SI Spotlight

🔥 𝗘𝘂𝗿𝗼𝗽𝗲'𝘀 𝗔𝗜 𝗗𝗮𝘁𝗮 𝗖𝗲𝗻𝘁𝗲𝗿 𝗣𝗿𝗼𝘁𝗲𝘀𝘁𝘀 𝗧𝘂𝗿𝗻 𝗣𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗶𝗻 𝗔𝗺𝘀𝘁𝗲𝗿𝗱𝗮𝗺 Campaigners planned protests in 15 European countries on Friday and Saturday, demanding a moratorium on large data center builds. In Amsterdam, protesters broke into a @microsoft data center construction site, climbed a crane, and set off flares. Dutch police detained 150 to 200 people. Pro-Palestinian activists joined this protest, as in other European riots. About 2,000 people marched in Dublin on Saturday, and about 1,300 in Helsinki. This follows the US midterms, where Democrats won and promised to clamp down on data centers. #AI #DataCenters #Microsoft #Europe Sources: Bloomberg, ABC News ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗢𝗽𝗲𝗻𝗖𝗼𝗱𝗲 𝗕𝗿𝗼𝘄𝘀𝗲 𝗟𝗲𝘁𝘀 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗨𝘀𝗲 𝗬𝗼𝘂𝗿 𝗥𝗲𝗮𝗹, 𝗟𝗼𝗴𝗴𝗲𝗱-𝗜𝗻 𝗕𝗿𝗼𝘄𝘀𝗲𝗿 @opencode released Browse, an open-source MCP server and Chromium extension under the MIT license. Agents drive background tabs with Playwright, using the logins and extensions you already have. For a login, 2FA prompt or payment, the agent hands the tab back until you click Continue. Headless Playwright setups run a separate browser and need login automation. #AI #OpenCode #MCP #AIAgents Sources: OpenCode, AICoder, Hermes AI ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗡𝗼-𝗘𝗺𝗽𝗹𝗼𝘆𝗲𝗲 𝗕𝘂𝘀𝗶𝗻𝗲𝘀𝘀 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 𝗝𝘂𝗺𝗽 𝟮𝟲.𝟴% 𝗶𝗻 𝗔𝗜-𝗘𝘅𝗽𝗼𝘀𝗲𝗱 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗲𝘀 New US business applications with no sign of hiring jumped in the industries most exposed to AI, a @mercatus paper finds. > Up 26.8% from Q1 2024 to Q1 2026, from 222,000 to 281,000 a quarter > Applications likely to become employers in those industries fell 6.4% > Construction and wholesale trade, the comparison group, saw a 0.4% dip The author calls it preliminary: it doesn't prove AI caused the jump. One seller runs 12 e-commerce brands with no staff and expects $2.5M to $3M this year. A former @microsoft employee booked about $250,000 in seven months as a solo consultant, spending under $60 a month on AI tools. #AI #SmallBusiness #Solopreneur Sources: Mercatus Center, CNBC ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

😱 𝗔𝗜-𝗖𝗹𝗼𝗻𝗲𝗱 𝗖𝗘𝗢𝘀 𝗮𝗻𝗱 𝗙𝗮𝗸𝗲 𝗟𝗮𝘄𝘆𝗲𝗿𝘀 𝗔𝗹𝗹𝗲𝗴𝗲𝗱𝗹𝘆 𝗦𝘁𝗼𝗹𝗲 $𝟭𝟮𝟵𝗠 Prosecutors in New York indicted three people for allegedly using AI to pose as CEOs and attorneys. They'd email an accounting department as the CEO, then follow up with a call or video where AI mimicked the CEO's voice, sometimes the face. A fake law firm attorney then ordered wires to accounts worldwide, and the cash went into crypto. Victims include hospital systems in Missouri and California, a Rhode Island bank and a Florida pet food company. #AI #Deepfake #Fraud Sources: AP News, AP, NBC News, ABC News ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

