Search the archive

STABILITY AI ANNOUNCES PAID MEMBERSHIP FOR COMMERCIAL USE OF ITS MODELS

Stability AI is now offering a subscription service that standardizes and changes how customers can use its models for commercial purposes.

OUTBUNDLY

AI-powered cold email outreach.

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

Download your free copy here 📩

INTEL UNVEILS NEW AI CHIP

Intel has launched its 5th Gen Xeon processors, focusing on AI capabilities for cloud, network, and edge use cases. These chips are part of Intel's strategy to compete in the AI accelerator market dominated by Nvidia.

BYTEDANCE IS SECRETLY USING OPENAI’S TECH TO BUILD A COMPETITOR

ByteDance, TikTok's parent company, has been secretly using OpenAI's technology against its terms of service to develop a competing language model named Project Seed. Internal communications suggest plans to "whitewash" evidence of this misuse.

ATLASSIAN WELCOMES AI TO THE TEAM

Atlassian has launched its first set of AI-powered services across its cloud platform. Leveraging 20 years of data insights, Atlassian Intelligence aims to optimize teamwork by automating routine tasks, accelerating individual productivity, clarifying company jargon, and empowering data-driven decisions. Major features available include Generative AI in the editor, AI-powered summaries, and natural language automation.

GRAPH NEURAL NETWORKS FOR BETTER NODE CLASSIFICATION

Researchers have developed a new method called DSF to improve spectral graph neural networks. By introducing node-specific filter weights, DSF can better handle complex networks like the World Wide Web.

EVALUATING AND MITIGATING DISCRIMINATION IN LANGUAGE MODEL DECISIONS

This article discusses a method for proactively evaluating the potential discriminatory impact of language models. The method includes generating a wide array of potential prompts for various decision scenarios and demographic information variations. Strategies for mitigating both positive and negative discrimination primarily involve careful prompt engineering.

SMOOSEG: SUPERIOR SEMANTIC SEGMENTATION

This study introduces SmooSeg, a new approach that utilizes the principle that nearby features in images often share similar meanings.

OPEN SOURCE MEDIA GENERATION TOOLKIT

Amphion is a new tool kit for media generation. It supports music and speech and the team is working towards adding more models quickly.

FASTER DIFFUSION MODELS

A study on UNet encoders in diffusion models reveals their stable behavior compared to decoders. This insight led to a new encoder propagation scheme, significantly speeding up the process for tasks like text-to-image and text-to-video generation.

GOOGLE BIG VISION ADDS WEIGHT SHARDING

A hidden gem of AI, the Big Vision project is a great library for research and production of state-of-the-art vision systems. They just added weight sharding to make it easier to train on multi-node systems.

THOUGHTS ON JAXTYPING

Shape errors in ML are challenging to debug and are often only caught when you try running the model. By checking shapes as types you can eliminate much of this challenge and develop more quickly.

LEGALTECH X AI: THE LIGHTSPEED VIEW

Lightspeed’s take on the legaltech landscape as it relates to AI is interesting and transparent from the view of what’s-good-for-VCs-is-good-for-everyone take. There are a few interesting insights and only time will tell if they are on the right track or not.

NEW SEQUENCE MIXERS

The creators of Mamba (and many more models) have released a cool blog post outlining basic sequence mixing architectures that give pretty dramatic speed ups relative to standard Transformers.

MICROAGENTS

Agents capable of self-editing their prompts and Python code.

&FACTS

AI-powered product research for DTC brands.

THE COMPETITION IS COMING FOR NVIDIA

While things have been frothy in the AI space, we’re still barely a year into this new world order and the really interesting silicon has yet to hit. However, the competitive landscape is shaping up to be cutthroat and everyone is gunning for Nvidia.

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

Get started today.

META RELEASES HUGE FIRST-PERSON VIDEO DATASET

Video understanding is an old field that is just now starting to find success. The main driver of success is quality datasets. This work from Meta includes many sensor data sources and hand poses. It provides baseline performance for existing methods.

SNAPCHAT+ SUBSCRIBERS CAN NOW CREATE AND SEND AI-GENERATED IMAGES

Snapchat is introducing new AI-powered features for Snapchat+ subscribers, including an AI image generator that creates images from text prompts and a Dream selfie feature that creates fantastical images with friends. A new AI extend tool allows users to automatically expand and fill in the background of photos. These features are rolling out regionally to over 7 million Snapchat+ subscribers.

GOOGLE BRINGS GEMINI PRO TO VERTEX AI

Google has launched Gemini Pro, a lightweight version of its advanced AI model Gemini Ultra, for public preview in Vertex AI, offering text and image processing capabilities and customizable contexts for various use cases.

