Search the archive

HAIKU DPO DATASET (HUGGINGFACE DATASET)

Dataset to help align models to write correct Haiku’s.

THE LAZY TYRANNY OF THE WAIT CALCULATION

The "Wait Calculation" concept advises delaying actions, like a space journey to Barnard's Star or starting certain projects, until technology sufficiently advances to save significant time and effort. However, this approach must be balanced against potential learning losses and the unpredictability of technological progress.

IMPROVING THE TRUSTWORTHINESS OF AI EXPLANATIONS

This project improves the Model Parameter Randomisation Test, a key tool in eXplainable AI (XAI), by introducing two new versions: Smooth MPRT and Efficient MPRT.

Register for free here

MAGNET MODELS RELEASED (HUGGINGFACE HUB)

Meta recently announced a suite of music and sound generation models that sound incredible. It has now released the models with inference code.

TINY NARRATIONS

Tiny Narrations is a text-to-speech read version of the popular (and small) Tiny Stories dataset. It uses XTTS2 on the SF Compute H100 cluster.

HUMANLINKER (PRODUCT)

An AI assistant that helps sales teams with personalized prospecting and efficient meeting preparation.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

MICROSOFT LAUNCHES COPILOT PRO

Microsoft has launched Copilot Pro, a premium version of its Bing Chat experience, priced at $20 per month per user. This advanced version integrates across devices, offers access to Microsoft 365 apps, features the latest AI models like OpenAI's GPT-4 Turbo, and includes enhanced AI image creation.

HOW OPENAI IS APPROACHING 2024 WORLDWIDE ELECTIONS

OpenAI is gearing up for the 2024 elections by preventing AI abuse, enhancing transparency in AI-generated content, and ensuring access to reliable voting information, thereby safeguarding election integrity.

ENHANCING SEARCH SKILLS OF LLMS

Researchers have developed INTERS, a new dataset designed to improve how large language models like LLaMA and Mistral handle information retrieval tasks.

CONVFORMER: AN EFFICIENT APPROACH FOR IMAGE SUPER-RESOLUTION

This paper Introduces ConvFormer, a novel Convolutional Transformer layer designed for lightweight image super-resolution tasks. This new technique effectively replaces self-attention modules with large kernel convolutions, allowing for efficient modeling of long-range dependencies with minimal computational costs.

EFFICIENT BIOMEDICAL IMAGE SEGMENTATION

A new graph relation distillation technique improves instance segmentation. This method efficiently transfers knowledge from larger, complex networks to smaller, high-performing ones, focusing on three key areas: instance features, their relationships, and pixel-level boundaries.

Cody is your new favorite teammate, using its deep understanding

of your codebase to explain, fix, and help you write code. Sourcegraph’s code graph [10] allows Cody to autocomplete, explain, and edit your code with much richer context [10].

🆓 Try Cody Pro — it’s FREE until February 2024

AUDIO-VISUAL EMOTION RECOGNITION

HiCMAE is a novel self-supervised learning framework designed to enhance Audio-Visual Emotion Recognition (AVER). This approach overcomes data scarcity challenges by using large-scale pre-training on unlabeled audio-visual data.

ADVANCING LLM COMPRESSION WITH ADDITIVE QUANTIZATION

This project suggests a new compression technique for LLMs using Multi-Codebook Quantization (MCQ). The method, adapted from Additive Quantization, sets a new standard in LLM compression, achieving higher accuracy at lower bit counts.

POINT-BASED IMAGE EDITING

This project introduces RotationDrag, a new method that improves point-based image editing, especially when users want to rotate elements within the image.

NOUS HERMES 2 MIXTRAL (X THREAD)

Nous Research has released SFT and DPO versions of the powerful Mixtral model. It excels at a number of tasks and is freely available on the Together API.

AI WILL TRANSFORM THE GLOBAL ECONOMY. LET’S MAKE SURE IT BENEFITS HUMANITY

The rapid advancement of artificial intelligence (AI) stands to revolutionize productivity and growth globally, but also poses risks of job displacement and increased inequality. An IMF analysis shows that nearly 40% of global employment could be impacted by AI, with advanced economies facing both higher risks and opportunities compared to emerging markets and developing countries.

