Search the archive

MAMBA CHAT

Mamba is a non-transformer-based model with amazing performance. It has been chat-tuned with SFT and seems to have substantially better out-of-the-box retrieval ability than similarly sized transformers.

LLAMA IN MLX

Just a day after the MLX framework from Apple was released, folks have already enabled Llama 2 models to run on the new framework.

MISTRAL MLX EXAMPLE

An example of generating text with Mistral using MLX.

SINGLE-IMAGE 3D VIEW SYNTHESIS

Free3D is an innovative approach that creates 3D views from just one image, without the need for slow and bulky 3D models.

ML SYSTEM DESIGN: 300 CASE STUDIES TO LEARN FROM

This site contains a database of 300 case studies from over 80 companies that share practical machine learning use cases and learnings from designing ML systems. The database can be filtered by industry or ML use case. Tags based on recurring themes are available.

OPENAI COO BRAD LIGHTCAP TALKS ABOUT CHATGPT LAUNCH, DEV DAY, AND HOW SAM ALTMAN THINKS

OpenAI's achievements over the past year were recently eclipsed by a rollercoaster couple of weeks that involved Sam Altman being ousted and reinstated as well as a new board. This post contains an edited interview with OpenAI's COO Brad Lightcap where he discusses ChatGPT, OpenAI's Dev Day, and Altman.

activity. Start here

GOOGLE’S BEST GEMINI DEMO WAS FAKED

Google's recent "Hands-on with Gemini" video, which showcases the capabilities of its new Gemini AI model, has drawn criticism for being misleading. While the video demonstrates Gemini responding to various multimodal inputs, it was later revealed that the interactions were not live but rather carefully curated and prompted using still images and text, differing significantly from the real-time, intuitive interactions implied in the video.

UIDRAW

Draw and build a website from your phone.

SEO GPT (PRODUCT)

Boost your website’s SEO instantly right inside ChatGPT.

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

GOOGLE GEMINI

Google's flagship model Gemini is natively multimodal with text capabilities matching GPT4 while also being able to operate on many other data sequences. It also trained Alpha Code 2, which is an 85th percentile coder on code forces. Alpha Code 2 will be available in Google’s API after December 13th.

META’S AI IMAGE GENERATOR IS AVAILABLE AS A STANDALONE WEBSITE

Meta has launched 'Imagine,' a standalone image generator with its own website featuring generative AI that creates images from text prompts.

MICROSOFT’S COPILOT IS GETTING OPENAI’S LATEST MODELS

Microsoft is enhancing its Copilot service with new features, including OpenAI's latest GPT-4 Turbo model, an upgraded DALL-E 3 model, a new code interpreter feature, and a 'Deep Search' function in Bing.

DESCRIBE DIFFERENCES BETWEEN SETS OF IMAGES

This paper uses natural language to describe what is different between two collections of images. This is a novel problem and a challenging one. The solution involves captioning and reranking the individual images and summarizing them with a language model.

3D OBJECT DETECTION WITH DIFFUSION MODELS

Researchers have developed Diffusion-SS3D, a new method to improve semi-supervised 3D object detection using diffusion models. This approach involves adding noise to object size and class label distributions in 3D space and then using the diffusion model to denoise and generate better bounding box outputs.

A BENCHMARK TO TEST LARGE MULTIMODAL MODELS

This project introduces BenchLMM, a benchmark designed to test the robustness of Large Multimodal Models like GPT-4V and LLaVA against various image styles.

APPLE’S NEW ARRAY FRAMEWORK FOR MACS

Apple quietly released a cool Array framework that is accelerated on Mac devices because of unified memory. It is simple and clean and provides some GPU support as well.

LIGHTNING FAST DIFFUSERS INFERENCE

Stable Fast is an inference engine for HuggingFace diffusion models.

INTRODUCING THE OPENAI SWITCH KIT

Switch your project to open source in just a few lines of code.

EXPLORING LEAP-OF-THOUGHT ABILITIES IN LLMS

This study introduces the use of the Oogiri humor generation game to explore the Leap-of-Thought capabilities in multimodal LLMs. This approach challenges LLMs to exhibit non-sequential thinking, a skill crucial for generating creative and humorous responses to various types of multimodal information.

OPTIMIZING LLMS FOR REAL-WORLD APPLICATIONS

Lightspeed shares insights from Google and TitanML about the ins and outs of optimizing LLMs whether via prompting or fine-tuning.

SLOW-BURN AI: WHEN AUGMENTATION, NOT AUTOMATION, IS THE REAL THREAT

While automation will likely be the way of the future, today, it’s all about augmenting your existing processes/self without having to rip and replace everything. Don’t fall into the trap of just augmenting yourself to do your job better - thinking through what and how to automate things and viewing your role holistically will allow you to find the best use of AI for your needs.

USING META’S NEW EFFICIENTSAM (JUPYTER NOTEBOOK)

Segment Anything is a great model, but it is slow to run. Efficient Segment Anything is a new model that uses a number of optimizations to run significantly faster.

ELON TOLD OPENAI TO MOVE FASTER RIGHT BEFORE HE LEFT THE COMPANY

Right before leaving the company, Elon Musk implored OpenAI to move faster, a departure from his typically more skeptical comments toward AI.

COPILOTKIT

Build in-app AI chatbots and AI-powered Textareas into React web apps.

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

AI ALLIANCE LAUNCHES AS AN INTERNATIONAL COMMUNITY FOR SAFE AND OPEN AI

The AI Alliance, led by IBM and Meta, unites over 50 global organizations to foster open, responsible AI development. It focuses on setting AI standards, advancing hardware, and promoting AI education and skills. Members include major tech firms, universities, and research institutions. The Alliance emphasizes safety, diversity, and equitable access to AI innovation.

