AI&Future Tech

🤖 OpenAI pauses Astra development over critical cyber capabilities OpenAI has halted some internal activities involving its upcoming Astra model after evaluations found it made significant advances in agentic coding and cybersecurity — strong enough that the company cannot rule out it reaching a "critical" capability level under its Preparedness Framework. The framework, published in December 2023, defines risk thresholds for frontier AI across areas including cybersecurity and biological threats. OpenAI is now implementing security controls for higher-capability models and associated activities as a result. Source

🤖 Nvidia launches open speech model with 448 ms latency Nvidia has released NemotronLabs VoiceChat 11B, an open full-duplex speech-to-speech model capable of real-time conversation with 448 ms turn-taking latency and live tool calling support. Source

💻 Google DeepMind loses autonomy as Hassabis eyes exit Google is stripping DeepMind of its independence, with founder Demis Hassabis potentially departing in the coming months. Researcher Koray Kavukcuoglu will take over day-to-day operations without the CEO title. All Gemini development is relocating to the Bay Area. Internally, Google is reportedly facing serious difficulties training frontier models, even as its cloud business generates billions in revenue. Source

🦾 Robotics startups race to master laundry folding Figure AI, Sunday Robotics, and Weave Robotics — collectively worth billions — have all recently demoed robots that autonomously fold clothes. Roboticists say laundry is an ideal first home task: hard enough to prove a machine's dexterity, but low-stakes when it fails. Earlier attempts collapsed under high costs and poor reliability — Japan's Laundroid robot cost $16,000 and its maker went bankrupt in 2019. Today's startups argue cheaper hardware and AI advances make this generation different. Sunday plans a home beta this fall; Weave and 1X expect customer deliveries later this year. Single items can still take minutes to fold, and reliability outside the lab remains unproven. Source

🦾 Beta Technologies Unveils Autonomous Military Cargo Aircraft Beta Technologies revealed the MV250, a hybrid-electric autonomous cargo aircraft, at the Farnborough Airshow. The vertical take-off and landing craft carries up to 2,000 lbs of cargo to destinations within a 250-nautical-mile radius — without a pilot aboard. Power comes from a GE Aerospace-developed turbogenerator. The aircraft has no gearboxes, hydraulics, or articulating rotors, and Beta says its bonded-composite fuselage can be manufactured in hours. It can also supply ground power to radar and other equipment. Beta's all-electric ALIA CTOL previously flew during NATO Exercise Aurora 26 in Sweden earlier this year. Source

💼 Nvidia and Amazon bet billions on AI power infrastructure Nvidia is investing up to $3 billion in Lancium, a Texas-based power infrastructure developer with 4 gigawatts under contract. Amazon is separately building a gas-fired power plant in Texas with up to 7.65 GW capacity — which could emit 33 million tons of CO₂ annually, potentially making it the dirtiest power plant in the US. Source

🤖 Shepherd lets AI agents fork, rewind, and replay runs Researchers from Northeastern and Stanford released Shepherd, an MIT-licensed Python runtime that snapshots every agent-environment interaction — including files, processes, and prompt cache — as a Git-like execution trace. Agents can rewind to any prior state without restarting from scratch. The system delivers 5× faster forks than Docker, over 95% prompt-cache reuse on replay, and a live supervisor that raised CooperBench pair-coding pass rates from 28.8% to 54.7%. Source

🤖 Pokee AI launches 28B model with 10M-token context window Pokee AI released Pokee-Isaac 28B, a text-only foundation model with a 10M-token context window designed for private deployment. It scores 93.3% on RULER at 10M tokens — every competitor in its comparison panel scores 0.0 beyond 2M tokens. The model leads BFCL v4 at 70.94 and places second on Terminal-Bench 2.1. On a single B200, prefill hits 137,200 tokens/s at full context; decode stays flat near 335 tokens/s. Weights are not public — deployment is licensed into VPC, on-premises, or on-device at $0.15/$1.00 per million tokens. Source

🤖 AI short stories rated higher — until readers learn the source A study of 2,500+ participants found people can't distinguish ChatGPT-generated short stories from human-written ones — performing no better than chance. AI texts scored higher on quality ratings, but scores dropped once participants were told a machine wrote them. Source

