Search the archive

GOOGLE'S €25M AI SKILLS TRAINING INITIATIVE IN EUROPE

Google launched a skills training plan for Europe. It offers free courses and startup opportunity credits.

SGLANG

A structured generation language designed for LLMs.

GALILEO AI (PRODUCT)

Generate interface designs at lightning speeds.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

REKA HAS A NEW POWERFUL MULTI-MODAL MODEL

AI startup Reka has announced a 21B parameter model that was trained completely from scratch. It is natively multimodal and surpasses Llama 70B and Mixtral on a number of key benchmarks. Importantly, given where the founders of the company used to work, it also matches Gemini Pro. The model is available for use in Reka's playground.

APPLE RELEASES MGIE, A REVOLUTIONARY AI MODEL FOR INSTRUCTION-BASED IMAGE EDITING

Apple has launched MGIE, an open-source AI model for image editing that uses multimodal large language models to interpret text instructions and execute pixel-level edits.

DEEPMIND FRAMEWORK OFFERS BREAKTHROUGH IN LLMS’ REASONING

DeepMind and USC researchers have developed SELF-DISCOVER, a new framework that significantly improves LLM reasoning capabilities, outperforming traditional methods by up to 32% on tasks with models like GPT-4. The two-stage framework allows LLMs to form and utilize reasoning structures autonomously. This advancement represents a major step toward more sophisticated AI problem-solving abilities.

CONTROLLING REAL WORLD OBJECTS WITH FLUID IN RL

"Box o Flows" introduces a novel experimental system for testing reinforcement learning (RL) algorithms in dynamic real-world settings, addressing the challenge of simulating complex fluid dynamics for RL applications. It showcases the ability of model-free RL algorithms to generate complex behaviors from simple rewards and enhances data efficiency via offline RL and paves the way for broader RL applicability in complex systems.

WEBLINX CONVERSATIONAL WEB SCRAPING DATASET

Weblinx is a dataset of 100k web-based interactions in a conversational format. It was released to improve the study of language model guided web-based navigation.

3D AVATARS WITH IMPLICITDEEPFAKE1

This paper introduces ImplicitDeepfake1, a new approach that combines deepfake technology with Neural Radiance Fields (NeRFs) and Gaussian Splatting (GS) to create highly realistic 3D avatars.

during the breaks. Sign up or miss out (it’s 100% free)

GUIDING INSTRUCTION-BASED IMAGE EDITING VIA MULTIMODAL LARGE LANGUAGE MODELS

Apple released code that uses multimodal language models to improve human-provided natural language edits to images.

LAG-LLAMA: TOWARDS FOUNDATION MODELS FOR PROBABILISTIC TIME SERIES FORECASTING

Lag-Llama introduces a breakthrough in time series forecasting with a transformer-based model that excels in zero-shot and few-shot generalization, pre-trained on a broad spectrum of time series data. It surpasses existing deep learning methods in state-of-the-art performance through fine-tuning on minimal data, marking a significant advancement in foundation models for time series analysis.

LEARNING TO FLY

A repository for training end-to-end control policies using deep reinforcement learning. The training is done in simulation and is finished within seconds on a consumer-grade laptop. The trained policies generalize and can be deployed on real quadrotors.

MASSED MUDDLER INTELLIGENCE

The concept of Massed Muddler Intelligence (MMI) represents a shift from traditional monolithic AI scaling towards a model based on distributed, agent-based systems that learn and adapt in real time. Grounded in principles of embodiment, boundary intelligence, temporality, and personhood, MMI advocates for AI development that emphasizes scalable, interactive agents with a degree of autonomy and mutual governance, moving away from the current focus on accumulating larger datasets and computational resources.

AI COULD ACTUALLY HELP REBUILD THE MIDDLE CLASS

PACKING FOR LLM TRAINING

Packing in training models is a way to improve training efficiently by concatenating examples. If done incorrectly, contamination can occur between examples because the attention doesn’t know where to stop. The community has found that EOS is often sufficient, but there still can be problems. This repository provides an implementation in Hugging Face for common models to properly pack input data.

AI FORM ROAST (PRODUCT)

Free AI tool to audit and optimize online forms.

META WILL START LABELING AI-GENERATED IMAGES FROM OTHER COMPANIES

Meta plans to combat the spread of fake images on its platforms by detecting and labeling AI-generated content, even if produced by rival services.

AXOLOTL ADDS MPS SUPPORT

Axolotl is a tool used by many groups to fine-tune language models. This PR adds support for Mac M devices via Pytorch and MPS.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

LAION RELEASES ASSISTANT BUD-E

The Laion research group has released an open assistant that uses natural voice and extremely optimized language models that can run on a gaming laptop. The project aims to provide a low-resource assistant that is easy to install and powerful for personal use.

