The world’s largest aircraft breaks cover in Silicon Valley.

November 10, 2023

An annoyed cat

DataPanthy #8

DeveLearn

2,764 followers

November 10, 2023

Editor's Note :

Welcome to the 8th edition of DataPanthy. We hope that the Diwali spirit has reached you and wish you a safe and happy break ahead. As usual, we continue to look at and share the most impactful in Tech News and Industry Developments. Also, we’re experimenting with a new format of Breaking Down research for you as compared to our usual Featured Job listing.

A quick Summary of this week,

We take you on a journey through the skies as "A Giant Takes Flight." The remarkable emergence of the Pathfinder 1 Airship in the heart of Silicon Valley, where innovation knows no bounds.
Samsung has joined the Generative AI race with their alternative to ChatGPT - Gauss.
Lastly, we venture into the medical world, where groundbreaking technology offers hope to Parkinson's patients. The potential for spinal implant technology to restore mobility in individuals living with this challenging condition.

We hope you find this edition enlightening and inspiring, and we look forward to continuing this journey together in the editions to come. We will be taking a break next week on account of the festivities. See you on Edition #9.

Thank you for your continued support, and happy reading!

Headliner :

The world’s largest aircraft breaks cover in Silicon Valley.

With the first glimpse of Pathfinder 1, a prototype electric airship is being unveiled to the public as dawn breaks over Silicon Valley, its creator LTA Research intends to speed the humanitarian endeavors of its backer, Google co-founder Sergey Brin, and usher in a new era of climate-friendly air travel.

News of the Week :

Samsung unveils ChatGPT alternative Samsung Gauss.

Samsung has introduced a new generative AI model, Gauss, its answer to OpenAI’s GPT and Google’s LaMDA at its AI Forum. The AI model can generate text, images, and even code. This will be integrated into the upcoming Galaxy S24 phones and is said to run locally on devices.

InstaScoop :

Instagram is finally testing a feature to let you turn off read receipts.

In a message on his broadcast channel, Instagram head Adam Mosseri announced that the company is testing a new feature that allows users to turn off read receipts or the ability to be "seen" on Instagram DMs.If users turn this feature off, they can prevent others from seeing if they've read a particular message. Of course, this doesn't work for everyone - if you're an anxious texter and require read receipts, you can keep the feature on.

DoorDash’s Latest Safety Innovation :

DoorDash adds new safety tools for its delivery people, including ‘driving insights.

DoorDash, a leading food delivery platform, has unveiled three innovative in-app safety tools designed to enhance the safety of its delivery personnel. These tools aim to provide valuable insights and promote safe driving practices, underlining the company’s commitment to the well-being of its workforce.

Revolutionising HealthCare :

This spinal implant tech could restore mobility to Parkinson’s patients.

Researchers have released their findings into an experimental spinal neuroprosthesis device that’s designed to help people with Parkinson’s to walk properly. A spinal electrode implant, installed via surgery, has shown promise in helping people with mobility issues from the disease.

The State of Academia

In this first iteration, we’re breaking down an important but underrated breakthrough in the realm of LLM (Large Language Model) Research in this peculiar paper titled : Who is Harry Potter? Approximate Unlearning in LLMs

The What : Microsoft Research has published a technique for Unlearning a subset of the training data from an LLM without having to retrain the model from scratch. They demonstrate their technique by having a model (Llama2-7b) unlearn the Harry Potter books from its knowledge base.

The Why : Large Language Models are the technology that have defined the current AI trend. They are the core tech that lies under ChatGPT and other Text based AI models, and so a huge focus has gone into making them learn more and learn faster. But little thought has been given to what it should NOT know. Information that may be Biased or Harmful like False News, or even protecting Copyrighted or Confidential content from reaching the public. Such issues will come to the forefront as use of LLM powered AI products become more mainstream.

Full Article : Microsoft Research Blog

The How (in layman terms): Normally to ensure that an LLM does not include certain information, we would have 2 options : Option 1 - Simply exclude the topic from the training data. This would make the model neutral to predicting answers to questions on the topic. Option 2 - (Re)train such that the model cost function punishes correct predictions on the topic to be excluded (Negative Reinforcement).

Retraining from scratch would usually be unthinkably expensive in time and cost whereas method 2 would take long and results are not guaranteed. Here comes the proposed method, that provides the model with a plausible alternative to the token that would have otherwise been correct for the topic. They call this alternative generic prediction and achieve this via a process called reinforcement bootstrapping.

Read the full paper : Arxiv

Editor’s Note : I particularly find the results of this research amusing in that not only have they provided a novel way to achieve something important, but also that the foundational logic behind the proposed solutions coincides with what is found to be effective in our lives as well. The best method for “Unlearning Habits” as realized by psychologists is much similar to the method used here, in that the best way to get rid of a bad habit, is not to try to “Unlearn” it, but to Replace it with better habits.. Reinforce alternative behavior in the face of the triggers that lead to the habits you wish to get rid of. It seems that Nature has always known the keys to navigating the intelligence of our biological minds, and we are simply re-discovering them in the context of mechanical intelligence as we go along.

Resource Spotlight :

This week's Resource Spotlight showcases datasets commonly employed in responsible ML evaluations and repositories of intriguing and crucial data sources, providing a valuable resource for practitioners and researchers in the field of machine learning.

Data Repositories:

🔥 Don't Miss Out! Subscribe Today for FREE:

Datapanthy is a fully free NewsLetter curated by DeveLearn as part of its core mission and business of training young graduates & professionals alike in key skills of Analytics, Data Science & Web technologies. Join a growing community of data enthusiasts who have already unlocked the power of data with “DataPanthy.” Add Datapanthy to your weekly reading digest and embark on a journey of continuous learning, innovation, and success in the world of data science!

📩 Subscribe Here: SUBSTACK.COM/@DEVELEARN

🌐 Learn more at https://www.develearn.in/

📷Instagram: https://instagram.com/develearn?igshid=MzRlODBiNWFlZA==

🛜Facebook: https://www.facebook.com/develearntechnologies?mibextid=LQQJ4d

📌 LinkedIn: https://www.linkedin.com/company/develearn/

🎥 YouTube : https://youtube.com/@thedevelearninstitute

🚀Elevate your data game with "DataPanthy” – Where Data Science Meets Inspiration! 🚀