01News
The new generation TPU chip 'Xuyu' delivers single-chip mixed-precision floating-point compute of 896TFLOPS, 3 times the performance of the previous generation, with 8-bit inference compute of 1792TOPS. The simultaneously released 'Taize 2.0' intelligent computing platform delivers compute of 7.168P, with energy consumption only 80% of traditional GPU servers, and has completed adaptation of dozens of mainstream large models.
Source links can be verified on the original siteRead original ↗02News
An OpenAI engineer revealed that the company successfully reduced AI model inference costs by more than 50% through underlying optimizations such as improving server resource utilization. This means fewer Nvidia chips are needed to run AI, and the saved costs can be used to lower API pricing or increase user usage limits.
Source links can be verified on the original siteRead original ↗03News
Proton's Lumo 2.0 is dropping this week, giving users a broader variety of capabilities.
Source links can be verified on the original siteRead original ↗04News
OpenAI has fixed a fault that caused Codex users' quotas to be consumed abnormally fast. The engineering lead said the main causes were excessive backend computation and abnormal operation of features such as automated review. The company has fully reset user quotas and added monitoring mechanisms. As AI programming tools become widespread, computing resource and cost control have become challenges faced jointly by service providers and users.
Source links can be verified on the original siteRead original ↗05News
Republican and Democratic campaigns now run on AI at nearly every step, from vetting opponents to micro-targeting voters, according to a New York Times report. But the technology is still a political minefield, and Europe is taking a very different approach.
Source links can be verified on the original siteRead original ↗06News
The current environment is abnormal. After completing verification, you can continue to access. Go to verification
Source links can be verified on the original siteRead original ↗07News
Arena's AI Evaluations enterprise service has been online for only 8 months, with ARR exceeding $100 million, validating the feasibility of its business model. The platform originated as an academic project, accumulating users through free strategies, and now provides enterprises with model evaluation services based on human feedback. In contrast, another evaluation platform, Yupp, closed due to failing to find product-market fit.
Source links can be verified on the original siteRead original ↗08News
Republican and Democratic campaigns now run on AI at nearly every step, from vetting opponents to micro-targeting voters, according to a New York Times report. But the technology is still a political minefield, and Europe is taking a very different approach.
Source links can be verified on the original siteRead original ↗09News
Meta reportedly had hundreds of contractors pose as minors and send suicide, sex, and drug-related prompts to chatbots from OpenAI, Google, and Character.AI. In a single testing round, more than 45,000 prompts were sent. The companies being tested had no idea.
Source links can be verified on the original siteRead original ↗10News
Meta announces Brain2Qwerty v2 research results, using AI to reconstruct natural language from magnetoencephalography signals to help disabled people express themselves. The system requires no surgical implantation, with an average word recognition accuracy of about 61%, and up to 78% at best. #BrainComputerInterface# #AIMedicine# But the device still relies on large laboratory instruments for now, and is still some distance from practical use.
Source links can be verified on the original siteRead original ↗11News
Former Google engineering lead Hugh Williams, with the help of Claude Code and without hand-writing a single line of code, successfully developed Zettair, a search engine that can index 1.5 million Wikipedia entries. He emphasized that the best results from AI tools require combination with deep expertise, and experienced engineers are the best 'instructors'. #AI Programming#
Source links can be verified on the original siteRead original ↗12News
Taiwanese authorities have raided the offices of Super Micro Computer and several local partner companies.
Source links can be verified on the original siteRead original ↗13News
Musk transferred dozens of core engineers from SpaceX's Starlink and Starship programs to the AI team to fully iterate on the Grok large language model. The latest version, Grok 4.5, has begun internal testing, and SpaceX plans to launch a newly trained large model every month. This marks that after completing its $85 billion IPO, SpaceX is leveraging massive capital and engineering resources to fully bet on the AI niche.
Source links can be verified on the original siteRead original ↗14News
OKX is bringing together payments, identity, and reputation into a marketplace for AI agents.
Source links can be verified on the original siteRead original ↗15News
Tidal states that it absolutely will not tolerate any AI-generated music that exploits real people's reputations to deceive listeners or lowers service quality. The platform will impose higher integrity requirements on GenAI content.
Source links can be verified on the original siteRead original ↗16News
Xiaomi Pengpai OS 3 was officially announced to receive an upgrade today. Super Xiaoai, Xiaomi Super Island, and widgets launch practical match-watching tools to grasp first-hand event info anytime; album AI tools integrated for easier editing.
Source links can be verified on the original siteRead original ↗17News
Firefly Aerospace plans to deploy Nvidia Jetson edge AI platform to lunar orbit for the first time in the "Blue Ghost 2" mission launching at the end of 2026. Its self-developed Ocura imaging system will run AI algorithms directly in orbit, extract key information and transmit back to Earth nearly in real time, greatly shortening the traditional weeks-long data processing cycle and reducing communication costs.
Source links can be verified on the original siteRead original ↗18News
With surging demand for AI, foundry capacity is shifting toward AI-related products, altering the supply-demand structure of mature processes. Due to factors such as production cuts by major manufacturers, emerging applications squeezing capacity, and raw material inflation, foundry prices for eight-inch and twelve-inch mature processes have begun to rise, with increases expected to continue until 2027.
Source links can be verified on the original siteRead original ↗19News
Deepseek's new DSpark framework boosts per-user response speed by 60 to 85 percent. A small model proposes token candidates that the larger model checks in batches, squeezing more performance out of fewer chips. That could further reduce China's dependence on US high-end hardware.
Source links can be verified on the original siteRead original ↗20News
Meta AI just introduced Brain2Qwerty v2. It decodes natural sentences from non-invasive brain recordings in real time. The system reads magnetoencephalography (MEG) signals while a person types. It reconstructs what they typed, with no implant and no surgery. This is the follow-up to Brain2Qwerty v1, released in February 2025. Meta is also releasing the full training code for both versions. The pipeline combines a convolutional encoder, a transformer, and a character-level language model.
Source links can be verified on the original siteRead original ↗