News Summaries

Scientists Create Unlimited Supply of Cancer‑Fighting Immune Cells

Researchers at the Keck School of Medicine of USC have discovered a way to mass‑produce a special type of immune‑cell precursor that can be turned into powerful cancer‑killing agents. By using a stem‑cell‑inspired method, they can grow huge numbers of granulocyte‑monocyte progenitors—early‑stage cells that normally develop into several kinds of white blood cells. These progenitors can be genetically edited to carry chimeric antigen receptors (CARs), giving them the ability to recognize and attack tumor cells. In mouse experiments, the engineered cells not only shrank tumors but also helped restore normal immune function, suggesting they could act as a durable, off‑the‑shelf therapy that doesn’t require customizing each treatment for individual patients. Because the cells can be produced in large batches and stored, the approach could dramatically cut the cost and time needed for cell‑based immunotherapies. The findings, published in *Cell*, open the door to a new class of readily available, customizable immune weapons that may one day complement or replace existing CAR‑T treatments for a variety of cancers and immune disorders.

AI Daily Wrap: GPT‑5.6 Drops, Anthropic Tops Unicorn List, and Robots Get Real‑World Training

The China Intelligent Agent Developer Community’s June 28 roundup is packed with headline‑making AI news. OpenAI surprised everyone by launching GPT‑5.6, instantly out‑performing Claude Fable 5 in benchmark tests and shaking the “base‑model throne.” Meanwhile, Hurun’s 2026 Global Unicorn List crowned Anthropic as the world’s most valuable AI unicorn at ¥6.6 trillion, nudging OpenAI to second place and putting DeepSeek among the 308 new unicorns. In personnel news, Noam Shazeer—the co‑inventor of the Transformer architecture and former Google VP—has jumped to OpenAI as its new architecture research lead, hinting at fresh breakthroughs. On the tech front, NVIDIA unveiled NeMo AutoModel, delivering a 3.7× speed boost for MoE fine‑tuning, while UC San Diego’s DFlash paper showed block‑diffusion drafting can accelerate LLM inference up to 15×. Google’s Gemini 2.5 Pro Deep Think topped programming benchmarks, and China’s Ministry of Industry and Information Technology teamed up with the State‑owned Assets Supervision Commission to launch a nationwide humanoid‑robot training program. Domestic AI hardware startup Deep Machine Intelligence raised two massive funding rounds in two months, underscoring the hot‑bed of embodied‑intelligence investment. The Summer Davos Forum in Dalian highlighted the “China Solution” as a buzzword, and Anthropic opened a Seoul office, partnering with TCS and DXC to embed Claude in regulated sectors. Finally, analysts note a “dense period” of model releases—Claude Fable 5, DiffusionGemma, Kimi K2.7 Code, GLM‑5.2—and a rapid maturation of AI agents moving from labs to real‑world workflows, fueling the rise of the one‑person company model.

Boosting AI Chat Speed by 85%: Inside DeepSeek’s DSpark Breakthrough

DeepSeek has just released the code and model checkpoints for its new DSpark system, along with a companion repository called DeepSpec that contains three speculative‑decoding algorithms: Eagle3, DFlash and DSpark itself. The goal? To make large‑language‑model (LLM) inference dramatically faster without sacrificing answer quality. DSpark tackles two long‑standing problems in speculative decoding. First, the "draft" stage: parallel drafts are lightning‑quick but often guess wrong, while traditional autoregressive drafts are accurate but slow. DSpark adopts a semi‑autoregressive approach that adds only about 1 % extra latency but improves the length of accepted drafts by roughly 30 %, giving the best of both worlds. Second, the verification stage: static verification lengths choke under heavy traffic. DSpark uses a dynamic scheduler that allocates verification resources in real time based on current hardware load, pushing the performance frontier outward. In offline tests on models such as Qwen‑3‑4B/8B/14B and Gemma‑4‑12B, DSpark outperformed the strongest autoregressive baseline (Eagle3) by up to 31 % and the top parallel baseline (DFlash) by up to 18 %. Real‑world deployment shows an 85 % speed‑up in high‑concurrency settings, thanks to a greedy‑sorting, early‑stop verification algorithm that keeps computational cost low while preserving accuracy. The release also includes confidence‑threshold tuning that dramatically improves acceptance rates for chat tasks, turning over‑confident predictions into well‑calibrated outputs. For engineers looking to run big models at scale, DSpark offers a practical, open‑source toolkit that blends clever algorithm design with real‑system considerations.

China’s Qingzhou Space Lab Shows Off New Tech After Successful Orbital Test

China’s Qingzhou experimental spacecraft, nicknamed the “White Elephant,” has finished its second round of on‑orbit testing and the results are impressive. Launched in March, the tiny lab has proved four breakthrough ideas that could change how we work in space. First, ultra‑precise sensors measured the spacecraft’s tiny shape changes down to a single micron, while a new “chip‑gyroscope” tracked its spin with the accuracy of a navigation system, all without bulky equipment. Second, health‑monitoring tools for future astronauts are taking shape. A muscle‑watching device, built with a tiny neural chip, can continuously record faint muscle signals and send them back to Earth, eliminating the need for astronauts to manually run tests. A handheld blood‑cell analyzer does the same for medical samples, freeing crews from large, ground‑based labs. Third, a low‑cost, reusable life‑science cabin showed it can keep plants and tiny organisms alive in space, using a simple gas‑liquid mixer that replaces expensive imported parts. A spider‑inspired sticky arm grabbed floating objects without creating debris, pointing toward future space‑debris cleanup. Finally, a new space refrigerator kept food and experiments cold in microgravity, solving a long‑standing cooling problem. The Qingzhou lab will keep delivering experiments, with a full cargo version slated for launch in early 2027, promising to support China’s space station and future deep‑space missions.

China Mobile Unveils 6G‑Powered Smart Factories at 2026 Shanghai MWC

At the 2026 Mobile World Congress in Shanghai, China Mobile gave a glimpse of a future where 6G connects everything—from smartphones to robots—turning ordinary devices into intelligent collaborators. The company showcased a new 6G test hub that blends communication, sensing, computing and AI, and highlighted its role in shaping the first 6G standards within the 3GPP framework. In a live demo, a fleet of robotic arms worked together in real time, using edge‑computing to share tasks and adapt instantly to complex production lines, dramatically boosting precision and flexibility. China Mobile also introduced an “Agent Communication” prototype that lets smart glasses and robot dogs talk to each other through voice‑driven intent parsing, enabling on‑the‑fly object recognition and transport. Beyond factories, the carrier is building a three‑dimensional coverage network that spans space, air, land and sea. This includes a low‑orbit test satellite, a layered maritime coverage system, and a 5G‑ATG platform for high‑altitude aviation. Together, these innovations aim to break the limits of today’s networks, laying the groundwork for a truly connected, intelligent world powered by 6G.