Search the archive

THE CASE FOR REALITY CENTRIC AI

THE FUTURE OF CHATBOTS IN SEARCH BY ANDREW NG

Microsoft and Google's attempts to integrate large language models (LLMs) into chatbots for search have faced criticism due to their tendency to make up facts and produce toxic output. However, this article remains optimistic about the future of chat-based search, emphasizing the potential of retrieval augmented generation to improve factual accuracy and the development of techniques to reduce toxic behaviors. The author believes that with continued research and innovation, chatbots will evolve to better meet user needs and expectations.

SCIENTISTS USE AI TO CREATE NEW ENZYMES

SPOTIFY LAUNCHES AI-POWERED PERSONAL DJ

BEDROCK LAUNCHES THEIR 4TH FUND

RIVAL TRAINING

BAIN AND OPENAI PARTNERSHIP

DOUBT THE WORTH OF YOUR WORK

This article discusses how the prevalence of impressive AI projects on social media platforms like Instagram can make other developers feel inadequate about their own work. The author encourages AI developers to judge their projects based on their own standards and not to be swayed by what's currently trendy. The field of AI is constantly evolving, with new technologies emerging all the time, and it's important to keep learning and stay up to date. The article reminds developers that the emergence of a new hot technology doesn't mean their current project is any less valuable.

TALKING ABOUT LARGE LANGUAGE MODELS

Technology and Philosophy have found a fun intersection with Large Language Models (LLMs). We often use anthropomorphising words such as “knows”, “understands”, or “believes” when referring to these systems. This paper advocates the community to take care when talking about these extremely capable systems and use more fitting language (e.g., encodes, stores, contains).

ENCODING IMAGES IN TEXT, UNSUPERVISED TEXT-IMAGE ALIGNMENT

This paper introduces LQAE which uses a pre-trained language model and encodes images as a sequence of text tokens, then they train a decoder with the BERT masking objective which learns to represent similar images with similar clusters of text tokens. They train the decoder to reconstruct the original image from the predicted text token embeddings. This serves to align the two modalities in an unsupervised manner. A fun side effect is the ability to compress images as text (just a few kb) and reconstruct a lossy version via a single forward pass.

PIX2PIX3D: 3D-AWARE CONDITIONAL IMAGE SYNTHESIS

This work proposes pix2pix3D, a 3D-aware conditional generative model for controllable photorealistic image synthesis. Given a 2D label map, such as a segmentation or edge map, the model learns to synthesize a corresponding image from different viewpoints. To enable explicit 3D user control, the authors extend conditional generative models with neural radiance fields. Given widely-available monocular images and label map pairs, the model learns to assign a label to every 3D point in addition to color and density, which enables it to render the image and pixel-aligned label map simultaneously.

TURN YOUR DATAFRAME INTO A UI FOR VISUAL ANALYSIS

PyGWalker (pronounced Pig Walker) is a graphical tool for exploratory analysis of your data. It allows you to quickly inspect relationships, correlations, and distributions. Examples and screenshots are available in the repository readme.

UNIVERSAL AUDIO SYNTHESIS MODEL

Universal vocoder for generating high fidelity audio with the ability to condition on specific features. This is the first large scale GAN model of its kind. Models, code, and samples are available.

ADDING EXTRA CONDITIONS

Researchers have introduced a neural network structure called ControlNet, which can control pretrained large diffusion models to support additional input conditions. This new method allows for task-specific learning in an end-to-end fashion and is capable of learning with a small training dataset. The training is as fast as fine-tuning a diffusion model and can be done on a personal device, or it can scale to large amounts of data using powerful computation clusters. The researchers demonstrate that ControlNets can enable conditional inputs such as edge maps, segmentation maps, and key points for large diffusion models, which can benefit various related applications.

BIOGPT

This GitHub repository contains the implementation of BioGPT, a generative pre-trained transformer for biomedical text generation and mining. It outperforms GPT-3 with just a fraction of the parameters.

“THOUGHT”

MAGAZINE PAUSES SUBMISSIONS DUE TO SURGE OF AI-GENERATED CONTENT

GENERATIVE AI IS COMING FOR THE LAWYERS

OUR SOCIETY IS NOT PREPARED FOR THIS AI AWESOMENESS

ROBLOX IS INCORPORATING GENERATIVE AI

Popular online game Roblox is bringing Generative AI to its game universe, allowing players to use its code-writing ability to make its digital worlds even more customizable. The tool lets Roblox users create items such as buildings, terrain, and avatars; change the appearance and behavior of those things; and give them new interactive properties by typing what they want to achieve in natural language rather than complex code.

AI BEING USED IN CALL CENTERS

The AI in the workplace battle has begun, and its first battleground is call centers, where AI is being used to automate basic tasks and instruct the remaining humans how to do their jobs. For workers, the technology promises to eliminate the drudgery of dull, repetitive tasks such as data processing and password resets, while synthesizing huge amounts of information that can be accessed instantly.

SPEED UP CODE 2.5X WITH PERFORMANCE IMPROVEMENT EDITS

Compilers are excellent for optimizing code, but can only do so much. This work releases a dataset and models that edit code to improve run-time performance. The intuition behind this work is that language models may have better semantic understanding of the code which provides advantages over standard static analysis.

