Search the archive

EDITANYTHING

A project that aims to Edit and Generate Anything in an image, powered by Segment Anything, ControlNet, BLIP2, and Stable Diffusion.

44 OF THE MOST PROMISING GENERATIVE AI STARTUPS OF 2023, ACCORDING TO VCS

Generative AI exploded onto the scene in 2022, bringing with it buzzy startups and funding rounds. Insider asked investors to name the most promising generative-AI startups of 2023. Notable companies include foundational models like Open AI and Stability AI, next layer tooling like LangChain and Fixie, and early stage startups like Cursor and DeepScribe.

OPENAI PLANS JAPAN EXPANSION

Sam Altman, co-founder and chief executive officer of OpenAI, said the organization is looking at opening a Japan office and expanding Japanese language services after meeting with Prime Minister Fumio Kishida.

BAIDU LAUNCH SERVICES

REVISITING THE CLASSIC: SCALING RNNS OVER SEQUENCE LENGTH

Context length is a hot topic in AI. We're trying to scale text models to longer and longer sequences. This work scales an RNN to a million context tokens. It's not directly applicable, but there are some lovely nuggets here.

GENERATIVE AGENTS: INTERACTIVE SIMULACRA OF HUMAN BEHAVIOR

Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools. In this paper, the authors introduce generative agents--computational software agents that simulate believable human behavior. In an evaluation, these generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentine's Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time.

PANOPTIC SEGMENTATION

Researchers from Google propose Video-kMaX, a unified approach for online and near-online Video Panoptic Segmentation, comprising a within-clip segmenter and a cross-clip associater. By leveraging clip-kMaX and HiLA-MB for these components, this method establishes a new state-of-the-art on KITTI-STEP, VIPSeg, and VSPW datasets, bridging the gap between online and near-online segmentation approaches.

A GUIDE TO PROMPT ENGINEERING

LIBRARY FOR AI ON TPUS

A challenge in training large text models is fitting them on your hardware accelerators. This is often easier on TPUs than GPUs due to Pjit capabilities of Jax. This library (which was used for the imminently popular FlanT5 models) is easy to use and seems to scale quite well.

TURBO PILOT CODING ASSISTANT

A faux pilot based local coding assistant for vs code that runs local with 4GB ram. It uses the Salesforce 6B codegen model.

REDDIT MODS BRACE FOR CHATGPT SPAM APOCALYPSE

BIDEN ADMINISTRATION WEIGHS AI REGULATION

OPENAI TO OFFER REMEDIES TO RESOLVE ITALY’S CHATGPT BAN

ALIBABA TO ROLL OUT IT’S CHATGPT RIVAL

AI-EQUIPPED EYEGLASSES CAN READ SILENT SPEECH

OPENAI INTRODUCES BUG BOUNTY PROGRAM

OpenAI will start paying people as much as $20,000 to help the company find bugs in its artificial intelligence systems.

QUORA INTRODUCES POE: THE UNIVERSAL AI MESSAGING CLIENT

AMAZON TELLS EMPLOYEES IT ISN’T FALLING BEHIND ON AI

Despite Amazon's absence in the generative AI race, executives maintain that the company is actively involved in the AI space. Swami Sivasubramanian, Amazon's VP of database, analytics, and machine learning, said that Amazon's 100,000+ machine learning customers are interested in using the company's technology for personalization, search engine results, and call center automation. While Amazon Web Services has partnered with AI companies like Stability AI and Hugging Face, the company has not launched a consumer-facing generative AI product, leading to speculation that it may be lagging behind its competitors.

CHATSPOT 3, NOW WITH GPT-4

AGENTS ARE THE NEW PARADIGM

What happens if you use language models to simulate an entire town? They ask each other out on Valentine's dates. This paper explores what happens if you give language models open-ended tasks and shows that you can observe remarkable human-like social capabilities from these agents, such as keeping schedules, planning activities, and reminiscing about previous interactions. This is one step in a broader trend of language models moving more towards agent-like paradigms, which dramatically changes the safety considerations for true model alignment.

A TUTORIAL INTRODUCTION TO REINFORCEMENT LEARNING

The paper provides a brief survey of Reinforcement Learning (RL), with a focus on Stochastic Approximation (SA) as a unifying theme. The scope includes Markov Reward Processes, Markov Decision Processes, Stochastic Approximation algorithms, and widely used algorithms like Temporal Difference Learning and Q-learning.

LLMMAPS - A VISUAL METAPHOR FOR STRATIFIED EVALUATION OF LLMS

The paper introduces LLMMaps, a visualization technique that allows for the stratified evaluation of Large Language Models (LLMs) based on Q&A datasets, providing detailed insights into their knowledge capabilities and enabling comparative analysis of multiple LLMs. The approach was tested on state-of-the-art LLMs and evaluated through user feedback.

