Search the archive

KRUT AI (PRODUCT LAUNCH)

Krut AI is an AI co-pilot built for e-commerce that helps with product shots and content creation.

META DROPS AI BOMBSHELL: MULTI-TOKEN PREDICTION MODELS NOW OPEN FOR RESEARCH

Meta has released pre-trained models that leverage a novel multi-token prediction approach that forecasts multiple future words simultaneously, promising enhanced performance and drastically reduced training times.

GOOGLE DEEPMIND'S AI RAT BRAINS COULD MAKE ROBOTS SCURRY LIKE THE REAL THING

DeepMind and Harvard University developed a virtual rat with AI neural networks trained on actual rat movements and neural patterns to probe the brain circuits responsible for complex motor skills. This bio-inspired AI has the capacity to generalize learned movement skills to new environments, offering insights into brain function and advancing robotics. The research demonstrates that digital simulations can effectively mimic and decode neural activity related to different behaviors.

AN OPINIONATED LIST OF MECHANISTIC INTERPRETABILITY PAPERS

A post by Neel Nanda, a Research Engineer at Google DeepMind, about his favorite papers to read in Mechanistic Interpretability.

CONTROL IN VISUAL AI GENERATION

PartCraft is a new method in generative visual AI that allows users to select visual concepts by parts, moving beyond traditional text or sketch-based methods.

AUTONOMOUS AI AGENTS

AriGraph is a novel method that helps AI agents build a memory graph integrating both semantic and episodic memories.

SAE

EleutherAI trained SAEs for Llama 3 8B and released weights and the library code.

MINFERENCE

Microsoft's Minference dramatically speeds up supported models' inference with a number of system improvements.

MICRO AGENT

An AI agent that writes and fixes code for you.

SUPER-RESOLUTION FOR ANY SCALE

AnySR is a new approach to single-image super-resolution (SISR) that enhances efficiency and scalability. Unlike traditional methods, AnySR supports the 'Any-Scale, Any-Resource' implementation, reducing resource needs for smaller scales without additional parameters.

3D POSE ESTIMATION FROM VIDEOS

Researchers have developed a new approach to estimating category-level 3D poses from casual, object-centric videos without human supervision.

RESEARCHERS LEVERAGE SHADOWS TO MODEL 3D SCENES, INCLUDING OBJECTS BLOCKED FROM VIEW

PlatoNeRF is an AI technique developed by MIT and Meta researchers that uses shadows and single-photon lidar to create 3D models of scenes, including obstructed areas. This method could enhance the safety of autonomous vehicles and improve AR/VR experiences. PlatoNeRF outperforms traditional methods and has promising practical applications with lower-resolution sensors.

SENSEVOICE

A speech foundation model with multiple speech understanding capabilities, including automatic speech recognition, spoken language identification, speech emotion recognition, and audio event detection.

HOW GOOD IS CHATGPT AT CODING, REALLY?

OpenAI's ChatGPT has varied performance in code generation, with success rates ranging from less than 1% to 89% depending on factors like task difficulty and programming language.

DOTI (PRODUCT LAUNCH)

Doti is an AI-powered assistant for managing health and life goals.

A HACKER STOLE OPENAI SECRETS, RAISING FEARS THAT CHINA COULD, TOO

A hacker gained access to OpenAI's internal messaging systems early last year. They accessed details from discussions in an internal online forum but did not get into the systems where the company houses and builds its artificial intelligence. Details of the incident were revealed to employees in April 2023. Executives decided not to share the news publicly because no information about customers or partners was stolen. It is believed that the hacker was a private individual with no known ties to a foreign government. OpenAI did not inform the FBI or anyone else in law enforcement about the incident.

FIGMA PULLS AI TOOL AFTER CRITICISM THAT IT RIPPED OFF APPLE'S DESIGN

Figma's "Make Designs" AI feature was temporarily pulled amidst concerns it generated designs too similar to Apple's apps. CEO Dylan Field attributes the similarity to third-party models and a design system the company commissioned, rather than training on Figma's own content. Figma plans to ensure future AI features align with quality standards and user privacy choices.

AI RESURRECTS DECEASED ACTORS' VOICES TO READ AUDIOBOOKS

AI company ElevenLabs has launched a Reader app featuring voice-overs by digital versions of deceased celebrities, stirring questions of morality and legality in AI's use in media. While partnerships with estates enable monetization, concerns about authenticity and ethical implications persist. The emergence of such technologies heralds a potential expansion in markets such as audiobooks, where celebrity voices drive demand.

OPTIMIZING SUB BILLION PARAMETER MODELS

Thorough and interesting paper by Meta that shows how to train small models to be as performant as possible.

CAPPA TRAINING REPORT

The CapPa vision model trained by Craiyon shows state-of-the-art performance on a number of challenging vision benchmarks.

MEDICAL IMAGING

Researchers have developed a new generative model to improve MRI image translation between different sequences without requiring paired samples.

SDXL TRAINED FROM SCRATCH WITH A FULL LLM EMBEDDER (HUGGING FACE HUB)

A great and open alternative diffusion model with strong text understanding.

POINT CLOUD COMPLETION

EGIInet introduces a new approach to point cloud completion by effectively combining two modalities through geometric task guidance.

REAL-TIME 3D RECONSTRUCTION FOR SURGERY

An SfM-free method for 3D reconstruction of surgical scenes has been introduced, addressing the limitations of previous approaches that struggled with minimal textures and photometric inconsistencies.

FUN AUDIO GENERATION AND UNDERSTANDING

The Tongyi speech team released extremely strong audio understanding and generation models.