ALIGNMENT THROUGH THE LAW

It is infeasible to specify all the possible desired outcomes to a language model agent. Alignment studies this problem through many lenses, one new lens is that of legal contracts. Legal contracts face a similar issue in that they cannot predict every “if-then” scenario, but require counterparties to reason about the spirit of the agreement, conditioned on precedence. This paper explores SOTA language model’s ability to reason similarly about fiduciary duties in contracts. The latest OpenAI models are much better than previous versions, indicating potential positive progress for alignment!

A BENCHMARK FOR FEW-SHOT REGION-AWARE MACHINE TRANSLATION

RUNNING LARGE LANGUAGE MODELS ON A SINGLE GPU

FlexGen is a high-throughput generation engine for running large language models with limited GPU memory. It improves performance via “lightning fast” offloading when in limited memory environments. Benchmarks show 1 token per second on a T4 for OPT175B!

RUST BASED DEEP LEARNING FRAMEWORK

If you're looking for a well written, new, deep learning framework in Rust, burn seems like a good choice. Their Tensor crate is good to use standalone. This also is a fun library if you're looking to contribute to an open source project.

This work proposes the use of T2I-Adapters to control the generation more granularly by aligning internal knowledge in T2I models with external control signals. This approach can achieve rich control and editing effects and has practical value due to its composability and generalization ability. The proposed T2I-Adapter shows promising generation quality and a wide range of applications, according to extensive experiments.

MY CLASS REQUIRED AI. HERE'S WHAT I'VE LEARNED SO FAR.

AI THAT IMPROVES ITSELF THROUGH DARWIN-STYLE EVOLUTION

MEDIA IS UPSET THAT OPENAI IS USING THEIR ARTICLES TO TRAIN CHATGPT

Major media outlets, including the Wall Street Journal and CNN, are upset that OpenAI is using their articles to train ChatGPT without paying them.

NASA’S NEW BALLOON-BORNE TELESCOPE WAS DESIGNED WITH AI

AI was used to quickly design NASA’s new balloon-borne telescope, a significant development considering how long it usually takes humans to design new telescopes.

HOW AI CAN BE HELPFUL IN DISASTER RESPONSE

This article dives into how AI is being used by humanitarian teams in Turkey and Syria, specifically pointing out that AI is able to quickly scope out earthquake damage and strategize rescue efforts.

GITHUB COPILOT GETS AN UPGRADE

HOW SHOULD AI SYSTEMS BEHAVE, AND WHO SHOULD DECIDE?

1B MULTIMODAL MODEL OUT PERFORMS 175B COUNTERPART

Narrow models can still, unsurprisingly, beat larger general purpose models. This paper is another example, the authors use a small multimodal model to beat GPT3 by 16% on the visual / text benchmark ScienceQA. It's only one benchmark, but it's good to see improvements on more reasonably sized models.

THE STABLE ENTROPY HYPOTHESIS

This paper submits an hypothesis that “human-like” text generation lies in a narrow region of low entropy. Degenerate text which has repeated words, incoherent sentences, and grammatical errors lie away from this region. They provide supporting evidence and present an entropy aware decoding scheme.

MORAL SELF-CORRECTION IN LARGE LANGUAGE MODELS

Large language models (LLMs) exhibit harmful biases that can get worse with size. Reinforcement learning from human feedback (RLHF) helps, but not always enough. The authors show that simple prompting approaches can help LLMs trained with RLHF produce less harmful outputs.

NAVIGATING THE MANIFOLD OF LATENTS IN STABLE DIFFUSION

LANGCHAIN FOR THE WEB

Similar to other recently released libraries, Langchain now has a JavaScript library. Use this to build LM powered web apps. Importantly, interfacing with the python version is a huge focus specifically around serialization of objects.

TALK-TO-CHATGPT

Talk-To-ChatGPT is a Google Chrome extension that allows users to talk with ChatGPT using their voice and listen to the AI’s answer with a voice.

SCALING MEDIA ML AT NETFLIX

HOW TO FINE-TUNE THE MOST POWERFUL OPEN SOURCE LLM

THE PHYSICS PRINCIPLE THAT INSPIRED MODERN AI ART

SUMMARY OF CREATIVE AI RESEARCH AT SONY

A list of papers from the Sony creative AI research group. They focus on deep generative modeling, music, and cinematic AI. Lots of great links here to put on your reading list if you want to understand the state of the art for music.

AI CAN HELP DESIGN OPIOID DRUGS

AI BREAST CANCER DIAGNOSTICS

AI BECOMES SILICON VALLEY’S NEXT BUZZY BANDWAGON AS CRYPTO BOOM FIZZLES

Technologists agree that the generative AI that powers systems like ChatGPT has the potential to change how we live and work, despite the technology’s clear flaws. But some people see signs of froth that remind them of the crypto boom that recently fizzled.

HOW SHOULD AI SYSTEMS BEHAVE, AND WHO SHOULD DECIDE?

(AI LAWYER)