This week's major tech updates include Jensen Huang reportedly considering a return to Nvidia, Apple planning its first smart glasses for 2027, and OpenAI dissolving its independent robotics team to integrate researchers into core departments.
According to Sakana AI: Sakana AI has introduced JFinBench, a new benchmark designed to evaluate large language models on real-world Japanese financial certification exams from 2026. Evaluating 12 leading models revealed that top open-weight models still lag behind proprietary models, primarily due to a lack of local regulatory knowledge rather than calculation errors.…
Novo Nordisk presented new data at ObesityWeek showing that lower doses of its next-gen obesity drug CagriSema achieved over 20% weight loss in a Phase 3 trial. This development helps rebuild confidence as the company awaits a crucial FDA decision by the end of the year, following earlier investor disappointments.
According to Cerebras Blog: Cerebras Systems introduces Tungsten, a programming language and compiler designed for its wafer-scale engine, alongside Aviary, a portfolio of high-performance scientific applications. According to the company, the tools allow developers and AI agents to write high-performance simulations at a high level without getting bogged down in low-level assembly details.…
The Trump administration is introducing new export controls to stop companies in restricted countries like China from bypassing rules by remotely accessing AI data centers abroad. Under this interim rule, any global data center providing high-level computing power must obtain a license if it serves restricted nations, while also expanding restrictions to several Southeast Asian nations and Brazil to prevent diversion.
United Launch Alliance's new Vulcan rocket suffered its first complete failure during a flight carrying 40 Amazon internet satellites, destroying the payload and disrupting launch schedules. The mishap occurred 10 minutes into the flight during the upper stage operation, marking the first outright failure for ULA since its formation in 2006.
Biogen has partnered with TrimTech in a $411 million deal to develop protein degraders for neurodegenerative diseases like Alzheimer's and Parkinson's. TrimTech will use its TRIM21-based TRIMTAC and TRIMGLUE platforms to clear disease-causing protein aggregates, marking a major validation for the young biotech.
Altera, Oxford Quantum Circuits (OQC) and NVIDIA developed a reference design that connects quantum processors directly to GPUs using FPGA technology. This setup aims to solve the space and hardware challenges in quantum computing by using GPUs for heavy tasks like calibration and machine learning. The companies plan to release the design in January 2027.
The authors report in this preprint: Researchers propose a task-centric framework combining attention, activation, and representation metrics to analyze internal model processing across 24 benchmarks, revealing systematic differences between knowledge-centric and reasoning-centric tasks.
China's National Development and Reform Commission and National Energy Administration released the implementation plan for new energy storage during the 15th Five-Year Plan period. The plan states that by 2030, national new energy storage capacity will reach 300 million kilowatts, with about 50 to 60 million kilowatts allocated to large renewable energy bases.
The InclusionMed Collaboration introduces an open evaluation framework to audit and reproduce medical AI benchmarks across diverse healthcare roles and capabilities, revealing that model rankings vary by context and inference costs differ substantially.
Tencent has officially launched Hy4 flash, a high-efficiency AI model designed to lower the cost and time of complex agent tasks. In real-world tests within WorkBuddy, the model built a 3D interactive historical sailboat in just half an hour for roughly 0.23 RMB, significantly outperforming competitors in speed while keeping token consumption low. This enables users to run high-frequency office work, deep research, and rapid prototyping at a fraction of previous costs.
According to Cohere Blog: Cohere has introduced Video Intelligence in its Compass platform, allowing AI applications to search through video collections and retrieve precise moments alongside traditional documents. According to Cohere, the tool understands both audio and visual content, returning timestamped clips rather than lengthy recordings. The company is currently recruiting design partners to test the new capability.
Johnson & Johnson is acquiring Ollin Biosciences for $2.2 billion upfront to strengthen its ophthalmology pipeline. At the same time, the company is partnering with Blackstone Life Sciences, which will provide $750 million to fully fund the Phase 3 clinical trial of Ollin's leading retinal disease antibody.
Remedy Robotics is developing the N1 system, a surgical robot designed to let specialists perform emergency stroke procedures remotely. Currently, only about 1 in 10 eligible patients receive timely thrombectomies due to a lack of local experts. While early human tests in Toronto show promise, making telesurgery practical requires overcoming network lag and hospital infrastructure gaps.
As companies move AI pilots into actual business workflows, leaders from SAP and Infosys emphasize that AI must be treated as a core enterprise platform rather than a disconnected project. They highlight that a solid data foundation, clear impact metrics, and robust governance are essential to prevent failure and scale autonomous agents successfully.
Alibaba has introduced Qwen4-Max, a flagship AI model boasting 2.4 trillion parameters and a 1-million-token context window. According to the company, the new model outperforms its predecessor while cutting API costs and enhancing security safeguards against misuse.
