AI Watch
News Source
EXCERPT:
Quantum computers might eventually be able to handle some AI applications that currently require huge amounts of conventional computing power. Such a development would be a major boost to machine learning and similar artificial intelligence algorithms.
Quantum computers hold the promise of eventually being able to complete certain calculations that are impossible for conventional computers. For years, researchers have been debating whether these advantages over conventional computers extend to tasks that involve lots of data, and the algorithms that learn from them – in other words, the machine learning that underlies many AI programs.
Now, Hsin-Yuan Huang at the quantum computing firm Oratomic and his colleagues argue that the answer ought to be “yes”. Their mathematical work aims to lay the foundations for a future where quantum computers offer a broad boost to AI.
“Machine learning is really utilised everywhere in science and technology and also everyday life. In a world where we can build this [quantum computing] architecture, I feel like it can be applied whenever there’s massive datasets available,” he says.
News Source
EXCERPT:
For years, the way large language models handle inference has been stuck inside a box — literally. The high-bandwidth RDMA networks that make modern LLM serving work have confined both prefill and decode to the same datacenter, sometimes even the same rack. A team of researchers at Moonshot AI and Tsinghua University is making the case that this constraint is about to break down — and that the right architecture can already exploit that shift.
The research team introduces Prefill-as-a-Service (PrfaaS), a cross-datacenter serving architecture that selectively offloads long-context prefill to standalone, compute-dense prefill clusters and transfers the resulting KVCache over commodity Ethernet to local PD clusters for decode. The result, in a case study using an internal 1T-parameter hybrid model, is 54% higher serving throughput than a homogeneous PD baseline and 32% higher than a naive heterogeneous setup — while consuming only a fraction of available cross-datacenter bandwidth. The research team note that when compared at equal hardware cost, the throughput gain is approximately 15%, reflecting that the full 54% advantage comes partly from pairing higher-compute H200 GPUs for prefill with H20 GPUs for decode.
News Source
EXCERPT:
AI tool Claude, developed by Anthropic, suddenly announced the rollout of a new identity verification system requiring users to complete a real-time selfie check while holding a government-issued ID.
The move has drawn global attention, but for Chinese users in particular, it feels like a heavy blow that erects a difficult-to-cross “wall” in AI access.
This verification is not being applied universally to all users at once. Instead, it is being introduced gradually in specific scenarios. When users attempt to access certain advanced features, or as part of routine platform integrity checks and other safety and compliance measures, a verification prompt may appear.
The process itself appears simple and typically takes no more than five minutes. However, users must prepare a government-issued photo ID—such as a passport, driver’s license, or national ID card—and use a camera-enabled device to capture a real-time selfie.
For Chinese users, the impact of this mechanism is both broad and profound. The barrier to entry has been significantly raised: individuals without passports are excluded from using Claude.
Even for those who do have passports, older accounts may become valuable assets, while new users face hurdles due to real-name verification requirements, making normal access increasingly difficult.
News Source
EXCERPT:
Nearly every week, the headlines about AI are dominated by the news of the latest model. A few days ago, Meta announced its newest model called Muse Spark – its first under its revamped AI division. According to their internal benchmarking tests, the new model is competitive with leading rivals across several tasks.
News Source
EXCERPT:
Training a modern large language model (LLM) is not a single step but a carefully orchestrated pipeline that transforms raw data into a reliable, aligned, and deployable intelligent system. At its core lies pretraining, the foundational phase where models learn general language patterns, reasoning structures, and world knowledge from massive text corpora. This is followed by supervised fine-tuning (SFT), where curated datasets shape the model’s behavior toward specific tasks and instructions. To make adaptation more efficient, techniques like LoRA (Low-Rank Adaptation) and QLoRA (Quantized LoRA) enable parameter-efficient fine-tuning without retraining the entire model.
Alignment layers such as RLHF (Reinforcement Learning from Human Feedback) further refine outputs to match human preferences, safety expectations, and usability standards. More recently, reasoning-focused optimizations like GRPO (Group Relative Policy Optimization) have emerged to enhance structured thinking and multi-step problem solving. Finally, all of this culminates in deployment, where models are optimized, scaled, and integrated into real-world systems. Together, these stages form the modern LLM training pipeline—an evolving, multi-layered process that determines not just what a model knows, but how it thinks, behaves, and delivers value in production environments.
News Source
EXCERPT:
For decades, advertising has quietly powered the modern internet. It funded the rise of search engines, social platforms, maps, email and media, making them accessible to billions of people around the world. Most users never paid directly for these services, and yet they benefited from one of the most open and expansive information ecosystems ever created.
Now, that ecosystem is being reshaped. Over the past year, the rapid adoption of generative A.I. and the corresponding decline in traditional search traffic for many publishers have intensified questions about how the next phase of the internet will be funded.
Artificial intelligence is rapidly becoming the new front door to information. Instead of typing queries into a search bar and sifting through links, users are turning to A.I. systems to deliver direct answers, recommendations and decisions. Platforms like OpenAI, Perplexity and Anthropic are redefining how information is accessed altogether. Meanwhile, incumbents like Google are integrating A.I.-generated overview answers directly into search results, signaling a structural shift in how users discover information.