OPENAI HITS $2B REVENUE MILESTONE

Microsoft-backed OpenAI hit the $2 billion revenue milestone in December and it believes it can more than double this figure in 2025.

AI AGENTS AS A NEW DISTRIBUTION CHANNEL

AI agents are emerging as a new distribution channel, potentially leveling the playing field between startups and incumbents by making purchasing decisions on behalf of consumers. As this trend progresses, businesses must optimize their products for AI preferences rather than human ones, changing the traditional dynamics of product discovery, evaluation, and purchase. AI's advancements suggest a future where agent-driven commerce could transform how products are marketed and purchased.

THINKING ABOUT HIGH QUALITY HUMAN DATA

This post is about the role of humans in data generation. It also talks about collecting preference data, annotating, labeling, and more.

SKILL SET OPTIMIZATION

Transferring skills from one RL agent to another has proved challenging thus far. This work optimizes for a set of skills that can be used across environments. It shows promising generalization performance.

SELF-PLAY FINE-TUNING CODE

Allowing a language model to generate its own training data is a challenging but promising area of research. SPIN is a method that has shown promise. The code has been released but is reportedly challenging to work with.

RLX

A reinforcement learning framework based on MLX.

LLMWARE

Llmware is a unified framework for developing LLM-based application patterns including Retrieval Augmented Generation.

POINT TRANSFORMER V3 FOR EFFICIENCY AND SCALE

Point Transformer V3 (PTv3) is a simple and efficient model for processing 3D point clouds. It achieves faster processing speeds and better memory efficiency by focusing on scaling up and efficiency rather than on intricate design details.

AI WILL CHANGE ART FOREVER

AI will fundamentally change how we produce and consume art and ultimately the type of art we create. This article explores this point in more depth and highlights the companies driving the change.

BRAIN2MUSIC (7 MINUTE VIDEO)

Research talk from Google about reading music from a person’s brain.

AI PCS WILL MAKE UP NEARLY 60% OF TOTAL PC SHIPMENTS BY 2027

Shipments of AI PCs equipped with hardware and software for local generative AI tasks are predicted to hit nearly 50 million units this year, with potential growth to triple in a few years.

NEURALINK IS LIVE

According to Elon Musk, Neuralink has conducted its first FCA-approved surgical test of a brain implant. Techno accelerationism in action.

PHIDATA

A toolkit for building AI Assistants using function calling.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

GEMINI ADVANCED AND A MOBILE APP

Google launched a new service that allows you to finally interact with Gemini Ultra 1.0 - it will charge a monthly subscription for access to the model. There is also a companion mobile app.

1X ROBOTICS DEMONSTRATION

1X is a robotics company that has made strides in video-to-control models. The company has showcased its robot performing many tasks, all driven by neural networks that emit 10hz control signals from video input.

BOOSTING AI AGENTS' GAMEPLAY WITH TASK GUIDANCE

This paper explores the development of a generalist AI agent capable of understanding and following gameplay instructions, a step towards "read-to-play" abilities. The researchers enhance the agent's multitasking and generalization skills by integrating multimodal game instructions into a decision transformer.

TRANSFORMER-BASED DECISION TREE

MetaTree is a transformer-based decision tree algorithm. It learns from classical decision tree algorithms for better generalization capabilities.

FUNSEARCH: MAKING NEW DISCOVERIES IN MATHEMATICAL SCIENCES USING LARGE LANGUAGE MODELS

FunSearch, a new AI-powered method that combines Large Language Models with evaluative algorithms, has made verifiable discoveries in mathematical sciences, including solutions for the longstanding cap set problem and more efficient algorithms for the bin-packing problem. It introduces evolutionary approaches to generating and evaluating code that offers human-interpretable outputs, representing a significant leap in AI-driven scientific discovery.

Register now, it’s free!

requirements, trust of output, and debugging. Join 70+ speakers

from companies like Netflix, Google, Databricks, Perplexity AI, and Hugging Face. Attend hands-on workshops [10] for builders. Enjoy musical improv, games, and guided meditation during the breaks.

LOCAL RAG COOKBOOK

You can build a sophisticated and powerful RAG system that runs on your hardware using Ollama, pgvector, and local data.

HUGGINGFACE LIGHTEVAL LIBRARY

HuggingFace released a lightweight evaluation library for language model training based on HELM and Eluther AI’s evaluation harness.

A VISION-LANGUAGE MODEL WITH ENHANCED VISUAL REASONING

CogCoM, a new general vision-language model, incorporates a unique Chain of Manipulations mechanism. This allows it to handle multi-turn visual reasoning by actively adjusting input images.