Search the archive

AI GRANT SECOND BATCH ANNOUNCED

The increasingly popular and successful AI grant has announced its second cohort of start-ups. There are many notable groups included, such as GGML.

ADEPT AI RELEASES AN OPEN SOURCE LLM

Persimmon, originally Adept, has released Persimmon-8B, an extremely performant 8b parameter model. It has also released inference code. There are a number of niceties - one particularly interesting tidbit is 70k unused embeddings, which allows for multimodal extensions.

CLAUDE GETS A PAID PLAN

DIFFUSION

Neural fields use a multilayer perceptron (MLP) to represent a 3D scene. It turns out you can apply diffusion directly to the weights of the MLP and generate a novel 3D scene.

ONE WIDE FEEDFORWARD IS ALL YOU NEED

The feedforward layers of the transformer make up a significant portion of the total weights in the network. However, if you use a single large feedforward layer shared between the encoder and decoder, you can dramatically improve inference time with minimal drops in performance.

USING LLMS TO PREDICT STOCK TRENDS

This paper introduces a way to use large language models to quickly analyze long Annual Reports from companies, making it easier to understand their financial health and even predict stock prices.

GENERATIVE IMAGE RESTORATION

Powerful upscaling and restoration models that use diffusion to reduce blur and sharpen image details. Model checkpoints included.

A BETTER UNDERSTANDING OF 3D SCENES OVER TIME

ResFields is a new kind of neural network that's really good at understanding complex 3D scenes as they change over time. By adding something called 'temporal residual layers,' it can deal with a lot more information and still stay accurate.

ONPREM.LLM

OnPrem.LLM is a simple Python package that makes it easier to run large language models on non-public or sensitive data and on machines with no internet connectivity.

INSIDE ELON’S STRUGGLES FOR THE FUTURE OF AI

USING AI TO CREATE BREAST CANCER TREATMENTS

IBM ROLLS OUT NEW GENERATIVE AI FEATURES AND MODELS

THE AI DRAKE “GHOSTWRITER” IS BACK WITH A NEW SONG

MICROSOFT SAYS IT WILL TAKE THE HEAT IF COPILOT USERS GET SUED

APPLE IS POURING MONEY INTO SIRI IMPROVEMENTS WITH GENERATIVE AI

Apple has boosted its budget for developing artificial intelligence, emphasizing creating conversational chatbot features for Siri — allegedly spending millions of dollars daily on research and development.

FALCON 180B MODEL RELEASED

The Falcon models from the UAE have long been the best open models available. The newest 180B parameter model slightly outperforms Llama 2 70B and has a 2k context window. The models have historically been quite tunable. However, given the resource requirements, it's not clear whether the community will adopt this newest model.

OPENAI’S FIRST DEVELOPER CONFERENCE

OpenAI will hold its inaugural DevDay conference in San Francisco on November 6, 2023. This event will preview new tools and foster idea exchange, attracting hundreds globally. Currently, over 2 million developers use tools like GPT-4 and DALL·E through OpenAI's continuously updated API.

REDUCING PPO MEMORY USAGE IN RLHF

PPO is a pain in RLHF because you need to have 3 models floating around. However, they all stay close and only require small updates from one another. Enter LoRA. If you instead use adapters you can dramatically reduce memory costs without losing performance. Beautiful idea in its simplicity.

SIMULATING AN ORDER BOOK IN JAX

Financial institutions use something called a limit order book to maintain all of the transaction information for trades on their platforms. This is useful, but it usually runs on the CPU, which makes running RL on them hard. This paper talks through creating an order book that runs on the GPU written in JAX.

ENHANCING SPEECH-DRIVEN 3D FACE ANIMATION

This study delves into the complexities of 3D face animation driven by speech, highlighting two important aspects: global factors that change how the face moves over time (composite nature) and how different parts of the face move independently based on local muscles (regional nature).

TOKENFLOW

Using a pre-trained text-to-image model to edit videos produces dream-like and strange results. TokenFlow is much smoother and maintains many semantic and structural features of the original video. It almost seems to outperform Runway Gen-2.

COMGRA

A library for use with PyTorch that makes it easier to inspect the internals of your neural networks.

MAKING VIDEO AVATARS ADAPT TO DIFFERENT LIGHTS AND BACKGROUNDS

ReliTalk is a cool tech breakthrough that lets us make video avatars look natural even when you change the lighting or background. It uses a single video and the sound of your voice to create a 3D face model.

CAN LLMS LEARN FROM A SINGLE EXAMPLE?

It appears that AI models were able to rapidly memorize examples from the dataset after seeing them just once. This astonishing feat contradicts most prior wisdom about neural network sample efficiency.

A NEW WAY TO TELL IF TWO SIMILAR PICTURES ARE REALLY THE SAME IN 3D

Doppelgangers is a new tech tool that helps figure out if two pictures that look almost the same are actually showing the same 3D object or not. It's smart enough to avoid mistakes that even people might make.

HOW THE AI REVOLUTION WILL RESHAPE THE WORLD

An argument that the impending technological wave, largely driven by AI, will herald a historic redistribution of power.

