Welcome to the 4th issue of the AstroFeather AI newsletter! This week has been full of exciting events. While much-needed discussions about regulatory guidelines for AI platforms continue, the open source and research communities are giving us a glimpse of what is possible with generative AI!

As always, if you missed any of these discussions, feel lost in the whirlwind of AI news, or just want a recap to share with your network, this update is for you.

Be sure to check out astrofeather.com (the companion site to this newsletter) for daily trending news and updates!

In today’s recap (9 min read time):

• Bringing AI to Films

• Autonomous AI (Auto-GPT, BabyAGI, and More).

• Welcome to a Westworld-like Town Called Smallville.

• AI Regulation Talks Continue.

Must-Read News Articles and Updates

1. Bringing AI to Films.

The Venice Film Festival is the oldest film festival in the world, founded in 1932 to celebrate the art and diversity of cinema. Today, there are more than 3,000 film festivals around the world, covering a variety of themes and topics, such as documentary, animation, horror, or comedy.

Given the meteoric rise in popularity of generative AI (GenAI) tools and their continued use in the creation of films, it was only a matter of time before a festival was created to highlight AI films.

Recently, Runway ML completed screenings in San Francisco of the 10 finalist films from its inaugural AI Film Festival (AIFF), which the company describes as "a celebration of the art and artists making the impossible possible at the forefront of AI filmmaking."

The winning films were selected from hundreds of submissions and covered a range of topics, including human evolution, the history of intelligent machines, obsession, and distorted realities. The top 3 are highlighted below

Generation: Riccardo Fusetti's short film Generation is described as "a brief journey through the human experience as seen through the eyes of an artificial intelligence. The film depicts a humanoid creature dancing through the stages of life as it evolved on Earth. To create the film, Fusetti used a combination of rotoscoping, animation, and AI-generated art. Generation was also selected as a "Short of the Week" on YouTube. Grand Prix AIFF Award Winner.

Checkpoint: In this short film, Áron Filkey & Joss Fong provide a brief historical overview of how generative AI platforms fit into the evolution of creative technologies. The film highlights how AI and machine learning have enabled a new age of creativity, where machines can learn from the past and offer unexpected ideas, allowing for more personalized and innovative solutions. This technology offers the attributes of a good collaborator and allows for more creative ownership of the outcome. However, it is important to note that human creativity still plays a vital role in innovation. Gold AIFF Award Winner.

Given Again: In Jake Oleson's Given Again, the viewer is taken on a journey through the life of a painter and her obsession with a pattern in nature that disrupts her sense of what's real. As her obsession grows, her perception of reality deteriorates and her world collapses into a singular temporal point, revealing reality in its simplest form - pure energy. Created with Luma AI, the film is filled with stunning visuals and is complemented by a captivating soundtrack. Silver AIFF Award Winner.

Additional Links for “Bringing AI to Films”:

• AstroFeather AI Films and Headlines - Read summaries.

• AstroFeather AI Video News Headlines - Read summaries.

2. Autonomous AI (Auto-GPT, BabyAGI, and More).

Several exciting (and, as usual, somewhat concerning) experimental applications have recently emerged that allow large language models (LLMs) to autonomously complete tasks such as conducting market research or building websites all by themselves, without the need for constant human supervision or input. Some refer to these new applications as AutoGPTs (or "autonomous agents"), and in the roughly 2 weeks since their public appearance on GitHub, Twitter, and YouTube, they have gained worldwide popularity.

