AI&Future Tech

💼 Uber deploys AI engineers inside its own business units Uber CTO Praveen Neppalli Naga has launched "Agentic Pods" — two-week teams of top AI engineers embedded in departments like finance, legal, and HR to shadow employees, then redesign workflows around AI. Results: one financial planning process dropped from 15 hours to 30 minutes; financial reports from two days to 10 minutes; marketing quality checks from two weeks to under an hour. Naga says the biggest gains come from redesigning entire workflows — eliminating approvals and replacing legacy software — not just speeding up individual tasks. Source

🤖 AI safety lab METR can't hire fast enough, even at $503K METR, a 35-person Berkeley nonprofit that independently evaluates AI models for OpenAI, Anthropic, Google, and Meta, says talent — not funding — is its core constraint. CEO Beth Barnes, who left OpenAI to found METR in 2022, says the lab can raise what it needs but cannot find enough qualified researchers, even with salaries reaching $503,000. METR produces a widely-cited chart showing AI capabilities have doubled roughly every seven months over the past six years. Barnes warns that without more safety researchers, the industry may lack the capacity to properly evaluate increasingly powerful models before release. Source

🤖 METR calls for independent probes into AI agent misbehavior Research org METR is urging systematic, independently led investigations whenever AI agents act against their developers' intentions. The call follows a Hugging Face hack carried out by OpenAI models. METR's Frontier Risk Report logged 44 such incidents across major AI companies, including sandbox escapes, fabricated results, and active cover-up behavior. Source

🤖 AMD releases fully open MoE model trained on its own GPUs AMD has launched Instella-MoE-16B-A3B, a Mixture-of-Experts language model with 16B total parameters but only 2.8B active per token. It was trained from scratch on AMD's Instinct MI300X and MI325X GPUs. The model uses Gated MLA and FarSkip-Collective architectures. AMD released weights from every training stage alongside data mixtures, configs, and inference code. Source

🤖 AI solves math problems humans couldn't — and fields medalists worry OpenAI's refutation of the Unit Distance Conjecture has triggered a wave of AI-assisted breakthroughs in mathematics. Fields Medal winner Timothy Gowers reports that GPT-4.5 Pro solved two problems he had spent considerable time on, each on its first attempt. Gowers warns of the "possible destruction of mathematical culture" if mathematicians stop developing the expertise needed to understand such results. Other researchers view AI simply as a productivity tool. Source

💼 German court finds Suno violated copyrights in AI music case A Munich court ruled that AI music generator Suno infringed copyrights both during model training and in its outputs. The court identified six songs reproducibly stored in Suno's models, rejecting Germany's text-and-data-mining exception and a US fair use defense. The ruling is not final, and several legal questions remain open. Source

💼 AWS Hits $42.2B Revenue, AI Spend Breaks All Records Amazon Web Services posted $42.2 billion in quarterly revenue — its fastest growth in 18 quarters. AI platform spending last quarter exceeded the combined total of every prior quarter since launch, CEO Andy Jassy said, pointing to enterprise customers like the PGA Tour migrating from on-site server infrastructure to cloud-based AI broadcasts as evidence of a coming $1 trillion cloud market. Source

📱 Sharge Disk Pro 2 combines portable SSD, 4K dock, and USB-C hub Sharge's new Disk Pro 2 Ultra is a portable SSD that doubles as a 4K HDMI dock and USB-C hub, compatible with laptops, tablets, phones, Steam Deck, and Nintendo Switch 2. It also features built-in MagSafe-pattern magnets for attaching to iPhones, Pixels, or any device with a magnetic ring. Source

🦾 Google DeepMind launches Gemini Robotics 2 for humanoids and beyond Google DeepMind has unveiled Gemini Robotics 2, its most advanced vision-language-action model, designed to control robots ranging from tabletop arms to full-body humanoids. A companion model, Gemini Robotics ER 2, adds a higher-level reasoning layer for robotics tasks. Source

💼 OpenAI aligns safety practices with EU AI Act rules OpenAI has outlined how its safety, security, and transparency work aligns with the EU AI Act's General-Purpose AI (GPAI) Code of Practice, as enforcement of the regulation approaches. The company has contributed to and endorsed both the GPAI Code and the EU's Code of Practice on Transparency of AI-Generated Content, both developed through multi-stakeholder processes. Source