BUZZY AI STARTUP FOR GENERATING 3D MODELS USED CHEAP HUMAN LABOR

Kaedim, a 3D generation startup, often used human artists to make models, at one point using workers to produce the 3D design wholecloth themselves.

OPENAI WILL HOST ITS FIRST DEVELOPER CONFERENCE ON NOVEMBER 6TH

OpenAI will host a developer conference, OpenAI DevDay, on November 6, the company announced today.

HUGGING FACE TRAINING CLUSTER AS A SERVICE

Hugging Face launched a new service that allows you to simply train a model without managing the complicated process internally.

KINDO RAISES $7M FOR AI ENHANCED PRODUCTIVITY TOOLS

Led by Riot Ventures, Kindo aims to create a central set of tools for managing complex enterprise AI processes.

MODELS

X has updated its privacy policy to collect biometric, job, and education data from users. Notably, X will use this data and other public info to train its AI models. Alex Ivanovs suggests Elon Musk might leverage X's data for his AI venture, xAI. Musk clarifies only public data will be used.

ROBUST FINE-TUNING OF ZERO SHOT MODELS

Zero shot models are AI models that have robust performance on a broad range of distributions. However, if you tune them on your narrow task, they tend to lose that ability while simultaneously getting much better at what they're trained for. This is good but could be improved. This paper presents a method to tune such models, like CLIP to maintain their generalizability while improving specific performance.

A SMARTER WAY FOR SELF-DRIVING CARS AND ROBOTS TO SEE MORE CLEARLY

The authors have created a new tool called SQLdepth that helps self-driving cars and robots understand their surroundings in great detail.

IMPROVING MATH SKILLS IN BIG LANGUAGE MODELS

This research looks into how well LLMs can solve math problems and how different factors affect their abilities. They’ve found that a method called Rejection sampling Fine-Tuning (RFT) helps these models get better at math, especially when we use samples from several models.

BATCH LORAS ON THE SAME GPU

LoRAs are small parameter adapters that change the behavior of the underlying model by plugging into certain layers. You can batch generation for the base model to improve performance. Additionally, you can batch LoRA-modified versions of the model and serve models with different performance to different users.

MAKING AI BOTS BETTER AT REAL-LIFE CONVERSATIONS

Many AI chatbots struggle with having natural wide-ranging chats like humans do, so this project introduced 'Topical-Chat', a new dataset from human conversations on 8 big topics, to teach AI how to talk more like us without fixed roles.

UNRAVELING THE ETHICAL COMPLEXITIES OF AI

An exploration of the central ethical dilemmas posed by artificial intelligence, with the goal of illuminating how AI may shape our collective destiny and the power we have to guide it down a path aligned with human values.

BAN OR EMBRACE? COLLEGES WRESTLE WITH AI-GENERATED ADMISSIONS ESSAYS

Colleges grapple with the rise of AI-generated admissions essays. With chatbots like ChatGPT able to produce college essays, officials fear a potential for generic or plagiarized content. While some schools discourage AI use, Georgia Tech advises students to use it responsibly for brainstorming and refining, but not for direct content creation.

ALL 50 STATES CALL ON CONGRESS TO ADDRESS AI-GENERATED CSAM

US attorneys general from all 50 states urge Congress to create a commission investigating AI's role in child exploitation. They're concerned about AI being used to generate child sexual abuse material (CSAM) and deepfakes. Led by South Carolina's Attorney General Alan Wilson, the push aims to expand CSAM restrictions to cover AI-generated content.

1.6B PARAMETER MODEL REACHES 32% ON HUMANEVAL

Best in class and lightning-fast code generation model. Trained from scratch on 1.2T tokens of half code half language data.

SAM.CPP

Segment anything from Meta running in pure C++ powered by GGML.

TEACHING WITH AI BY OPENAI

OpenAI has released a guide for educators on incorporating ChatGPT into classroom settings, highlighting its utility in simulating challenging conversations, aiding in curriculum design, assisting non-English speakers, and fostering critical thinking skills. The guide offers prompts and methodologies to maximize its educational benefits, ensuring teachers remain critical of the tool's outputs.

IDEOGRAM LAUNCHES AI IMAGE GENERATOR WITH IMPRESSIVE TYPOGRAPHY

New AI startup Ideogram, backed by former Google Brain researchers, debuts with $16.5M in seed funding from a16z and Index Ventures. Ideogram sets itself apart with its reliable text generation within images, a common challenge for many AI image generators. The platform offers diverse style presets, with "typography" drawing notable attention.

OPENAI’S MOONSHOT: SOLVING THE AI ALIGNMENT PROBLEM

An interview with Jan Leike, OpenAI’s head of alignment research.

TINY LLAMA 1.1B PARAMETERS

The Tiny Llama project aims to train a 1.1B parameter Llama2 model for 3T tokens. According to scaling laws, this is a dramatic overtraining for compute optimality, but it brings significant gains for deployment optimality.

SMARTER WAY TO LEARN MULTIPLE THINGS ONE AFTER ANOTHER

This paper proposes a system that helps models learn new tasks without forgetting the old ones.