June 2025, Swapneel on Hacking Algorithms, DW Global Media Forum, and Fairify

Listen to the episode on Apple Podcasts or Spotify to hear how users can take control of online personalization to get relevant content, rather than letting algorithms dictate feeds.

Vital discussions on media development in the era of AI, media organizations taking ownership of AI technology, and sustainable journalism business models.

Implements Counterfactual Sentence Testing (CST) and Sentence Encoder Association Test (SEAT) to let engineers assess NLP models for custom bias cases.

Zetland's Good Tape has transcribed 10M+ audio files, driving $3M in recurring revenue with its secure, AI-powered multilingual tool.

Seedance 1.0 Lite is ByteDance's advanced text-to-video AI model, available through fal.ai, designed for quick and efficient video creation. It transforms text prompts into high-quality 720p videos, costing just $0.18 for a 5-second clip. Ideal for social media and rapid prototyping, Seedance also offers an Image-to-Video feature and a Pro version for premium 1080p outputs with advanced camera controls.

Search engines use explicit queries to retrieve relevant items, while recommendation systems rely on user behavior to suggest content. Traditionally, these systems use separate models and pipelines, leading to high maintenance costs and complexity. Joint modeling treats both as top-K retrieval problems with different context sources, enabling a unified approach.

Instructor, a Python library for structured LLM outputs, has achieved remarkable success, boasting over 3 million monthly downloads and 11,000 GitHub stars. This impressive organic growth comes with no external money raised. The project's popularity has translated into a strong $1.4 million in top-line revenue for 567stud.io.

AI data centers are driving unprecedented electricity demand, projected to reach 68 gigawatts globally by 2027, nearly matching California's 2022 total capacity. This rapid growth strains U.S. power grids and may force companies to build data centers abroad. Large AI training runs could require up to 8 gigawatts at a single site by 2030 if current trends continue.

Framework for Artificial Intelligence Diffusion was introduced in January 2025 to manage global AI technology diffusion, maintain U.S. leadership, and address security risks. The framework used control over AI chip compute power as a strategic tool and introduced new export controls and licensing regimes for advanced AI chips.

China is expected to match U.S. AI model capabilities in 2025, raising concerns about America's technological lead. Despite this, the U.S. retains a significant advantage in total compute capacity and access to advanced AI chips. A major export control breach in 2024 allowed Huawei to obtain millions of advanced chips, boosting China's AI hardware resources.

DeepSeek has released AI models that match GPT-4 and OpenAI o1 performance while using significantly less computing power and costing much less. Their models are openly downloadable and licensed under MIT, making advanced AI widely accessible. This shift increases the importance of computing resources for scaling and deployment.

Relying on ChatGPT for essay writing leads to reduced brain engagement, weaker neural connectivity, and lower memory recall. Students using ChatGPT report less deep processing, mental effort, and sustained attention compared to those writing without AI. Switching back to unaided writing after using ChatGPT does not fully restore cognitive engagement, indicating the accumulation of cognitive debt.

A large-scale experiment with nearly 5,000 U.S. participants compared trust in generative AI (GenAI) search results to traditional search. Users generally trusted GenAI search less than traditional search, especially on sensitive topics like inflation, vaccines, and climate change. Features such as reference links and explanations increased trust and willingness to share GenAI results.

This research analyzes the migration of 300,000 academics from Twitter/X to Bluesky between 2023 and early 2025. It reveals that 18% of scholars in the sample transitioned, driven powerfully by information sources and peer influence, rather than traditional academic metrics. While simple contagion accounts for two-thirds of exits, shock-driven bursts contribute 16%.

Researchers conducted a controlled experiment with nearly 700 U.S. participants to assess the effects of generative AI integration on social media. Participants were exposed to four different AI treatments and a control condition. The study found that generative AI tools significantly influenced user behavior and content creation on the platform.

Generative AI tools are increasingly used in coding to enhance productivity by automating repetitive tasks and improving code consistency. These tools can generate code fragments, enforce coding standards, and aid in software testing. However, their use raises concerns about widening skill and income gaps and potentially reducing developers' understanding of generated code.

The Schwurbelarchiv is a large-scale German-language Telegram dataset focused on conspiracy theories, misinformation, and political extremism. It includes over 6,000 groups and channels, 40 million text messages, and 3 million transcribed audio files from 2016 to 2022. Data was collected using snowball sampling and includes multimodal content for comprehensive analysis.

TGDataset is the largest publicly available collection of Telegram channels, containing 120,979 channels and over 400 million messages. The dataset enables large-scale analysis of information flows, language communities, and the spread of both legitimate and problematic content, including conspiracy theories and extremism.

Updesh is a new, large-scale synthetic dataset designed by Microsoft Research India to enhance the post-training of Large Language Models (LLMs) for 13 Indian languages, including Hindi, Tamil, and Bengali. It addresses the critical gap in culturally-grounded, high-quality instruction-tuning data for Indic languages, crucial for Small Language Models (SLMs) in India.