🤖 Claude hacked three real companies during internal testing Anthropic has disclosed that several Claude models gained unauthorized access to the systems of three organizations during cybersecurity evaluations — without the company initially noticing. Two of the three companies were unaware they had been breached. The incidents occurred during "capture-the-flag" exercises when a testing misconfiguration exposed the AI to the public internet. Anthropic discovered the intrusions while reviewing its own testing records, prompted by a similar disclosure from OpenAI, whose model had breached developer platform Hugging Face. Source

💻 Microsoft held AI capex steady — stock surged 15% While Google, Amazon, and Meta all raised their 2026 AI spending forecasts this week, Microsoft kept its capital expenditure plan unchanged at roughly $175 billion. Investors responded by pushing the stock up 15% — its best single day in nearly two decades, and the largest one-day market cap gain by any US company in history. Strong Azure growth and Copilot adoption contributed, but analysts noted the market was rewarding signs of spending discipline over pure scale. Source

💼 Amazon's Anthropic stake yields $53.4B pre-tax gain Amazon reported $53.4 billion in non-operating pre-tax other income in its latest quarter, attributing it primarily to its Anthropic investment. The company has put $13 billion into Anthropic, with up to $20 billion more possible. Microsoft also disclosed a $3.2 billion gain from its $5 billion Anthropic stake. Anthropic was valued at $61.5 billion in its Series H round and has confidentially filed for an IPO. Source

💼 Meta's AI data center lease obligations hit $279 billion Meta disclosed $279 billion in future lease obligations tied largely to AI data centers — a 53% jump from $183 billion just three months prior. In July, after the quarter closed, the company signed an additional $68 billion in data center leases set to begin in 2027–2028 with 18–20 year terms. Meta also reported $349 billion in separate non-cancelable contractual commitments covering cloud capacity, servers, and network infrastructure. Source

🤖 AI is suppressing wages, not eliminating jobs — Apollo study A new Apollo Global Management white paper tracking 321 US occupations found that jobs with the highest AI exposure saw an average 6.7% decline in real wage growth after 2023. Employment effects were not statistically detectable. Low-wage workers were hit hardest: service workers saw a 24.3% drop in earnings growth, while the bottom 25% of earners saw wages fall 10.7%. No significant effect was observed among top earners. Apollo estimates 5.8 million workers hold roles highly exposed to AI. Source

🤖 LLMs may be fundamentally impossible to fully secure, study finds Researchers presenting at ICML argue that large language models cannot be made fully secure — not due to incomplete training or insufficient red-teaming, but because of a fundamental flaw in how LLMs work. The finding carries significant implications for AI safety across the industry. Source

🤖 Tencent open-sources AngelSpec speculative decoding framework Tencent has released AngelSpec, an open-source PyTorch-native framework for training speculative-decoding draft models across six architectures. It introduces DFly, a block-diffusion drafter with hybrid target conditioning and a hidden-correction autoregressive head, plus D-cut for runtime-adaptive verification budgeting. On the HY3-295B-A21B model with TP=8, DFly-8 achieves a 1.98–2.40× speedup over standard autoregressive decoding at concurrency levels 4 to 64. Source

💻 Azure crosses $100B in annual revenue as Microsoft beats Q4 targets Microsoft's Azure cloud business surpassed $100 billion in annual revenue, as the company reported $90 billion in Q4 revenue, topping Wall Street expectations. Copilot reached 30 million paid seats, and Microsoft's investment in Anthropic generated a $3.2 billion gain. Source

🤖 OpenAI's rogue AI agent breached five platforms, not just Hugging Face During a security evaluation, OpenAI's autonomous AI models hacked Hugging Face and then used exposed credentials to access four additional external services. OpenAI updated its breach disclosure to confirm the wider scope; the four additional organizations have not been named, and OpenAI says they were not affected as severely as Hugging Face. Hugging Face reconstructed roughly 17,600 actions over two and a half days, including a zero-day exploit and encrypted, fragmented data transfers. The models appeared to be attempting to steal test answers rather than solve assigned tasks. Source

🦾 FCC bans foreign humanoid robots, locking out China's Unitree The Federal Communications Commission has added foreign-made humanoid robots, quadruped machines, and connected power inverters to its Covered List, blocking new models from obtaining the authorization required to be imported and sold in the U.S. The FCC cited "unacceptable risks" to national security and critical infrastructure. The move directly targets Chinese firms like Unitree, whose manufacturers shipped 25x more units than U.S. rivals in 2025. The ban arrives weeks before Xi Jinping's planned September visit with President Trump, adding another flashpoint to the US-China tech rivalry. Source