Search the archive

IMPROVED VIDEO MOMENT RETRIEVAL AND HIGHLIGHT DETECTION

UVCOM is a new framework that better handles the distinct requirements of Video Moment Retrieval (MR) and Highlight Detection (HD).

DEEPSEEK LLM EVALUATION

Deepseek coder was an amazing model that reached amazing code synthesis performance. It now has a 67B parameter chat LLM that dramatically outperforms Llama 2 70B.

NEW LIGHT-BASED CHIPS OFFER 3X MORE COMPUTER PER WATT THAN GPUS

This recent paper published in Nature details a CNN trained for classification that processes each frame in just 72ns and uses far fewer watts than traditional computer systems. With the recipe of AI continuing towards more FLOPs, more data, and larger models - these chips might be necessary to reach future levels of performance.

➡️_YOU CAN SIGN UP FOR FREE HERE…_

LLAMAFILE, BUILD ONCE RUN ANYWHERE

llamafile is a combination of llama.cpp and Cosmopolitan libc. This means that you can target essentially any CPU on any OS and bundle your model in a single executable file.

GETTING STARTED WITH LLAMA

This is a blog post from Meta outlining how to use its llama language model. The guide includes fine-tuning and deployment advice.

PREDICTING DRIVING SCENARIOS IN 3D

OccWorld, a self-supervised AI model, is designed to predict future driving scenarios by understanding and forecasting the movements of surrounding agents and map elements in 3D.

CHATGPT’S TRAINING DATA CAN BE EXPOSED VIA A “DIVERGENCE ATTACK”

Research reveals that large language models, including ChatGPT, can memorize and inadvertently reveal specific training data, highlighting privacy concerns, particularly in larger models.

WHAT YOU’RE FEELING IS SHINY OBJECT SYNDROME

The hype around OpenAI's rumored AI breakthrough, Q*, highlights the tech community's tendency to rapidly shift focus and speculate on the next big advancement in AI, often with limited information, exemplifying the "Shiny Object Syndrome."

IN THE AGE OF AI, GOOGLE EXPERIMENTS WITH BOLD CHANGE TO SEARCH

Google Search is being upended by AI and it can only blame itself for sitting on tech it has had for years. That aside, integrations with the next generation of AI tooling are springing up across Google’s suite of tools and are set to make a significant impact.

DATAIKU ANNOUNCES DATABRICKS AS LATEST LLM MESH PARTNER

Dataiku and Databricks are collaborating to further the reach and capability of Dataiku's LLM Mesh Partner Program. The partnership aims to create a simpler way to achieve generative AI-driven business transformations while allowing the enterprise to capitalize on the immense potential of LLMs.

WINN.AI (PRODUCT)

Winn.AI handles all of your sales busywork so you can type less and win more.

AI TRANSLATION TOOLS HELP SINGERS BREAK DOWN BORDERS

YouTube, Mr. Beast, and a South Korean label are using AI to dub video content into various languages, though the cost is seen as high by some creators.

MLCOMMONS ALGOPERF TRAINING BENCHMARK COMPETITION

META AI INTRODUCES CICERO

Meta’s CICERO is the first AI agent to achieve human-level performance in the strategy game Diplomacy.

STABILITY AI RELEASES SDXL TURBO

Stability AI has released a new model that can generate high resolution images in 207ms. It doesn’t use LCM, which is what most of the field has moved towards. Instead, it uses a novel distillation technique called Adversarial Diffusion Distillation. Its weights and code have been used for research.

ENHANCING PRIVACY IN SIGN LANGUAGE VIDEOS

DiffSLVA is an innovative method for anonymizing sign language videos while preserving linguistic content. This technique, which uses pre-trained diffusion models and a specialized module for facial expressions, overcomes previous limitations by not requiring precise pose estimation.

ENHANCING DATASET DISTILLATION WITH GENERATIVE DIFFUSION TECHNIQUES

A new method using generative diffusion techniques has been developed to create surrogate datasets that are more representative and diverse while requiring significantly less computational resources.

