RadixArk Cuts AI Post-Training GPU Costs by 57% With DaisyRose-27B
RadixArk used an asynchronous reinforcement learning setup across 144 H200 GPUs to train its new DaisyRose-27B model, more than doubling benchmark task performance.
Pick the topics you care about. Your homepage shows them first. No account needed.
Get new stories on your phone
Keep your saved stories, streak and topics on every device. Free. No password.
New here? The same steps create your free account. By signing in you agree to our Terms and Privacy policy.
Check your email. We sent a sign-in link to . It works for 20 minutes.
or wait a minute and send it again.
377 stories, newest first
RadixArk used an asynchronous reinforcement learning setup across 144 H200 GPUs to train its new DaisyRose-27B model, more than doubling benchmark task performance.
Microsoft CEO Satya Nadella warned that vital systems connected to the internet hold unfixable flaws that artificial intelligence can exploit.
Target delivered its fastest sales growth in over five years, driven by digital orders and store visits. The retailer also lifted its full-year earnings expectations.
Venture funding for voice AI startups reached $9.5 billion in 2026, surging past last year's total even as overall deal counts dropped.
Spatial analytics startup Wherobots launched Waldo, an AI agent that lets engineers build and run geographic data pipelines from plain-text prompts.
A study shows electron pairs persist at room temperature, revealing that synchronizing their motion is the main barrier to zero-loss power transmission.
Computer models show that deep mantle movements lifted and then sank ancient continents, eroding up to a billion years of geological history before new rocks formed.
With over 130,000 units shipped, Pudu Robotics is leveraging real-world commercial scale to feed its shared AI architecture for quadrupeds and humanoid machines.
Semiconductor leaders highlight a shift toward co-design, 3D chiplets, and secure data links as AI workloads outgrow traditional hardware methods.
IBM's time-series foundation models now run natively in Confluent via Apache Flink, allowing real-time forecasting and anomaly detection on CPUs without costly GPUs.
European regulators say online marketplace Temu is failing to stop the sale of dangerous goods, opening the company up to heavy fines.
Federal Reserve Governor Christopher Waller wants aggressive interest rate hikes, suggesting a 50-basis-point increase in December to tackle persistent inflation.
Investors are digging past simple revenue run rates as AI startups mix subscriptions with usage fees, making headline numbers far less reliable.
Dell is packaging liquid-cooled networking and storage into single racks while targeting stranded energy sources to handle massive AI inference demands.
Researchers at the University of Córdoba have used vitamin B12 as a catalyst to build a rechargeable zinc-air battery that lasts up to 250 cycles.
NASA's Voyager 1 probe has become the first human-made object to travel one light-day from Earth, reaching a distance of 16.1 billion miles.
California regulators have halted Waymo's driverless freeway operations following incidents in active construction zones and near police stops.
Semiconductor experts warn that coordinating layers of AI design agents will demand huge budgets and institutional data that startups simply lack.
Acting CEO Colette Kress announced Nvidia will end H200 chip sales to China amid tighter export rules and the departure of co-founder Jensen Huang.
At Shanghai's AI Engineer event, practical deployment on factory floors and lightweight smartphone models took center stage over Silicon Valley's existential dread.