Writer RAG tool: build production-ready RAG apps in minutes

• Writer RAG Tool: build production-ready RAG apps in minutes with simple API calls.

• Knowledge Graph integration for intelligent data retrieval and AI-powered interactions.

• Streamlined full-stack platform eliminates complex setups for scalable, accurate AI workflows.

Learn more about our production ready RAG tooling here.

Today:

• Google DeepMind unveils AlphaFold3's code

• Qwen2.5-Coder Rivals GPT-4o Capabilities

• Google Introduces GenSelfie for Stylized Portraits

• FrontierMath Reveals AI’s Math Limits

• ARI Backs Musk for AI Leadership

Google DeepMind unveils AlphaFold3's code

Google DeepMind has released the full code for AlphaFold3, its advanced AI model for protein prediction, fulfilling a promise made after facing criticism. Initially, DeepMind shared only limited access via an online portal, causing concern among researchers who sought full transparency and usability. 

Now, the complete model is available on GitHub for noncommercial use, enabling scientists to explore protein structures and molecular interactions for various research applications, including drug discovery. Despite some lingering dissatisfaction over the six-month delay, scientists are eager to test and build upon AlphaFold3’s advanced capabilities in understanding biological structures.

Qwen2.5-Coder Rivals GPT-4o Capabilities

Qwen2.5-Coder, an open-source AI coding model series by Qwen Team, introduces models across six sizes, including the flagship 32B model, which rivals GPT-4o in coding capabilities. Its diverse capabilities cover 40+ programming languages and excel in tasks like code generation, repair, and reasoning, achieving state-of-the-art (SOTA) results on benchmarks such as EvalPlus and LiveCodeBench. 

This release caters to various developers, from small to large-scale needs, supporting scenarios like code assistants and visual Artifacts. Licensed mostly under Apache 2.0, Qwen2.5-Coder aims to empower the coding community with robust, flexible, and scalable tools.

Google Introduces GenSelfie for Stylized Portraits

Google’s new zero-shot model, GenSelfie, allows users to transform selfies into artistic portraits in a variety of styles while preserving facial likeness. By using image and control adapters, the model captures unique facial features and expressions, ensuring that outputs retain resemblance to the original selfie. 

Users can prompt styles like 3D cartoon, watercolor, anime, and pencil sketch and modify expressions, such as smiling or frowning. Accessible on Imagen in Vertex AI, this innovation offers a personalized, stylized image generation experience. Google plans to extend the model’s capabilities to include body pose personalization, opening up new possibilities for artistic self-expression.

FrontierMath Reveals AI’s Math Limits

Epoch AI’s new FrontierMath benchmark reveals significant challenges for AI in advanced mathematics, where models like GPT-4o and Gemini 1.5 Pro can solve less than 2% of complex problems. Designed in collaboration with top mathematicians, FrontierMath includes unique, research-level math problems requiring deep reasoning and creativity, unlike standard benchmarks where AI performs well due to data familiarity. 

This test highlights AI's current limitations in multi-step logical reasoning and precise problem-solving. While FrontierMath’s hurdles illustrate AI’s limits, its future use could track advancements toward more sophisticated, human-like intelligence in mathematical reasoning.

ARI Backs Musk for AI Leadership

A nonprofit group, Americans for Responsible Innovation (ARI), is pushing for Elon Musk to become President-elect Donald Trump’s special adviser on AI. ARI argues Musk's AI expertise and safety advocacy could help the U.S. lead responsibly in AI development. The petition emphasizes Musk’s history with AI safety, including his cofounding of OpenAI, calls for a moratorium on advanced AI development, and support for AI safety legislation. 

While Musk hasn’t detailed specific AI policies, ARI believes his evolving stance could positively shape AI policy with the right checks in place. ARI aims to gather 10,000 signatures in support.

🧠RESEARCH

LLM2CLIP: Powerful Language Model Unlock Richer Visual Representation

LLM2CLIP, a method to enhance the CLIP model's visual representation by integrating advanced language models (LLMs). By using LLMs’ extensive text comprehension, LLM2CLIP improves CLIP’s ability to interpret complex captions, resulting in better performance on cross-modal tasks and a more robust visual-textual representation.

StdGEN: Semantic-Decomposed 3D Character Generation from Single Images

StdGEN is a new tool for creating high-quality 3D characters from single images, useful in VR, gaming, and films. It quickly generates detailed, decomposed characters (body, clothes, hair) in just three minutes. Using an innovative transformer-based model, StdGEN produces customizable 3D characters with superior geometry and texture compared to existing methods.

RetrieveGPT: Merging Prompts and Mathematical Models for Enhanced Code-Mixed Information Retrieval

RetrieveGPT addresses the challenge of finding relevant information in code-mixed conversations, particularly Roman-transliterated Bengali mixed with English. By combining GPT-3.5 Turbo prompts with a mathematical model, RetrieveGPT effectively identifies relevant responses in multilingual online discussions, improving information retrieval in complex, mixed-language digital text environments.

SG-I2V: Self-Guided Trajectory Control in Image-to-Video Generation

SG-I2V is a new approach for controllable image-to-video generation, enabling zero-shot adjustments in video elements like object motion or camera angles without fine-tuning or extra data. By using a pre-trained diffusion model, SG-I2V achieves high-quality, flexible control over video generation, matching the performance of supervised models without additional resources.

DELIFT: Data Efficient Language model Instruction Fine Tuning

DELIFT is a new algorithm that streamlines fine-tuning of large language models by optimizing data selection. It efficiently enhances model training across instruction, task-specific, and continual tuning stages. Using a utility metric to prioritize informative data, DELIFT reduces required data by up to 70%, saving resources while maintaining high performance across tasks.

🛠️TOP TOOLS

Anotherwrapper - Have AI build fully functional AI wrappers, incl. backend API routes.

Texta - AI Generates High-Quality Articles That Convert Visitors to Customers

FlowScraper - Make web automation easy, efficient, and flexible with flow builder and customizable AI actions.

Zappit AI - AI agents specifically trained to fix technical errors and to generate high-ranking SEO optimized content

Kombai - A new model trained to code email & web designs like humans

📲SOCIAL MEDIA

🗞️MORE NEWS

• Social platform X is testing free access to its AI chatbot Grok, previously limited to paid users. Free users in select regions can now try Grok with limited queries, aiming to expand feedback and user engagement.

• AI-driven startup 11x, which creates bots to assist sales teams with routine tasks, has raised $50 million in a funding round led by Andreessen Horowitz, bringing its valuation to $320 million. This round, which also saw participation from Hubspot Ventures, follows a $24 million investment led by Benchmark in September. 

• Google's new AI flood forecasting model improves global accuracy and lead time, predicting floods up to seven days in advance. Expanded to 150 countries, it aids disaster management, especially in data-scarce regions, fostering research and safety.

• The Vatican, in collaboration with Microsoft and Iconem, launched AI-enabled virtual tours of St. Peter’s Basilica, enhancing visitor experiences and accessibility ahead of the 2025 Jubilee. This initiative aims to preserve and modernize cultural heritage.