In general, AutoGPTs can give themselves (and other LLMs) tasks to work on until a goal is achieved, have Internet access, the ability to read and save files, and memory (both long-term and short-term) to "remember" what's been done. While the most popular AutoGPTs are BabyAGI, Auto-GPT, and JARVIS / HuggingGPT, others such as Robo-GPT and the “Jarvis” GPT-4 Coding Assistant have grown in popularity:

BabyAGI: Widely regarded as a landmark project in autonomous agent development. BabyAGI is a Python script-based system that allows a collection of LLMs, including all OpenAI models and Meta's LLaMA, to systematically complete, add, and reprioritize tasks. [Twitter | GitHub]

Auto-GPT: Described as "an autonomous GPT-4 experiment", this is one of the first examples of GPT-4 running fully autonomously. It can prompt itself, debug recursively, and improve itself. The project has many features, including file storage, Internet access for information gathering, the ability to connect to a database for memory management, and the ability to generate and summarize text. [Twitter Demo | GitHub]

Microsoft JARVIS (HuggingGPT): Uses ChatGPT as a brain that can delegate tasks to other models (hosted on HuggingFace) to produce a final output. For example, you can ask it to generate an image of a battle between alien civilizations and then write a news article about it. ChatGPT would then analyze the request, select the correct models to complete each task, compile the completed results from each of the models, and generate an image and news article. [GitHub | HuggingFace]

Robo-GPT: Demonstrates how to extend Auto-GPT with text-to-speech capabilities. In a Twitter demonstration, Robo-GPT can be seen searching the web for top digital nomad destinations, consolidating the information into a CSV file, and narrating each of the steps it took. [Twitter Demo | GitHub]

"Jarvis" GPT-4 Coding Assistant: Demonstrates how to extend Auto-GPT with speech-to-text capabilities, allowing users to issue voice commands to autonomous agents. A Twitter demonstration showed the "Jarvis" GPT-4 Coding Assistant building a basic social network (with a user interface, backend, and working database), uploading the codebase to GitHub, and then deploying the finished application to Vercel. All using only voice commands. [Twitter Demo]

Additional Links for “Autonomous AI (Auto-GPT, BabyAGI, and More)”:

• AstroFeather ChatGPT Coverage - Read summaries

• AstroFeather GPT-4 Coverage - Read summaries.

3. Welcome to a Westworld-like Town Called Smallville.

Westworld is a dystopian science fiction story about a futuristic amusement park of the same name, inhabited by lifelike humanoid robots known as "hosts" that have been preloaded with personal identities and a backstory (or narrative). Each day, the Hosts act out their pre-programmed narratives by interacting with each other and with the park's human guests. To the human guests, these hosts are remarkable in their ability to hold convincing conversations, make plans, and respond dynamically to their environment.

If you were to recreate Westworld, on a smaller scale of course, how would you go about it? Well, for starters, you would need to create hosts and a world for them to inhabit. Since you're on a budget, the hosts would almost certainly have to be simulated characters (like those in games such as The Sims) that live in a virtual world and have the following characteristics: a personal identity, memory, the ability to create new thoughts, plan future actions, and interact with their world.

Using ChatGPT, this is roughly what a team of Stanford and Google researchers accomplished in a virtual setting.

In what many are calling groundbreaking work, Stanford and Google researchers introduced the concept of "generative agents"-virtual AI characters based on ChatGPT that have their own identities, talk to each other, throw parties, form friendships, and coordinate group activities. The researchers found that over time, the generative agents continued to simulate human behavior, characterized by emergent individual and group behavior.

The research team created 25 AI characters with the ability to communicate with each other, interact with their environment, store and recall experiences, reflect on current and past actions and events, and make plans for each day. They then preloaded the 25 AI characters with a backstory (narrative) that included:

• A personal identity (name, occupation, interests, priorities)

• Information about the other characters in the virtual world

• Defined relationships with other characters

• A starting description of how to plan their day

These characters were then instantiated in a sandbox world called Smallville, complete with a cafe, bar, park, school, dorm, houses, and shops. The houses and dorms were furnished with amenities such as beds, desks, closets, bathrooms, and kitchens. The researchers then hit "play" and watched - the results were incredibly fascinating!

Observation 1: Once together, the AI characters began to socialize and share information with each other.

Sam Runs for Mayor:

• One of the characters, named Sam, decided to run for mayor. He meets Tom at the grocery store and they chat about Sam's candidacy.

• Later that day, Tom and John (who heard the news from another character) discuss Sam's chances of winning the election. Over time, Sam's desire to become mayor becomes the talk of the town!

