Figma AI

Figma AI introduces a suite of AI-powered features designed to enhance designers’ creativity and efficiency. These features include Visual Search for finding and reusing designs, AI-enhanced Asset Search for locating components, and AI tools for text editing, image generation, and prototyping. The AI features, currently in beta and free through 2024, aim to streamline tedious tasks and foster creativity by automating busywork and providing realistic content.

Read more

AI Revolutionized Protein Science, but Didn't End It

AlphaFold2 by Google DeepMind revolutionized protein science by achieving over 90% accuracy in predicting 3D structures from amino acid sequences, significantly outperforming its competitors. This breakthrough has transformed the field, inspiring new AI algorithms and biotech companies, and catalyzed protein structure research and drug development. However, AlphaFold2 does not simulate protein dynamics over time or account for their contextual environment within cells. Its successor, AlphaFold3, advances multi-molecule modeling but maintains some existing limitations. Overall, while AlphaFold2 has solved the structure prediction problem, deeper biological questions and complete mechanistic insights remain for future exploration.

Read more

Researchers Prove Rabbit AI Breach by Sending Email as Admin

Hackers reverse-engineering the Rabbit R1 AI assistant found exposed API keys that allowed access to all device responses and enabled misuse of various services. They demonstrated this breach by sending emails using internal admin addresses. The device, extensively criticized for poor design, runs numerous off-the-shelf APIs like ElevenLabs' text-to-speech. The discovery was made by Rabbitude, a community that focuses on jailbreaking and improving the Rabbit R1.

Read more

Insights into LLM Long-Context Failures: When Transformers Know but Don't Tell

LLMs show positional bias, having trouble with information from the middle or end of long contexts. By examining hidden representations, it's clear that although LLMs encode target information positions, they often don't use this effectively in responses. This highlights a "know but don't tell" issue. The study also links extraction time to final accuracy, shedding light on transformer model mechanics.

ArXiv

Towards Robust Detection of AI-Generated Videos

Current detectors excel at spotting GAN-generated Deepfakes but falter with diffusion-generated videos created using tools like SORA and Runway Gen-2. A new framework using a CNN + LSTM architecture was proposed to better detect such videos by handling temporal features between frames. A benchmark dataset for diffusion-generated videos was also introduced. Evaluation showed the method achieved 93.7% accuracy for in-domain videos and increased out-domain accuracy by up to 16 points.

ArXiv

Indications of superconductivities in blend of variant apatite and covellite

Heavily doping sulfur into an apatite framework results in a new blend of variant apatite and covellite (CuS). Magnetic measurements reveal significant diamagnetism around 260 K and a sharp drop below 30 K, suggesting two distinct superconducting phases. The upper critical magnetic field exceeds 1000 Oe at 250 K. Electric measurements show non-linear current-voltage curves and a zero-resistance effect, with a critical current around 50 μA at 140 K. These findings indicate that variant apatite and covellite each induce superconducting phases at near-room and low temperatures, respectively.

ArXiv

Evaluating Reasoning by LLMs Using the New York Times Connections Word Game

The study evaluates the performance of LLMs, specifically GPT-4o, on 200 New York Times Connections games compared to both expert and novice human players. GPT-4o fully solves only 8% of the games, underperforming compared to both human groups, with experts far exceeding its results. Analysis identifies associative, encyclopedic, and linguistic knowledge as major challenge areas for LLMs. This positions the Connections game as a valuable benchmark for assessing abstract reasoning in AI and humans.

ArXiv

R2R V2 – A open source RAG engine with prod features

R2R is a state-of-the-art Retrieval-Augmented Generation (RAG) system that bridges local LLM experimentation with production-ready scalability. It supports multimodal file ingestion, hybrid search combining semantic and keyword techniques, automatic knowledge graph extraction, and a comprehensive API for client-server interactions. R2R features rich document and user management, is highly configurable, and can be extended with a builder + factory pattern. Additionally, it includes an open-source React+Next.js dashboard for intuitive interaction.

Repo

Code to run Gemini (Nano) locally on desktop/Chrome

This project lets you run Google's Nano LLM locally on your desktop via Chrome Canary. After setting up the necessary configuration flags in Chrome Canary, you can test out the model by opening an HTML file from the repo and entering prompts. It operates without internet connectivity and provides local responses, though they may be slow and unreliable.

Repo

Agentpanel – Universal LLM API Written in Rust

Agent-panel is a Rust-based control and observability platform for managing LLM/AI agents, particularly focusing on optimizing multi-agent systems' performance. It introduces an AI gateway that provides seamless access to over 100 LLMs across 20+ platforms, using an API-first approach for easy integration. The platform supports various LLMs, including OpenAI, Gemini, Claude, Mistral, Cohere, Perplexity, and more.

Repo

WASM AI: run LLMs natively in the browser

WASM AI is a quickstart template for running 7B LLMs natively in the browser. It includes compiled models like Dolphin 2.2.1 and OpenHermes-2.5, allowing for tasks such as text-to-SQL search, creative writing, and NLP with high privacy and no cloud costs. The project supports switching between local and cloud models easily, uses web workers for a smooth UI, and integrates Whisper for local voice transcription. It is designed for easy iteration and distribution as part of larger applications.

Repo