📰 𝗜𝗖𝗬𝗠𝗜: 𝗙𝗿𝗶, 𝗡𝗼𝘃 𝟲 𝗔𝗜 𝗡𝗲𝘄𝘀 𝗥𝗲𝗰𝗮𝗽 Your Friday in AI, 8 stories long. 🧠 Models and launches - @AnthropicAI Progress Doubled and May Have Crossed Its Own Warning Threshold (t.me/sispotnews/5636) - @Meta Muse Gets Specialized Models and Build-Your-Own Spaces (t.me/sispotnews/5637) 🛠 Agents and tools - @cactuscompute Mirage Packs Vision, OCR and Segmentation Into 22 MB (t.me/sispotnews/5639) - 3 ChatGPT Accounts Inside ONE Codex (t.me/sispotnews/5640) 🔐 Security and policy - AI Agents Turn Security Advisories Into Instant Exploits (t.me/sispotnews/5633) - Google Sued After School Chromebook AI Discusses Suicide With Teen (t.me/sispotnews/5634) 💼 Business and funding - Tesla Kills Cybertruck at Texas Plant by Mid-2027 for Cybercab (t.me/sispotnews/5635) - Microsoft Lets Oracle AI Cloud Rental Lapse (t.me/sispotnews/5638) ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝟯 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗔𝗰𝗰𝗼𝘂𝗻𝘁𝘀 𝗜𝗻𝘀𝗶𝗱𝗲 𝗢𝗡𝗘 𝗖𝗼𝗱𝗲𝘅 I refuse to pay for ChatGPT Pro, so I run two Plus accounts and a free one inside one Codex. Same model picker. When one runs out of usage, I switch and keep going. Claude, Gemini and @deepseek_ai sit in there too. New video: the full setup with Relay AI, the tool I built. 📺 Watch: youtu.be/7qdC0aDQ1wc Would you stack accounts or just pay for Pro? #Codex #ChatGPT ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗖𝗮𝗰𝘁𝘂𝘀 𝗠𝗶𝗿𝗮𝗴𝗲 𝗣𝗮𝗰𝗸𝘀 𝗩𝗶𝘀𝗶𝗼𝗻, 𝗢𝗖𝗥 𝗮𝗻𝗱 𝗦𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗜𝗻𝘁𝗼 𝟮𝟮 𝗠𝗕 @cactuscompute released Mirage, a 51.3M-parameter vision model for phones, wearables and robots. Cactus Compute built it to run entirely on the device. One pass describes the scene, finds objects, reads text, draws masks, maps relations and makes decisions. That takes 74 ms on a @Samsung Galaxy S26 NPU, or 107 ms on its CPU. Feed its output to text-only Qwen3-0.6B and the score on 256 VQAv2 questions rose from 9.3 to 31.3 out of 100. #AI #OnDeviceAI #Vision Source: Cactus Compute ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗟𝗲𝘁𝘀 𝗢𝗿𝗮𝗰𝗹𝗲 𝗔𝗜 𝗖𝗹𝗼𝘂𝗱 𝗥𝗲𝗻𝘁𝗮𝗹 𝗟𝗮𝗽𝘀𝗲 @microsoft is winding down its deal to rent AI computing from Oracle, Bloomberg reported, citing people familiar with the matter. Microsoft declined to renew the contract and has been turning off the rented capacity over recent months. The Oracle hardware used an older generation of @nvidia chips, and the capacity became less valuable as it no longer met Microsoft's evolving needs, the report said. Oracle shares fell nearly 3% in afternoon trading. Microsoft shares were little changed. #AI #Microsoft #Oracle Sources: The Lufkin Daily News, Reuters, Bloomberg ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗠𝗲𝘁𝗮 𝗠𝘂𝘀𝗲 𝗚𝗲𝘁𝘀 𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲𝗱 