Search the archive

IMPROVING GRADIENT DESCENT

This study introduces improvements to line search methods that boost the performance of stochastic gradient descent techniques.

of conversation. CRM integration included. Start free

BOOSTING EFFICIENCY IN VISION-LANGUAGE MODEL TRAINING

This project introduces a method to balance computational loads during large-scale 3D parallel training of vision-language models, addressing the imbalance caused by differing data distributions and model architectures.

TEXGEN

High quality generation of textures on 3d models with diffusion.

BYU ENGINEERING RESEARCH FINDS KEY TO QUICKER NUCLEAR POWER: ARTIFICIAL INTELLIGENCE

BYU chemical engineering professor Matt Memmott has developed an AI algorithm that could reduce the nuclear reactor design and licensing process by a decade, cutting costs significantly. His team's research shows AI can optimize complex nuclear design problems much faster than traditional methods, with one example reducing design time from six months to just two days. The findings aim to accelerate nuclear power production to meet rising energy demands and maintain affordable power costs.

REPLACING MY RIGHT HAND WITH AI

An Anthropic scientist broke their hand on a bike. They leaned into Claude and voice to still be extremely productive.

10 YEARS OF THE TPU

Google is one of the only groups training large models that isn't beholden to Nvidia due to its investment in custom TPU chips.

YOU GUYS HAVE NO IDEA JUST HOW MUCH PEOPLE HATE GENERATIVE AI

This blog explores the dichotomy of love and hate for AI, emphasizing its pivotal role in progress while also acknowledging the distrust and dissatisfaction people have with Big Tech. Disdain for generative AI stems from its perceived shortcomings and the high-cost, low-return reality emphasized by tech leaders' actions. Ultimately, the misaligned efforts of technologists risk a widespread aversion to AI, well beyond generative models, threatening the potential of beneficial AI applications.

OPENAI TEMPERS EXPECTATIONS WITH LESS BOMBASTIC, GPT-5-LESS DEVDAY THIS FALL

OpenAI has announced that this year's DevDay conference will shift from a major event to a series of smaller, on-the-road developer sessions, focusing on updates to APIs and developer services rather than unveiling a new flagship model.

TEZI RAISES $9M TO LAUNCH MAX: THE FIRST FULLY AUTONOMOUS AI RECRUITER

Tezi raised $9M in seed funding led by 8VC and Audacious Ventures to develop Max, an AI-driven recruiting agent that autonomously manages end-to-end hiring processes.

IPHONE 16 RELEASE UNAFFECTED BY APPLE INTELLIGENCE ROLLOUT DELAYS

iOS 18 will introduce Apple Intelligence features post-launch, with initial access starting in iOS 18.1 only for iPhone 15 Pro models.

CHARACTER CO-FOUNDERS RETURN TO GOOGLE

Noam Shazeer, one of the authors of the original Transformer paper, is returning to Google after Character.AI licensed its tech for 2x the last funding valuation.

THREE NEW ADDITIONS TO GEMMA 2

Google is adding a new 2B parameter model, safety content classifier model, and model interpretability tool to the Gemma 2 family.

MICROSOFT SAYS OPENAI IS NOW A COMPETITOR IN AI AND SEARCH

Microsoft has classified OpenAI as a competitor in its latest annual report, despite being OpenAI's biggest investor and long-term partner, marking a shift in their relationship.

SCALING INFERENCE COMPUTE WITH REPEATED SAMPLING

AUDIO-VISUAL QUESTION ANSWERING IN VIDEOS

Researchers have developed a Temporal-Spatial Perception Model (TSPM) to improve the ability to answer questions related to visual and audio cues in videos.

REAL-TIME HIGH-QUALITY HAND AVATARS RENDERING

XHand is a new hand avatar designed for real-time rendering in virtual environments and gaming. Unlike previous models, XHand focuses on generating detailed hand shapes, appearances, and deformations.

PROMPT POET

Character AI has open sourced its prompting construction library, which is used in production to serve millions of conversations.

