The week in AI: Govt eyes AI tool for news monitoring, experts prep AI doomsday exam

Meanwhile, Lenovo started manufacturing AI servers at its Puducherry facility in India earlier this week. The target? 50,000 enterprise AI rack servers and 2,400 graphics processing units (GPU) per year. The electronics hardware giant has also set up a new Bengaluru-based lab to focus on studying AI servers.

Features and finds

This week saw the launch of SocialAI, a new social media platform where you’re the only “real” user and other “users” are AI bots.

In new features, Google announced its search engine will be upgraded with a new technical standard called Content Credentials by 2025. It is meant to help users verify if images and videos in search results have been modified using AI.

Speaking of, Google-owned YouTube said it is bringing an in-house AI video model to its Shorts feature as part of a broader set of rollouts aimed at unlocking “new forms of expression for creators.”

AI startup of the week: Supermaven

AI coding startup Supermaven created a lot of buzz among VCs this week after it managed to secure $12 million in its first outside funding round led by Bessemer Venture Partners. It saw participation from high-profile angel investors such as the co-founders of OpenAI and Perplexity.

What’s behind all this interest? Supermaven has developed an in-house generative AI coding assistant called Babble that boasts of a one million-token context window and lower latency thanks to a new neural architecture built from scratch, according to a report by TechCrunch.

Quirk bot

Tech experts are putting together a final exam with the “hardest and broadest set of questions ever” in order to test the highly sophisticated AI systems being developed today. Dubbed as “Humanity’s Last Exam,” the questions for the test are being crowdsourced by the organisers. Though, there is one topic that won’t be part of the exam syllabus: weapons.

Moving on, a Bloomberg piece about the evolution of generative AI makes an obvious point that’s actually worth mentioning: the way we interact with AI chatbots has changed since 2022. How? AI models are becoming increasingly multimodal, which means that people are not only using them to generate text and images but also create songs and short videos in a few minutes.

Catching heat

LinkedIn drew flak this week for scraping user-generated content on the platform to train its AI models – something that the company had not mentioned in its privacy policy until recently. Since all LinkedIn users are signed up by default to permit AI training on their posts and personal data, here’s how you can opt out.

Microsoft-backed OpenAI also came under fire for reportedly threatening to ban users trying to probe the reasoning capabilities of its latest o1 model. They claimed that using certain words like ‘reasoning trace’ triggered a warning email from OpenAI that asked them, nicely, to stop. Another thing, OpenAI boss Sam Altman stepped back from the AI startup’s internal safety committee as it gets more independent.

The AI Rulebook

A UN body has made seven recommendations to address AI-related risks and gaps in governance. They include setting up a panel to provide impartial and reliable scientific knowledge about AI as well as creating an AI standards exchange, a global AI capacity development network, a global AI fund, a global AI data framework, and more.

Finally, California Governor Gavin Newsom signed eight new AI bills into law this week. Most of them look to tackle the flood of AI-generated deepfakes while others are focused on identifying AI-generated content through watermarks and setting standards on use of AI in Hollywood. Around 30 AI-related bills in California are still TBD, including the big one – SB 1047.