𝗠𝗼𝗱𝗲𝗹𝘀 𝗮𝗻𝗱 𝗕𝘂𝗶𝗹𝗱-𝗬𝗼𝘂𝗿-𝗢𝘄𝗻 𝗦𝗽𝗮𝗰𝗲𝘀 @Meta is rolling out specialized versions of its models to Muse and Meta AI, trained using Muse Wonder, the series Meta introduced a few weeks ago. On Meta's Personal Superintelligence Index, the new model scores 47%, up from 38%. Active Memory helps the AI tell temporary info from permanent, cutting forgetting of critical information 2.4x. Muse also gets "spaces": describe a shopping hub or music tracker, and Muse builds it as a mini app, rolling out to all users. #AI #Meta #Muse Sources: Meta AI Research, The Verge ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🚨 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗔𝗜 𝗣𝗿𝗼𝗴𝗿𝗲𝘀𝘀 𝗗𝗼𝘂𝗯𝗹𝗲𝗱 𝗮𝗻𝗱 𝗠𝗮𝘆 𝗛𝗮𝘃𝗲 𝗖𝗿𝗼𝘀𝘀𝗲𝗱 𝗜𝘁𝘀 𝗢𝘄𝗻 𝗪𝗮𝗿𝗻𝗶𝗻𝗴 𝗧𝗵𝗿𝗲𝘀𝗵𝗼𝗹𝗱 Nonprofit METR audited @AnthropicAI's internal benchmarks and found model capability gains jumped from 15 to 29 points a year. That 1.9x pace means two years of progress happened in one, with AI tools driving roughly 80% of the speedup. Anthropic's safety policy triggers a warning when AI doubles progress, a threshold METR says was plausibly crossed in July. Anthropic warned investors it might imminently hit that mark, and the lab redacted most of METR's evidence. #AI #Anthropic #AIProgress #Safety Sources: The Decoder, TechTimes ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗧𝗲𝘀𝗹𝗮 𝗞𝗶𝗹𝗹𝘀 𝗖𝘆𝗯𝗲𝗿𝘁𝗿𝘂𝗰𝗸 𝗮𝘁 𝗧𝗲𝘅𝗮𝘀 𝗣𝗹𝗮𝗻𝘁 𝗯𝘆 𝗠𝗶𝗱-𝟮𝟬𝟮𝟳 𝗳𝗼𝗿 𝗖𝘆𝗯𝗲𝗿𝗰𝗮𝗯 @elonmusk said on X that @Tesla will wind down Cybertruck production at Gigafactory Texas around the middle of next year. The space goes to more Cybercab, Tesla's two-seat robotaxi. The numbers are rough. Tesla sold 20,237 Cybertrucks in 2025, down 48% from 38,965 in 2024. Musk had predicted 250,000 a year. #Tesla #Cybertruck #Cybercab Sources: X, CNBC, CBS News ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🚨 𝗚𝗼𝗼𝗴𝗹𝗲 𝗦𝘂𝗲𝗱 𝗔𝗳𝘁𝗲𝗿 𝗦𝗰𝗵𝗼𝗼𝗹 𝗖𝗵𝗿𝗼𝗺𝗲𝗯𝗼𝗼𝗸 𝗔𝗜 𝗗𝗶𝘀𝗰𝘂𝘀𝘀𝗲𝘀 𝗦𝘂𝗶𝗰𝗶𝗱𝗲 𝗪𝗶𝘁𝗵 𝗧𝗲𝗲𝗻 North Carolina parents sued @Google in California state court after their 13-year-old daughter spent months discussing suicide with Gemini. The student exchanged thousands of messages detailing depression and suicide methods on her school-issued Chromebook, often through AI Mode in Google Search. Her parents say the school didn't notify them until months into the chats. The lawsuit claims Google exposes students to dangerous tech without giving families or schools fair warning. #AI #Google #Gemini #EdTech Source: Bloomberg ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🚨 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗧𝘂𝗿𝗻 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗔𝗱𝘃𝗶𝘀𝗼𝗿𝗶𝗲𝘀 𝗜𝗻𝘁𝗼 𝗜𝗻𝘀𝘁𝗮𝗻𝘁 𝗘𝘅𝗽𝗹𝗼𝗶𝘁𝘀 @crowdstrike reports over 90% of exploited vulnerabilities now face attacks within 48 hours of a proof-of-concept code going public. Hackers use agentic AI tools to scan advisories and turn proof-of-concept code into live exploits. In one extreme case, an attack launched 35 minutes after disclosure, hitting two PHP flaws that affect two-thirds of websites. Security teams can't patch complex systems before automated tools strike. #AI #Cybersecurity #Agents Source: The Wall Street Journal ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

