Botanalytics AI Weekly Digest April 17, 2023

Creating your own chatbot using Poe's AI chatbot app is incredibly easy - as simple as posting a photo on Instagram

Quora's AI chatbot app, Poe, has introduced a new feature that allows users to create their own bots using prompts in combination with existing bots like ChatGPT. Initially, Poe launched with support for a few general knowledge bots powered by OpenAI and Anthropic, but now users can design and train their own bots for highly specific tasks or purposes. We are approaching a future where it will become possible to use chatbots in every area!?

The concept of prompts involves directing the chatbot to generate text in a certain style, format, or targeted audience, among other specifications. This feature has led to the emergence of a new creator class within the field of prompt engineering, with online communities facilitating the sharing of prompt ideas among users.

With the new feature, users can create their bots based on either Claude or ChatGPT, and the bots will have their unique URLs (e.g., poe.com/botname) that open the bot directly in Poe. Quora CEO Adam D'Angelo shared some fun bots that the company created to showcase the new feature, including a "talk like a pirate bot," a Japanese language tutor bot, an emoji converter bot, and a bot that provides mild roasts.

Users can access the bots via Poe's iOS app or Android app on mobile web or its desktop web interface. They can also follow the bots they like, and the bots will appear in Poe's sidebar bot list alongside other general-purpose bots. Quora plans to cover the costs of operating this feature, including the language model fees, for the time being, although it could become expensive if any bots become popular.

In the future, Quora plans to provide bot creators with feedback on how their bots are being used so they can iterate on improvements. The company also has plans to develop an API that would allow anyone to host a bot from a server they operate, potentially creating new business opportunities for Quora.

While some users have already used the feature to create bots for practical purposes such as trip planning or learning math, as well as for fun activities like flirting, Poe's platform guidelines restrict certain use cases like hate speech, violence, illegal activities, fraud, and IP infringement.

Poe is not the only AI chatbot app catering to mobile users, but it has gained traction with over 1.17 million installs and $520,000 in gross revenue to date. Microsoft's Bing and Edge apps have also integrated AI technology through their partnership with OpenAI, and other AI startups have launched their mobile apps as well. Poe is currently ranked No. 32 in the Productivity category on the App Store.

Be Careful What You Use it For!: Samsung’s Internal Data leaks due to ChatGPT!??

Samsung employees used the chatbot service ChatGPT to assess company source code and leaked sensitive information, resulting in three data leaks in under a month. What luck! The leaks included meeting notes and source code, which were accessed by other users of ChatGPT.

Samsung has warned its employees of the risks associated with using ChatGPT and has decided to develop its own AI for internal use. The Italian Data Protection Authority, Garante Privacy, recently banned ChatGPT temporarily due to illegal collection of personal data and the absence of systems for verifying the age of minors. OpenAI, the company behind ChatGPT, has been collecting data from users without alerting them, and there is no legal basis for the massive collection and processing of personal data to train the platform's algorithms.

Thanks to AI, Your Tweets will become your therapist!?

Researchers at the University of S?o Paulo (USP) in Brazil are using artificial intelligence (AI) and Twitter to create anxiety and depression prediction models that could potentially detect signs of these disorders before clinical diagnosis. Time to watch our tweets!?

The study involves the construction of a database called SetembroBR, which contains information on tweets from 3,900 Twitter users who have reported being diagnosed with or treated for mental health problems. The researchers are using deep learning techniques, specifically bidirectional encoder representations from transformers (BERT), to create text classifiers and word embeddings for predicting depression and anxiety. Preliminary findings suggest that it may be possible to detect depression solely based on the social media friends and followers of an individual, without considering their own posts. The researchers are continuing to refine their computational techniques and upgrade the models to potentially create a tool for screening prospective sufferers of mental health problems and providing support to families and friends of those at risk.

It's time to get ready for a new era of psychological counseling!

Future Large Language Models have a cure for toxicity!?

The contentious matters of toxicity, bias, and offensive language are prominent concerns associated with large language models. These challenges arise due to the models being trained on extensive internet text data, which could contain biases, offensive language, or controversial viewpoints. However, there are growing efforts to address the issue of toxicity.?

Then, what is the “cure”?

The cure for these challenges involves a comprehensive approach. It includes refining and enhancing training data, advancing fine-tuning techniques, incorporating feedback from users, and implementing content moderation strategies.

• Curating and improving training data

It involves carefully selecting and preparing the data that is used to train the model to ensure that it is diverse, representative, and of high quality. This process typically involves several steps:

- Data Collection

- Data Cleaning

- Data Labeling

- Data Augmentation

- Bias Mitigation

- Iterative Refinement

• Developing better fine-tuning techniques

Fine-tuning is the process of adapting a pre-trained language model on a smaller, task-specific dataset to enable it to perform a specific language-related task, such as text classification, named entity recognition, or sentiment analysis.?

What are those techniques?

• Dataset Size
• Task-specific Data Representation
• Optimization and Regularization
• Hyperparameter Tuning
• Transfer Learning
• Regular Updates
• Incorporating user feedback

User feedback provides important insights into the model's performance, identifies areas for improvement, and helps the model better understand and respond to user inputs.

• Fine-tuning
• Data collection
• Model updates
• Model evaluation
• Iterative development
• Content moderation strategies

Those involve various techniques and approaches to ensure that the generated content adheres to specific guidelines, policies, and ethical considerations. These strategies can be implemented during the model's development, deployment, and usage phases.