Search the archive

I CLONED MYSELF WITH AI. SHE FOOLED MY BANK AND FAMILY

CAN CHATGPT MAKE BUSINESS DECISIONS?

US SUPREME COURT REJECTS AI PATENT CASE

ITALY RESTORES CHATGPT

The ChatGPT chatbot was reactivated in Italy after its maker OpenAI addressed issues raised by Italy's data protection authority, the agency and the company confirmed on Friday.

GPT4FREE

This repository offers reverse-engineered third-party APIs for GPT-4/3.5.

WONNX

Wonnx is a GPU-accelerated ONNX inference run-time written 100% in Rust, ready for the web.

PINECONE’S $100M SERIES B

Pinecone seems to be one of the go to choices for vector databases. This is useful when you need a language model to interact with external data. You embed that data as a vector and store it in a database like this. One of the key reasons you’d do this is because of the context length problem - pinecone says this is “extending the long term memory of language models”. This certainly is exciting and many people use pinecone in production.

ELON MUSK RAMPS UP A.I. EFFORTS, EVEN AS HE WARNS OF DANGERS

LANGUAGE CONTROLLED MEMORY

SCM is a new system that allows LLMs to process long inputs. It is composed of three key modules: the language model agent, the memory stream, and the memory controller. SCM has been shown to be effective in processing long inputs, and it has the potential to be used for a variety of tasks.

SPARSEFORMER: SPARSE VISUAL RECOGNITION VIA LIMITED LATENT TOKENS

SparseFormer, a novel method inspired by human sparse visual recognition, represents images using a limited number of tokens in the latent space, achieving competitive performance with lower computational costs. This approach offers better accuracy-throughput tradeoffs and can be easily extended to video classification, potentially inspiring further research on sparse neural architectures.

SHOULD CHATGPT BE BIASED?

This article examines the challenges and risks of biases in large-scale language models like ChatGPT, discussing their origins, ethical concerns, and potential mitigation strategies. It emphasizes the need for a multidisciplinary, collaborative effort to develop more equitable, transparent, and responsible AI systems, encouraging thoughtful dialogue within the AI community about the role of biases and the pursuit of ethical AI.

LEARNING AGILE SOCCER SKILLS WITH REINFORCEMENT LEARNING

In this study, researchers use Deep Reinforcement Learning (Deep RL) to train a low-cost, miniature humanoid robot with 20 actuated joints to play a simplified 1v1 soccer game, resulting in the development of robust and dynamic movement skills and basic strategic understanding. The combination of high-frequency control, targeted dynamics randomization, and perturbations during simulation training enabled successful transfer to real robots, exhibiting improved performance and agility compared to a scripted baseline.

TEXTDEFORMER: GEOMETRY MANIPULATION USING TEXT GUIDANCE

The study introduces a technique for automatically deforming a triangle mesh based on a text prompt, leveraging differentiable rendering and pre-trained image encoders such as CLIP and DINO. By using Jacobians for mesh deformation representation, the method achieves smooth deformations and avoids noisy gradients, enabling large shape changes and fine details while maintaining coherence from all 3D viewpoints.

PHOTO-REALISM IN MULTI-VIEW IMAGE GENERATION

This research focuses on multi-view image generation for 3D-related applications such as image viewpoint editing. The proposed ray conditioning method generates multi-view images using a geometry-free approach, conditioning a 2D GAN on a light field prior, which offers improved photo-realism, identity consistency, and explicit viewpoint control while minimizing geometry artifacts and loss of fine-scale details.

PEOPLE

AN INTERVIEW WITH OPENAI CTO MIRA MURATI

A LOOK INSIDE THE DISCORD WITH THOUSAND OF AI MUSIC PRODUCERS

CHATGPT IS A SOUGHT-AFTER SKILL

A new survey found that 91% of employers are looking to hire workers with ChatGPT experience.

INSIDE THE SECRET LIST OF WEBSITES THAT MAKE CHATBOTS SOUND SMART

This article dives into the dataset used to train some of the more popular LLMs.

RESEARCH AT CHI 2023

WASMGPT

WasmGPT is a ChatGPT-like chatbot in the browser using ggml and emscripten.

NEW WAYS TO MANAGE YOUR DATA IN CHATGPT

ANTHROPIC PARTNERS WITH SCALE

STABILITY AI UPSCALE MODELS

LANGUAGE MODELS FOR PLANNING

Language models are great at producing plausible text and in certain cases reasonable plans. However, in general, they fail to generate long horizon plans. Without changing the model at all, even via fine-tuning, with special prompts we can dramatically improve the success of both plan generation and plan execution. Specifically, language models generate a plan language (fed a few-shot example) and that plan is executed by an actual planner then the resulting plans are translated back to natural language. The improvement is massive.