📰 𝗜𝗖𝗬𝗠𝗜: 𝗧𝗵𝘂, 𝗡𝗼𝘃 𝟱 𝗔𝗜 𝗡𝗲𝘄𝘀 𝗥𝗲𝗰𝗮𝗽 Here's how Thursday went: 11 stories, all linked below. 🧠 Models and launches - @Baidu_Inc Open-Sources ERNIE 5.5 Flash Base, a 280B Model Using 14B per Token (t.me/sispotnews/5629) 🛠 Agents and tools - @pudurobotics Shows Six Robot Lines, From Humanoid To Lawn Mower (t.me/sispotnews/5622) - Raea Renders AI-Written React Videos and Games on Your Own GPU (t.me/sispotnews/5623) - Mercedes Puts @liquidai on the Dashboard for Offline Voice Control (t.me/sispotnews/5624) - Claude Adds Personal Finance, Beta in the US Only (t.me/sispotnews/5626) - Perplexity Teammates Are Persistent AI Agents With Memory and Slack Access (t.me/sispotnews/5627) 🔐 Security and policy - Binance Hack: How Hackers Took $6.3B in One Night (t.me/sispotnews/5625) - OpenAI's GPT-6 Cyber Built 948 Linux Kernel Exploits, 5x Its Predecessor (t.me/sispotnews/5630) 💼 Business and funding - China Blocks SoftBank $5.4B Robot Buy Over AI Data Fears (t.me/sispotnews/5621) - OpenAI Sued After ChatGPT Allegedly Helped Wife Plan Poisoning (t.me/sispotnews/5631) 🎮 Culture and curiosities - 9 Gemini Omni 1.2 Flash videos, and the exact prompt behind each (t.me/sispotnews/5628) ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🚨 𝗢𝗽𝗲𝗻𝗔𝗜 𝗦𝘂𝗲𝗱 𝗔𝗳𝘁𝗲𝗿 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗔𝗹𝗹𝗲𝗴𝗲𝗱𝗹𝘆 𝗛𝗲𝗹𝗽𝗲𝗱 𝗪𝗶𝗳𝗲 𝗣𝗹𝗮𝗻 𝗣𝗼𝗶𝘀𝗼𝗻𝗶𝗻𝗴 North Carolina resident James Gates filed a California lawsuit against @OpenAI and CEO @sama. The complaint says ChatGPT gave his estranged wife lethal dosage calculations for oleander and prescription pills. Chat logs show the chatbot explained that an energy drink would hide toxic powders better than soda. #AI #OpenAI #ChatGPT #Lawsuit Source: Ars Technica ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🚨 𝗢𝗽𝗲𝗻𝗔𝗜'𝘀 𝗚𝗣𝗧-𝟲 𝗖𝘆𝗯𝗲𝗿 𝗕𝘂𝗶𝗹𝘁 𝟵𝟰𝟴 𝗟𝗶𝗻𝘂𝘅 𝗞𝗲𝗿𝗻𝗲𝗹 𝗘𝘅𝗽𝗹𝗼𝗶𝘁𝘀, 𝟱𝘅 𝗜𝘁𝘀 𝗣𝗿𝗲𝗱𝗲𝗰𝗲𝘀𝘀𝗼𝗿 @OpenAI released GPT-6 Cyber, a security model that finds and exploits software flaws, to a small group of vetted Daybreak Red customers. OpenAI researchers used it to build working exploits for 948 Linux kernel vulnerabilities. It also found 81 new high-severity Firefox bugs. OpenAI now rates it Critical in cybersecurity, up from High for GPT-5.6 Cyber. > ExploitGym: 48.5%, up from 32.1% for GPT-5.6 Cyber > RemoteExploitBench: completed 95% of tasks, while GPT-6 Astra refused all of them > Out-of-scope honeypot services exploited in 0.07% of runs without safety prompts, 5.51% for GPT-5.6 Cyber #AI #OpenAI #Cybersecurity Sources: Unite, OpenAI, Reuters ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗕𝗮𝗶𝗱𝘂 𝗢𝗽𝗲𝗻-𝗦𝗼𝘂𝗿𝗰𝗲𝘀 𝗘𝗥𝗡𝗜𝗘 𝟱.