202 | 🤔␡ Helping AI Forget Things

In today’s Brainyacts:

1. Machine Unlearning

2. Meet DuckDuckGo for web and AI chat privacy

3. Antitrust investigations for OpenAI, Nvidia, & Microsoft? and make decisions and other AI model news

4. Short on new you can use

   👋 to new subscribers!

To read previous editions, click here.

Lead Memo

🤔 ␡Machine Unlearning?

An AI concept that is gaining significant attention is "machine unlearning." This idea, while seemingly simple, addresses complex challenges associated with AI models, particularly in the context of generative AI.

What is Machine Unlearning?

Machine unlearning refers to the ability of an AI model to selectively forget or remove specific pieces of information that it has learned during training. This is in contrast to the typical process of training an AI, where vast amounts of data are ingested and used to build the model's capabilities. When problematic or undesirable data is included in the training set—such as copyrighted material, personal information, or violent content—it can be challenging to remove this data without starting the training process from scratch.

Traditionally, removing such data meant retraining the entire model, which is time-consuming and resource-intensive. However, machine unlearning aims to offer a more efficient solution by allowing specific data to be purged from the model without needing to rebuild it entirely.

Why is Machine Unlearning Important?

Machine unlearning is crucial for several reasons:

1. Compliance with Copyright and Privacy Laws: AI models often train on vast datasets scraped from the internet, which can include copyrighted material and personal information. Unlearning helps ensure compliance with copyright laws and privacy regulations by removing such data upon request. For instance, if an AI model used copyrighted images from a news article without permission, unlearning could remove those images without affecting the model's overall performance.

2. Ethical AI Development: Ensuring that AI systems do not perpetuate harmful content, such as violence or hate speech, is critical. Machine unlearning allows developers to remove this content, making AI applications safer and more ethical.

3. Commercial Viability: For AI to be commercially viable, particularly in sectors like healthcare, finance, or content creation, it must adhere to strict ethical and legal standards. Machine unlearning provides a mechanism to meet these standards without incurring the high costs associated with retraining models.

Potential Harm and Challenges

While machine unlearning offers significant benefits, it also comes with potential risks and challenges:

1. Incomplete Unlearning: Ensuring that all instances of the undesired data are completely removed is challenging. If traces of the data remain, it could still influence the model's outputs.

2. Model Integrity: Removing data from a trained model without affecting its overall performance is complex. Poorly executed unlearning could degrade the model's accuracy and reliability.

3. Ethical Dilemmas: Deciding what content should be unlearned can be subjective and controversial. Balancing free expression with the need to remove harmful content requires careful consideration and transparent guidelines.

Current State

To accelerate unlearning, we have to know more about how these models ‘think.” Two of the leading models (along with scored of independent scientists and researchers) are working on this. OpenAI just shared some of their research. Anthropic recently did as well.

Spotlight

🦆🦆💥 DuckDuckGo: Intro and Private AI Chats

DuckDuckGo has launched “Private AI chats”, which allow users to use AI models like—OpenAI’s ChatGPT—anonymously, meaning models can’t store chats and use them for training. I’ll show you how in a bit, first though you may not know what DuckDuckGo is. It is my default web search tool – has been for years – and it is great. It doesn’t snoop on you or sell your data to advertisers.

What is DuckDuckGo?

DuckDuckGo is like Google or Bing, but with a key difference: it doesn't keep tabs on you. When you use DuckDuckGo, it doesn't save what you search for, which websites you visit, or any other personal details.

Key Features:

1. Privacy: The main selling point of DuckDuckGo is its commitment to user privacy. It doesn't store your search history or any personal data. This means you won't see personalized ads based on your search habits.

2. Unbiased Search Results: Since DuckDuckGo doesn't track you, the search results are the same for everyone. You're not stuck in a "filter bubble," where you're only shown content that aligns with your previous searches or preferences.

3. Simplified Experience: DuckDuckGo offers a clean, straightforward search experience without the clutter of personalized ads or distractions.

The Company:

DuckDuckGo was founded by Gabriel Weinberg in 2008. The company is based in Paoli, Pennsylvania, and its mission is to protect the privacy of internet users. It makes money through ads, but these ads are not personalized. Instead, they are based on the search terms you use, similar to looking up a topic in a library and seeing relevant ads on the bulletin board.

Why Use DuckDuckGo?

1. Privacy Concerns: If you’re worried about your online activities being tracked, DuckDuckGo is a great choice.

2. Simplicity: It’s user-friendly and offers a straightforward search experience.

3. No Filter Bubble: You'll get the same search results as anyone else using the engine, which can provide a more unbiased view of information.

How to Use It:

Using DuckDuckGo is simple. Just go to duckduckgo.com and type in your search query, just like you would with any other search engine. You can also set it as your default search engine in your browser settings.

Private AI Chats:

First, know that these chats do not use the latest models - this is fine for many beginners and simple use cases.

Go to: https://duckduckgo.com/aichat

Choose the model you want to chat with.

Notice the privacy language.

Have fun.

AI Model Notables

► FTC and DOJ opening antitrust investigations into Microsoft, OpenAI, and Nvidia.

► Meta faces call in EU not to use personal data for AI models.

► Because Meta is using public posts on Instagram to train its models (only EU users can opt-out), artists are leaving and flocking to new anti-AI social media site.

► New terms and conditions from Adobe (Photoshop, Illustrator, etc.) give it the right to access user’s content through “manual and automated methods.”

► Google's NotebookLM gets big upgrade and expands to over 200 countries and territories around the world.

► Microsoft’s Recall takes a screenshot every five seconds. Cybersecurity researchers say the system is simple to abuse—and one ethical hacker has already built a tool to show how easy it really is.

► Anthropic shares how it is testing and mitigating election-related risks.

► Thirteen former OpenAI employees said in an open letter that AI companies have retaliated against people who speak out about safety concerns. Here is the open letter.

► Bloomberg Law’s AI-driven Clause Adviser assesses which side of a deal – buyer, seller, or neutral – an individual clause favors and allows a user to quickly modify the favorability of the clause.

► FoondaMate, an AI-powered study aid for students in emerging markets, has integrated Meta's Llama language models to enhance its conversational capabilities. The "study buddy" chatbot, accessible through WhatsApp and Messenger, provides personalized educational support to millions of students who often lack individual attention from teachers or access to textbooks.

News You Can Use:

➭ Big Tech launches campaign to defend AI use - Chamber of Progress, in a campaign called "Generate and Create," says it will showcase "how AI lowers barriers for producing art."

➭ Scientists used an AI algorithm to mine ‘the entirety of the microbial diversity’ on Earth, speeding up antibiotic resistance research.

➭ Want a job working at Harvey.ai? Here are the openings.

Who is the author, Josh Kubicki?

Some of you know me. Others do not. Here is a short intro. I am a lawyer, entrepreneur, and teacher. I have transformed legal practices and built multi-million dollar businesses. Not a theorist, I am an applied researcher and former Chief Strategy Officer, recognized by Fast Company and Bloomberg Law for my unique work. Through this newsletter, I offer you pragmatic insights into leveraging AI to inform and improve your daily life in legal services.