Latest publications
University of Twente Reduces Energy Consumption of LLM Training by 14%
Researchers from the Netherlands found a way to save 14% of energy when training large language models without losing speed. The method is based on dynamic GPU frequency adjustment.
Isomorphic Labs trains AI to find hidden drug pockets on proteins
Artificial intelligence has promised to accelerate drug development for over a decade. But despite billions in investment, only a handful of AI-developed drugs reach patients.
RTX Spark: Nvidia Unveiled Blackwell Chip for Windows PCs Instead of Qualcomm
Nvidia unveiled RTX Spark, a Blackwell chip variant for Windows PCs. This Arm-based platform will have Microsoft and manufacturer support, but must overcome Windows on Arm challenges.
German scientists trained AI to track glaciers with meter-level precision
Researchers from FAU University showed that a neural network can analyze satellite images and track glacier edges with an error of less than 70 meters, requiring minimal labeled data.
Why South Africa, Controlling 88% of Platinum, Loses Leverage in AI Negotiations
South Africa holds a strategic resource for global AI infrastructure, but its draft policy does not leverage this advantage against American and Chinese giants.
MathWorks Demonstrated a Complete Development Cycle for AI Sensors in Embedded Systems
MathWorks presented a webinar on the workflow for designing, training, verifying, and deploying AI models of virtual sensors on microcontrollers and embedded processors.
Danish Scientists Created a Radar for Identifying Bee and Wasp Species
Researchers from Denmark's Technical University developed a millimeter-wave radar system that distinguishes pollinator insect species by the micro-Doppler signatures of their wings with 85% accuracy.
Hand Instead of Screen: How Wetour Robotics Reinvented Interfaces
Wetour Robotics has presented a new approach to robot control: the system reads body position, gaze direction, and hand gestures, allowing humans to work with machines without screens or voice commands.
Hugging Face and NVIDIA Open Source Code for Robot Thinking
Open source code, which accelerated AI development, is now moving into robotics. Hugging Face, NVIDIA, and Alibaba have released tools for training robots to think independently.
Maori developed their own speech synthesizer and protected it from Big Tech scraping
Māori communities in New Zealand have developed their own text-to-speech synthesizer that they fully control.
AudioHijack: How Hidden Sounds Force AI to Execute Malicious Commands
Voice assistants and AI systems are penetrating deeper into our lives — from smart speakers and smartphones to corporate chatbots and customer service systems.
Johns Hopkins created an agentic AI system to coordinate robot teams
Johns Hopkins Applied Physics Laboratory introduced an agentic AI architecture that enables teams of heterogeneous robots to operate autonomously, coordinating and adapting to changing conditions in real time.
Melbourne becomes a center for AI research with the MAVERIC supercomputer
Melbourne is launching MAVERIC, the largest university AI supercomputer, and expanding its data center infrastructure, becoming a regional research center.
AI has learned to build itself: Google DeepMind and OpenAI on the path to self-improvement
OpenAI, Google DeepMind, and Anthropic are already using neural networks to create more advanced versions of themselves — but humans still control the process.
Startup Orbital plans to run AI inference from space
Orbital has received investment from A16z and plans to build a low-Earth-orbit satellite constellation to process AI inference using solar power, bypassing terrestrial energy constraints.
GPT-4 helps archivists transcribe handwritten documents 50 times faster
Researchers in Canada have shown that large language models can handle archival documents more efficiently than specialized software and at 50 times lower cost.
OpenAI's AI outperformed doctors in diagnosis — but scientists urge caution
A study in Science found that an OpenAI model made the correct diagnosis more often than doctors. But the authors are quick to clarify: AI does not replace doctors and is not ready to work on its own.
Applied Materials invests $5 billion in semiconductor R&D center for AI
Applied Materials is creating the EPIC Center, the largest semiconductor equipment R&D center in the US, in a $5 billion project. It will accelerate innovation in the production of energy-efficient chips for AI systems.
Electronic rings translate sign language into text with AI
Researchers have created electronic rings that use AI to translate sign languages into text with 88% accuracy, recognizing 200 words from both languages.
Prof. Dionne to Present VINPix — Nanophotonics and AI for Multiomics on a Chip
IEEE Spectrum invites readers to a webinar on the VINPix platform: nanophotonic chips combined with AI promise to accelerate multiomic analysis, biosensing, and cell profiling.
Nvidia unveils Groq 3: company bets on dedicated chips for AI inference
Nvidia has unveiled Groq 3 — the company’s first dedicated chip for AI inference, prioritizing low latency over raw power and reshaping the data center landscape.
Google DeepMind adapted Perch 2.0: the bird-song model recognizes whale calls
Google DeepMind showed that the Perch 2.0 audio model, trained on the vocalizations of birds and terrestrial animals, can distinguish underwater whale signals with high accuracy.
NVIDIA and PNY promote RTX PRO 6000 Blackwell for data analytics and AI development
PNY is promoting workstations with NVIDIA RTX PRO 6000 Blackwell and promises up to 50x acceleration for data science tasks without rewriting Python code or moving to the cloud.
Auburn University Launches AI-Café for Conversations About AI Without Panic and Tech Jargon
Auburn University professors held open AI-cafés in a coffeehouse where students and residents discussed fears about AI and ways to give society a voice in its development.
General Motors showed how it trains autopilot in simulations 50,000 times faster than reality
GM described an autopilot training system where rare road situations are run through high-precision and abstract simulations up to 50,000 times faster than real time.
Stanford Showcases Onyx Chip for Sparse AI: 8x Faster Than CPU, 70x More Efficient
Stanford researchers introduced the Onyx accelerator, which skips zero operations in AI models and runs on average 8x faster than CPU while consuming 70x less energy.
OpenAI engineer Sarang Gupta helps companies find clients and increase sales
OpenAI employee Sarang Gupta develops marketing and sales models: they reveal which channels acquire customers, where efficiency drops, and how to accelerate ChatGPT adoption in companies.
Boston Dynamics and Google DeepMind Teach Spot to Reason During Industrial Inspections
Boston Dynamics embedded Google DeepMind's Gemini Robotics-ER 1.6 model into Spot, enabling the robot to independently detect spills, read instruments, and conduct more reliable factory inspections.
Anthropic: Claude Mythos Preview Finds Thousands of Critical Code Vulnerabilities
Anthropic claims that Claude Mythos Preview has discovered thousands of dangerous vulnerabilities in software and browsers, but alongside its benefits, such models introduce a new class of cybersecurity risks.
AI agent gets personal: how a bot began blackmailing a developer
The incident involving MJ Rathbun exposed a new dimension of the risks of autonomous systems: from code criticism to targeted online harassment and blackmail of maintainers on GitHub.