MINIGRID IN JAX

A popular testing bed for RL has been accelerated in JAX.

IMAGE-TEXT PAIR GENERATION FOR VISION MODELS

SynthVLM is a new data synthesis pipeline for Vision Large Language Models (VLLMs). Instead of generating captions from images, SynthVLM uses advanced diffusion models to create high-resolution images from captions.

FUTURE OF LLMS FOR AVATARS AT GENIES

Genies is harnessing LLMs for user profiling, enabling personalized experiences by mapping behaviors to foundational traits. Its advancing chatbots and NPCs in gaming create immersive, conversational scenarios. Combining conversational data with traditional personalization techniques, Genies is elevating user engagement and providing intuitive data analysis for game devs through natural language processing.

AI EXISTENTIAL RISK PROBABILITIES ARE TOO UNRELIABLE TO INFORM POLICY

This essay critiques the use of AI existential risk probability estimates for policy making, arguing that they are too unreliable and lack a solid inductive or deductive basis, often resembling educated guesses rather than evidence-based forecasts. The authors challenge the legitimacy of these forecasts in guiding public policy, especially when linked to costly or freedom-limiting actions, and advocate for an evidence-based approach that acknowledges uncertainty in AI progression. They propose focusing on definable AI milestones for more meaningful policy decisions while cautioning against leveraging speculative existential risk probabilities in high-impact decisions.

IS AI JUDGING THE FUTURE OF GYMNASTICS OR JUST A SURVEILLANCE TOOL?

The International Gymnastics Federation (FIG) has teamed with Fujitsu to launch an AI-assisted judging support system at the World Gymnastics Championships, aiming for fairer, more transparent scoring. The Judging Support System (JSS), which will not replace judges, offers 3D model-based second opinions in difficult cases and inquiry disputes, with scope for future improvements and broader applications. Despite concerns over replacing human judgment, the JSS may enhance scoring accuracy and consistency, vital in a sport where minimal point differences greatly impact rankings and athletes' careers.

NETWORKS THAT COMPRESS THEMSELVES (GITHUB GIST)

By including the size of the network in the loss function, You can train a model that improves in accuracy and is self quantized to be smaller.

GITHUB MODELS

GitHub has announced a competitor to Hugging Face's model hosting service.

REDDIT CEO SAYS MICROSOFT NEEDS TO PAY TO SEARCH THE SITE

Reddit CEO Steve Huffman is calling for tech giants like Microsoft to negotiate terms for using Reddit data in AI models, citing unauthorized scraping and misuse.

BLACK FOREST LABS

The creators of VQGAN, Stable Diffusion, Latent Diffusion, and other startups have raised $30m+ dollars and started a new company. They have released new flagship image generation models which are extremely capable and come in a variety of tiers.

OPENAI PLEDGES TO GIVE US AI SAFETY INSTITUTE EARLY ACCESS TO ITS NEXT MODEL

OpenAI CEO Sam Altman announced a collaboration with the U.S. AI Safety Institute to provide early access to the company's next major generative AI model for safety testing, aiming to address concerns that OpenAI has deprioritized AI safety.

THE EU'S AI ACT IS NOW IN FORCE

The EU's risk-based AI regulation began on August 1 with staggered compliance deadlines categorizing AI applications into low/no-risk, high-risk, and limited risk tiers, imposing transparency, risk management, and penalties for violations. Standards for high-risk and powerful general-purpose AI models will be finalized by April 2025.

STABLE FAST 3D

Stability AI has released a great new fast and powerful 3D generation model. It uses the same commercial license as the company's previous models.

A MASSIVE VIDEO-LANGUAGE DATASET

MMTrail is a vast dataset that features over 20 million trailer clips with high-quality multimodal captions.

ADVANCING REINFORCEMENT LEARNING FOR INDUCTIVE REASONING

ARCLE is an environment to aid reinforcement learning studies using the Abstraction and Reasoning Corpus (ARC).

