Search the archive

CHOOSE YOUR MODEL FOR GITHUB COPILOT

You can now use Claude, Gemini, o1, and other models in GitHub Copilot.

APPLE RELEASES FIRST BATCH OF APPLE INTELLIGENCE FEATURES

Apple launched new AI features, called Apple Intelligence, on recent devices for text processing and photo editing. The revamped iMac is powered by the M4 chip, which features a Neural Engine and boasts triple the AI performance of past chips. Future AI updates will enhance Siri and integrate with ChatGPT for more complex inquiries.

A LITTLE HELP GOES A LONG WAY: EFFICIENT LLM TRAINING BY LEVERAGING SMALL LMS

DeepMind shows in this work how you can use a small language model to provide soft supervision labels and select informative/challenging data points for pretraining. This small helper model can dramatically speed up pre-training time.

DECODING BRAIN ACTIVITY INTO SMOOTH VIDEO

The new NeuroClips framework advances the reconstruction of continuous videos from fMRI brain scans by decoding both high-level semantics and low-level perceptual details.

AI TACKLES VEHICLE ROUTING PROBLEMS

Researchers introduce a Proactive Infeasibility Prevention (PIP) framework aimed at improving neural networks' performance on Vehicle Routing Problems (VRPs) with tough constraints.

customer support, and data analysis. Get the guide.

UNLEASHING THE POWER OF AI ON MOBILE

PyTorch has done excellent work on ExecuTorch, which allows for edge deployment of AI models. This deployment can dramatically improve a variety of end systems.

AN OPEN-SOURCE EVALUATOR FOR LLMS

CompassJudger-1 is the first open-source, all-in-one judge model designed to improve the evaluation of LLMs.

MINT-1T: SCALING OPEN-SOURCE MULTIMODAL DATA BY 10X

MINT-1T, a massive open-source multimodal dataset, has been released. Featuring one trillion text tokens and 3.4 billion images, including content from HTML, PDFs, and ArXiv papers, the dataset is about 10 times larger than previous datasets. The resource is aimed at advancing research in large-scale multimodal machine learning.

HOW GPU ACCESS HELPS STARTUPS BE AGILE

Andreessen Horowitz's Oxygen program addresses GPU shortages by providing startups in its portfolio with more affordable and flexible access to GPUs, helping it circumvent price spikes and supply constraints. This initiative supports AI startups in focusing on product development without the burden of long-term capital expenditure commitments. The program highlights the importance of equitable resource access in an increasingly competitive AI landscape.

THE MASK COMES OFF: AT WHAT PRICE?

OpenAI is nearing its transition to a Public Benefit B-Corporation, which has implications for investor relations and its partnership with Microsoft. This move raises questions about control and valuation, especially concerning the nonprofit's stake, which could be significant given OpenAI's potential in AGI development. The future profitability and strategic direction of OpenAI hinge on the successful creation of safe AGI, a prospect that holds immense potential value.

2034. TLDR readers get early access.

HOW ADVEX CREATES SYNTHETIC DATA TO IMPROVE MACHINE VISION FOR MANUFACTURERS

Advex AI is solving data shortages in AI training using generative AI to produce synthetic imagery for computer vision systems.

A NEW TOKENIZATION FOR VIDEO GENERATION

LARP is a new video tokenizer that aims to improve video generation for autoregressive (AR) models by focusing on global visual features rather than isolated patches.

COFRAME RAISES $9 MILLION FOR WEBSITES THAT OPTIMIZE THEMSELVES USING AI

AI startup Coframe raised $9.3 million in seed funding to enhance its platform, which uses generative AI for website optimization and personalized marketing experiences.

🎟️ Use code TLDR at checkout for 30% off All Access Passes

(individual or corporate), granting entry to KubeCon + CloudNativeCon North America + 15 CNCF-hosted co-located events. Claim your offer

here

GOOGLE EXPANDS VOICE TECHNOLOGY SUPPORT TO 15 MORE AFRICAN LANGUAGES

Google has introduced voice recognition for 15 additional African languages across its platforms, including Voice Search, Gboard talk-to-type, and Translate dictation. This update empowers approximately 300 million more Africans to interact with digital content in their native languages.