MACHINE TRANSLATION NEEDS ATTENTION

The Structured State Spaces for Sequences (S4) model, successful in various tasks, is applied to Machine Translation (MT), but lags behind the Transformer by 4 BLEU points and struggles with long sentences. This gap is attributed to S4's inability to summarize full source sentences in a single hidden state, which can be remedied by introducing an attention mechanism.

HOW TO TRAIN LLMS TO BE COMPETENT AND ETHICAL

One negative side effect of training agents to maximize reward is that they often become power seeking, meaning they try to increase their resources and access to information. This is a novel finding and somewhat hard to measure, this paper introduces a benchmark to help understand the power seeking behavior of these language models. Additionally, they give insights on how to train models to both act with high utility and according to moral guidelines. The world needs more high quality benchmarks and this is an excellent step in that direction.

DEEPFLOYD IF CODE RELEASED

The next image model from Stability AI has a code release. The model weights, blog, inference, and tooling will be coming in a few days. It is a multi-stage model with a frozen T5 text encoder and two super resolution models. One big win here is the ability to render extremely high quality text. The claim is that this model will likely be for research only at first, then open after some community feedback.

POSEVOCAB: LEARNING POSE EMBEDDINGS FOR HUMAN AVATAR MODELING

PoseVocab, a novel pose encoding method for human avatar modeling, constructs key poses and latent embeddings to effectively encode dynamic human appearances and achieve pose generalization and temporal consistency. The method outperforms state-of-the-art baselines in synthesis quality through the use of feature lines, a compact 3D representation, and a hierarchical query strategy for interpolating pose embeddings.

HOW AI IS UPENDING THE FREELANCE WORLD

WHY DO WE NEED RL FOR LANGUAGE MODELS?

A LOOK AT LAION

The debate over how to regulate LAION, a database with over 5 billion images commonly used to train AI models, is becoming more contentious.

PALANTIR SHOWS OFF AN AI THAT GOES TO WAR

TIKTOK IS TESTING GENERATIVE AI AVATARS

COPILOTFORXCODE

A version of GitHub Copilot designed for Xcode.

INPAINT-ANYTHING

Inpaint-Anything allows users to easily use Meta’s Segment Anything AI model.

GENERATIVE AI

SPOTIFY CEO SAYS AI IS “COOL AND SCARY”

SELF SUPERVISED LEARNING COOKBOOK

It could be argued that a lot of the recent success with deep learning has come on the back of self-supervised tasks. There's not a clear boundary between unsupervised and self-supervised, but where it does exist this paper is a really nice exploration of the methods and all of the various hyperparameters that need to be taken into consideration. Self supervised learning is a really powerful technique and this is a cool way to learn more

FULLY SPARSE FUSION FOR 3D OBJECT DETECTION

This paper investigates the integration of image modality into fully sparse architecture for efficient long-range 3D detection, using instance queries to fuse 2D instance segmentation with LiDAR-based 3D segmentation. The resulting framework achieves state-of-the-art results on nuScenes and Argoverse 2 datasets, while being 2.7 times faster than other multimodal 3D detection methods.

INTERACTIONS FROM SINGLE MONOCULAR VIDEOS

HOSNeRF is a 360° free-viewpoint rendering method that reconstructs neural radiance fields for dynamic human-object-scene interactions from a single monocular video, addressing challenges like complex object motions and varying interactions. The method significantly outperforms state-of-the-art approaches on two datasets, with a 40%-50% improvement in LPIPS, and provides compelling 360° renderings from single videos.

TRACK ANYTHING

Based on segment anything and XMem, track anything is a flexible and scalable way to track objects and videos. You can then use the masks for automatic video editing, style transfer, or video recognition tasks.

SEGMENT ANYTHING IN 3D

SA3D is a novel framework for one-shot 3D segmentation using neural radiance fields, enabling users to segment target objects in rendered views with minimal manual prompting. The iterative process uses 2D masks, density-guided inverse rendering, and cross-view self-prompts for accurate 3D mask learning, completing segmentation in about 2 minutes without optimization.

HEALTHGPT

HealthGPT is an experimental iOS app that allows users to interact with their health data stored in the Apple Health app using natural language.

GPT-4 COST $43 MILLION TO TRAIN?

While we don’t know exactly how much it cost to train the newest flagship model from OpenAI - this report breaks down a few potential pieces of information to try and guess how much it might have cost. The amount is mind-blowing (it’s not a lot in absolute terms, but an expensive model nonetheless). There are other nice tidbits in the report around compute usage and an estimate of when we’ll run out of data!

WHAT AMERICANS THINK ABOUT THE IMPACT OF AI ON WORK

OPENAI IMPROVES CHATGPT PRIVACY

REPUBLICANS RELEASE AI-GENERATED POLITICAL AD