Search the archive

SUPERAGI

Infrastructure to build, manage, and run useful Autonomous AI Agents.

CHATGPT-LIKE AI MODEL

South Korean internet giant Naver plans to provide foreign governments with tailored versions of its latest ChatGPT-like AI model, focusing on localization to address political and cultural needs. The company is also set to launch HyperClova X, an upgrade to enhance its domestic services.

STATEMENT ON AI RISK

A statement from leading AI researchers imploring society to treat the risk of extinction from AI similarly to pandemic and nuclear war.

GENERATE IMAGES WITH A LANGUAGE MODEL

Multimodal training of more than one data type has been gaining popularity recently. Usually, the model architecture requires complicated and somewhat brittle solutions. Gill is a cool new project that interleaves text tokens with image tokens to allow for truly multimodal input and output. The reported results are compelling although there are still limitations.

CONTROLVIDEO: EXTENDING CONTROLNET TO VIDEOS

ControlVideo is a new way to edit videos based on written instructions, making the video more faithful and consistent with the text while maintaining the original video's structure. It achieves this by combining advanced techniques. ControlVideo improves upon previous methods in terms of accuracy and realism. It offers a lot of flexibility in how much of the original video information to use.

BOOSTING LANGUAGE MODEL PERFORMANCE WITH SELF-NOTE TAKING METHOD

Large language models sometimes have trouble remembering things and thinking through multiple steps. A new method lets these models take notes on their own, helping them remember better and solve more complex problems.

LARGE LANGUAGE MODELS ARE NOT FAIR EVALUATORS

It is challenging to compare the performance of multiple models to one another. Traditional automated measures aren't robust and human evaluation isn't scalable. Recently people have been turning to GPT4 as a proxy for human preference. After all, it was trained with RLHF. However, doing this naively results in miscalibration. This work proposes a few simple fixes that aligns GPT4 evaluations more closely with those of humans.

GORILLA: ENHANCING THE POWER OF LLMS IN GENERATING API CALLS

Gorilla, a refined model based on Large Language Models (LLMs), significantly outperforms existing models like GPT-4 in writing API calls. It helps language models use tools more effectively. Combined with a document retriever, Gorilla adjusts well to updated documents, minimizes inaccurate creations, and maintains the reliability of outputs, as demonstrated using a new comprehensive dataset called APIBench.

ENHANCING TEXT-TO-3D GENERATION WITH PROLIFICDREAMER

ProlificDreamer is a new method for improving text-to-3D generation. It solves problems like over-saturation and low-diversity found in the earlier Score Distillation Sampling (SDS) approach by treating the 3D parameter as a random variable. This Variational Score Distillation (VSD) framework can produce higher quality and more diverse results, allowing the creation of highly detailed and photo-realistic 3D images from text, including complex effects such as smoke and drops.

CORPORATE VCS RIDE THE AI WAVE

STABILITYAI’S LETTER TO THE SENATE

CHATGPT SHARED LINKS

EDTECH GROUPS INSIST AI IS A FRIEND

Edtech groups such as Duolingo insist that AI will enhance their business, despite the possibility that cheap platforms undercut their business.

THREESTUDIO

TINYCORP RAISES $5M TO BUILD DEEP LEARNING SOFTWARE FOR AMDS

SAM ALTMAN SHARES HIS OPTIMISTIC VIEW OF AI

NEW SUPERBUG-KILLING ANTIBIOTIC DISCOVERED USING AI

AI was used to discover "abaucin," a new antibiotic effective against the superbug Acinetobacter baumannii. The AI, trained with thousands of known drugs, shortlisted potential antibiotics from over 6,000 untested compounds in 90 minutes. Although promising, abaucin requires further tests and clinical trials. It is expected to be available around 2030.

A NEW METHOD FOR CORRECTING ERRORS IN CONVERTING TEXT TO SQL

This paper focuses on improving how computers understand and convert everyday language into SQL, a language used for managing and manipulating databases. It introduces a new method that corrects errors by looking at whole sections of code instead of individual words, and uses a new representation for SQL queries that's closer to how pre-existing code language models work, leading to a significant improvement in accuracy.

A HYBRID MODEL COMBINING THE STRENGTHS OF TRANSFORMERS AND RNNS

This research introduces a new model called Receptance Weighted Key Value (RWKV), which brings together the efficient parallelized training of Transformers and the efficient operation of Recurrent Neural Networks (RNNs). With its unique approach of utilizing a linear attention mechanism, RWKV can be used like a Transformer or an RNN during different stages, proving equally effective as Transformers, but with more efficient use of computing resources.

IMPROVING QUESTION ANSWERING SYSTEM THROUGH USER FEEDBACK

Researchers explore how a system that extracts answers to questions can be constantly improved by feedback from human users. Through experiments involving thousands of user interactions, we found that user feedback significantly improved the system's performance over time, including its ability to adapt to new topics or areas of knowledge.

MODELS

Researchers have developed the first dataset for testing AI language models using 800 question-answer pairs based on over 350 STEM (Science, Technology, Engineering, and Mathematics) theorems. This dataset offers a new benchmark to measure an AI's ability to solve complex, university-level problems using these theorems.

