Large Language Models

A Large Language Model (LLM) is a language model notable for its ability to achieve general-purpose language understanding and generation.

LLMs acquire these abilities by learning statistical relationships from text documents during a computationally intensive self-supervised and semi-supervised training process.

It uses neural networks following a transformer architecture and deep learning to model language in a holistic way, capable of performing multiple NLP tasks without task-specific modifications.

List of LLM

LLM featured on this post:

  • OpenAI ChatGPT
  • Google Bard
  • Google Gemini
  • Google Gemma
  • Microsoft Bing AI
  • Microsoft Phi3
  • Meta LLaMA
  • Alibaba Qwen2
  • Mistral AI
  • Aya

OpenAI ChatGPT

ChatGPT, developed by American company OpenAI, it is open worldwide to the general public.

Official website

Google Bard

Bard, developed by American company Google, is only available in some countries as of 2023.

Official website

Google Gemini

Gemini is developed by American company Google.

Official website

Google Gemma

A lightweight LLM from Google DeepMind.

Microsoft Bing AI

Bing AI is Microsoft’s chatbot. As of 2023, it is open only to testers.

Official website

Microsoft Phi3

Phi3 are lightweight LLMs from Microsoft.

Meta LLaMa

Language Model Meta AI (LLaMA) is developed by American company Meta.

LLaMa 3 is an openly available LLM from Meta

Official website

Alibaba Qwen2

Qwen2 is a new series of LLMs from Alibaba.

Mistral AI

Mistral AI is a 7B model from Mistral AI..

It is backed by a French start-up known and it is one of the major players of an European AI.

Mistral AI official website

Aya

Aya is multilingual models in 23 languages.

Running LLMs on a Local Machine

Solutions to run LLMs on a local machine:

  • Ollama

Ollama

Ollama is a free and open source software to run LLMs locally. It is compatible with many other LLMs.

You can read this TheNewStack artcile about Ollama.

LLM Security

LLM Security Standards

OWASP Top 10 for LLM Applications

OWASP Top 10 for LLM Applications official website

LLM Security Tools

LLM security tools:

  • Vigil
  • LLM Guard

Vigil

Vigil is a LLM security scanner.

It is an security scanner that detects prompt injections, jailbreaks, and other potential threats to Large Language Models (LLMs).

It is free and open source (FOSS).

Vigil code repository

LLM Guard

LLM Guard is a toolkit designed to fortify the security of Large Language Models (LLMs). It is designed for easy integration and deployment in production environments.

It is free and open source (FOSS).

LLM Guard code repository

LLM Security Concerns

Summary of AI Chatbot Security Concerns

  • Upload sensitive information
  • Data Protection
  • Children Protection
  • Malicious use of the application
  • Misinformation

Upload Sensitive Information

Users or employees may upload sensitive information to the website.

Depending on the chatbot conditions, this information could be visible to the support team of the tool.

For example when using ChatGPT’s API, the conversations with the chatbot are not visible to OpenAI’s support team and are not used to train the company’s models.

This does not happen for the general public version. The user input is visible to support team and it may be used to train ChatGPT.

Data Protection

Data protection was among the reasons why ChatGPT was banned in Italy in 2023. ChatGPT added a European Union’s GDPR-compliant form before it was readmitted in this country.

Children Protection

Children protection was among the reasons why ChatGPT was banned in Italy in 2023. ChatGPT added an age verification before it was readmitted in this country.

Risks related to Internet Exposure

Chatbots may have security issues that may compromise privacy. For example, ChatGPT bug temporarily exposed AI chat histories to other users, as it can be read on this external link.

This risk is also shared with any cloud tool that is exposed to the internet, like social networks, online banks, etc.

Malicious use of the Application

This aspect is not exclusive to ChatGPT. In fact, any tool that could be used be used for malicious intents (e.g., an e-mail account) presents a risk.

The main concern is that a tool with so much potential as ChatGPT means both potential benefits and potential misuse.

https://hbr.org/2023/04/the-new-risks-chatgpt-poses-to-cybersecurity

Misinformation

Users could be misinformated by chatbots, in many aspects: technical, political, ethical, etc. This could be deliberate or because of errors in the chatbot.

Take into account that this risk exists on any other tool, like media, newspapers, social networks, books, etc.

Organizations that have restricted the use of LLMs

Organizations that have restricted the use of chatbot:

  • Countries
    • Italy
  • JPMorgan
  • Software
    • Samsung
  • Banking
    • Bank of America
    • Citigroup
    • Deutsche Bank
    • Goldman Sachs
    • Wells Fargo

An article about chatbox user restriction in Italy can be read on this external link.

An article about chatbot use restriction on Samsung can be read on this external link.

An article about chatbot use restriction on JPMorgan can be read on this external link.

An article about the restriction on banks can be read on this external link.

You might also be interested in…

External References

Leave a Reply

Your email address will not be published. Required fields are marked *