News Summaries

Why Today’s Chatbots Miss the Mark: Yann LeCun Says Real AI Needs a ‘World Model’

Yann LeCun, the chief AI scientist at Meta, argues that large language models (LLMs) like ChatGPT are hitting a dead‑end on the road to artificial general intelligence. While LLMs have smashed language benchmarks by predicting the next word, they still operate as autoregressive pattern‑matchers and struggle with anything that requires true physical reasoning or long‑term planning. When the same approach is tried on video, the predictions quickly turn blurry, exposing a fundamental flaw: the models lack a global causal understanding of the world. LeCun proposes a different direction—building a “world model” using joint‑embedding prediction architectures (JEPA). Instead of trying to reconstruct raw pixels, JEPA encodes inputs into compact abstract vectors and learns to predict how those vectors will change after an action. This method, exemplified by the Barlow Twins and V‑JEPA systems, has already outperformed traditional supervised models on ImageNet, achieving over 73 % accuracy without any labeled data. The key breakthrough is a new loss function that forces the embeddings from two networks to be highly correlated for the same image while remaining uncorrelated across different features, avoiding the “representation collapse” that plagued earlier attempts. LeCun believes that only by giving AI the ability to simulate physical laws and plan ahead can we move beyond clever text generation toward true AGI.

The Next Evolution of AI: From Smart Bots to Self‑Learning Digital Lifeforms

A new wave of research is turning today’s large language models (LLMs) into something far more dynamic than static text generators. The journey begins with classic “offline” training—feeding massive datasets into a single model. Researchers then add a second step called Model Online Adaptation (MOA), where the model is fine‑tuned after deployment using techniques like supervised fine‑tuning, Low‑Rank Adaptation (LoRA) or Reinforcement Learning from Human Feedback (RLHF). These tweaks let the AI better follow instructions and align with human preferences, but they still only adjust the model’s internal parameters. When tasks become too complex for one model, the third phase—Multi‑Agent Orchestration (MAO)—kicks in. Multiple AI agents start working together, exchanging messages or debating to break down problems into manageable pieces. So far, the ways they talk and cooperate are hand‑crafted by researchers. The cutting‑edge fourth phase is Multi‑Agent Self‑Evolving (MASE). Here, a whole population of agents learns to improve not just their answers but also their prompts, memory structures, tool‑use strategies, and even how they connect with each other—all automatically, based on feedback from the environment. This creates a closed‑loop, lifelong learning system that behaves more like a digital organism than a simple program. In short, AI is moving from a “black‑box tool” toward a self‑evolving digital life form capable of continuous growth and adaptation.

China’s Taiji Space Mission Hits Milestone: New Instrument Can Spot Tiny Ripples in Space

China’s ambitious Taiji Program, a home‑grown effort to detect gravitational waves from space, has taken a big step forward. The project plans to launch three satellites spaced about 3 million kilometres apart, forming a giant laser interferometer that can sense the faint ripples in spacetime caused by massive cosmic events. After the successful test flight of the Taiji‑1 satellite, engineers have now built a fully functional optical platform – the heart of the detector – and put it through rigorous ground testing. The new platform uses an innovative layout that keeps temperature swings from messing up the measurements, achieving a precision of just a few picometres – roughly one‑ten‑thousandth the width of a human hair. Test results show that instrument noise has been slashed and measurement stability improved tenfold, meeting all the demanding criteria for a space‑based observatory. With these results, the technology is moving from laboratory prototypes to real‑world engineering use. The breakthrough has been documented in the international journal *Research*, providing vital technical support for the next phase of China’s space‑based gravitational‑wave hunt. This achievement underscores China’s growing capability in frontier astrophysics and signals that the Taiji mission is on track to join the global effort to listen to the universe’s most violent collisions.

China’s Chip Materials Race: New ESG Report, Factory Moves, and Breakthroughs Fuel Fast‑Track Innovation

On April 21, Xian Yicai (688783) published its first ESG report since joining the Sci‑Tech Innovation Board, highlighting its push for greener, more innovative 12‑inch silicon wafer production. The company now ranks as China’s top 12‑inch wafer maker and the world’s sixth‑largest, using ESG goals as a long‑term competitive edge. A week earlier, Tianjin welcomed Nantong Zhanding Materials Technology, which signed a deal to relocate its headquarters and manufacturing to the Tianjin Port Free‑Trade Zone. Zhanding supplies high‑purity fluorinated chemicals used in chip fabs, AI‑center cooling, and semiconductor equipment, and has earned certifications from leading global wafer manufacturers. In research news, a joint team from the National University of Defense Technology and the Chinese Academy of Sciences announced a breakthrough in wafer‑level growth and controllable doping of two‑dimensional semiconductor materials, a key step toward post‑Moore‑era chips. Their findings appeared in *National Science Review*. Meanwhile, Qingyi Optoelectronics is scaling up mask production, now offering 180 nm and 150 nm process nodes and planning 28 nm development, while also collaborating with domestic quartz and equipment suppliers. Domestic photoresist maker Jiuri New Materials has launched 35 new products, with 17 already in the market, and ANJI Technology reported a 63.7 % revenue jump to 453 million yuan, driven by functional wet‑chemical sales and advances in CMP slurry for 3D/2.5D ICs. At a recent industry briefing, executives from seven semiconductor‑material firms stressed that parallel pushes for localization and rapid technology iteration are reshaping China’s chip supply chain, ensuring stable material supplies and keeping pace with fast‑moving process nodes.