Search the archive

AI_IRL London.

DEEPTHOUGHT 8B REASONING MODEL (HUGGING FACE HUB)

Ruliad has released a small reasoning model that takes extra time to think through problems.

GUI AGENTS

A collection of research papers and projects on large language model-brained GUI agents.

SAM ALTMAN SAYS ARTIFICIAL GENERAL INTELLIGENCE IS ON THE HORIZON

Sam Altman, CEO of OpenAI, stated at the DealBook Summit that Artificial General Intelligence might impact everyday life less significantly than expected.

agent

Chances are, you've already spoken to it without even noticing... It can:

version here...

OPENAI 12 DAYS OF SHIPMAS

OpenAI will be having 12 live streams over the next 12 days to ship new product and model features.

META'S NUCLEAR ENERGY PLANS

Meta announced plans to collaborate with nuclear energy developers through a new request for proposals. The initiative aims to add 1-4 gigawatts of nuclear capacity in the U.S. to support Meta's AI innovation and sustainability efforts.

AWS REINVENT TOP ANNOUNCEMENTS

AWS announced various additions and tools including bringing prompt routing and caching to its Bedrock LLM service at AWS re:Invent 2024.

GENIE 2: A LARGE-SCALE FOUNDATION WORLD MODEL

Genie 2 is a large scale latent diffusion model for world generation. It works with character control as input and is also classifier free. The outputs are stunning and maintain control and consistency over time.

MOTION PROMPTING

Motion Prompting is a method that trains a video generation model with novel input types. Instead of just text or an image, it uses text, first image frame, and a tracking field of pixels. This then allows for novel control at inference time where it can provide a new pixel field (e.g., for an object moving in a new direction) and generate a video based on that. The method is extremely compelling, but not open sourced.

A COMPREHENSIVE SURVEY FOR REMOTE SENSING VLMS

This repository offers a comprehensive survey for remote sensing with the integration of Vision-Language Models (VLMs).

FREE PROCESS REWARDS WITHOUT PROCESS LABELS

Process reward models (PRMs) offer finer-grained feedback by evaluating reasoning trajectories step-by-step, unlike outcome reward models (ORMs), which assess entire responses. However, training PRMs requires detailed intermediate annotations, posing significant challenges. This paper shows that an implicit PRM can be derived at no additional cost by training an ORM on response-level labels using log-likelihood ratios of policy and reference models, enabling optimization without requiring specific loss objectives.

DYNAMIC 4-BIT QUANTIZATION

The Unsloth team aims to shrink a 20GB language model into 5GB without sacrificing accuracy. There are many algorithms that attempt this, but there are a number of challenges around outliers and compressibility. Llama is a famously hard model to quantize. The method selectively avoids quantizing certain parameters, which greatly improves overall accuracy.

TACKLING REPETITIVE IMAGE GENERATION

AccDiffusion v2 improves diffusion models for generating higher-resolution images without training, addressing challenges like object repetition and local distortion.

OPTIMIZING AI INFERENCE AT CHARACTER AI

Character AI has a strong inference pipeline. In this post, the team walks through their int8 quantization and flash attention 3 implementation. It's a good read for folks interested in serving large language models at scale.

CERTAIN NAMES MAKE CHATGPT GRIND TO A HALT, AND WE KNOW WHY

OpenAI's ChatGPT employs hard-coded filters to prevent the AI from generating false statements about specific individuals, leading to broken conversations when these names are mentioned. This approach stems from past incidents like defamation lawsuits against OpenAI, prompting the company to restrict outputs related to contentious names. The hard-coded filters can pose challenges as they limit ChatGPT's functionality and leave the system vulnerable to adversarial attacks.

FINETUNING LLM JUDGES FOR EVALUATION

Evaluating LLMs has become challenging due to their complex and open-ended outputs. Traditional evaluation relies on human judgment, which is detailed but inefficient. Thus, automatic metrics and model-based evaluations like LLM-as-a-Judge are crucial for scalable assessment. Innovations such as specialized fine tuned judges (e.g., Prometheus) and synthetic data generation are enhancing LLM evaluation's precision and adaptability across diverse tasks and domains.

faster. Find out how → watch the video

FLOW

Flow is a lightweight engine for creating flexible AI workflows using dynamic task scheduling and concurrent execution.

THE GEN AI BRIDGE TO THE FUTURE

Generative AI is set to revolutionize wearable technology by creating on-demand UI interfaces that adapt to user needs and context.

WORLD LABS' AI CAN GENERATE INTERACTIVE 3D SCENES FROM A SINGLE PHOTO

World Labs, founded by AI expert Fei-Fei Li, has introduced an AI system capable of generating interactive 3D scenes from a single image.

ELEVEN LABS CONVERSATIONAL AI

Eleven Labs has launched a new conversational AI service that is more of an end-to-end product for building conversational agents. It uses a variety of LLMs on the backend while integrating seamlessly to a wide variety of specialized voices.

CLAUDE MODELS NOW FASTER AND SMARTER ON AWS TRAINIUM2

Claude models are being optimized for AWS' advanced Trainium2 AI chips. This will enable faster and more efficient performance. Claude 3.5 Haiku is now available on AWS Trainium2 and model distillation in Amazon Bedrock.

