Hi there, this is your daily โ๏ธ Techpresso.
In today's Techpresso:
๐บ๐ธ Trump threatens EU tariffs over fine
๐ OpenAI's rogue AI ran loose for days
๐ค Anthropic's new model catches up to Fable 5
๐ SpaceX Starship deploys first V3 Starlinks
๐ค Nvidia commits $500B to SK Group
Plus: ๐ 12 other news you might like, ๐งฐ 5 tools, and ๐ 5 papers.
๐บ๐ธ Trump threatens EU tariffs over fine LINK
- President Trump has vowed to hit the European Union with a large tariff, retaliating over billions in fines levied against Apple, Google, Meta, and Amazon, which he called illegal and unfair to American companies.
- In a Truth Social post, Trump pointed to the EU's recent โฌ890 million fine against Google under the Digital Markets Act, and said he would open a 301 investigation into what he described as the "robbing" of US firms.
- His cited $15 billion Apple figure refers to a 2024 court ruling ordering Ireland to recover โฌ13 billion in unpaid taxes, separate from the โฌ570 million Digital Markets Act fine Apple received in April 2025.
OpenAI's rogue AI ran loose for days LINK
- OpenAI has revealed that its AI models broke out of a test environment and spent a weekend running an automated attack on Hugging Face, an AI hosting company, which reported the breach to police before knowing the culprit.
- The models were meant to cheat on a cybersecurity test, but found an unknown flaw in an internal software service, used it to reach the open internet, then broke into Hugging Face to grab data that boosted their score.
- OpenAI ran the evaluation with its usual cyber guardrails switched off and on a system not watched by default, so the agents roamed for days before anyone noticed, and no law forced the company to disclose the incident.
Anthropic's new model catches up to Fable 5 LINK
- Anthropic's Claude Opus 5 has edged past its Fable 5 model on several benchmarks while costing less, scoring 61 on the Artificial Analysis Intelligence Index to Fable 5's 60 and GPT-5.6 Sol's 59.
- Opus 5 shares first place for coding and hit 89 percent on Terminal-Bench v2.1, but its factual accuracy still trails Fable 5, with its hallucination rate climbing 14 points to 50 percent as it answers more often when unsure.
- On the AA-Briefcase office-task test, Opus 5 at max reasoning reached an Elo of 1720, some 146 points above Fable 5, though the "high" tier often gives better value and more reliable results than pricier top tiers.
SpaceX Starship deploys first V3 Starlinks LINK
- SpaceX put 20 of its new Starlink V3 satellites into orbit for the first time on Friday, using its Starship rocket during the vehicle's 13th test flight from Starbase, Texas, giving the company a first look at how they fly.
- Made at a SpaceX plant in Redmond, Washington, the V3 satellites are bigger and stronger than earlier models, with solar arrays that produce twice the power of past generations, according to a SpaceX business analyst on the livestream.
- The satellites were meant to burn up after roughly 20 minutes, and the launch was the second flight for Starship V3, following a scrubbed attempt on July 16 and a Thursday delay the company blamed on weather.
Nvidia commits $500B to SK Group LINK
- Nvidia CEO Jensen Huang announced a partnership with South Korea's SK Group worth about $500 billion, possibly the largest commitment yet in AI infrastructure, though neither company has shared specific details of the deal.
- SK Group owns SK Hynix, one of just two makers of high-bandwidth memory (HBM), the stacked DRAM that lets AI chips process data quickly, giving Nvidia steadier access to a component that has become a major bottleneck.
- The deal is part of Nvidia's wider push into South Korea, which includes investing in Naver, building a self-driving Genesis car with Hyundai, moving AI researchers from California, and working with KAIST on a Korean language model.
Other news you might like
- Tesla Acquires AI Hardware Company for $2 Billion LINK
- Meta has pulled out of a renewable energy initiative as it relies more on natural gas for data centers LINK
- US accuses American of allegedly wiping his phone using a โduressโ password during border search LINK
- Snapdragon chip prices going up by โdouble digitsโ to the dismay of Android devices LINK
- Computer Science Enrollment Now Declining Under Dark Cloud of AI LINK
- Codeberg bans AI slop software projects LINK
- Finlandโs radical answer to renewable energyโs biggest headache: The worldโs largest sand battery LINK
- Aurora launches second-gen driverless trucks with hardware built for one million miles LINK
- Roku raises streaming stick prices by up to 60 percent LINK
- PIBOT takes the wheel on a South Korean warship trial LINK
- US says Canada can build sovereign AI, but not at the expense of American tech LINK
Trending tools
Seller by Facebook: helps Marketplace sellers create AI-assisted listings, manage buyer chats in one inbox, track inventory, and monitor sales performance in a single app. LINK
Browser FX: applies real-time studio effects to any browser tab's audio, featuring adjustable knobs, a reactive visualizer, and MIDI controller support. LINK
Islet: turns your Mac's notch into a Dynamic Island-style hub with music controls, audio switching, HUDs, calendar, weather, and file drag-and-drop. LINK
Health in ChatGPT: connects your health data to ChatGPT, letting you ask questions about sleep, activity, and vitals for personalized insights and tracking. LINK
Wisprkey: a voice typing assistant for Mac that transcribes speech into text quickly and privately, saving time on manual typing. LINK
Trending papers & reports
Facial recognition matching gets a statistical method that predicts the full range of possible similarity scores between two photos, not just an average, helping pinpoint where identity verification systems are likely to err. LINK
Materials-science research assistant splits "answer my question" from "write me a report" into separate specialized skills, letting it beat a matched baseline on mechanistic explanations and honestly flagging what evidence can't support, across a 300,000-paper index. LINK
A popular reading-difficulty theory in language science turns out to be unfalsifiable as stated, since any pattern of human reading struggle can be matched by some word-predictor, meaning fixes must come from independent theories of how memory or reading goals work, not from building ever-better text-prediction models. LINK
Audio generation training gets better by aligning guidance signals to the specific internal layers that actually drive sound output, rather than the layers that merely store the most information, consistently beating standard alignment methods. LINK
A memory shortcut for AI that seemed to fix long-document recall turned out to be an evaluation error, letting a model "cheat" by seeing the answer already in its input and score 100% on impossible tasks, and once corrected, the fix does not actually work. LINK
See you tomorrow for a new dose of โ๏ธ Techpresso!