TLDR AI 2025-05-14

TikTok has launched AI Alive, a new feature that animates static images within Stories. It uses intelligent editing tools to turn photos into short-form videos with dynamic effects.

Amazon's audiobook service, Audible, is partnering with publishers to expand its catalog with AI-narrated audiobooks, offering over 100 voices in English, Spanish, French, and Italian. Spotify made similar updates earlier in the year in partnership with ElevenLabs.

WizardLM, a former Microsoft AI group, has joined Tencent's AI division, Hunyuan. It quickly released a new AI model, Hunyuan-TurboS 0416, claiming it outperforms Google's offerings. WizardLM will continue to develop AI models at Tencent.

Using a classifier that detects distinctive AI word choices (like "delve") in academic papers, researchers found ChatGPT usage was higher in countries where it's prohibited by OpenAI. 22% of Chinese academic preprints included AI-generated content versus 11% from legal-access countries by August 2023, suggesting geographic restrictions can be easily circumnavigated. While ChatGPT use correlated with increased views and downloads, it showed no impact on citations or journal placements.

Duolingo announced a major shift to become an AI-first company, integrating AI tools across products and workflows. The company's new tenets include starting every task with AI, dedicating time for learning, careful experimentation, and preserving technical excellence. Leadership emphasized that while AI raises expectations, it's about working smarter, not harder.

While chat interfaces provide an accessible medium to LLMs for first-time users, they're a design dead-end that forces users to accommodate to the model rather than the reverse. The next generation of assistants will use more adaptive UIs and proactively communicate their capabilities.

Japanese AI company Sakana developed a new type of AI model that allows individual neurons to remember their previous actions and coordinate their activity based on timing patterns. While it underperforms compared to traditional models, it does allow researchers to understand how the model reasons, and similar to recent reasoning models like o3, its answers improve when given more time to “think”.

EAR introduces a continuous visual autoregressive generation method that avoids quantization by leveraging strictly proper scoring rules like the energy score. This enables direct generation in continuous data spaces without probabilistic modeling.

UCGM provides a shared framework for training and sampling across multi-step and few-step continuous generative models.

Microsoft plans to host Elon Musk's Grok AI on Azure AI Foundry despite potential tensions with OpenAI. This move signals Microsoft's intent to diversify AI partnerships and strengthen its position as a leading AI platform provider. The plans to host Grok could be announced at Microsoft's Build developer conference on May 19.

Hugging Face has released a new Whisper endpoint that delivers up to 8x faster transcription. The endpoint enables one-click deployment of optimized, cost-effective models for speech tasks through its Inference Endpoints.

Meta's first LlamaCon Hackathon hosted 238 participants who built projects using the Llama 4 suite of tools. The winners, selected from 44 submissions based on innovation and technical execution, have now been announced.

Over 250 CEOs, including Microsoft's Satya Nadella and Uber's Dara Khosrowshahi, urged integrating AI and computer science into K-12 education to maintain U.S. competitiveness.

Ai2's Olmo 2 1B is a 1-billion-parameter model that outperforms similar-sized models from major tech companies on key benchmarks.

YC's Summer 2025 investment themes will focus on AI agents that replace or augment traditional roles across industries, including voice assistants, healthcare automation, personalized tutoring, and tools that enable workers to build their own custom agents.