Topic
Explore English articles about Generative AI.
On July 23, 2026, the UK AI Security Institute and the US CAISI jointly evaluated Kimi K3's cyberattack capabilities. It scored only 32% on ExploitBench (compared to roughly 76% for the average US model), but asymmetric testing conditions and the timing of the announcement complicate how the results should be interpreted.
SK and NVIDIA have combined AI factory construction with long-term HBM4 supply into an LOI exceeding $500 billion. Initial operation of up to 2GW is planned for 2027, but the cost breakdown and construction site remain undisclosed.
A technique using "Lossless Scaling" to split rendering and frame generation across separate GPUs is gaining attention. Unlike legacy technologies such as SLI, this method allows GPUs from different generations to be combined to distribute rendering load, offering a promising new approach to affordable performance gains by repurposing older hardware.
Etched announced a $300M raise that included participation from SK hynix, and has begun initial production of inference clusters combining low-voltage compute with shared memory. Supply agreements and comparable performance figures have not yet been disclosed.
AMD layers Skills, a CLI, and Hyperloom's autonomous optimization on top of the ROCm Core SDK, connecting everything from environment setup to performance tuning through agents. The current implementation is a Tech Preview, with constraints remaining on supported GPUs and ROCm versions.
At the 2026 International Mathematical Olympiad in Shanghai, multiple AI models achieved a perfect score—solving all six problems correctly—for the first time in history. AI algorithms that have evolved from language processing to rigorous "machine reasoning" now stand shoulder to shoulder with the top one percent of human mathematicians, marking a new stage in the automation of scientific research.
AMD and Cerebras plan to launch a division-of-labor AI inference platform in the second half of 2026, with Helios handling input processing and WSE handling token generation. The claimed up-to-5x figure comes from internal modeling comparing against WSE alone, and disclosure of measured latency and pricing will be the next thing to watch.
AMD has fleshed out the Instinct MI455X, equipped with 432GB of HBM4, and the Helios system with a 72-GPU configuration. After shipping begins in the second half of 2026, the question will be whether theoretical performance can be translated into real-world speed and operating costs.
The new MCP spec, slated for publication on July 28, 2026, eliminates initialization and the protocol session. While it lightens cloud deployments, authorizing state handles and designing for old-new compatibility become challenges for implementers.
Amazon has cut jobs in part of its organization referred to as AGI. The company continues roughly $200 billion in capital expenditures while narrowing down model development priorities and reshuffling staff allocation.