ADVANCED VIDEO SUPER-RESOLUTION THROUGH DIFFUSION MODELS

StableVSR is a novel approach to video super-resolution (VSR) that leverages Diffusion Models and a Temporal Conditioning Module to enhance the quality of upscaled videos.

EFFECTIVE PROMPTING LIBRARY FOR LARGE MULTIMODAL MODELS

Set of Mark prompting is a clever way to dramatically improve vision+language model performance. This library from Roboflow gives you similar control and makes it easier to build robust vision pipelines with these models. It even supports open models like LLaVa.

TANUKI

Easily build LLM-powered apps that get cheaper and faster over time.

CREATING REALISTIC 3D HUMAN AVATARS FROM VIDEOS

This project proposes "Animatable Gaussians," a novel technique that combines 2D CNNs and 3D Gaussian splatting to create more lifelike and detailed human avatars from videos.

A DATA-DRIVEN LOOK AT THE RISE OF AI

This is a slide-heavy post from the Cerebral Valley AI Summit that takes a data approach to looking at the growth of AI. We all hear it constantly, but the data backing it up is pretty immense. Most interesting is the growth and subsequent collapse of developer interest.

HOW MUCH DOES IT COST TO USE AN LLM?

Reviewing the details on common queries to an LLM across the different models reveals that it is really cost-bound at the moment. With the latest OpenAI announcements and advancements with the open source models, that may be set to change.

THE 10-YEAR “OVERNIGHT” SUCCESS STORY OF CASETEXT (17 MINUTE AUDIO)

Casetext sold to Thomson Reuters for $650M in what looks like an acquisition swept up in the AI craze. However, this was no overnight success, with the company having started in 2023. A walk down memory lane along with what it took to persevere.

OPENAI UNLIKELY TO OFFER BOARD SEAT TO MICROSOFT

SPORTS ILLUSTRATED PUBLISHED ARTICLES BY AI-GENERATED AUTHORS

Sports Illustrated has secretly been publishing AI-generated articles by fake AI authors.

OPENAI DATA PARTNERSHIPS

OpenAI launched Data Partnerships to collaborate with organizations to produce datasets to train AI models. The initiative already includes partners like the Icelandic Government and the non-profit Free Law Project.

AMAZON AND SALESFORCE EXPAND PARTNERSHIP TO ADD NEW AI CAPABILITIES

Salesforce and Amazon Web Services have significantly expanded their partnership to make it easier for customers to manage data across both platforms and integrate generative AI technologies into their applications and workflows.

TESLA STARTS RELEASING TO EMPLOYEES FSD V12

Tesla has started releasing its Full Self-Driving v12 update to employees. The biggest difference with the update is how vehicle controls will be taken over by neural nets rather than being hard-coded by engineers. Tesla may be able to push the update to customers by the end of the year. Elon Musk has previously said that he believes that Tesla will achieve true self-driving capability by the end of the year.

INTRODUCING CLAUDE 2.1

Claude 2.1, the updated AI model from Claude, now offers a 200K token context window and a 2x decrease in false statements. The model also introduces a beta feature, tool use, allowing Claude to integrate with users' existing processes and APIs. The update seeks to increase trust, reliability, interoperability, and cost efficiency in AI applications.

STARLING-7B: INCREASING LLM HELPFULNESS & HARMLESSNESS WITH RLAIF

A group of researchers at Berkeley trained a new state-of-the-art 7B parameter model using synthetic preference data. This blog talks about the novel challenges of training reward models (e.g., the position of the example in the list changes its ranking) and how they overcame them. The resulting model is available along with the trained reward model.

360-DEGREE PANORAMIC IMAGING USING CIRCULAR BLENDING TECHNIQUE

Researchers have devised a unique circular blending method to address the longstanding challenge of creating seamless 360-degree panoramic images. This technique is a key feature in their innovative models for generating panoramic images from text and single images.

IMAGE-TO-PAINTING

Researchers have developed a new technique for transforming images into paintings that mimic human-like quality and style.

