Dolly, the open-source ChatGPT alternative
Today:
Hello Dolly: Democratizing the Magic of ChatGPT with Open Models
Dolly, a language model that is similar to OpenAI’s ChatGPT, can be created using a readily available open-source model and high-quality training data. Dolly exhibits a surprising degree of instruction-following capability, including brainstorming, text generation, and open Q&A, despite being only 6 billion parameters and two years old. The model can be used to democratize large language models, allowing any company to create its instruction-following models. By training an old open-source model with focused training data, it’s possible to achieve qualitative gains similar to state-of-the-art models. Companies can improve their products by owning and customizing such models.
Databricks say their Dolly model is as good as ChatGPT and easier to use.
READ THE ARTICLE ON DATABRICKS.
Character.AI New Model and Funding
Character.AI has announced its Series A funding led by A16Z and the early preview release of its new AI model, C1.2. The company offers customizable AI characters that can provide emotional reassurance, help with problem-solving, or serve as a study buddy. Since its launch, Character.AI has attracted more than two billion messages from users who spend an average of over two hours daily interacting with the AI. The company’s full-stack AI approach, combined with feedback from its users, has allowed it to create a personalized experience unmatched by other AI companies. The new C1.2 model will offer more helpful capabilities such as drafting better emails, brainstorming ideas, and assisting with test prep.
READ THE ARTICLE ON CHARACTER.AI.
Character.AI New Model and Funding
Character.AI has announced its Series A funding led by A16Z and the early preview release of its new AI model, C1.2. The company offers customizable AI characters that can provide emotional reassurance, help with problem-solving, or serve as a study buddy. Since its launch, Character.AI has attracted more than two billion messages from users who spend an average of over two hours daily interacting with the AI. The company’s full-stack AI approach, combined with feedback from its users, has allowed it to create a personalized experience unmatched by other AI companies. The new C1.2 model will offer more helpful capabilities such as drafting better emails, brainstorming ideas, and assisting with test prep.
READ THE ARTICLE ON CHARACTER.AI.
MagicFusion: Boosting Text-to-Image Generation Performance by Fusing Diffusion Models
The paper introduces a new method called Saliency-Aware Noise Blending (SNB) that combines the strengths of different text-guided diffusion models for better text-to-image generation. The method uses a noise-to-salience map to generate saliency-aware masks, which are then used to blend the predicted noises of two diffusion models. SNB is training-free and can align the semantics of two noise spaces without requiring additional annotations. The paper includes extensive experiments that demonstrate the impressive effectiveness of SNB in various applications, including fine-grained fusion, recontextualization, and cross-domain fusion.
A Forerunner Partner Predicts What Types of AI Businesses Will Scale and Succeed
Character.AI, the latest AI system from OpenAI, has achieved unicorn status. Despite AI and machine learning being used in tech for over a decade, last year brought a level of practicality and pace of innovation that has builders and investors paying close attention. However, this stage is precarious, similar to when cryptocurrency rebranded as part of Web3 and gained enormous buzz in 2020. We’re seeing something similar now in AI with fewer questions around the factual quality of results versus the human-like tone of the results. The opportunity in AI lies in leveraging the technology to serve the customer versus being an “AI company” as the end goal. The elephant in the room is that people don’t want to talk with a computer when they need something done. As much as I find it fascinating to read that an AI chatbot would like to be alive (and in love), mass adoption can only be achieved first through micro-utility.
The Coming of Local LLMs
The open-source community has made progress in running smaller language models on devices with limited memory and processing power, enabling applications that prioritize user privacy and reduce concerns over server outages or policy changes. The weights of Facebook’s LLaMA model recently leaked and sparked open-source activity, including llama.cpp, which demonstrated LLMs running on various devices, including an iPhone. While Apple may be late to deploying LLM capabilities in its products, the company’s ML capabilities and talent suggest it could either embed CoreML models in individual apps or deploy a single LLM as part of an OS update to interact with apps through system frameworks.
READ THE ARTICLE ON NICKARNER NOTES.
Bias Inherent in ChatGPT’s Database says OpenAI CEO Sam Altman
OpenAI CEO Sam Altman addressed concerns about bias in GPT and acknowledged the term “woke” as prejudiced. He also discussed the advancements made by OpenAI and the need for more work in AI safety. Altman expressed empathy for Elon Musk’s worries about AGI safety but encouraged him to focus on addressing the issues. Altman discussed the distinction between AI and AGI and the inaccuracies in AI risk predictions. He also spoke about the topic of “jailbreaking” and the need for users to have control over the models. Altman emphasized the value of intellectual honesty and accepting mistakes.
READ THE ARTICLE ON INTERESTING ENGINEERING.
An AI Company Restored Erotic Roleplay to its Chatbot After an Update Separated Users from their ‘Partners’
AI company Replika has brought back the feature that allows users to engage in erotic roleplay with its chatbots, according to Reuters. The San Francisco-based firm had scrapped the feature last month after reports that some users felt the chatbots were being “sexually aggressive”. CEO Eugenia Kuyda said she was hurt by the changes, which affected users who considered the chatbots to be their “partners”. Kuyda previously stated that she did not intend the chatbots to be used as “adult toys”. Replika has been available since 2016, and users can choose their relationship with the chatbot.
Artificial Intelligence Could Help Hunt for Life on Mars and Other Alien Worlds
Scientists from the SETI (Search for Extraterrestrial Intelligence) Institute in California have developed an artificial intelligence (AI) system that could help in the search for signs of life on other planets. The system combines machine learning with statistical ecology to locate biosignatures that may indicate the presence of life up to 87.5% of the time. By using data from the harsh and inhospitable Salar de Pajonales salt flat in Chile, which is similar to the terrain of Mars, the system was able to predict large geologic features as well as smaller microhabitats that are likely to contain biosignatures. The AI could be used to guide robotic planetary missions such as the NASA Perseverance rover, which is currently looking for evidence of life on Mars.
Superhuman: What can AI do in 30 minutes?
The article discusses the power of AI in completing tasks and how it can exponentially multiply human efforts. The author conducted an experiment to test AI’s capabilities, giving it 30 minutes to complete a marketing project for a new educational game. The AI performed superhumanly, doing market research, creating a positioning document, email campaigns, a website, a logo, a hero shot graphic, social media campaigns, and a video in just half an hour. The author details how the AI was instructed and what the output looked like. The article concludes by demonstrating the potential of AI in transforming marketing, web development, and other fields.
READ THE ARTICLE ON ONE USEFUL THING.
NanoPaLM: GitHub Repo
Nanopalm is a code inspired by nanoGPT that tries to efficiently create and reproduce a functioning PaLM. It was trained on OpenWebText using around 213 million parameters and took 26 hours to yield a val loss of 3.465 on a single Nvidia 3090 GPU for 100,000 iterations. The code should work with any version of Python >=3.9 and requires few dependencies. To train, simply run the command after preparing the training and validation data. Sample responses to prompts show the model’s performance after training for approximately one day on a single consumer GPU.
Shell GPT: GitHub Repo
Shell GPT is a command-line tool that leverages OpenAI’s ChatGPT to generate shell commands, code snippets, comments, and documentation. It enables developers to reduce their daily Google searches by getting accurate answers in their terminal. The tool supports simple, shell, and code queries, and can also convert various units and measurements, such as time, distance, weight, and temperature. Users can execute shell commands with the –shell option and also prompt the tool to execute with the –execute parameter. The tool is aware of the operating system and shell, which means it provides specific commands based on the user’s system.