ELON’S AI FIRM XAI FILES TO RAISE UP TO $1B

Elon Musk's artificial intelligence startup xAI has filed with the U.S. securities regulator to raise up to $1 billion in an equity offering.

GOOGLE’S NEW AI EXPERIMENT COMPOSES ABSTRACT MUSICAL CLIPS INSPIRED BY INSTRUMENTS

Google's Instrument Playground allows users to create unique music pieces inspired by over 100 global instruments using MusicLM. Users can input an instrument name with an adjective for style and the AI generates a 20-second audio clip, which can be further customized with options like "Ambient," "Beat," and "Pitch" in a sequencer, allowing the creation and download of personalized musical compositions.

LANGUAGE MODEL SELF-TEACHING FOR DOMAIN ADAPTATION

When trying to use language models in domains that require niche knowledge, you can either fine-tune or use some retrieval techniques. Both have drawbacks. This novel method uses self-generated synthetic data to better learn the information at test time. It shows marked improvements in standard adaptation benchmarks over both fine-tuning and RAG.

DEEP DIVE: RECENT ADVANCES IN INSTRUCTION TUNING

TRACKING WITH 3D GAUSSIAN SPLATTING

Gaussian Splatting is an amazing method for 3D reconstruction. In this case, the scientists are working to remove COLMAP from their visual mapping and tracking systems. The results are compelling, though there is still some heavy computation that needs to happen.

DIFFUSION VISION TRANSFORMERS FOR HIGH-RESOLUTION IMAGE GENERATION

This project explores the effectiveness of vision transformers in diffusion-based generative learning, unveiling the Diffusion Vision Transformers (DiffiT). This model combines a U-shaped encoder-decoder architecture with a novel time-dependent self-attention module.

3D-CONSISTENT IMAGES FROM A SINGLE VIEW WITH DIFFUSION MODELS

This project presents Zero123++, a model that generates consistent multi-view images from just one input image using diffusion principles. By utilizing pretrained 2D models, Zero123++ addresses challenges like texture quality and alignment issues.

ADVANCING SALIENT OBJECT DETECTION IN RGB-D VIDEOS

STYLE ALIGNED TEXT TO IMAGE GENERATION

Really stunning work by Google built on SDXL that shares attention for generations to maintain consistent styles. This method crucially doesn't require any fine-tuning.

RESPONSIBLE INNOVATION LABS LAUNCHES FIRST PRO-INNOVATION RESPONSIBLE AI PROTOCOL FOR STARTUPS

Nonprofit Responsible Innovation Labs (RIL) has launched a voluntary industry-driven protocol for responsible AI, particularly for startups and their investors. The protocol, supported by 35 leading venture capital funds, is designed to ensure startups integrate responsible AI practices as they scale. This move aligns with the Biden Administration's vision for responsible AI.

TWO TITANS ON THE FUTURE OF AI (WITH REID HOFFMAN AND VINOD KHOSLA)

A good recap of the 45+ minute talks between Reid Hoffman and Vinod Khosla with each diving into a variety of topics from AI to “techno-optimist” manifestos. Both titans have great perspectives on things to come and how the tech industry should be regulated.

NEW REPORT ILLUMINATES WHY OPENAI BOARD SAID ALTMAN “WAS NOT CONSISTENTLY CANDID”

Sam Altman's brief ousting as OpenAI CEO was linked to accusations of manipulating board members to dismiss Helen Toner, who had critiqued OpenAI's AI safety approach, a situation variously viewed as manipulative or a clumsy attempt at control.

SOLVE INTELLIGENCE HELPS ATTORNEYS DRAFT PATENTS FOR IP ANALYSIS AND GENERATION

Solve Intelligence is an AI-native platform that helps draft high-quality patents efficiently. Since their launch in July, more than 25+ IP firms globally are now using them, with customers reporting 60-90% efficiency improvements. The company graduated from Y Combinator and recently announced their $3M seed round.

AIRBNB ACQUIRE GAMEPLANNER.AI

Airbnb has acquired AI company GamePlanner.AI, a company led by Siri co-founder Adam Cheyer and AI expert Siamak Hodjat. The team will focus on accelerating selected AI projects and integrating their tools with Airbnb's platform. Airbnb and GamePlanner.AI share a commitment to using AI to enhance human connection.

VECTOR ART (PRODUCT)

Create vector images with AI.

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

Start with 100,000 free monthly monitoring predictions

Tired of guessing whether your model is good enough? Openlayer

lets you build AI products with the rigor of traditional software — with testing, evaluation, and observability [5]. Manage models, identify edge cases, and detect production errors. It only takes. one line of code [5] to get started:

OPENAI AGREED TO BUY $51 MILLION OF AI CHIPS FROM A STARTUP BACKED BY CEO SAM ALTMAN

During his tenure, OpenAI CEO Sam Altman planned to purchase $51 million worth of AI chips from Rain AI, a company in which he personally invested. Despite his firing and reinstatement, OpenAI expressed possible future collaboration with Rain. This aspect highlights Altman’s intertwining roles as a vested investor and CEO, which contributed to his firing.

NOUS RESEARCH RELEASES HERMES 2.5 VISION MODEL

Built on top of the best 7B language model with SigLIP integration, this vision model is a powerful new open source text and vision model that can run on consumer hardware. One of the cool innovations here is the integration of function calling. The model is still in alpha because of a hallucination issue.