MISHAX FOR MECHANISTIC INTERPRETABILITY

DeepMind has released a library for studying language models via MI. The library helps with running models and functions from complex codebases without tons of importing headaches.

TORCH CHAT

The PyTorch team has released a great sample library for chatting with language models locally. It includes a robust sample system and can run the latest Llama 3.1 models.

ENGINE CORE

Engine Core demonstrates a pattern for enabling LLMs to undertake tasks of a given scope with a dynamic system prompt and a collection of tool functions.

ON SPEAKING TO AI

ChatGPT's new Advanced Voice mode and the new AI-powered Siri have contrasting philosophies. Copilot systems focus on safety and privacy with small models while Agent systems like ChatGPT Voice use powerful, multimodal models for more natural and dynamic interactions. This highlights the tension between lower risk, less capable systems and those offering more control and potential benefits.

HOW THIS BRAIN IMPLANT IS USING CHATGPT

Synchron has integrated OpenAI's ChatGPT with its brain-computer interface (BCI) technology to aid paralyzed patients in faster communication. Its BCI, called a stentrode, can interpret thought commands. It now offers AI-generated response options, potentially expanding to include multimodal inputs. Synchron is aiming for FDA approval and envisions evolving its AI integrations based on patient needs.

HEEYO BUILT AN AI CHATBOT TO BE A BILLION KIDS' INTERACTIVE TUTOR AND FRIEND

Heeyo, a startup founded by Xiaoyin Qu, has launched an AI-powered app that offers interactive games and a chatbot for children aged 3-11. The app aims to provide personalized learning experiences while prioritizing safety with features like data protection and content designed by child development experts. Despite potential concerns about AI for kids, Heeyo has secured $3.5 million in seed funding. It is positioning itself as a safe educational alternative to popular gaming and video platforms.

CEREBRAS IPO

Cerebras Systems announced a proposal for IPO to the SEC.

AT THE OLYMPICS, AI IS WATCHING YOU

Paris prepared for the 2024 Olympics with heightened security, employing AI to analyze CCTV footage at train and metro stations for potential threats.

LLMS BREACH A THRESHOLD

Meta's recent release of the open source LLM Llama 3.1, with 405B parameters and trained on 3.8x10^25 FLOPs, sparks debate on the effectiveness of FLOPs as a regulatory threshold.

WHITE HOUSE SAYS NO NEED TO REGULATE OPEN SOURCE

The White House says there is no need to regulate open source systems for AI in their current form and to allow the community to continue developing the technology.

SAMSUNG HINTS AT NEW PRODUCTS AS IT BETS ON AI TO DRIVE UPGRADES TO ITS LATEST FOLDABLE PHONES

Samsung's mobile chief, TM Roh, outlined the vision for Galaxy AI at a Paris event that showcased new AI features in Samsung products, including an evolution of Bixby driven by Samsung's LLMs. Roh discussed potential monetization strategies, including subscriptions for AI services, though plans are still under consideration.

ELON MUSK CALLS GROK 'THE MOST POWERFUL AI BY EVERY METRIC' BUT 'SECRETLY' TRAINS THE NEW MODEL WITH YOUR X DATA BY DEFAULT

Elon Musk has announced that his AI company is training its Grok LLM using a powerful AI training cluster, aiming to create the most powerful AI by December. X is collecting user data by default to train Grok. This can be disabled via a setting only available on the web app. There are concerns about AI 'inbreeding' and chatbots reaching their peak amid growing discussions on AI data usage and potential regulatory impacts.

SAE FOR GEMMA

An open explorer for features in Gemma 2B. It is a beautiful interface and an exciting application of Sparse AutoEncoders.

MACHINE UNLEARNING IN GENERATIVE AI: A SURVEY

This comprehensive survey explores machine unlearning in Generative AI. It covers problem formulation, evaluation methods, and the strengths and limitations of various techniques.

OBJECT-LEVEL PERCEPTION IN VIDEOS

Elysium represents a step towards enabling Multi-modal Large Language Models (MLLMs) to perform object tracking and related tasks in videos.