MIT'S GENERAL-PURPOSE ROBOT TRAINING

MIT researchers have introduced Heterogeneous Pretrained Transformers (HPT), a new model architecture inspired by large language models, to train versatile robots using data from various domains and modalities.

COHERE RELEASES STATE-OF-THE-ART MULTIMODAL AI SEARCH MODEL

Cohere announced that its Embed 3 AI model is now multimodal, enabling fast and accurate search across important enterprise image data sources like graphs, charts, product catalogs, and design files. With this upgrade, Embed 3 is the most generally capable multimodal embedding model on the market.

VIEW SYNTHESIS WITH MINIMAL 3D INDUCTIVE BIAS

Novel view synthesis has long been thought to require strong 3D inductive biases. This work shows that with scale and some weak inductive bias, you can dramatically outperform previously thought limits.

CONTINUOUS SPEECH SYNTHESIS USING PER-TOKEN LATENT DIFFUSION

Auto regressive models still dominate in many cases. However recent work on diffusion heads for images has inspired continuous AR diffusion. This work extends the per token diffusion idea to variable length outputs.

CHANGE DETECTION IN REMOTE SENSING

This paper introduces a change description instruction dataset to finetune LMMs for better remote sensing change detection.

FLUX IC LIGHT

IC Light is the current best way to relate an image with a pretrained text-to-image backbone. This discussion is the beginning of extending that functionality to the strong Flux models.

SCENE LANGUAGE FOR 3D SCENE GENERATION

Generating 3D scenes from scratch is challenging for a number of reasons, such as data constraints. This work introduces a programming-like language to describe 3D scenes and shows that Claude Sonnet can generate extremely compelling scenes despite not having been explicitly trained on this task.

3D SEMANTIC SEGMENTATION

FtD++ is an approach for cross-modal learning that improves unsupervised domain adaptation for 3D semantic segmentation.

OPEN SOURCE REPLICATION OF CROSSCODER ON GEMMA 2B

Anthropic recently released two works that highlight its new interpretability method. This post is an open replication of the crosscoder on Gemma 2B.

A COLLECTION OF OUT-OF-DISTRIBUTION GRAPH LEARNING APPROACHES

This repository lists papers on graph out-of-distribution learning, covering three primary scenarios: graph OOD generalization, training-time graph OOD adaptation, and test-time graph OOD adaptation.

THINKING LIKE AN AI

LLMs function as sophisticated autocomplete systems, predicting the next token based on training data and current input. Tiny changes in input affect predictions, leading to different outputs even for the same question. Understanding token prediction, training data context, and memory limitations can help refine AI use.

users. Start for free

MULTIMODAL WEB AGENTS

OpenWebVoyager provides tools, data, and models for developing multimodal web agents capable of navigating and learning through real-world web interactions.

AUTOMATED COLORIZATION FOR ANIMATION

Researchers have unveiled a novel inclusion-matching approach that addresses the challenges of automated colorization especially in animations where occlusions and wrinkles can disrupt traditional segment matching.

LOFI MUSIC DATASET (HUGGING FACE HUB)

A dataset of music clips and detailed text descriptions generated by a music generative model.

Download now → _The Definitive Guide to Onboarding_

WAYMO RAISES $5.6B

Waymo's driverless taxi service is well-loved. The company has raised extra capital to expand beyond the several cities and millions of miles already in operation.

PIXTRAL 12B WEIGHTS RELEASED

Pixtral 12B base model weights have been released on Hugging Face.

AI JEWELRY

Arcade AI has built a generative product platform for creating unique and quality jewelry items from text prompts. The best part is, you can buy what you prompt.

EVALUATING FEATURE STEERING

How do you evaluate the ability and efficacy of feature steering? Anthropic experimented with Golden Gate Claude earlier this year. It found that many features, especially similar safety ones, share common ranges for control.

CROSSCODERS

Crosscoders are an extension of sparse autoencoders that can be used to better understand internal mechanisms of language models.

RETRIEVAL-AUGMENTED TIME SERIES DIFFUSION MODEL

chatbot - start a free trial & experience the difference.

NOTEBOOKLLAMA

Meta has released a quick start guide to recreate a basic version of the very popular NotebookLM system from Google.

CHART REASONING IN MLLMS

Code-as-Intermediary Translation (CIT) is a novel method for enhancing visual reasoning in multimodal language models (MLLMs) that uses code to translate chart visuals into text.