STABILIZING TRAINING BY PREVENTING ENTROPY COLLAPSE

Transformers, as amazing as they are, are often hard to train, with their many instabilities. One of the main issues is when an Attention matrix’s entropy collapses. This work is a way to prevent that with a simple reparametrization.

ESTIMATING LIGHTING IN IMAGES WITH DIFFUSION MODELS

This project introduces a novel technique for estimating lighting in images, overcoming the limitations of current methods that rely on HDR panorama datasets. By using diffusion models to render chrome balls into standard images, the technique reveals a unique relationship between chrome balls and diffusion noise.

LANGUAGE MODELS FOR AUTONOMOUS DRIVING

DriveMLM is a new framework that uses large language models to enhance autonomous driving. This system, which combines language decisions with vehicle controls, not only integrates with existing autonomous driving systems, but also shows better performance in simulations.

data LLM-ready

hard to find, hard to work with, and hard to clean. Unstructured

extracts and transform complex data (webpages, docs, images…) for use with every major vector database and LLM framework [16]. “Unstructured has solved the most difficult part of building an LLM application.” — LangChain Co-founder Harrison Chase. Make your

COFRAME COFFEE

Coffee, launched last week, helps streamline frontend development with AI by plugging into existing codebases. The project focuses specifically on a first-class DX based on learnings the Coframe team has gathered from their own experience in generating 80%+ of their frontend with AI.

FRAMEWORK FOR LLM APP EVALUATION

DeepEval is an all-in-one evaluation framework for LLM apps that allows users to easily identify and iterate on unsatisfactory LLM outputs. It currently offers 14+ evaluation metrics such as hallucination, summarization, G-Eval, RAGAS, etc., for users to evaluate entire datasets in a Pytest-like or component-based fashion.

TROUBLESHOOTING PYTORCH MEMORY ISSUES

One of the biggest challenges for practitioners using Pytorch has been figuring out why things run out of memory. Some amazing new tools in Pytorch 2.1 give insight into memory usage. It can even be used to classify the usage into familiar buckets (e.g., activations, gradients).

AMD COMMUNITY PERFORMANCE BENCHMARKS

MI300X is AMD’s flagship AI accelerator that, with ROCm 6, is now close to NVIDIA performance for inference workloads. This is great news for the community, as it opens up alternative chips for emerging AI companies.

FINE TUNING MISTRAL 7B ON MAGIC THE GATHERING DRAFTS

An experiment with fine-tuning to enhance GPT-4's capabilities using Magic: The Gathering draft as a test case led to a fine-tuned 7B parameter model that closely matched human-level performance.

THE WHERE, WHEN, AND HOW OF AI

Leading founders and venture investors talk about what's happening in AI, from the application of LLMs to enterprise innovation. If the pace of “things you should know about AI” is somewhat overwhelming, this is a good quick recap.

MMLU PROMPT TEMPLATES

Medprompt+ from Microsoft is the current best prompting method for MMLU. Microsoft has released the template along with many other chain-of-thought style templates that are standard across the evaluation community.

PRICEGPT

Analyze your (or a competitor’s) pricing page.

AGILITY IS USING LLMS TO COMMUNICATE WITH ITS HUMANOID ROBOTS

Generative AI and large language models are transforming robotics. Agility's bipedal robot Digit can now perform complex tasks based on natural language commands, signaling a shift towards more human-like, intuitive robotic interactions.

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

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

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 [6] allows Cody to autocomplete, explain, and edit your code with much richer context [6].

DEEPMIND’S AI UNCOVERS NOVEL MATHEMATICS

DeepMind published a paper in Nature about FunSearch, which uses LLM-guided function search to find solutions to open problems in mathematics and computer science. It tackled the capset problem and discovered some of the largest capsets to date.

OPENAI SUPERALIGNMENT GRANTS

OpenAI is distributing a number of grants for students and researchers hoping to work on super alignment. Applications are due Feb 18th.

FAST AI FOUNDER LAUNCHES NEW FOR PROFIT RESEARCH LAB

Answer AI is a new lab that isn’t aiming to build new models, but instead trying to figure out truly productive and valuable uses for existing models. It aims to do fundamental research for organizations to help them unlock AI-enabled use cases.

WEAK TO STRONG GENERALIZATION

This new result from the OpenAI super alignment team (with code) suggests that you can use much weaker supervisor models to steer or align a far more powerful model, an analog to weak humans aligning super intelligent models. They used GPT-2 to recover much of the alignment performance of GPT-4. One important point they note is that this method has a tractable path for dramatic improvement, unlike RLHF-like methods.