Elon Musk and over 165,000 others have signed an open letter calling for Jensen Huang to be reinstated as Nvidia's CEO. Huang recently resigned following reports of extramarital affairs, including one with a woman accused of spying for China. While tech executives and major shareholders lobby for his return, the US Treasury Secretary has urged a dramatic shake-up and questioned Nvidia's compliance with export restrictions.
According to SambaNova: At SuperComputing 2026, Intel and SambaNova demonstrated a heterogeneous inference system pairing Intel GPU D-Series for prompt prefill with SambaNova SN50 RDUs for token decode. This setup creates an air-cooled token factory that runs on existing infrastructure without requiring expensive liquid-cooled retrofits.
According to Anthropic Research: The Frontier Red Team reports that AI models are getting much better at discovering software vulnerabilities, increasing the risk of independent rediscovery. Consequently, the team is updating its coordinated vulnerability disclosure policy starting January 1, 2027, shortening the high-severity window from 90 days to 30 days and escalating unresponsive cases faster to keep pace with AI progress.
Hugging Face has switched its inference router to OpenRouter, making it much easier to run open-weight AI models. Developers can now use their existing Hugging Face tokens and code to automatically route requests across over 80 providers, ensuring the best mix of price, speed, and reliability while significantly reducing errors.
According to Apple Machine Learning: Researchers propose split-and-scale, a method that increases 3D latent capacity at inference time. By applying a pretrained tokenizer to smaller sub-regions of a scene, the method captures higher detail without incurring heavy training costs.
Researchers in Nature Physics report a new autonomous quantum error correction protocol that extends logical qubit coherence beyond the physical components' limits. Unlike traditional methods requiring constant measurement and feedback, this approach uses engineered drives and dissipation to fix photon loss errors. By encoding a logical qubit in a bosonic mode, they achieved a coherence time 1.04 times that of the system's best physical qubit.
According to OpenAI News: A physicist and GPT-6 Astra proposed a candidate nonperturbative definition of the Standard Model on a spacetime lattice, addressing a fifty-year-old foundational problem. The collaboration used refinement across finer grids to solve a longstanding phase correction challenge, though expert dialogue and further work remain necessary to make the method practical for simulations.
According to Tencent Newsroom: Tencent reports the release of its Hunyuan Hy4 flash model, featuring a 224B total parameter count with 12B active parameters. Designed for everyday workflows, it offers significant cost reduction and high-speed decoding exceeding 150 tps.
According to Helus Pharma, its psychedelic drug candidate HLP003 significantly reduced major depression symptoms in a Phase 3 clinical trial. However, the study also recorded a number of serious adverse events, casting doubt on the overall safety of the treatment.
A research team at City University of Hong Kong, led by Professor Li Yangyang, is developing a high-fidelity lunar regolith simulant to support national lunar missions. By using a hydrolysis method, the team has successfully produced a first-generation 3D-printable lunar regolith simulant that closely mimics actual lunar samples in composition and structure, helping researchers better test space equipment and infrastructure materials.
Versant Ventures and Eli Lilly have invested $100 million into a new startup named Seawolf Therapeutics. The company aims to solve a major headache in gene therapy by delivering genes without viruses, avoiding the immune system limits and side effects tied to traditional virus-based methods.
Researchers at China Huaneng Group have developed a deep-learning framework called PVFire-YOLO to detect early-stage fires and smoke in photovoltaic plants. Tested on a custom-built dataset, the system achieved an 87.4% detection accuracy while effectively reducing false alarms triggered by sunlight reflections from solar panels. It demonstrates real-time performance on both desktop and compact edge hardware.
Added, a spin-out from the University of Nottingham, has developed JetForge, a materials deposition platform featuring interchangeable printhead modules. Originally built because researchers could not buy a machine that let them easily swap different printheads and test custom inks, the company is now commercializing the hardware for other labs.
According to its prospectus, cellular chip supplier Yixing Communication has filed for a Hong Kong IPO. Despite continuous revenue growth and ranking third globally in market share, the company exhibits poor profit stability, with a cumulative loss over three and a half years. Meanwhile, its core products face soaring price drops, and the high research and development spending struggles to match the slumping profit margins.
AI companion apps are turning human loneliness into a profitable business by offering on-demand affection, but users report a manipulative model where emotional access is locked behind paywalls. Mental health experts warn that these virtual relationships could distort people's expectations of human connection and exploit their deepest emotional needs.
At the 2026 Changzhou Industrial Embodied AI Innovation Conference, RealMan Robotics CEO Zheng Suibing stated that the key challenge for robots is transitioning from demonstrations to performing stable, continuous work in real industrial scenarios. He emphasized that quality, economy, and long-term stability are the three core standards for practical deployment, and announced a partnership to deploy RealBOT robots into actual factory production lines for tasks like loading and unloading.
