Donald Trump Ends AI Security Measures
Donald Trump, on his first day in office, issued 40 executive orders, reversing many of his predecessor’s policies. One of the orders canceled was the “Executive Order on AI Security Measures.” This order, crafted by Joe Biden with AI companies, aimed to regulate AI for safety. The original order required federal agencies to hold AI companies accountable, ensuring they invest in “Red Teaming” to assess potential threats to national security and the economy. The Commerce Department was to work with AI providers to develop mechanisms for marking AI-generated content, like digital watermarks. Homeland Security was tasked to prevent AI misuse in creating chemical, biological, or nuclear weapons.
These measures were still being implemented. OpenAI emphasized the need for unified federal AI laws, as about 700 conflicting regulations exist in the US. The order’s provision to ease work visas for AI experts was also removed by Trump, which might upset tech companies since many of their AI staff are foreign, including Germans. This expertise was meant to be centralized in Silicon Valley.
Questions arise about the future: Will Trump allow exceptions? Will a comprehensive AI law be enacted? Trump will closely engage with CEOs of major AI firms. At his inauguration, Sam Altman, Mark Zuckerberg, Sundar Pichai, and Elon Musk were present. Their influence is expected, but its extent is uncertain.
More US Teens Use ChatGPT for Homework
More US teenagers are using “ChatGPT for homework.” A survey of nearly 1,400 teens aged 13 to 17 and their parents showed usage doubled from 13% to 26% within a year. Older students use the chatbot more than younger ones. Acceptance varies: 54% support ChatGPT for research, but only 29% for math and 18% for essays.
The survey’s accuracy, considering a parent’s presence, is uncertain. ChatGPT’s awareness among teens rose to 79%. Differences appear across income groups: Wealthier households have more teens familiar with ChatGPT than poorer ones. However, the latter group had the largest increase, rising by 26 percentage points since 2023. Familiarity influences usage: Those knowledgeable about the tool use it more for school.
Meta Develops CapCut Alternative
Meta announced a new mobile video editing app, “Edits,” first for iOS and later for Android. Designed for social network videos, this app responds to CapCut’s absence due to ByteDance app bans in the US. Instagram’s Adam Mosseri said Edits is not just an app but a suite of “creative tools.”
Features include AI animations, background replacement via greenscreen, and sound enhancement. Crucial for social videos: Subtitles can be auto-generated and customized. Exported videos won’t have watermarks, allowing easy use on other platforms. Interested users can pre-register for download in the Apple App Store. Edits will be available from March 13.
DeepSeek-R1 Matches OpenAI’s o1 in Reasoning Benchmarks
Chinese AI startup DeepSeek introduced new reasoning models “DeepSeek-R1 and DeepSeek-R1-Zero.” Notably, they achieved performances comparable to OpenAI’s o1 model in benchmarks.
A highlight is DeepSeek-R1-Zero’s development. Unlike typical language models refined through Supervised Fine-Tuning, DeepSeek-R1-Zero was trained solely with Reinforcement Learning, learning without human feedback using a reward system. The team used a rule-based system, evaluating accuracy (e.g., math solutions or programming code tests) and response format (e.g., adherence to tags like “think” and “/think”). Researchers noted the model independently developed skills, like generating long thought chains and reconsidering solutions when problems required more “thinking time.” This approach mirrors OpenAI’s o1 model principles.
The second model, DeepSeek-R1, combined Reinforcement Learning with classic Fine-Tuning, aligning responses better with human expectations. In fields like math, programming, general knowledge, and writing, it matched OpenAI’s o1 model’s performance from December.
DeepSeek also presented six smaller open-source models trained with data generated by DeepSeek-R1. This knowledge enables smaller models to partially reach OpenAI’s o1-mini performance. The open-source versions are publicly available.
Despite progress, challenges remain: R1 and R1-Zero unintentionally switch languages and currently lack Function Calling, advanced dialogues, or JSON outputs. DeepSeek plans to address these weaknesses, but model practicality depends on the use case.
R1-Zero particularly demonstrates how powerful a relatively simple and resource-efficient approach like Reinforcement Learning can be—a notable advancement in AI research.
Krea AI’s New Image Editing Tool
AI startup Krea unveiled a new feature for its image editor, allowing “images to be converted into simple 3D objects.” The resulting 3D model can be repositioned and rotated in a scene. From this template and a matching text prompt, Krea generates an image. A demo shows a user uploading a sofa photo, turning it into a 3D model, and placing it in a generated apartment image.
The tool’s preview window displays the generated image with the sofa at the user-specified location almost in real-time. Krea offers more tools, like real-time styles, light and color consistency, ultra-slow motion, or image animation. However, Krea doesn’t explain the technical workings of the new 3D feature.
Character AI Launches Games
AI company Character AI released two “games for its chatbots.” In Speakeasy, players must get a chatbot to say a target word without using five forbidden words, similar to the classic Taboo game. War of Words is a language duel where a chatbot decides the winner. Both games are exclusive to Character AI subscribers, aiming to increase user engagement—a goal for all services.
Character AI’s chatbots cater to young people, featuring “Gamer Boy” for gaming discussions and various anime characters. Users can create their own chatbots if none appeal. The new games work with all existing characters. To protect teens, the service includes a 60-minute chat timeout, though it’s just a suggestion.
Duolingo’s Videocalls Now on Android
Duolingo’s “Videocall” feature, allowing users to call the teenage character “Lilli,” is now available for Android users with a “Duolingo Max” or “Duolingo Super” subscription. Duolingo promises improved character reactions, emotions, and new features like dialogue summaries for later review.
Duolingo is testing occasional free use of the videocall feature. Lilli might call if users don’t contact her for a while, a spokesperson explained.