Search the archive

a scoop of the action.

ADEPT ACQUIHIRE

Adept's founders are going to Amazon and licensing some of their tech as well.

OPENAI STRIKES A DEAL WITH TIME

Time has signed a licensing deal with OpenAI, allowing the company's AI to train on 101 years of its journalism and access real-time content for answering user queries about breaking news. In return, OpenAI will cite Time and link back to its source material.

AMAZON INVESTIGATING PERPLEXITY AI OVER POTENTIAL DATA-SCRAPING VIOLATIONS

Amazon Web Services is investigating Perplexity AI for allegedly scraping data from websites without consent.

DEEPDIVE: TRAINING MOES

The Mosaic team has partnered with PyTorch to write about scaling their MoE models to thousands of GPUs.

TEXT2BRICK

Open Sora video model trained to make Lego movies by Lambda Labs on its 1-click cluster.

MUMU VISION MODEL

Generative image model that can interleave text and images for stronger conditioning and higher performance than pure text conditioned models.

your free account in just a few clicks

LAPLACE NEURAL OPERATOR

The Laplace operator is a neural network based architecture for approximating PDEs.

LLAMA AGENTS

Llama-agents is an async-first framework for building, iterating, and productionizing multi-agent systems, including multi-agent communication, distributed tool execution, human-in-the-loop, and more.

IMPROVING AI'S WRITING SKILLS

Suri is a dataset of 20,000 long-form texts paired with complex instructions. It is designed to improve AI's ability to follow detailed writing constraints. The team that developed Suri has introduced Instructional ORPO (I-ORPO), an alignment method that uses synthetically corrupted instructions, for feedback.

CAMBRIAN-1

Fully open and high performing vision model from NYU with massive ablations to text encoders and data mixes.

EXPRESSIVE TEXT-TO-SPEECH

DEX-TTS is a new model for expressive text-to-speech (TTS) that uses reference speech to improve style representation and model generalization.

HELPFUL GUIDE TO DEBUGGING PYTORCH

PyTorch is a great tool for building models. However, there are some common bugs that can dramatically reduce model performance. This list is useful to review and can be helpful when debugging your model code.

APPLE COULD ANNOUNCE A GOOGLE GEMINI DEAL THIS FALL

Apple plans to announce a deal to integrate Google Gemini with its devices this fall.

INVESTING IN THE AGE OF GENERATIVE AI

Despite being in a state of funding "euphoria", the Generative AI market is already experiencing some weakness.

VISION AGENT

Vision Agent is a library that helps you utilize agent frameworks to generate code to solve your vision tasks.

algolia.com

SNAP LENSE STUDIO 5.0

Snap has released the GenAI suite in Lens Studio 5.0 - this is a great advancement and is massively useful for developing AR applications.

INSTAGRAM LAUNCHING AN AI STUDIO

Instagram's new "AI Studio" lets creators develop AI chatbot versions of themselves. It is currently rolling out as an early test in the US.

APPLE, MICROSOFT SHRINK AI MODELS TO IMPROVE THEM

LLM COMPILER

Meta has released two language models that can compile code to assembly and decompile to LLVM IR. They were trained on 546 Billion tokens of high-quality data then further instruction tuned. They are able to achieve 77% of optimized assembly performance and have 45% round trip disassembly performance.

CRITICGPT

OpenAI has trained a model based on GPT-4 that can catch GPT-4 mistakes and help humans spot mistakes during the RLHF process.

A BENCHMARK FOR ACTIVE LEARNING PIPELINES

ALPBench is a tool designed to standardize benchmarks for active learning query strategies.

and securely. Get started or book a demo

Need to deploy AI /ML at the edge? Deliver fast AI applications

with high throughput and ultra-low latency anywhere in the world with Gcore [12]. Automatically adjust resources based on demand (only pay for what you use), deploy and scale effortlessly, process data locally

FAL GIGAGAN UPSCALER

AuraSR is a super resolution image upscaler recently open sourced by FAL. It requires a single forward pass and can upscale by 4x even over repeated applications. AuraSR works well on generated images.