INSANELY FAST WHISPER WITH SPEAKER DIARIZATION

A somewhat opinionated CLI tool for transcribing audio files. It also now includes speaker segmentation and diarization (e.g., recognizing speaker 1 vs speaker 2). The claim is they can transcribe 2.5 hrs of audio in 100s, even on some of the largest models available.

AGENCY

Library designed for developers to explore the potential of LLMs and other generative AI through a clean, effective, and Go-idiomatic approach.

IMPROVING INSTRUCTION QUALITY BY COACHLM

CoachLM introduces a novel AI technique to refine instruction datasets for LLMs. This method uniquely improves datasets by revising rather than removing low-quality samples, significantly elevating the effectiveness of instruction-following in LLMs.

EXPLORING THE GROWING CONVERGENCE BETWEEN BLOCKCHAIN AND AI

This report, which polled 600+ IT executives from around the world, digs into the primary drivers behind the ever-growing appetite for AI, the most popular use cases, and the primary hurdles hindering broader adoption. Notably, it also debunks the notion that blockchain and AI are not complementary technologies.

GOD HELP US, LET’S TRY TO UNDERSTAND AI MONOSEMANTICITY

Researchers from Anthropic have developed a method to interpret the complex workings of AI by simulating a larger AI within a smaller one, revealing that AI neural networks can encode information in a sophisticated manner, akin to superposition in physics. By applying this technique to a simple AI, they discovered that it could represent distinct concepts, like "God", in individual features, and they speculate that the same approach might reveal deeper insights into both artificial and biological neural systems, potentially leading to more efficient and safer AI development.

REBUILDING ORGANIZATIONS FOR AI

The integration of AI in organizations is transforming traditional work processes, enabling teams to work more efficiently by automating tasks and decision-making. As AI evolves rapidly, it's crucial for organizations to adapt by encouraging team experimentation with AI, preparing for future advancements, and acting swiftly to stay competitive.

ALIGN API OFFERS PLUG AND PLAY GUARDRAILS FOR YOUR AI APP (PRODUCT)

The Align API proactively identifies NSFW/Sensitive prompts before they get sent for inference. This saves unnecessary costs, protects your brand, and prevents you from getting banned by platforms like OpenAI, Apple’s App Store, or Google’s Play Store.

INTRODUCING STABLE VIDEO DIFFUSION

Stability AI highlights the addition of Stable Video Diffusion to its portfolio of open-source models. Its suite of AI models expands across various modalities including image, language, audio, 3D, and code.

CHATGPT WITH VOICE

ChatGPT with voice is now available to all free users.

WHAT IS Q*?

A recent leak that stated the OpenAI leadership debacle was caused by an internal breakthrough that combined search and RL. This post outlines one theory explaining what might actually be happening with this new method.

INSIDE OPENAI, A RIFT BETWEEN BILLIONAIRES AND ALTRUISTIC RESEARCHERS UNRAVELED OVER THE FUTURE OF ARTIFICIAL INTELLIGENCE

An all-encompassing recap of the OpenAI fiasco. This will be studied for generations. If you haven’t followed the entire saga, this summary is a must read including all of the interesting angles and fall-out.

ANTHROPIC LOWERS PRICES AS AI COMPETITION HEATS UP

Anthropic included a price cut for tokens generated with the release of the newest version of Claude. This comes from both open and closed model pressure.

ZIPLORA

LoRAs are low rank learned matrices that change model behavior but are cheaper than regular fine-tuning. This work proposes an effective way to merge LoRAs together while maintaining the novel information from each.

IMPROVING MEDICAL IMAGE SEGMENTATION

The introduction of SegVol, a universal model for medical image segmentation, marks a new era in clinical analysis. Trained on a comprehensive collection of CT scans, SegVol excels in segmenting an extensive variety of anatomical categories.

VISUAL IN-CONTEXT PROMPTING: A NEW ERA IN AI VISION TASKS

This innovative approach accommodates a range of prompts and contexts, greatly boosting performance in segmentation tasks and showcasing impressive results in open-set challenges.