LEARNING MINECRAFT WITH GPT4 WITHOUT TRAINING

GPT-4 is widely regarded as the most powerful language model to date. However, because it is closed behind an API, we can’t train it for novel tasks. Voyager uses code as the “policy representation” and uses the environment errors and performance as natural language feedback. The set of code functions then allows GPT4 to iteratively improve at mining diamond and other tasks.

ANSE

Anse is a fully optimized UI for AI chats.

THE HARD STUFF NO ONE TALKS ABOUT BUILDING ON LLMS

LAWYER CITES FAKE CASE INVENTED BY CHATGPT

OPENAI HAS NO PLANS TO LEAVE EUROPE

OpenAI has no plans to leave Europe, CEO Sam Altman said on Friday, reversing a threat made earlier this week to leave the region if it becomes too hard to comply with upcoming laws on artificial intelligence.

INTERVIEW WITH BEN GOERTZEL

Ben Goertzel, renowned AI researcher, shares his views on future AI developments and their societal implications. He talks about the possibility of conscious AIs, suggesting it may be unethical to withhold rights from them. However, he also highlights potential risks of superintelligent AIs.

GPT YOUR GDRIVE WITH LANGCHAIN

Accessing personal information dramatically reduces hallucinations for various tasks. This integration lets you connect to your Google Drive and ingest data to give to language models.

OPENAI MAY LEAVE THE EU

OpenAI CEO Sam Altman said on Wednesday the company might consider leaving Europe if it could not comply with the upcoming AI regulations by the European Union.

MICROSOFT CALLS FOR AI REGULATION

Microsoft Corp. is calling for a new US agency to regulate artificial intelligence and licensing requirements to operate the most powerful AI tools, company President Brad Smith said Thursday.

ENHANCING DIFFUSION MODELS WITH REINFORCEMENT LEARNING

This paper investigates the use of reinforcement learning to optimize diffusion models for downstream objectives like image quality. The proposed approach, Denoising Diffusion Policy Optimization (DDPO), proves effective in adapting text-to-image models to challenging objectives and improving prompt-image alignment without additional data or human annotation.

RECURRENCE

This project introduces RECURRENTGPT, a language-based recurrence mechanism incorporated into large language models like ChatGPT. By simulating the Long Short-Term Memory mechanism, RECURRENTGPT enables arbitrary-length text generation while preserving long-term memory. It demonstrates potential for interactive writing systems and personalized interactive fiction, showcasing the utility of borrowing ideas from cognitive science and deep learning.

GOAT: SMALL MODELS BEAT GPT4 ON ARITHMETIC

To be fair, calculators also beat GPT4 on arithmetic. The interesting discovery here is that the llama tokenizer (which splits numbers into individual digits) provides benefits for arithmetic even over other similarly powerful base models fine-tuned in the same task.

TWIST: IMPROVING SPEECH LANGUAGE MODELS WITH TEXTUAL WARM-START

This work introduces TWIST, a method for training Speech Language Models (SpeechLMs) using a warm-start from pretrained textual models. TWIST outperforms starting from scratch, with empirical analysis emphasizing the significance of model and dataset scale. The study presents the largest SpeechLM to date and introduces spoken versions of a textual benchmark for improved evaluation and future research advancements.

GORILLA

Gorilla enables LLMs to use tools by invoking APIs. Given a natural language query, Gorilla comes up with the semantically- and syntactically- correct API to invoke.

HOW TO FINETUNE LLMS ON CUSTOM DATASETS

GRIMES REVIEWS AI GRIMES SONGS

AI-music proponent Grimes takes some time to review AI-generated Grimes songs, which she found quite impressive.

NVIDIA STOCK EXPLODES

CHATGPT PLUGINS NOW SUPPORT POSTGRES & SUPABASE

GOOGLE SEARCH STARTS ROLLING OUT AI RESULTS

TIKTOK TESTING AI CHATBOT

TikTok is testing an AI chatbot called Tako that can converse with users about short videos and help them discover content.

IX

Ix is an experimental platform for designing and deploying semi-autonomous LLM agents.

ELON WANTS TO CHALLENGE GOOGLE AND MICROSOFT

Elon Musk said he sees the need for an artificial-intelligence business to rival Google and Microsoft that could involve different parts of his corporate empire, including Twitter.

META RELEASES MEGABYTE

MANAGEMENT SOLUTION

QLORA 65B PARAMETER MODELS FINE-TUNED ON 48GB GPU

LoRA is a way to fine-tune models much more cheaply than full fine-tuning. It works by just updating a small, carefully selected, subset of the model. However, even then it is often too expensive to fine-tune large (more than 13B parameter) models on commodity hardware. Quantization reduces the precision of parameters in the model which in turn takes less space. More great open source progress.

SEAHORSE: A TOOL FOR EVALUATING MULTILINGUAL SUMMARIZATION SYSTEMS

SEAHORSE is a dataset created to assess the quality of summarization systems in multiple languages. It contains 96,000 summaries that have been rated by humans based on six important aspects: clarity, repetition, grammar, attribution, main ideas, and conciseness. This dataset not only serves as a benchmark for evaluating the performance of automated metrics but also provides a valuable resource for training such metrics, helping researchers advance the field of multilingual summarization evaluation.