SEGMENT ANYTHING CHROME EXTENSION

Since the Segment Anything Model from Meta is so lightweight, it can run directly in the browser. As a result, we can copy items from images directly to the clipboard - in the wild. It works just like the demo from Meta, but instead of displaying the segmented object, that object is copied to the system clipboard for you to paste as needed.

OPENPLAYGROUND

Openplayground is an LLM playground you can run on your laptop.

UPTRAIN

UpTrain is an open-source framework to observe ML applications built for engineers.

GO SMALL OR GO HOME

A GUIDE TO NATURAL LANGUAGE PROCESSING WITH PYTHON USING SPACY

EXPERTS WARN OF EFFECTS OF AI ON ELECTIONS

GERMANY CONSIDERS BANNING CHATGPT

MICROSOFT RELEASING EDGE’S AI IMAGE GENERATOR TO EVERYONE

GOOGLE CEO SUNDAR PICHAI SAYS SEARCH TO INCLUDE CHAT AI

Google plans to add conversational artificial-intelligence features to its flagship search engine, Chief Executive Officer Sundar Pichai said, as it deals with pressure from chatbots such as ChatGPT and wider business issues. Advances in AI would supercharge Google’s ability to answer an array of search queries, Mr. Pichai said in an interview with The Wall Street Journal. He dismissed the notion that chatbots posed a threat to Google’s search business, which accounts for more than half of revenue at parent Alphabet Inc.

GPT-4 IS A REASONING ENGINE

SEGMENT - DETECT - IN PAINT

A RECIPE FOR TRAINING LARGE LANGUAGE MODELS

ML PAPERS OF THE WEEK

Another great list of papers to refer to regularly. There is lots of hype circulating online at the moment and so it is good to get back to the source sometimes and understand the trends. There are many lists like this one, but it seems high quality and thoughtfully curated.

FULL PROOF GENERATION WITH PROOF REPAIR

One common problem with generative language models is their propensity to hallucinate. Folks have pointed out that sometimes code these models generate is subtly wrong, but looks plausible. The act of verifying is sometimes more important than generating in this case. Verification is hard though, especially for code. This paper studies proof generation which can be checked by a formal system. There are lots of great insights in this work, my favorite is that when given error messages, the system can prove 1.5% more proofs.

AUTOMATICALLY GENERATE LANGCHAIN TOOLS OR CHATGPT PLUGINS

[/%E2%80%8B%E2%80%8Bhttps://github.com/hey-pal/toolkit-ai?utm_source=tldrai]

LLAMA.CPP PROJECT REVOLUTIONIZES META'S LANGUAGE MODEL ACCESS

Georgi Gerganov's llama.cpp project optimizes LLaMA performance, boasting 100x faster load times and halved memory usage, enhancing the user experience on edge and personal computer devices.

INST-INPAINT: INSTRUCTING TO REMOVE OBJECTS WITH DIFFUSION MODELS

Introducing Inst-Inpaint, a novel inpainting framework that seamlessly removes objects from images based on natural language input, eliminating the need for time-consuming binary masks. Using the GQA-Inpaint dataset, the framework outperforms GAN and diffusion-based baselines, demonstrating quantitative and qualitative improvements.

PRETRAINING GAUSSIAN PROCESSES

IN AI RACE, MICROSOFT AND GOOGLE CHOOSE SPEED OVER CAUTION

STABILITY AI IS ON SHAKY GROUND

PYTHON OPTIMAL TRANSPORT UPDATE

Optimal Transport theory is a field with a rich history that has fully embraced computation. One of the standard frameworks has received an upgrade which improves speed, adds examples, and increases general reliability. If you want to compare two distributions, this is likely a good place to start.

SAVE MONEY BY REDUCING YOUR GPT-4 PROMPT

[/%E2%80%8B%E2%80%8Bhttps://www.promptreducer.com/?utm_source=tldrai]

AI SHELL

AI Shell is a CLI tool that converts natural language to shell commands.

TABBY

Tabby is an open-source alternative to GitHub Copilot.

CHROMA’S $18M SEED FUNDING

Another big funding announcement for an open source tool for AI. This time the fast growing embeddings database company raises cash. They’re hoping that folks will use their tools to add state and memory to different LLM applications. The goal is for this to reduce hallucinations and increase factuality. They’re growing quickly, it’ll be interesting to see how all the different LLMOps tools come together.

FIXIE RAISES $17M SEED FUNDING

HOW DOES TRAINING A MODEL WITH RLHF ACTUALLY WORK?

Training a model on an unsupervised dataset is so last year. Using human feedback data is all the rage and while it sounds easy in theory - it turns out to be tricky in practice. The HuggingFace folks have written another great blog showcasing StackLLaMA which answers StackExchange-like questions.