Isabella Plans a Valentine’s Day Party:

• On another side of Smallville, Isabella is initialized with the intention of planning a Valentine's Day party at the Hobbs Cafe on February 14th. She spends the afternoon of the 13th decorating, and word of the party soon spreads.

• By the end of the simulation, several characters knew about the party. Just like us humans, many of them didn't show up, either because they had "other plans" or simply didn't want to.

Observation 2: The characters coordinate with one another.

• When Isabella was planning the Valentine's Day party, her good friend Maria stopped by for a chat and eventually started helping Isabella. Maria's pre-loaded narrative mentions that she has a crush on Klaus.

• Without any input from the researchers, Isabella starts inviting people to the Valentine's Day party, and Maria takes the opportunity to invite her crush, Klaus, as well.

Observation 3: The characters make plans and form new relationships over time.

• When the simulation begins, Sam (the mayoral candidate) does not know another character named Latoya Williams. Over the course of the simulation, they meet in a park where Latoya mentions that she's working on a photography project.

• When Latoya and Sam meet again later in the simulation, Sam says, "Hi, Latoya. How's the project going?" and she replies, "Hi, Sam. It's going well!"

Observation 4: The characters can form and retrieve memories.

• Each of the Smallville residents has a memory stream filled with observations that are relevant (and irrelevant) to their current situation.

• They select the correct memories for a given situation based on relevance, importance, and recency.

Observation 5: The characters can reflect on memories.

• Klaus (Maria's crush) was observed to periodically review his memory stream, identify memories to reflect upon, and then form new insights. 

• Klaus has also been observed to reflect on previous reflections, resulting in higher level abstract thinking.

Additional Links for “Welcome to a Westworld-like Town Called Smallville”:

• Generative Agents: Interactive Simulacra of Human Behavior - [Abstract | Research Paper | Demo]

4. AI Regulation Talks Continue

In last week's newsletter, I updated you on the decision by the Italian Data Protection Agency (the "Garante") to temporarily ban OpenAI's ChatGPT service in the country, citing several violations of the European Union's General Data Protection Regulation (GDPR). Since then, Italy's Garante has followed up with a specific list of demands that OpenAI must meet to address the agency's concerns and allow the ChatGPT service to resume in Italy. In total, OpenAI must do the following:

• Publish an information notice on its website detailing its data processing practices.

• Clarify its claimed legal basis for processing users' personal data to train its AI models

• Provide a means for users and non-users to request the correction of disinformation generated about them by ChatGPT (or delete the data if correction is technically impossible).

• Provide an easy and accessible way for users to opt out of having their personal data used to train its AI models.

• Immediately implement age gating to prevent minors from accessing ChatGPT and implement an improved age verification system.

• Promote an information campaign on radio, television, newspapers, and the web to inform users that their personal data may be processed to train its AI models.

Spain: The Spanish data protection agency (AEPD, Agencia Española de Protección de Datos) has opened a preliminary investigation into OpenAI for possible violations of the GDPR. The agency has also asked the European Data Protection Board (EDPB) to assess privacy concerns related to ChatGPT.

European Data Protection Board (EDPB): In response to the Italian Garante's move to temporarily suspend OpenAI's ChatGPT service, the EDPB has reportedly launched a dedicated AI task force to develop privacy guidelines for AI.

Additional Links for “AI Regulation Talks Continue”:

• AstroFeather Regulations Coverage - Read summaries.

• AstroFeather Generative AI Coverage - Read summaries.

• AstroFeather GPT-4 Coverage - Read summaries.

Thanks for reading this issue of the AstroFeather newsletter!

Be sure to check out the AstroFeather site for daily AI news updates and roundups. There, you'll be able to discover high-quality news articles from a curated list of publishers (ranging from well-known organizations like Ars Technica and The New York Times to authoritative blogs like Microsoft's AI Blog) and get recommendations for additional news articles, topics, and feeds you might enjoy.

See you in the next issue!

Adides Williams (astrofeather.com)