3D SEGMENTATION WITH INTERACTIVE GUIDANCE

To meet the growing demand for detailed 3D data, researchers have introduced Point-SAM, a transformer-based model for 3D segmentation.

GENERATIVE INFORMATION RETRIEVAL

This survey delves into generative information retrieval (GenIR), a fresh take on finding information that moves away from traditional search methods to ones that create answers on the fly.

ARE TECH-ENABLED VERTICAL ROLL-UPS THE FUTURE OR THE PAST?

Roll up strategies depend on the ability to drive operational efficiencies to create excess cash flows. The emergence of AI possibly presents a new lever that unlocks the rollup strategy to a new degree. Are Vertical/SMBs rollups the future? This post presents two opposing viewpoints on this question.

GEMMA 2 MODELS

The Gemma 2 9B and 27B models are now available.

LANGUAGE MODELS FOR MATERIALS SCIENCE

MatText is a suite of benchmarking tools and datasets designed to evaluate language models' performance in materials science.

MAMBA 2 IMPLEMENTATION

A quick implementation of Mamba 2 that doesn't require associative scans

DUST RAISES $16M SERIES A

One of the first chaining and agents companies of the modern era, Dust raised additional funding after passing $1m ARR.

ELEVENLABS LAUNCHES ITS FIRST IOS APP

ElevenLabs has launched its first iOS app, "ElevenLabs Reader: AI Audio," designed to convert text files or web links into audio narration, allowing users to listen while on the go.

Don't miss your chance. Invest in a renewable future.

YOUTUBE TRIES CONVINCING RECORD LABELS TO LICENSE MUSIC FOR AI SONG GENERATOR

YouTube is negotiating with major record labels for licenses to use their music in AI song-generation tools. It is offering substantial upfront payments but facing resistance from artists worried about devaluing their work. The platform, which trialed a generative AI tool called "Dream Track," aims to expand its AI music capabilities, potentially integrating them into YouTube Shorts. Meanwhile, the music industry is pushing back against unlicensed AI usage, with leading companies suing AI startups Suno and Udio for copyright infringement.

EVOLUTIONARY SCALE RAISES $142M SERIES A

Evolutionary Scale, a biology startup, just emerged from stealth with big backers. It also announced the release of its foundation model, ESM 3, which is a 98B parameter model trained on 771B biology tokens for 10^24 Flops. It used the model to discover a novel fluorescent green protein unlike any known in nature.

WAYMO ONE IS NOW OPEN TO EVERYONE IN SAN FRANCISCO

Waymo One now allows anyone in San Francisco to hail a ride with its autonomous vehicles. The company is scaling up operations after offering tens of thousands of trips weekly. Its all-electric fleet supports the local economy and its sustainability goals. Citing enhanced safety, Waymo says its vehicles are involved in significantly fewer crashes compared to human drivers.

NAMED ENTITY RECOGNITION WITH RETRIEVAL AUGMENTED INSTRUCTION TUNING

Retrieval Augmented Instruction Tuning (RA-IT) improves information extraction using large language models.

DEEPDIVE: SPARSE AUTOENCODERS

Golden Gate Claude was a powerful example of steering and interpreting models with SAEs. This work provides an intuitive explanation of how it works and has some example code for training these models.

DATA CURATION VIA JOINT EXAMPLE SELECTION FURTHER ACCELERATES MULTIMODAL LEARNING

Actively selecting the next best batch is a challenging and open problem in pre-training. This work from DeepMind explores how to spend just 10% of the FLOPs and hard mine negative samples and still match SOTA for various tasks.

Use code TLDR60 for a 60% discount →

REAL-WORLD 3D SCENE GENERATION

Director3D is a framework designed to enhance real-world 3D scene generation and camera trajectory modeling. Utilizing a Trajectory Diffusion Transformer and a Multi-view Latent Diffusion Model, Director3D generates realistic 3D scenes from textual descriptions.