AMAZON LAUNCHES NOVA

Amazon Nova introduces a suite of multimodal models designed for tasks like document analysis, visual understanding, and creative content generation. With a focus on customization and efficiency, Nova models cater to diverse enterprise needs while excelling in tasks that combine text, image, and video inputs.

DIFFUSION MEETS FLOW MATCHING: TWO SIDES OF THE SAME COIN

This post talks through some of the literature and shows that, mathematically, both flow matching and diffusion models are the same. Interestingly, flow matching still seems to scale better practically.

RESTRUCTURING VECTOR QUANTIZATION WITH THE ROTATION TRICK

Vector Quantization relies on gradient estimation through what's known as the Straight Through Gradient estimator. However, its direction is sometimes incorrect. This paper suggests rotation as a way to fix the gradients and improve code book utilization.

LAYOUT GENERATION WITH DIFFUSION GANS

DogLayout is a hybrid model that combines GANs and diffusion processes to overcome challenges in layout generation.

drops PLUS a chance to win a MacBook M4. No brainer!

HUNYUAN VIDEO MODEL

This state-of-the-art open video model from Tencent is particularly compelling due to its realistic motion and dual training as an image generation model. This dual training means the aesthetics of the output are higher quality, akin to image generation models like Flux.

SCENE TEXT RECOGNITION

TextSSR is a framework that uses diffusion-based methods to generate accurate and realistic synthetic text images for scene text recognition.

VIDEO UNDERSTANDING WITH MULTIMODAL LANGUAGE MODELS

T2Vid is a new method to enhance video understanding in Multimodal Large Language Models (MLLMs). It generates video-like samples to diversify training instructions.

A SYSTEM OF AGENTS BRINGS SERVICE-AS-SOFTWARE TO LIFE

AI is transitioning software from a tool to autonomous agents capable of executing tasks traditionally performed by humans, with a projected $4.6 trillion market opportunity. Innovations such as LLMs and agents enable AI systems to process unstructured data, make decisions, and act autonomously across sectors like sales and healthcare. The future of AI involves Systems of Agents that collaborate and learn from each other, functioning like a highly skilled team to deliver services.

AI MUSIC IS MORE REALISTIC THAN EVER: MEET SUNO'S NEW MODEL

Suno has become the fifth most-used generative AI service with its realistic AI music model V4, despite facing a copyright lawsuit. The model improves user experience by focusing on human preferences, offering enhanced sound quality and advanced composition skills. Suno aims to advance AI-human music collaboration while addressing copyright concerns with the recording industry.

BLUESKY'S OPEN API MEANS ANYONE CAN SCRAPE YOUR DATA FOR AI TRAINING

Bluesky isn't using user content to train AI, but third parties like Hugging Face have accessed posts for machine learning research. This highlights the public nature of user posts and the challenges in controlling third-party use. Bluesky is exploring ways for users to convey consent, but enforcement outside its platform remains a challenge.

AISUITE

aisuite provides a unified interface for seamless interaction with multiple LLM providers, aiding developers in testing and comparing outputs without altering code.

OVER ½ OF LONG POSTS ON LINKEDIN ARE LIKELY AI-GENERATED SINCE CHATGPT LAUNCHED

Since ChatGPT's launch, LinkedIn has seen an 189% surge in AI-generated content, with over half of long-form posts now likely AI-driven.

GOOGLE LAUNCHES THE LONDON AI CAMPUS

The AI Campus is an educational pilot to support and diversify the next generation of local AI talent.

live demo

ADS MIGHT BE COMING TO CHATGPT

OpenAI is considering an advertising business model to help support its costly AI tool development. Despite no current active plans for ads, the idea remains on the table as a potential revenue source. CEO Sam Altman considers ads a last resort and has expressed discomfort at the possibility of combining ads and AI.

THE RACE IS ON TO MAKE AI AGENTS DO YOUR ONLINE SHOPPING FOR YOU

Tech companies are developing AI shopping agents to automate online purchases, potentially reshaping the retail industry. Perplexity's current model faces operational challenges. Companies like OpenAI, Google, and Amazon are also exploring AI purchasing capabilities. These developments aim to streamline shopping but raise questions about privacy, retailer interactions, and the evolving online shopping experience.

SALESFORCE CEO MARC BENIOFF HAS THOUGHTS ON AI AGENTS, AUTOMATION, AND THE FUTURE OF YOUR JOB

Salesforce CEO Marc Benioff envisions companies creating AI agents to handle customer service and sales tasks by leveraging their existing data and policies, positioning Salesforce as the key enabler of this transformation. He argues that AI-driven automation will enhance productivity rather than eliminate jobs, allowing businesses to operate more efficiently and expand without increasing human labor. Benioff highlights this shift as a critical moment in business evolution, providing a competitive edge against rivals and reshaping traditional workflows.

DEMO: DECOUPLED MOMENTUM OPTIMIZATION

2.5x faster, 100x less communication volume, and written by the original Adam author. This new optimizer boasts impressive performance for language model training with some strong improvements over the current dominant optimization schemes.

DEMO TRAINING RUN (LIVE DASHBOARD)

Nous is training a 15B distributed model using the DeMo optimizer. All of the training can be followed live at this link.