Search the archive

TXTAI

Txtai is an all-in-one embeddings database for semantic search, LLM orchestration, and language model workflows.

A UNIFIED BENCHMARK FOR TEXT DETECTION AND RECOGNITION

OpenOCR is a unified training and evaluation benchmark for scene text detection and recognition algorithms.

CONVERTING CODEBASES WITH LLMS

Mantle utilized a Gemini 1.0 Pro LLM with a one million token window to convert a prototype project into a production-ready codebase, reducing the workload by handling boilerplate code and repetitive patterns. This approach, which leveraged extensive context and iterative code generation, saved months of developer time by allowing the team to focus on refining the crucial 20% of the project.

AI PAID FOR BY ADS

OpenAI's new gpt-4o mini model offers extremely low costs, enabling AI-generated content supported by ads to become financially viable. By generating dynamic blog posts for $0.00051525 each and earning approximately $0.0026 per ad impression, publishers can achieve a net profit of $0.002 per page view. This may cause a potential shift towards AI-generated content in response to user queries.

A MULTI-HEADED MODEL FOR OBJECT DETECTION

The new CerberusDet framework offers a flexible solution for object detection by combining multiple task heads within a single model based on the YOLO architecture.

AI CODING INTERN

This extremely simple CLI, which uses Claude Sonnet 3.5, can be used to suggest code edits to improve an existing code base.

THE WORLD NEEDS MORE MENTORS. AI CAN HELP US BRIDGE THE GAP

People around the world want to learn new skills and master new technologies.

ACADEMIC AUTHORS 'SHOCKED' AFTER TAYLOR & FRANCIS SELLS ACCESS TO THEIR RESEARCH TO MICROSOFT AI

Taylor & Francis sold access to its authors' research to Microsoft for AI use without informing the authors or offering opt-out options, raising concerns over authors' rights.

✅ Access to technical training and hands-on labs

✅ Meet with dedicated startup experts

✅ Connect with startup communities

✅ Access Google-wide discounts

✅ Up to $350K in cloud credits over 2 years

Applying is easy and free. Start here

MISTRAL AND NVIDIA'S NEW MODEL

Mistral Nemo 12B is a multilingual model trained with a new tokenizer that shows powerful multilingual and English performance. It supports 128k contexts too.

GPT-4O MINI

GPT-4o Mini is a new, smaller, model from OpenAI aimed to replace GPT-3.5. It scores 82 on MMLU, which is reasonable for cheap models.

COHERE AND FUJITSU STRATEGIC PARTNERSHIP

Cohere and Fujitsu announced a strategic partnership to develop and provide enterprise AI services with industry-leading Japanese language capabilities. These services will be created with security and data privacy at the core, offering private cloud deployments to serve organizations in highly-regulated industries including financial institutions, the public sector, and R&D units.

LEAPFROGGING TRADITIONAL VECTOR-BASED RAG WITH LANGUAGE MAPS

Retrieval is a huge part of building a chat application over data. However, systems are often sensitive to the format of things being retrieved. Building a language map (e.g., Wikipedia style entry) of the content, and using that for retrieval dramatically improves chat-based performance. Mutable AI does this for code base question answering.

PATCH-LEVEL TRAINING TECHNIQUE ENHANCES EFFICIENCY OF LLMS

Researchers propose patch-level training for large language models to improve training efficiency.

REMOVING INAPPROPRIATE CONTENT FROM DIFFUSION MODELS

Reliable and Efficient Concept Erasure (RECE) is a novel method that erases inappropriate content from diffusion models in just 3 seconds without needing extra fine-tuning.

more, to the highest bidders - every day. The good news? Incogni

confronts brokers on your behalf, preventing bad actors from accessing your data.

TLDR60.

LLM2SH

LLM2sh is a command-line utility that leverages LLMs to translate plain-language requests into shell commands.

SYMBOLIC MUSIC TASKS WITH GRAPH NEURAL NETWORKS

GraphMuse is a new framework designed to enhance the use of Graph Neural Networks (GNNs) in symbolic music tasks.

UNIVERSAL MULTIMODAL EMBEDDINGS WITH LLMS

E5-V is a new framework that adapts Multimodal Large Language Models (MLLMs) to create universal multimodal embeddings. It bridges the gap between different input types using prompts, achieving impressive performance in multimodal tasks without fine-tuning.

STRATEGIZING YOUR PREPARATION FOR MACHINE LEARNING INTERVIEWS

Machine Learning interviews can be challenging. Understanding the spectrum of ML roles and tailoring your preparation to specific job responsibilities and specializations can significantly improve your chances. Focus on mastering fundamentals, researching company-specific technologies, and consistently tracking your progress to confidently tackle interviews.