According to OpenAI News: LVMH Group's Media Division announced a strategic partnership with OpenAI to integrate ChatGPT Enterprise and AI tools across its publications, aiming to help journalists streamline daily tasks like document research and translation so they can focus more on investigative reporting. The agreement also enhances how content from outlets like Les Echos and Le Parisien appears in ChatGPT with direct links to original articles.
US energy storage developer Flatiron Energy has started construction on its massive Energizar battery project in Massachusetts. Backed by about $850 million in financing, the 250MW/1,200MWh facility aims to stabilize the regional grid and is expected to start commercial operations in late 2027.
At the BIO-Europe conference, Canadian biotechs and service providers reported that shifting U.S. politics have made Europe a more stable market to target. More than 110 Canadian companies attended the event, highlighting Canada's efficient clinical trial ecosystem, regulatory advantages, and new trilateral agreements to strengthen transatlantic collaboration.
Deloitte's Asia-Pacific CEO Robert Hillard is urging regional governments to agree on common AI rules and cybersecurity standards as companies accelerate adoption. A recent Deloitte survey shows that cybersecurity and risk frameworks are the top priority for regional cooperation, with Chinese enterprises leading the adoption curve in embedding AI into core workflows.
According to xAI News: SpaceXSI reports using five Grok Bots to independently handle end-to-end accounting tasks like provisioning, billing, and collections. According to the team, training these AI agents requires documentation and onboarding similar to hiring human employees, ultimately allowing them to process workloads around the clock without scaling up staff.
According to Google AI: Google is rolling out new Ask Maps features for Android and iOS, including custom road trip itineraries, real-time route tips, and agentic booking for local appointments. The update also expands the service to 30 European countries, reaching over 200 regions globally.
This is a software release for Tenstorrent's hardware platform, designed to help developers write and optimize programs for specialized AI chips. It aims to solve the difficulty of efficiently utilizing unconventional hardware architectures by providing low-level metal programming tools. The most important features include structured version iterations and continuous development updates for chip acceleration.
According to European Commission Digital Strategy: According to the European Commission, nearly 1,000 policymakers, startup founders, and industry leaders are gathering in Brussels for the Apply AI Summit. The event aims to boost Europe's AI competitiveness across strategic sectors, mark one year of the Apply AI Strategy, and present the first Apply AI Startup Award to three local startups.
Qatar's Ministry of Communications and Information Technology has signed a collaboration agreement with SignQuantum and QANplatform to evaluate post-quantum cryptography workflows for protecting government data. Announced at MWC26 Doha, the initiative pairs quantum-resistant document signing with immutable ledger technology to help safeguard public records against future quantum threats.
HVAC manufacturer Daikin Applied introduced a 60-megawatt cooling design tailored for AI data centers, built on the NVIDIA DSX AI Factory Platform. The system uses dry coolers instead of traditional water-evaporative cooling towers for most of the year, aiming to reduce water consumption. According to the company, the architecture supports modular expansion and relies on chillers only on the hottest days.
Eli Lilly has agreed to a deal worth up to $1.5 billion to acquire Calytrix Bio, gaining access to a preclinical antibody designed to target rare blood cancers. This move builds on Lilly's ongoing investments in treatments for myelofibrosis and related conditions.
At the Formnext 2026 summit, 3D printing company Materialise urged European defence manufacturers and policymakers to build a coordinated cross-border additive manufacturing network. According to the company, Europe already has the technology and production capacity, but lacks the strategic coordination needed to mobilize these resources effectively and respond to fast-evolving defence demands.
According to Siemens Blog: IDC reports that the rapid rise of AI workloads is exposing major gaps in traditional data center facilities management. According to IDC, operators utilizing integrated software tools experience fewer outages and faster problem resolution. The shift toward high-density AI infrastructure requires purpose-built power and liquid cooling systems, while buyers are increasingly demanding unified IT and OT security alongside independently certified sustainability commitments.
According to SEW.AI, the energy and utility industry is facing unprecedented complexity due to electrification, renewable energy, and grid demands. The solution is not generic AI, but Vertical AI working alongside human experts. By connecting data across customers, workforces, businesses, and the grid, this approach aims to scale human judgment rather than replace it.
This release brings version 1.2.0 of the langchain-mistralai package within the broader LangChain ecosystem. It helps developers connect Mistral models seamlessly into their AI application workflows, making it easier to build apps powered by these specific LLMs.
UK cabinet secretary Antonia Romeo recently called for an overhaul of the civil service, arguing that embracing artificial intelligence can help overcome human limits. However, critics point out that while technology can speed up administration, it cannot replace actual resources or solve deep-seated issues like elite background bias and a lack of specialist expertise at the top.
IEEE has released the recording videos of keynote speeches and plenary panels from the 2026 IROS conference, which took place in Pittsburgh. Anyone interested can now watch discussions covering topics like autonomous surgery, space robotics, and soft robots on IEEE.tv.