Search the archive

PIANO PERFORMANCE GENERATION

This paper introduces a two-stage Transformer-based model for generating emotion-driven piano performances.

3D GENERATIVE MODEL FOR DYNAMIC SCENES

DynaVol-S is a 3D generative model that excels in learning object-centric representations from unsupervised videos.

INSTRUCTION-BASED OBJECT ADDITION

Add-SD is a tool that inserts objects into realistic scenes using simple text prompts. This method, unlike others, doesn't need detailed references like bounding boxes.

RAISING CHILDREN ON THE EVE OF AI

As transformative AI increasingly seems probable, this author, grappling with the timing and uncertainty of changes, ponders on how to prepare children for a future that could radically differ from today's norm. They also explore the ethical considerations of having children amidst AI-driven uncertainty, alongside pragmatic reflections on raising "AI-native" kids and parenting strategies that prioritize adaptability and joy over traditional career-focused paths. The author advocates for living a balanced life, preparing for various futures, and maintaining an open dialogue with children about potential risks.

FLOW MATCHING INSTEAD OF SCORE MATCHING

Diffusion models are powerful. Understanding them can be challenging. One way to look at them, theoretically, is via flow matching. This blog dives deep into the math of flow matching for diffusion.

YOUR NEW AI FRIEND IS ALMOST READY TO MEET YOU

Avi Schiffmann is developing "Friend," an AI companion encapsulated in a wearable pendant, intended to offer friendship and support rather than productivity enhancements. The device, which communicates via an app, is set for a preliminary run of 30,000 units at $99 each. It will ship in January without a subscription fee. Schiffmann envisions Friend evolving into a platform for digital relationships, distancing the product from task-oriented AIs and focusing instead on the emerging trend of forming meaningful connections with digital entities.

Start building for free

NVIDIA ACCELERATES HUMANOID ROBOTICS DEVELOPMENT

SIMPLY DRAW (PRODUCT LAUNCH)

AI feedback meets art - anyone can learn to draw.

US' FIRST ROBOT-ASSISTED DUAL KIDNEY TRANSPLANT PERFORMED IN OHIO

A 70-year-old Ohio woman underwent the first US robot-assisted dual kidney transplant at the Cleveland Clinic, leading to a quick recovery.

❌ Delete your personal data today

OPENAI HAS RELEASED A NEW CHATGPT BOT THAT YOU CAN TALK TO

OpenAI is launching a new AI voice chatbot powered by the GPT-4o model with features that include natural, fluent conversation capabilities. It is initially available to select ChatGPT Plus subscribers.

META'S NEW AI STUDIO HELPS YOU CREATE YOUR OWN CUSTOM CHATBOTS

Meta's new AI Studio tool will soon allow users without technical skills to create personalized AI chatbots for Instagram, Messenger, and WhatsApp. The tool will enable customized interactions with followers and full control over auto-replies.

PERPLEXITY WILL SOON START SELLING ADS WITHIN AI SEARCH

Perplexity will start paying publishers a revenue share from ad sales when their content is used to form AI search answers, marking a new business model for the AI search engine, which will allow brands to buy "related follow-up questions" appearing below user query answers. Publishers will earn a percentage of ad revenue and gain free access to Perplexity's large language models and Pro service tier.

THEIA: DISTILLING DIVERSE VISION FOUNDATION MODELS FOR ROBOT LEARNING

This paper introduces Theia, a vision foundation model for robot learning that integrates multiple existing vision models. Theia offers rich visual representations, enhancing robot learning with less training data and smaller model sizes. Experiments show that Theia outperforms its predecessors, and the authors suggest that higher entropy in feature norms improves performance. The code and models are available for public use.

SAULLM-54B & SAULLM-141B: SCALING UP DOMAIN ADAPTATION FOR THE LEGAL DOMAIN

This paper introduces SaulLM-54B and SaulLM-141B, two large language models designed for the legal sector that feature 54 billion and 141 billion parameters. Using the Mixtral architecture, the researchers employed large-scale domain adaptation through continued pre-training with a vast legal corpus, a specialized legal instruction-following protocol, aligning outputs with human legal interpretations. The models outperform previous open-source models on LegalBench-Instruct and offer state-of-the-art performance. Base, instruct, and aligned versions of these models have been released under the MIT License for reuse and collaborative research.

RECOMMENDER SYSTEMS WITH GRAPH CONVOLUTIONS

LightGODE is a new approach to improve the efficiency and scalability of recommender systems. It bypasses intensive computations during training by focusing on post-training graph convolution and utilizing a continuous graph ODE.

FACE IMAGE RESOLUTION

Researchers have developed a wavelet-based feature enhancement network to improve face super-resolution. The method enhances facial details without introducing distortions by decomposing input features into high and low-frequency components and using a full domain Transformer.

CHART QUESTION ANSWERING WITH MULTIMODAL MODELS

This project proposes a new method to improve chart question answering using multimodal large language models.

TRAJECTORY PREDICTION FOR AUTONOMOUS DRIVING

DGFNet is a new approach for predicting the trajectories of multiple traffic participants in autonomous driving. It enhances predictions by considering the difficulty differences among agents, capturing rich spatio-temporal features, and using a difficulty-guided decoder.