UNCENSOR ANY LLM WITH ABLITERATION

Llama models are heavily censored for safety, limiting flexibility. The "abliteration" technique uncensors them by identifying and removing the refusal mechanism, allowing models to respond to all prompts without retraining.

QUESTION-ANSWERING FOR SCIENTIFIC PAPERS

SPIQA is a QA dataset designed to help readers quickly find answers in scientific research articles by interpreting complex figures and tables.

OPENAI AND BROADCOM HELD DISCUSSIONS ABOUT PRODUCING AN AI CHIP

Broadcom has had discussions with OpenAI about building a new artificial intelligence server chip.

OVERCOMING THE LIMITS OF CURRENT LLMS

Large language models face challenges such as hallucinations, lack of confidence estimates, and lack of citations.

FLOW STUDIO (PRODUCT LAUNCH)

Flow Studio generates fully-produced, film-quality 3-minute-long videos with engaging plots, consistent characters, and automatically matched background music and sound effects.

Start building with AI Studio

META WON'T OFFER FUTURE MULTIMODAL AI MODELS IN THE EU

Meta will withhold future multimodal AI models from EU customers due to regulatory uncertainties, although a text-only version of Llama 3 will be available.

ANTHROPIC IS LAUNCHING A $100M AI STARTUP FUND

Anthropic, in partnership with Menlo Ventures, is launching a $100 million Anthology Fund to support AI startups. The fund will provide at least $100k per startup, access to Anthropic's AI models, and various other perks such as networking opportunities and workspace access.

PROVER-VERIFIER IMPROVES LLM LEGIBILITY

OpenAI trained a strong model to output more legible text for a weak model to grade accurately and found that this led to general purpose improvements in overall legibility.

DEPTH ESTIMATION WITH STEREO MATCHING

Researchers propose a new method for video stereo matching that enhances depth estimation by ensuring temporal consistency.

LIDAR SEMANTIC SEGMENTATION

SFPNet is a new method designed to generalize across different types of LiDAR technologies. By using sparse focal point modulation instead of traditional window-attention, SFPNet extracts and dynamically aggregates multi-level contexts.

PRAISONAI

Praison AI, which leverages other agent frameworks, is a low-code, centralized framework designed to simplify the creation and orchestration of multi-agent systems for various LLM applications, featuring ease of use, customization, and human-agent interaction.

VIDEO OBJECT SEGMENTATION WITH WORLD KNOWLEDGE

Reasoning Video Object Segmentation (ReasonVOS) is a new task that uses implicit text queries to generate segmentation masks. It requires complex reasoning and world knowledge.

ENHANCING CLASS LEARNING WITHOUT FORGETTING

This project introduces a background-class separation framework to improve Class-Incremental Semantic Segmentation (CISS).

MY TRIP TO THE FRONTIER OF AI EDUCATION

First Avenue Elementary School in Newark is integrating AI tools for education using Khanmigo, an AI-powered tutor and teacher assistant developed by Khan Academy. Classroom teachers use the tool to personalize learning and save time on tasks. There is ongoing work to improve inclusivity and responsiveness. This initiative, supported by the Gates Foundation, aims to help level the educational playing field by enhancing teacher-student engagement.

AI-DRIVEN BEHAVIOR CHANGE COULD TRANSFORM HEALTH CARE

OpenAI and Thrive Global are funding Thrive AI Health to develop a personalized AI health coach catering to daily behaviors essential to health, like sleep and diet. The mobile app and enterprise product will be powered by AI's hyper-personalization capabilities, integrating peer-reviewed science and individual data. The initiative aims to democratize healthy behavior change, address chronic disease management, and demonstrate efficient AI integration into healthcare with strong privacy measures.

GRAPHRAG ANALYSIS, PART 1: HOW INDEXING ELEVATES KNOWLEDGE GRAPH PERFORMANCE IN RAG

Analysis of Microsoft's GraphRAG paper reveals that knowledge graphs like Neo4j may not significantly outperform FAISS in context retrieval for RAG applications. While Neo4j without its indexing can achieve a higher answer relevancy, the marginal gains may not justify the investment given ROI constraints. However, the use of Neo4j's indexing shows a noticeable improvement in answer faithfulness, which reduces the risk of fabricated information.

Read the full report by Composable

ANTHROPIC DOUBLES OUTPUT TOKEN LIMIT

Anthropic has doubled the max output token limit for Claude 3.5 Sonnet from 4096 to 8192 in the Anthropic API.

GOOGLE VIDS (PRODUCT LAUNCH)

Google Vids is an AI-powered video creation app designed for work and deeply integrated with the Workspace suite.

AIXPLAIN SECURES $6.5M PRE-SERIES A TO UNIVERSALIZE AI AGENT DEVELOPMENT

aiXplain has announced a $6.5M pre-series A funding from Wa'ed Ventures, the venture arm of Saudi Aramco (a global top 10 company by market cap).

❌ Delete your personal data today