深入探索AI人工智能领域的最新进展、核心技术及应用场景，涵盖机器学习、深度学习与自然语言处理，助您把握智能时代脉搏。
Everyone assumes the DeepSeek story is about releasing a stronger model. In September 2026, the story quietly changed shape. In early September, reports emerged that DeepSeek plans to build a data center of roughly 1GW in Ulanqab, Inner Mongolia, deploying at least 160,000 of Huawei's next-generation Ascend 950DT accelerators. At the price leaked on September 10 — over 250,000 yuan per ac …
A Top Kernel Engineer Just Wrote His Own Obituary — That's the Signal A 3,000-character essay by a DeepSeek engineer went viral in both Chinese and English tech circles this week. The author, Liu Shengyu, works on GPU kernels — the lowest layer of AI infrastructure. Days earlier, he shipped the main Attention operator for DeepSeek V4.1 Flash: a head-dim-512 MQA attention that is a major reason the
While Sam Altman and Dario Amodei spent last week warning the world to slow down, Satya Nadella sat down on the All-In Podcast and placed a $175 billion bet on the opposite conclusion. Microsoft's CEO did not talk about how dangerous models might become. He talked about who will own the layer above them. His core claim deserves to be stated plainly: the value of AI will not stay at the model layer
Zhipu and MiniMax both grew revenue 300-400% while cutting sales spend. The go-to-market engine behind AI has flipped: the model sells itself, and the sales function moves to customer success.
Anthropic Fable 5.1 shuts down distillation by validating thinking blocks, forcing the race back from copying reasoning to building it.
AI moves fast again this week. GPT-6 sent the industry chasing a revived line of research, Claude locked down a landmark formal proof, and the robots are quietly getting smarter by unlearning. Here's what mattered on September 6. Top 3 1. GPT-6 revives the recurrent transformer — and Alibaba was already there. The architecture buzz is back, with the Chinese giant holding two top-conference papers
Researchers from UC Berkeley and MIT just open-sourced FreeToken, an inference engine that runs frontier Mixture-of-Experts (MoE) models on consumer hardware. On an 8GB RTX 4060 laptop it serves Qwen3.6-35B at roughly 39 tokens/second; on an RTX 5090 it handles DeepSeek-V4-Flash (284B). The idea: treat your PC as an elastically schedulable heterogeneous compute pool instead of a resource-starved d
Stripe pays about $7.5B for OpenRouter, an 82-person model-free firm, as AI value shifts from models to the default-traffic entry point.
As generation commoditizes, the durable AI moat migrates to the results layer: verifiability, traceability, and trustworthy delivery at predictable cost.
Big day in frontier AI. OpenAI officially launched GPT-6 Astra, calling it the most intelligent and intent-aligned model it has ever built; Nvidia agreed to acquire Hugging Face for $12.93 billion while keeping it an open platform; and ByteDance reportedly raised a $29.6 billion loan, partly to fund an AI infrastructure build-out approaching $70 billion this year. Here's what matters today. Top 3
Robot foundation models keep getting bigger, which means slower inference. Real deployments hide that latency with asynchronous inference: the robot keeps executing the current action chunk while the model computes the next one. It's smooth, it's necessary — and it quietly breaks online reinforcement learning. Here's the problem. With async execution, the actions the model planned and the actions
Ant Group OmniTable, VLDB 2026 best industrial paper, uses one logical wide table to cut LLM data prep from 14 days to 2.5 across 305B rows.