stop using both. YouTube Music's 'Beyond the Beat' interrupts for about 30 seconds between songs, and he describes a long commentary on Olivia Rodrigo's remix with Chappell Roan that was followed by the original track, with no Roan on it. The same commentary also misspelled and mispronounced Charli XCX's name, he says. On Spotify he played all eight English-language voices and found that several lean on filler words such as 'you know' and 'like.' Spotify says 80% of listeners come back to DJ week after week, but Rogers writes that he will not be one of them because the content of the commentary is empty.

Why it matters

Why It Matters

Rogers's complaint is about what the DJs say, not how they sound. The voices now pass for human, but getting facts about a song wrong or filling the gap with platitudes breaks the listening. Spotify's 80% return figure is the company's own claim, and this piece is one reporter's experience.

This AI Tamagotchi Turns Everyday Objects Into Needy Virtual Pets

Summary

If you skip skateboarding for a few days, the tiny avatar stuck to your board starts to look sad and lonely. OATs, short for Object Agent Twins, were built by Daehwa Kim, a PhD student in Carnegie Mellon's Future Interfaces Group: AI avatars on E Ink displays that attach to everyday objects. The avatar's expression changes with readings from moisture, temperature, humidity, light and motion sensors, and the sprites are generated with Google Gemini's Nano Banana image model. Kim presented it at ACM UIST in Detroit in early November and won two Honorable Mention awards. The main module costs about $20, and battery life of three to four days (one day with the sound sensor) keeps it a proof of concept, so she has not decided whether to sell it.

Why it matters

Why It Matters

Kim built OATs to move AI agents' bids for attention off the smartphone screen and onto physical objects. Testers used them in three ways: checking whether an object is in use, warning others nearby (such as a 3D printer giving off fumes), and tracking their own habits. With today's short battery it is a research demo, not a product, though Wired notes that companies increasingly lean on cuteness to spread AI devices.

Anthropic's Claude Fable 5.2 doesn't generate images, it paints them

Summary

Claude Fable 5.2 paints each picture as a Python program of 1,000 to 2,000 lines, and one piece used about 700,000 brushstrokes. Instead of generating pixels the way models like Midjourney do, Anthropic's model feeds brushstroke coordinates one at a time into a simulated paint engine that models bristle texture and how paint goes on. It uses no external image model and no pre-made source images. Strokes cannot be undone, so mistakes have to be painted over, and the model invents artist personas with fictional backstories to keep a consistent style. On Anthropic's internal robotics simulation benchmark RoboDojo Sim it scored 64.9% versus 29.8% for OpenAI's GPT-6 Astra, but when both used real robot arms to paint 'The Starry Night,' GPT-6 Astra was faster.

Why it matters

Why It Matters

Unlike one-shot image generation, each painting is thousands to hundreds of thousands of strokes that cannot be undone, so the model has to plan far ahead. The robotics score is from Anthropic's internal benchmark, and the same article notes that fine motor control is still a limit on real robots.

Australia's central bank finds AI hitting young workers first, especially in software and customer service

Summary

Australia's central bank, the Reserve Bank of Australia (RBA), finds that AI has barely displaced jobs in Australia so far, but that young workers in the most AI-exposed occupations are doing worse. In programming, clerical work and customer service, workers aged 15 to 24 and young graduates show weaker employment and higher unemployment. Since late 2022, employment for young people in these jobs has fallen while it has risen for older workers. The RBA says it has not established causation and that it measures theoretical exposure rather than actual AI use. A Stanford study covered in the same article found employment for 22- to 25-year-olds in exposed occupations fell 13% in relative terms since late 2022.

Why it matters

Why It Matters

The Australian data and Stanford's US payroll records both show entry-level jobs in exposed occupations shrinking first. A University of Pittsburgh study covered in the same article cuts the other way: it finds the trouble began in early 2022, before ChatGPT, and weighs interest rates and the post-pandemic hiring correction more heavily. Researchers cited in the article warn that fewer junior roles now could mean fewer experienced professionals later.

A top tax-accountant exam scorer admits AI cheating, and what exams should check

[기고] 채점이 편한 시험은 AI에도 편하다

Summary

The top scorer on this year's tax-accountant exam in Korea, identified only as A, confessed on the 15th that he cheated with generative AI and turned himself in to police. According to the confession, a tiny camera photographed the questions, generative AI produced the answers, and he checked them on a smartphone. The columnist notes that A had cited his 663 out of 800 on the essay-style second round as proof of innocence a day earlier, and he has admitted cheating on that round too. On the first round he answered 159 of 160 questions correctly, and the exam agency found nothing unusual in its automated grading. The columnist argues that the government's fixes, radio detectors and more CCTV, must be redone every time the device changes, and proposes an oral check in which top candidates explain their own answers.

Why it matters

Why It Matters

Generative AI beat the essay round as well as the multiple-choice one. An answer sheet records what was written but not who thought of it, so cracking down on devices alone leaves that gap open. The columnist proposes not a new exam format but an added step of questioning top candidates, and with 16,856 people sitting a single subject, that step can only come at the final stage.

Terms in this briefing

Past Briefings

Monthly Reports

Weekly Recaps