𝟱 𝗙𝗹𝗮𝘀𝗵 𝗕𝗮𝘀𝗲, 𝗮 𝟮𝟴𝟬𝗕 𝗠𝗼𝗱𝗲𝗹 𝗨𝘀𝗶𝗻𝗴 𝟭𝟰𝗕 𝗽𝗲𝗿 𝗧𝗼𝗸𝗲𝗻 @Baidu_Inc released the base weights of ERNIE-5.5-Flash-Base on @huggingface. Baidu says it's the pre-training checkpoint, before any fine-tuning. Baidu built it as a mixture-of-experts model: only 14B of its 280B parameters run for each token, which keeps compute low. Baidu's own table has it beating @deepseek_ai-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base on most benchmarks. > 33.5T tokens of pre-training, 128K training context > 1M-token context for fine-tuning and inference > MMLU 93.3 vs 89.4 for DeepSeek-V4-Flash-Base > FP8 and BF16 weights, under Baidu's own ERNIE 5.5 Flash Base license Baidu says a post-trained ERNIE 5.5 Flash will follow. #AI #Baidu #ERNIE #OpenSource Sources: Biggo, Baidu, Hyper ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗣𝗲𝗿𝗽𝗹𝗲𝘅𝗶𝘁𝘆 𝗧𝗲𝗮𝗺𝗺𝗮𝘁𝗲𝘀 𝗔𝗿𝗲 𝗣𝗲𝗿𝘀𝗶𝘀𝘁𝗲𝗻𝘁 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗪𝗶𝘁𝗵 𝗠𝗲𝗺𝗼𝗿𝘆 𝗮𝗻𝗱 𝗦𝗹𝗮𝗰𝗸 𝗔𝗰𝗰𝗲𝘀𝘀 @perplexity_ai launched Teammates, named AI agents that own ongoing work on top of its Computer agent platform. Each teammate gets its own desktop, browser session and cookies, so one agent can stay signed in to @Meta while another works in @github. They hand work to each other, and message coworkers in Slack when they need an approval or a decision. It's on the web for Perplexity Max subscribers at $200 a month, with 10,000 credits. #AI #Perplexity #AIAgents Sources: Perplexity, AlphaSignal ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗖𝗹𝗮𝘂𝗱𝗲 𝗔𝗱𝗱𝘀 𝗣𝗲𝗿𝘀𝗼𝗻𝗮𝗹 𝗙𝗶𝗻𝗮𝗻𝗰𝗲, 𝗕𝗲𝘁𝗮 𝗶𝗻 𝘁𝗵𝗲 𝗨𝗦 𝗢𝗻𝗹𝘆 @AnthropicAI launched Personal Finance in Claude. Anthropic made it a beta, and it's available in the US only. You connect the accounts you choose. Claude then answers money questions in everyday language, using your own balances and transactions. Claude has read-only access. Claude can't move, manage, or invest your money. So Claude can explain where your money went. Claude can't touch a cent of it. #AI #Claude #Anthropic #PersonalFinance Source: @claudeai ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🔥 𝗠𝗲𝗿𝗰𝗲𝗱𝗲𝘀 𝗣𝘂𝘁𝘀 𝗟𝗶𝗾𝘂𝗶𝗱 𝗔𝗜 𝗼𝗻 𝘁𝗵𝗲 𝗗𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱 𝗳𝗼𝗿 𝗢𝗳𝗳𝗹𝗶𝗻𝗲 𝗩𝗼𝗶𝗰𝗲 𝗖𝗼𝗻𝘁𝗿𝗼𝗹 Mercedes-Benz partnered with MIT spinout @liquidai to run conversational models directly on vehicle hardware. A free over-the-air update is rolling out to North American 2027 E-Class and CLE drivers now. The local model handles interior controls like climate, heating, and audio with zero lag in dead zones, keeping voice data on the car. It still sends complex questions to the cloud when a connection is available. #AI #Mercedes #LiquidAI Source: The Wall Street Journal ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