AI News Roundup – China builds “Manhattan Project” for production of advanced AI chips, Meta developing new generative AI image and video model, new “reservoir computing” chip could reduce AI power usage, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• • A recent Reuters investigation has unearthed key details about a massive technological undertaking in China as it seeks to rival the West in the production of AI chips. Chinese scientists in the major manufacturing city of Shenzhen have reportedly built a prototype extreme ultraviolet lithography machine, commonly known as EUVs. EUVs use beams of light to etch microscopic circuits onto silicon wafers, and are key to creating the most advanced semiconductors. The Dutch technology firm ASML currently dominates the market for EUVs, but its ability to sell to China has been hampered by U.S.-led sanctions. However, Reuters has discovered that China has reverse-engineered older ASML machines bought on secondary markets to build its protype EUV, which is reportedly generating light but has not yet produced working chips. The Chinese researchers worked in complete secrecy, and recruited recently-retired ASML employees to assist in the project. While the Chinese machine is only a prototype, it demonstrates that China is much closer to catching up to the West than previously thought in the field of advanced semiconductors, especially those that power AI models.
  • Meta is developing a new generative AI model focused on image and video outputs, according to The Wall Street Journal. The new model, codenamed “Mango,” will compete with OpenAI’s Sora video model and Google’s new Nano Banana image model, as companies seek to attract customers by highlighting the visual features of their AI systems. Nano Banana, introduced in Google’s most recent updates to its Gemini model, boosted the monthly users of that AI system from 450 million to 650 million. Meta is also developing a next-generation text-based AI model, codenamed “Avocado,” which, as Bloomberg reported last week, may not be open-source like Meta’s earlier Llama models, as the company has deprioritized open-source technology in recent months.
  • IEEE Spectrum reports on a new AI chip developed by a Japanese research team which could help reduce the energy impact of AI. Researchers at Hokkaido University and TDK Corporation have designed an AI chip that may be able to speed up predictive capabilities of AI at a fraction of the power usage and latency. The chip uses a technique known as “reservoir computing,” which can extract features from time-series data using a similar structure to neural networks, but without the “layers.” Rather, the artificial neurons are connected in a complicated web, and those connections do not change weights; rather, only the last connection between the “reservoir” network and the outputs are re-weighted during training. The idea debuted in the 1990s, but has received renewed interest in recent years as it has low power usage compared to more modern machine learning techniques. They are particularly effective at predictions using time-series data, such as weather. The researchers developed a chip that used analog circuits to represent the reservoir, which only consumed 20 microwatts of power, much less than other similar designs. While the prediction ability itself is comparable to current systems, one researcher said that “the power consumption, the operation speed, is maybe 10 times better than the present AI technology. That is a big difference.”
  • YouTube has taken down two large channels that used generative AI to create fake movie trailers, according to Deadline. The channels, Screen Culture and KH Studio, based in India and Georgia, respectively, together had over a billion views of fake movie trailers created using generative AI, often confusing viewers searching for legitimate trailers. Screen Culture in particular had created 23 versions of a trailer for Marvel Studios’ recent “Fantastic Four” movie, of which several outranked the official trailer in search results. YouTube stated that the channels violated the platform’s spam and misleading-metadata policies, thus resulting in their termination.
  • Bloomberg reports that DoorDash, a major food delivery service, has partnered with OpenAI to allow users to order groceries through ChatGPT. Users will be able to ask ChatGPT for recipe suggestions, generate a shopping list, and then can be redirected to DoorDash to place the order. The majority of DoorDash’s business is delivery from restaurants, but the company decided to focus on groceries as recipe planning is reportedly a top use of ChatGPT. Instacart, a rival grocery delivery service, also recently introduced a similar feature through the ChatGPT app, though Instacart users can complete the purchase through ChatGPT rather than being redirected to another app as with DoorDash.