• WeeklyDispatch.AI
  • Posts
  • Personality simulation is coming: what it will mean when an AI can accurately model you

Personality simulation is coming: what it will mean when an AI can accurately model you

Plus: First peek at WebAgent, Google DeepMind's tool that will soon perform web-related tasks for you

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, guides, technical analysis and exciting developments in open source.

In today’s Dispatch:

  • OpenAI has quietly introduced a method allowing website administrators to prevent the company’s dedicated web crawler, GPTBot, from scraping their data. Members of the AI community spotted the addition on Monday but it has come without an official announcement.

  • More data sovereignty issues: Zoom continues to face criticism for covertly updating its Terms of Service to state that it can train AI on user data without providing an opt-out option. The move has particularly impacted Zoom's reputation amid ongoing debates about privacy and trust.

  • What could some of the potential impacts be when language models are able to very closely mimic our unique personality traits in various aspects of daily life? Our lead news & opinion story covers this topic in a fascinating article from Every.

Plus: Navigating the complex world of government contracts through AI, a woman in Texas brings AI into local politics, trending tools and more.

Image: Every.to

The blog: This Every piece offers an introspective dive into some of the implications of AI modeling personalities, as seen through the lens of the author's struggles with an excessively agreeable disposition. It discusses the potential of language models to not only mimic human personality traits but also alter them to better manage time, interactions, and decision-making. The author emphasizes the potential power of this technology to revolutionize virtually all aspects of daily life: personal habits, business productivity, societal norms, etc.

More details:

  • Scientists at DeepMind have successfully taught LLMs to simulate and change their personality based on prompts, demonstrating that LLMs can adopt and consistently maintain a defined personality.

  • The author experimented with ChatGPT-4, observing how it performed (as him) on a personality test. The results were surprisingly accurate and he believes with further fine-tuning, it could mimic his personality almost perfectly.

  • The author discusses the potential of using LLMs in the workplace to simulate reactions to questions or scenarios, thereby reducing bottlenecks.

  • LLMs have limitations in understanding situational context; there will be a challenge to incorporate all the necessary contextual factors.

Takeaways: As AI advances to simulate personality traits more accurately, this technology could help boost business productivity, enhance self-understanding, and even facilitate personal growth. Although there are challenges in terms of the LLMs' ability to fully comprehend situational context, the progress made thus far suggests a promising direction for AI in personality simulation. The technology's implications extend beyond personal use - they’ll likely lead to shifts in interpersonal business operations and social dynamics. The author posits that there will likely be many ways to misuse this technology as well.

The interview: Semafor interviews Sweetspot, an AI-based search engine startup that aims to simplify the intricate process of searching for U.S. government contracts. The startup uses large language models to navigate the $650 billion govt contract market - not only assisting bidding businesses but also proving useful for lawmakers and journalists keeping tabs on federal tax expenditure.

More details:

  • Sweetspot's AI-based platform enhances the ease of parsing and discovering relevant contracts. For instance, users can type "facial recognition" into the search engine, and the software identifies any relevant contracts.

  • The AI preprocesses publicly available government data to allow more efficient search term filtration.

  • Approximately 80% of a contract's content consists of general amendments and best practices. Sweetspot's technology filters this out and focuses on vital information, thus reducing noise.

  • While GPT is used for some scenarios, Sweetspot fine-tunes its own models for the core product. The company plans to move into bid proposals where it will use an LLM to draft the proposal for a contract.

Takeaways: Sweetspot is another fascinating example of how AI technology will soon streamline complex administrative tasks. By reducing the tedious work associated with sorting through contracts, it can not only make businesses more efficient but also help ensure greater transparency in government spending. It's still early days for the startup, but Sweetspot's initial progress is promising.

“If they’re not listening to me – a person of some color and a woman – maybe they’ll listen to this guy,” Delsooz said. “I’m just experimenting.”

Yahoo! News • Kirsten Crow

California artist and musician Ilya Shkipin defended his use of artificial intelligence on Twitter, sharing examples before deleting them.

Decrypt • André Beganski

More News & Opinion:

From our sponsors:

Holi: Up to 60% off on Sustainable Brands.

Exclusive Savings - for you & the planet.

Find the best deals from top sustainable and ethical brands. And support their mission without hurting your wallet!

Meet WebAgent: Google DeepMind’s tool that will soon perform web-related tasks for you

The story: Google’s DeepMind team recently published a research paper on one of their in-development projects called WebAgent. This technology will make AI assistants better at understanding websites and then completing complex tasks on them. This could allow us to give AI agents natural language instructions for complex web-related tasks like planning and facilitating a multi-destination trip, applying to desired job openings automatically as they pop up, etc.

Breakdown of the paper:

  • DeepMind combined two AI systems - one focused on understanding web pages, the other on taking actions.

  • The web understanding system uses attention techniques to parse the hierarchical structure of HTML. WebAgent then breaks down instructions into steps, summarizes only the relevant parts of complex web pages, and generates Python code to take actions.

  • In tests, WebAgent completed tasks on real websites around 70% of the time (previous AI systems completed these tasks less than 20% of the time). Breaking down the task into a sequence of sub-problems and then assigning them to separate AI substantially enhanced overall task success.

  • There are still challenges mentioned around generalizing WebAgent's abilities across different websites and instructions. No release dates or indications of plans for commercialization are discussed.

Takeaways: This is a very first look into the creation of AI assistants that can help us with real-world digital tasks. The modular approach shows how effectively specialized AI systems can be combined to handle complex instructions and long web pages. If perfected, this technology could enable AI agents to assist us with all sorts of things like travel planning, shopping, booking services, and more by autonomously navigating and interacting with websites on our behalf. It's an important milestone towards more versatile real-world AI.

The number of programmers who are “replaced by AI” will be small. Here’s why, and here’s how the use of AI will change the discipline as a whole.

VentureBeat • Mike Loukides

More Open Source & Technical:

Social media/YouTube:

  • AI adds color to a 1946 clip with Albert Einstein [Reddit]

  • ChatGPT goes completely off the rails in the middle of a conversation with user (full conversation linked in Reddit comments) [Reddit]

  • Someone made a 1-click install script for Meta’s AudioCraft tools [X]

  • OpenAI’s Andrej Karpathy’s weekend hack: llama2.c [X]

  • Matt Wolfe does a video breakdown of Google’s Generative Search Experience [YouTube]

Did you know?

Four years ago, China instituted a law (now widely viewed as successful) limiting how much time persons under 18 could spend playing video games. Now, the country is mulling similar restrictions on smart phone usage.

Trending AI Tools & Services:

  • YouLearn: Learn about any YouTube video (or PDF, Google Doc, etc.) with an AI tutor copilot

  • guidde: guidde is the generative AI platform for businesses that helps your team create video documentation 11x faster

  • tl;dv: Record meetings for free with Zoom app or Google Meet Chrome Extension, take notes, transcribe into 20x languages - powered by GPT

  • Kadoa: extract targeted web data easily with AI

  • DreamyRooms: redesign the interior of any room in your house by applying various design styles

Thanks for reading, have a great Tuesday! We’ll see you tomorrow.