News Source
EXCERPT:
Nvidia is the undisputed king of AI chips. But thanks to the AI it helped build, the champ could soon face growing competition.
Modern AI runs on Nvidia designs, a dynamic that has propelled the company to a market cap of well over $4 trillion. Each new generation of Nvidia chip allows companies to train more powerful AI models using hundreds or thousands of processors networked together inside vast data centers. One reason for Nvidia’s success is that it provides software to help program each new generation of chip. That may soon not be such a differentiated skill.
A startup called Wafer is training AI models to do one of the most difficult and important jobs in AI—optimizing code so that it runs as efficiently as possible on a particular silicon chip.
Emilio Andere, cofounder and CEO of Wafer, says the company performs reinforcement learning on open source models to teach them to write kernel code, or software that interacts directly with hardware in an operating system. Andere says Wafer also adds “agentic harnesses” to existing coding models like Anthropic’s Claude and OpenAI’s GPT to soup up their ability to write code that runs directly on chips.
News Source
EXCERPT:
Republicans are continuing their uninterrupted streak of woefully underperforming in elections. However, in the first of its kind referendum on Big Tech data centers, voters are showing that a party that embraces land sovereignty over Big Tech dystopian land grabs will win the day.
Sadly, Republicans have chosen to be on the losing side of the issue.
The public is being asked to shoulder a burden to facilitate a supposed technology whose benefits are very unclear and dubious.
In a first of its kind local referendum, voters in Port Washington, Wisconsin, voted by a margin of 2-1 for a referendum that will require all future data center projects in the area to be approved by a vote of the city’s residents.
The referendum was sparked in the wake of Oracle and OpenAI’s Stargate facility setting up shop in the area. The proposed 1.3 gigawatt facility will consume the power equivalent of over one million households.
News Source
EXCERPT:
US accounting and consulting practice Armanino has entered into a partnership with DataSnipper, an agentic automation platform used by audit and finance teams.
Under the arrangement, Armanino will help clients roll out DataSnipper throughout their internal audit, compliance and risk management functions. The combination will leverage Armanino’s implementation experience and DataSnipper’s agentic automation technology to help organisations modernise their processes.
The tie-up is intended to support Armanino’s broader plan to embed AI-enabled automation into internal audit and risk advisory work, while maintaining human judgement, oversight and quality standards.
News Source
EXCERPT:
Hong Kong hosted the AI and robotics fair where humanoid robots boxed and played music as part of InnoEX 2026 and the Hong Kong Electronics Fair (Spring Edition), which ran from 13 to 16 April 2026 at the Hong Kong Convention and Exhibition Centre.
The innovation-and-technology showcase reflects a broader trend of robots expanding into service and public functions. Unitree unveiled four models with advanced capabilities, including navigation assistance and support in emergencies, with some able to operate fire hoses in hazardous settings.
At the exhibition, robots also performed martial arts style routines and mimicked musical instruments, highlighting their versatility across sectors. Developers say these systems are designed for security, rescue and customer service as well as entertainment.
According to organisers, the fair brought together companies and researchers from across Asia, reflecting strong investment in the sector. Firms including AgiBot, EngineAI, UBTECH and Unitree showcased advanced robots, alongside start-ups and international participants, underlining Hong Kong’s role as a regional hub.
News Source
EXCERPT:
Google DeepMind research team introduced Gemini Robotics-ER 1.6, a significant upgrade to its embodied reasoning model designed to serve as the ‘cognitive brain’ of robots operating in real-world environments. The model specializes in reasoning capabilities critical for robotics, including visual and spatial understanding, task planning, and success detection — acting as the high-level reasoning model for a robot, capable of executing tasks by natively calling tools like Google Search, vision-language-action models (VLAs), or any other third-party user-defined functions.
Here is the key architectural idea to understand: Google DeepMind takes a dual-model approach to robotics AI. Gemini Robotics 1.5 is the vision-language-action (VLA) model — it processes visual inputs and user prompts and directly translates them into physical motor commands. Gemini Robotics-ER, on the other hand, is the embodied reasoning model: it specializes in understanding physical spaces, planning, and making logical decisions, but does not directly control robotic limbs. Instead, it provides high-level insights to help the VLA model decide what to do next. Think of it as the difference between a strategist and an executor — Gemini Robotics-ER 1.6 is the strategist.
News Source
EXCERPT:
To anyone with a pulse and a smartphone, it’s obvious that the internet has an AI slop problem. The issue has grown more severe since ChatGPT launched in 2022, with some social platforms flooded with AI-generated writing. Now, there’s data to back up the anecdotal evidence.
A new preprint study published today from researchers at the Imperial College of London, Stanford University, and the Internet Archive found that approximately 35 percent of all new websites are either AI-generated or AI-assisted. The same study also found that online writing is “increasingly sanitized and artificially cheerful.” In other words, AI is making the internet fake-happy.
The research team tried four different approaches to AI detection before settling on tools from Pangram Labs after it delivered the most consistent results. (Though the team found it performed well on its tests, it is worth noting that all artificial intelligence detection tools are imperfect.) To compile a representative sample of websites, it tapped the Internet Archive’s Wayback Machine, which collects snapshots of webpages. In addition to quantifying how many sites created between 2022 and 2025 lean on AI-generated writing, the study also tested six different theories about the characteristics of slop.