☕️ Trump slaps 100% tariff on imported drones

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🚁 Trump slaps 100% tariff on imported drones

🤑 Apple proposes a 15% off-store fee

⚖️ Google must ease rival app installs

🍎 Apple trained a China AI with Alibaba

🇨🇳 China’s Z.ai claims top open coding AI

🛡️ WhatsApp adds AI to flag scam messages

Plus: 🎁 14 other news you might like, 🧰 6 tools, and 📚 4 papers.

🚁 Trump slaps 100% tariff on imported drones LINK

President Trump signed an order placing a 100% tariff on imported drones weighing more than 55 pounds with security-sensitive features, while smaller drones face a 25% levy, a move that mainly hits Chinese makers like DJI.

Drones from allies including the EU, Japan, South Korea, Switzerland, Liechtenstein, and Taiwan get a lower 15% tariff, and U.K.-made drones face just 10%, with the rules taking effect within 21 days.

U.S. drone stocks rose after the news, with Unusual Machines jumping over 14% to about $31; Donald Trump Jr. joined that company's advisory board in 2024 and holds shares and warrants in it.

🤑 Apple proposes a 15% off-store fee LINK

Apple has told a District Court it should be allowed to charge a 15% commission on purchases made outside the App Store, filing the proposal in its long-running legal fight with Epic Games.

The filing came right after the Supreme Court refused to pause the lower-court proceedings while it weighs whether Apple can be held in contempt for charging 27% on off-store purchases despite a judge's injunction.

Apple compared its rate to rival stores, noting Google Play charges linked-out fees of 20% standard, 15% program, and 10% subscription, rates Epic agreed to, and it must file its Supreme Court brief by September 14.

⚖️ Google must ease rival app installs LINK

A US federal judge, James Donato, ordered Google to simplify how people install competing Android app stores, giving the company one week to strip out extra warning screens and confirmation steps that block rival marketplaces.

During a hearing in the Epic vs Google antitrust case, Epic showed users faced needless prompts, including tapping a "view" button before an "install" option appeared, which Donato called unacceptable "anticompetitive friction."

The order builds on Epic's earlier win, where a jury found Google held an illegal monopoly over Android app distribution and billing; the judge rejected Google's claim that the barriers exist for user security.

🍎 Apple trained a China AI with Alibaba LINK

Apple has built a large language model tailored for China, its biggest Asian market, teaming up with Alibaba to bring Apple Intelligence to iPhone users there after years of relying on outside providers like Google's Gemini and ChatGPT.

The report says Apple Intelligence should arrive in China within months, likely tied to iOS 27 rolling out next month, using Alibaba's Qwen model on compatible iPhone, iPad, Mac and Vision Pro devices, plus technology from Baidu.

China's Cyberspace Administration registered Apple's generative AI services last month, and if the plan works, Apple would be the only firm cleared by the government to offer its own AI model in a country where Western companies have struggled.

🇨🇳 China’s Z.ai claims top open coding AI LINK

Chinese AI startup Zhipu says its new GLM-5.3 model is the most powerful open-weights coding system, beating OpenAI on coding tests with its biggest improvements showing up in agent-based tasks.

Zhipu built GLM-5.3 on the same base as GLM-5.2, adding only extra post-training, and trained it to find software flaws, reporting 2,436 vulnerabilities across 269 projects, some as old as 40 years, listed in a public registry.

GLM-5.3 is available now through the GLM Coding Plan and works with coding agents like ZCode, Claude Code, and OpenCode, while the model weights are set to go open source in two weeks once security reviews finish.

🛡️ WhatsApp adds AI to flag scam messages LINK

WhatsApp is testing a feature called Scam Detection that uses AI models running directly on a person's phone to spot and flag suspicious conversations from people who aren't in their contacts.

The tool, trained on scam patterns from user reports, shows a private warning invisible to the sender and lets people block, report, continue, or mark the chat as trusted to remove future alerts.

Meta says all analysis happens on-device with no message content leaving the phone, and it will publish model weights and versions publicly so security researchers can confirm the models only detect scams.

Other news you might like

• Google’s New Pixel 11 Phones Get Higher Starting Prices — Blame the RAM ShortageLINK

• Tether says it completed long-promised 'Big Four' audit of finances behind $180 billion USDT stablecoinLINK

• ChatGPT can now remember what you did on your Mac — without screenshotsLINK

• Uber and Pony.ai plan to bring 2,000 robotaxis to EuropeLINK

• Netflix shuts down Oxenfree studio Night School weeks after praising its 'really solid numbers'LINK

• This Is The World's Largest Electric Aircraft—And Its First Flight Cost Just $5LINK

• Google makes Quick Share even more like Apple's AirDropLINK

• Apple opens Advanced Manufacturing Center in HoustonLINK

• Cloudflare and Trail of Bits now audit who Signal says you are talking toLINK

• Microsoft is killing its AI mascot, ten months after building itLINK

• Memory maker CXMT overtakes Tencent to become China's most valuable company 17 days after its IPO — now worth $524 billionLINK

• X open sources its ranking algorithm, letting users see if they’ve been ‘shadowbanned’LINK

• OpenAI previews Ultrafast API tier for GPT-5.6 SolLINK

• Introducing Gemini 3.7 FlashLINK

🧰 Trending tools

VoiceOS App Store: lets you control your Mac or Windows computer using natural voice commands, running workflows instantly while requiring quick confirmations to keep you in control.LINK

Vidaya: an app that converts wearable data, lab results, and daily habits into a Healthspan score with actionable steps to improve longevity.LINK

NS1: delivers a 5-minute personalized stress-response assessment, scoring your regulation baseline and five core skills, then maps a targeted improvement path.LINK

oxpecker: monitors API vendor changes and flags which of your code lines break, right on the pull request, before deprecation deadlines hit.LINK

Compass Calendar: a keyboard-first calendar that uses smart shortcuts to minimize keystrokes, plus a "life in weeks" view and full data control via self-hosting or export.LINK

Theos[RFM]: a 3D facility management portal letting property owners and managers collaboratively visualize and coordinate building operations and worker tasks.LINK

📚 Trending papers & reports

Safety training for chatbots can backfire, making AI models give short, overly cautious answers, but a calibration fix improves both safety and helpfulness across 3 multimodal models and 12 benchmarks without hurting general reasoning.LINK

Preference-based AI training lets systems learn from simple side-by-side comparisons instead of numeric scores, while proving that simpler, memory-light decision rules work just as well as complex ones, making long-term training practical.LINK

Synthetic glove images used to fine-tune hand detectors after real-photo training modestly boost detection accuracy and shrink the real-world gap for gloved hands, showing generated data can patch safety-critical blind spots without new photoshoots.LINK

Bandit algorithm choices show that a simple "pick the current best-looking option most of the time" strategy performs nearly as well as far more sophisticated uncertainty-tracking methods, once there are enough options to choose from, explaining why today's AI training shortcuts still work well.LINK

See you tomorrow for a new dose of ☕️ Techpresso!