SUPERINTELLIGENCE—10 YEARS LATER

A decade after Nick Bostrom's influential book "Superintelligence," AI advancements have brought the possibility of AGI and its risks into mainstream discussions. The AI research community is now urgently addressing AI safety, with 2024 marking a pivot toward ensuring control and alignment with human values. The field grapples with ethical and safety challenges previously considered speculative as AI technologies rapidly progress.

GOOGLE SMART PASTE

Google's core code tooling team has trained a smart paste model that adjusts code pasted in based on the context. They found a 42% acceptance rate for 6% of all copied code.

QUALITY PROMPTS

This library consolidates a growing set of standard prompting tricks into a single place and can automatically apply them to your prompts to optimize them.

APPLE M5 CHIP'S DUAL-USE DESIGN WILL POWER FUTURE MACS AND AI SERVERS

Apple will reportedly use a more advanced System on Integrated Chip (SoIC) technology for its M5 chips.

APPLE INTELLIGENCE AND A BETTER SIRI MAY BE COMING TO IPHONES THIS SPRING

Apple's Siri is set to improve with the iOS 18.4 update releasing in Spring 2025, which will introduce new features like advanced photo searching and possibly AI-driven emoji and writing tools.

HOW BIG TECH IS SWALLOWING THE AI INDUSTRY

Amazon has hired most of Adept's team, another aspiring OpenAI competitor, and licensed its technology, following a pattern that mirrors Microsoft's acquisition approach with Inflection. This move reflects a trend where big tech companies absorb AI talent and innovations without formal acquisitions to navigate antitrust scrutiny. Such strategies may become commonplace as the AI industry faces consolidation and high development costs.

INTEL SHOWS OFF FIRST FULLY INTEGRATED OPTICAL COMPUTE INTERCONNECT, DESIGNED TO SCALE UP AI WORKLOADS

Intel has presented the industry's first fully integrated optical compute interconnect (OCI) chiplet designed to significantly boost high-bandwidth data interconnect performance. Essential for growing AI demands, the milestone represents a step forward in silicon photonics, combining integrated circuits with lasers to enhance data transfer rates and support AI infrastructure scaling. The OCI chiplet aims to overcome the limitations of current electrical I/O interconnects by enabling higher bandwidth and longer-distance data transmission.

LLM101N: LET'S BUILD A STORYTELLER

Andrej Karpathy announced a new course that guides learners in building a Storyteller AI Large Language Model from scratch. It covers the fundamentals of web app implementation using Python, C, and CUDA. The content spans language modeling, deep learning frameworks, and AI deployment including multimodal applications.

MAGIC INSERT

Method from Google to insert semantic objects into images with diffusion. Dataset and demo available.

CTR PREDICTIONS WITH A NEW TOKENIZATION

UIST is a novel approach that drastically improves click-through rate predictions by converting dense embeddings into compact, discrete tokens for user and item representations.

ENHANCED CAUSALITY UNDERSTANDING

CELLO is a new dataset with 14,094 causal questions to improve AI's grasp of causality beyond commonsense reasoning.

CAMERA TRACKING WITH MULTI-SESSION SLAM

Multi-Session SLAM is a new system that innovatively tracks camera movements across several unconnected video sequences under one universal frame of reference.

OPENSTREETVIEW-5M: A GLOBAL DATASET FOR LOCALIZING IMAGES

OpenStreetView-5M is a significant open-access dataset with over 5 million geo-tagged street images from 225 countries designed to push the limits of computer vision by testing image localization capabilities.

MAKING SAM MORE PRACTICAL WITH POST-TRAINING QUANTIZATION

PTQ4SAM is a new framework designed to reduce the memory and computation demands of the large-scale Segment Anything Model (SAM).

CONTRA ACEMOGLU ON AI

Economist Daron Acemoglu's working paper on AI suggests a modest 0.06% annual increase in TFP growth, challenging more optimistic forecasts. He distinguishes AI's productivity impact through four channels but discounts deepening of automation and creation of new labor-intensive products, arguably underestimating AI's full economic potential. Critics find his approach overly narrow, potentially misrepresenting the broader macroeconomic implications of AI advancements.

BOOSTING SMARTPHONE CAMERA CLARITY

This project introduces a method for enhancing image resolution on smartphones using a self-supervised learning model that improves reference-based super-resolution (RefSR).

SPEECH ENHANCEMENT WITH SEMAMBA

SEMamba is a new speech enhancement system that leverages the Mamba state-space model to improve the clarity of speech signals.

help get you started. Get the cookbook →

FLORENCE 2 ON WEBGPU

The small vision model runs entirely in the browser built on onnx and WebGPU.

OPENAI'S CHATGPT MAC APP WAS STORING CONVERSATIONS IN PLAIN TEXT

OpenAI's ChatGPT macOS app, until a recent update, stored user chats in plain text, making them vulnerable to unauthorized access.

LONG VIDEO GENERATION WITH NEW DIFFUSION MODEL

FlexiFilm is a diffusion model designed specifically for generating long videos over 30 seconds with high consistency and quality.

GROK 2 ANNOUNCED FOR AUGUST

xAI's next model is slated for August. It reportedly took 20k H100s to train. Grok 3 is rumored for the end of the year and potentially requires 100k H100s.

NVIDIA CEO JENSEN HUANG ADDRESSES RISING COMPETITION AT SHAREHOLDER MEETING AFTER HISTORIC STOCK SURGE

Nvidia's CEO Jensen Huang attributes the company's AI chip market dominance, maintaining an over 80% market share despite rising competition, to a decade-old strategic investment. Advocating for Nvidia's AI chips' cost-effectiveness and performance, Huang highlights the firm's transformation into a data center-focused entity and expansion into new markets.