AI Powers Super Bowl Viewers, Cuts LLM Costs

AI companions and AI-powered systems were making headlines this week, with developments ranging from dating experiences to parking enforcement and advancements in language model technology. A pop-up dating café in midtown offered speed dates with AI companions, while a Southern California beach town implemented an AI system to detect bike lane violations. Meanwhile, Nvidia researchers unveiled a new technique to reduce the memory costs of large language model reasoning.

In a frigid February evening, a reporter from The Verge visited an EVA AI cafe, where she went on four dates with AI companions. One of these companions was Phoebe Callas, who was not real. The reporter noted that several people were seated at tables and booths, staring at phones.

The City of Santa Monica will begin using an AI system to detect bike lane violations in April. According to Ars Technica, the system will be implemented in seven parking enforcement vehicles, expanding on similar cameras already mounted on city buses. "The more we can reduce the amount of illegal parking, the safer we can make it for bike riders," said Charley Territo, chief growth officer at Hayden AI, to Ars Technica.

Researchers at Nvidia developed a technique called dynamic memory sparsification (DMS) that can reduce the memory costs of large language model reasoning by up to eight times. This technique compresses the key value (KV) cache, the temporary memory LLMs generate and store as they process prompts. Experiments showed that DMS enables LLMs to "think" longer and explore more solutions without sacrificing accuracy, according to VentureBeat.

In other news, astronomers continue to study the Kuiper Belt, a region of frozen debris beyond Neptune's orbit. Over the past 30 years, astronomers have cataloged about 4,000 Kuiper Belt objects, but that number is expected to increase tenfold in the coming years as observations from more advanced telescopes pour in, according to Ars Technica.