AI IS A SERVICES REVOLUTION

This article discusses the impact of LLMs on the service economy, particularly on knowledge work sectors such as legal, healthcare, and education. While rapid AI advancements suggest significant automation potential, the adoption by enterprises is slow, with many still in the experimentation phase. Actual deployment is expected to be gradual. Specialized AI startups that leverage LLMs to improve industry-specific workflows will have a competitive advantage in the evolving market.

THE AI JOB INTERVIEWER WILL SEE YOU NOW

AI-powered job interviews, which leverage large language models for automation and potential bias reduction, are being increasingly adopted by firms globally. Companies like Meituan and multinationals are using AI avatars to conduct initial interviews, perceiving them as efficient despite concerns over potential embedded biases and transparency in decision-making. Skeptics advocate for clear reporting on AI tools' construction and bias mitigation. Some candidates are using AI to formulate responses, highlighting the evolving competition between AI systems and job applicants.

WHY BIG TECH WANTS TO MAKE AI COST NOTHING

Meta has open-sourced Llama 3.1, an LLM rivaling OpenAI's ChatGPT, allowing almost all companies to incorporate it freely. This strategy may commoditize LLMs and drive demand for corresponding complements like server space. As big tech releases similar models, AI startups might face challenges, with industry giants potentially outpacing smaller competitors in AI advancements.

CANVA BUYS LEONARDO

Generative image company Leonardo joins Canva to supercharge both groups' creative tools.

ANNOUNCING PHI-3 FINE-TUNING, NEW GENERATIVE AI MODELS, AND OTHER AZURE AI UPDATES

Microsoft has announced updates to Azure AI, including serverless fine-tuning for PHI-3 models, improvements in PHI-3-MINI performance, and the integration of models like GPT-4o mini and Meta's LLAMA 3.1 into Azure AI.

WHO WILL CONTROL THE FUTURE OF AI?

Sam Altman of OpenAI has presented a strategic imperative for the US and allies to lead a global AI initiative based on democratic values that suggests robust security, infrastructure investment, commercial diplomacy, and collaborative norms development to maintain AI supremacy over authoritarian regimes.

Start deploying enterprise today

META RELEASES SAM 2

Real time promptable segmentation for videos and images from Meta.

APPLE SAYS ITS AI MODELS WERE TRAINED ON GOOGLE'S CUSTOM CHIPS

Apple said that the artificial intelligence models underpinning Apple Intelligence, its AI system, were pre-trained on processors designed by Google.

ANTHROPIC FACES BACKLASH FOR EXCESSIVE WEB SCRAPING

Freelancer.com and iFixit CEOs criticized Anthropic for excessive and unauthorized web scraping by its ClaudeBot, which disrupted their sites. Anthropic claims to respect the robots.txt file and says it will investigate the issue.

SMALL MOLECULE OPTIMIZATION WITH LARGE LANGUAGE MODELS

Two new language models were able to achieve an 8% improvement on the Practical Molecular Optimization benchmark by collecting a dataset of 100m molecules (40B token equivalent).

THE LARGER THE BETTER? IMPROVED LLM CODE-GENERATION VIA BUDGET REALLOCATION

Using smaller models multiple times can improve code generation performance with a very similar inference cost.

SELF-DIRECTED SYNTHETIC DIALOGUES AND REVISIONS TECHNICAL REPORT

300k+ dataset of conversations and critiques to be used for open models. The dataset was mostly created using synthetics and is a powerful example of synthetic data working with open models.

ACCDIFFUSION

A method that dramatically improves high resolution image synthesis by diffusion models.

MOBILE AR WITH ROBUST DEPTH ESTIMATION

HYBRIDDEPTH is a depth estimation pipeline designed to tackle mobile AR challenges like scale ambiguity and hardware diversity.

ENHANCING VISION TRANSFORMERS WITH NON-CAUSAL STATE SPACE MODELS

The Visual State Space Duality (VSSD) model is a new approach that improves vision transformers by addressing their high computational demands.

APPLE FOUNDATION LANGUAGE MODELS

Apple has outlined the basics of its language models for its newly announced “Apple Intelligence” initiative.

CHALKBOARDS AND WHAT THEY CAN TEACH US ABOUT GENERATIVE AI

This article reflects on using generative AI as a teaching tool, arguing the technology should be analyzed beyond its technical aspects, with consideration for its alignment with educational values. The author, while open to using AI, is cautious about its impact and emphasizes the need to articulate why certain tools are implemented in the classroom. The discussion draws parallels between traditional tools like chalkboards and AI, considering both for their cultural significance and pedagogical implications.

THE EVOLUTION OF SAAS PRICING IN THE AI ERA

The traditional seat-based pricing model in SaaS is becoming obsolete with AI's capability to automate tasks. Companies are increasingly adopting work-based or outcome-based pricing models, which charge based on the amount of work AI performs or the outcomes achieved. While incumbents are still relying on seat-based pricing, startups are leveraging new models for competitive advantage and to reflect AI's value more accurately.

install --cask warp;_ or **download Warp for Mac

A NEW BENCHMARK FOR AUTONOMOUS AGENTS

AppWorld Engine is a sophisticated execution environment that features nine daily apps and 457 APIs.

CRASH COURSE IN DEEP LEARNING (FOR COMPUTER GRAPHICS)

This article covers the construction and use of multi-layer perceptrons (MLPs), a type of fully connected neural network used in deep learning.

AIRTABLE'S NEW COBUILDER UNLOCKS INSTANT NO-CODE APP CREATION

Airtable's Cobuilder is the fastest way to build no-code applications.

📥 Download your free copy