AIGC #003 0410-0416 Weekly
Auto-GPT, Generative Agents, Interactive graphical LLMs generated text, and no GPT-5 training is ongoing.
Good morning!
Every Monday, I will compile updates and collect new AI products that emerged in the AIGC field over the past week. If you're interested in staying up-to-date on the latest developments in AI, feel free to subscribe to my newsletter.
01 /
Generative Agents simulate belivable human behaviors.
One of the most remarkable and exciting papers this week 🤯
Researchers have built a sandbox virtual town inspired by "The Sims," and then added 25 generative agents to it. In the experiment, these agents simulate believable human behavior. For a pre-assigned Valentine's Day party task, agents independently spread the party invitation two days before the event, made new friends, invited them to the party, and coordinated their appearances at the party.
This experiment has the potential for significant impact on gaming and personal life, especially when connected to the next story. (Link)
02 /
Auto-GPT: an impressive automative GPT process
Generative AI has the potential to revolutionize many industries. If it were able to generate believable behavior and granted process automation capability, what would the potential developments be?
While OpenAI's plug-in service is still in invited-only testing, Auto-GPT gives everyone a chance to use GPT in an automated way. You can set goals and targets for GPT, then let it run on its own. The advantage of such a system is that it can break down requirements and goals into smaller, executable ones with GPT's ability, then complete tasks one by one, repeatedly, until the goals are complete.
However, it is important to note that it will quickly consume tokens, and since it only supports using your own APIs, you should carefully consider what you want the bot to do. (Link)
03 /
Visualizing LLM-Generated Text
Language models (LLMs) are powerful tools that can be used to generate text, translate languages, and answer your questions in an informative way. However, the large amount of text that LLMs can generate can sometimes be difficult to understand.
UCSD's Creativity Lab has launched two projects to help: Graphologue and Sensecape. Graphologue transforms text into interactive diagrams in real-time, and Sensecape enables users to spatially organize information obtained from GPT-4. (Link)
Other news
OpenAI's CEO confirms that they are currently not training GPT-5 and have no plans to launch relevant training programs in the near future (link).
Amazon AWS launches Amazon Bedrock service and CodeWhisperer. Users can use APIs to access models provided by AI21 Labs (Jurassic-2 multilingual model), Anthropic (Claude), Stability AI (Stable Diffusion), and Amazon (Titan) to build their own business. The newly launched Titan contains two models, one is a generative AI for output, and the other is an embedded AI that supports other similar search and personalized businesses. CodeWhisperer is a code assistant service with a 57% speed improvement in completing tasks according to internal test data (link).
Stability.AI launches Stable Diffusion XL beta, and enterprises can call image generation models through APIs (link). Access DreamStudio to experience the new model (link), or use Clipdrop's service to experience it (60 free generations per day) (link).
OpenAI launches a bug bounty program that rewards users up to $20,000 for reporting system vulnerabilities or security flaws (link).
Databricks launches Dolly 2.0, 12B LLM, open-source data sets, model weights, and training code (link).
Google collaborates with UC Berkeley to research on using LLM debugging code on its own (link).
Alibaba Cloud releases "Tongyi Qianwen" and opens invitation for testing (link).
Products and Tools
Cognosys.ai, best choice for experiencing Auto-GPT at now. (Link)
Course AI, a tool that automatically generates courses containing concepts, cases, summaries, and quizzes based on a specified topic. (Link)
Boring Report summarize and transform news into a more mundane tone, mocking itself as "boring news." It should attract users who are dissatisfied with the media's tendency to use sensational vocabulary. (Link)
Magic Copy is a one-click image extraction tool based on Meta SAM. (Link)
Myshell.ai is a chatbot that provides voice dialogue. (Link)
Mochi Diffusion is an open-source native Mac client for Stable Diffusion. (Link)
ChatGPT Box is an extension that deeply integrates ChatGPT into the browser. (Link)
TeamSmart AI help you builds your own AI team through a large number of high-quality prompts. (Link)
How-to and Best Practices
Snack Prompt, a community for users submit and share prompts (Link).
Learning Prompt, a project for prompt learning tips (Link).
ChatGPT Study Guide "The Ultimate GPT-4 Guide" (Link).
DeepSpeed Chat: Easy, Fast and Affordable RLHF Training of ChatGPT-like Models at All Scales (Link)
A Notion page collecting Auto-GPT related information (Link).
Misc.
A paper discusses the bias problem caused by training data issues in image generation engines, and the possibility of reinforcing these biases while providing generation services.(Link)
Someone has used WebGPU to implement LLM (Vicuna) that runs in the browser, requiring Chrome Canary version 113. (Link)
Meta has open-sourced a model called Animated Drawings that can animate a person's portrait using object detection models, pose estimation models, and image-based segmentation methods to quickly generate an electronic version of the painting. Additionally, specified joints can be animated to make the character move.(Link)
Gallery
Midjourney, mixed style of Kawacy, Krenz Cushart and Kaethe Butcher.
(fin)