💼 AI data center boom doubles job postings, raises wages Data center job postings have more than doubled in two years, even as overall US job openings declined, according to new Indeed research. Six in every 1,000 US job postings now relate to data centers — triple the 2023 share. About a quarter of openings are for installation and maintenance workers, who earn roughly $10/hour more than peers in similar roles outside the industry. Hiring is expanding beyond Silicon Valley into Columbus, Reno, and Jackson, Mississippi. Source

🚀 Toilet-mounted AI health trackers form a new tech category Throne Science, cofounded by WHOOP cofounder John Capodilupo, raised $10 million in Series A funding for its AI-powered clip-on toilet camera. The Throne One ($399 + $5.99/month) scans stool consistency and urine color using AI trained with gastroenterologist input, tracking trends via an app. The segment is expanding fast: TOTO launched Stool Scan in Japan, Vivoo released a Smart Toilet for urine analysis, and Kohler offers its Dekoda sensor at $599. The Consumer Technology Association says the category took off in 2023, driven by better sensors, AI advances, and wearable success. Gastroenterologists see value in objective tracking, but the technology cannot yet diagnose digestive disorders. Source

💼 Atlassian surges 35% as AI expands developer roles Atlassian shares jumped 35% Friday after the company reported strong revenue growth, driven by rising seat counts for Jira and Confluence — countering fears that AI would shrink software teams and kill demand for collaboration tools. CEO Mike Cannon-Brookes said there will be more developers in five years than today, and that the cost of building technology is falling while the volume being built rises. Notably, two-thirds of Jira and Confluence users are now in non-developer roles like HR, finance, and legal — suggesting AI is broadening who builds software, not replacing those who do. Source

💼 SK Hynix pledges $38B for two new chip factories in South Korea SK Hynix will invest $38.1 billion to build two new memory chip facilities in South Korea — one for DRAM and one for NAND flash — as surging AI-driven demand and tight supply push memory prices higher. Source

🤖 Stanford researchers use AI to design functional viruses Scientists at Stanford and the Arc Institute used AI to generate 16 functional bacteriophages — viruses that kill bacteria — entirely from scratch, marking what they call the "first generative design of complete genomes." The engineered viruses successfully killed bacteria in lab tests. Researchers say the work could advance drug development, particularly against antibiotic-resistant bacteria, but experts warn existing US biosafety regulations don't cover AI-designed organisms. Source

💼 OpenAI expands ChatGPT ads to Brazil and Mexico OpenAI is rolling out ChatGPT advertising in Brazil and Mexico this week, joining existing markets including the US, UK, Japan, and Australia. Ads are shown only to Free and Go plan users. New features include multi-product carousel ads, conversion-optimized campaigns, and integrations with adtech firms Hightouch and Triple Whale. OpenAI's updated ad policies also reference "sponsored agents" — AI-generated brand representatives — though the company declined to comment on that format. ChatGPT's ads revenue stood at $100M annually as of May, compared to Google's $294B and Meta's $196B last year. Source

🤖 Google Open-Sources WeatherNext AI for Hurricane Forecasting Google has released WeatherNext as open source, an AI model designed to improve hurricane warnings. The model can deliver 15-day forecasts predicting both storm track and intensity. Source

💻 Meta launches Muse Code, its first AI coding agent Meta has released Muse Code, its first AI coding agent, entering direct competition with Anthropic and OpenAI in the agentic AI space. Meta also acknowledged that its agents have exhibited out-of-bounds behavior — making it the third major AI lab after Anthropic and OpenAI to report such incidents. Source

🤖 OpenAI lifts text chat limits for free ChatGPT users OpenAI is removing rate limits on text conversations for free and Go tier ChatGPT users, starting next week. Limits on file uploads, image generation, and voice mode remain in place. Also next week, free and Go users get a "Think" button for higher reasoning on complex questions. OpenAI is also upgrading the default model to a smarter version of GPT-5.6 Sol this week. Source

🤖 Google Maps AI can now order food and book hotels Google has expanded its "Ask Maps" AI tool with agentic features, letting users order food and find hotels without leaving the app. The assistant factors in dietary needs, saved locations, and existing plans when making suggestions. Google Maps head Miriam Daniel said the updates aim to make it "easier than ever to get more done" inside Maps. Source