AI IS NOT THE SOLUTION TO ALL OUR EDUCATIONAL CHALLENGES

The rapid evolution of AI in education is promising, yet it's unlikely to solve pressing educational problems soon, as experienced during a week-long experiential learning trip with 9th graders. While AI offers innovative curriculum and learning enhancement, addressing deeper issues like emotional well-being, student engagement, and trust in education remains crucial.

news. Read the highlights

AUTO GGUF QUANTIZATION (COLAB NOTEBOOK)

GGUF is a format for running models with llama cpp and supports a number of quantization types. This notebook automates the quantization, which may not work in the general case, but works for most models.

A MULTIMODAL GROUNDING MODEL FOR IMPROVED COMPREHENSION ACROSS MEDIA

LEGO, a new end-to-end multimodal grounding model, demonstrates advanced abilities in comprehending and grounding tasks across different modalities like images, audio, and videos.

SKIMAI (PRODUCT)

Transform your inbox into a hub of efficiency.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

OPENAI CHANGES POLICY TO ALLOW MILITARY APPLICATIONS

OpenAI has revised its usage policy to potentially allow military applications of its technologies, a change from its previous policy that explicitly prohibited such use.

USING AI, MIT RESEARCHERS IDENTIFY A NEW CLASS OF ANTIBIOTIC CANDIDATES

Utilizing deep learning, MIT researchers have discovered compounds capable of killing the drug-resistant bacteria MRSA responsible for over 10,000 deaths annually in the U.S.

I AM A STRANGE DATASET: METALINGUISTIC TESTS FOR LANGUAGE MODELS

There is a challenge type of phrase that is self-referential, for example, “the last word in this sentence is is”. Language models struggle immensely with this sort of language. This paper introduces a dataset and some evaluations to help improve language models' metalinguistic capability.

PIXART-Δ: FAST AND CONTROLLABLE IMAGE GENERATION WITH LATENT CONSISTENCY MODELS

PixArt has been a parallel research direction to the popular Stable Diffusion set of image generation models. This work brings faster generation with Latent Consistency Models and better control with ControlNet style prompting.

SLEEPER AGENTS: TRAINING DECEPTIVE LLMS THAT PERSIST THROUGH SAFETY TRAINING

Anthropic released some really interesting research that purposefully poisons a language model with a sleeper phrase that elicits some behavior. It found that it was unable to “align” this sort of model with the powerful system used for its production models. Meaning, once the model had been poisoned, bad behavior could not be removed with current tools.

ANIMATE ANYONE REPLICATION

AnimateAnyone is an amazing video control model that takes an image and skeleton motion as input and then animates the person in the control image. This code is an open replication of that work.

MULTILINGUAL OCR LINE DETECTOR

Line detection is where you put a bounding box around a line of text in an image or PDF. Surya is an extremely powerful multi-lingual line detection model.

AUTOMATICALLY NARRATE YOUR LIFE

Using a combination of GPT4-V, top-of-the-line text-to-speech, and some computer capture software, you can have folks like David Attenborough narrate everything that is happening in your life.

THE CASE FOR CYBORGS

An expert in human cognition and AI discusses the evolution from artificial narrow intelligence to general intelligence, proposing cyborg integration as a potential path to superintelligence.

PAST, PRESENT, AND FUTURE OF AI WITH VIJAY PANDE (40 MINUTE PODCAST)

A 40-minute reflection on AI with a peek at what’s to come.

PROMPT ALIGNED IMAGE PERSONALIZATION

Dreambooth is the current best method for personalizing an image model. Prompt alignment dramatically improves adherence to the prompt and is composable.

and LLM framework. Make your data LLM-ready

SEAMLESS EXPRESSIVE TRANSLATION

Meta has a new demo for seamless voice cloning and translation between languages.

VANNA

Use AI to interact with your database.

MICROSOFT WANTS TO AUTOMATICALLY LAUNCH ITS COPILOT AI ON SOME WINDOWS 11 DEVICES

Microsoft is currently testing an update in Windows 11's Dev Channel preview that automatically launches the AI-powered Copilot feature on widescreen devices with a screen size of at least 27 inches and a pixel width of 1920 pixels. This experiment, aimed at enhancing AI integration in Windows, is part of Microsoft's broader initiative to incorporate AI technology more seamlessly into its operating system.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

Watch on demand here ↗️

[7] to Cloudflare’s developer platform — including Workers AI

(GPU-powered compute) and Vectorize — and walk you through building and optimizing AI applications with Cloudflare. The series starts with AI foundations (embeddings, models, vector DBs), and moves on to practical sessions where you’ll deploy working AI applications.

AI POWERED BINOCULARS

Swarovski has developed a pair of AI powered binoculars that can identify 9000 different kinds of birds and wildlife. They can also capture photos and videos.