GenAI Weekly Updates - 6 April 2024
DBRX, 16+ Indian Languages, Drug Discovery and Ban on GenAI
Introduction
The world of large language models (LLMs) is constantly evolving, with new models, applications, and challenges emerging at a rapid pace. I track a lot of articles, white papers, open-source tools, and LLMs. I wanted to give back to the community by summarizing Generative AI (GenAI) from my perspective. I hope you'll enjoy it and find it helpful. Happy reading!
You are going know about DBRX, LLM on 16+ Indian Languages, Drug Discovery, Ban on GenAI, and Open source Tools.
This article dives into the exciting world of LLMs, exploring recent developments and key considerations for responsible AI development. We'll delve into advancements like the open-source DBRX model and the Indian-focused Indic Gemma. We'll also explore the roadblocks faced by AI, such as security concerns that led to the banning of certain AI assistants.
This article also examines the growing focus on AI regulations, with the EU AI Act and the US's Blueprint for an AI Bill of Rights setting the stage for a future of safe and responsible AI development. Finally, we'll explore the world of open-source tools designed to test and secure LLMs, along with online demos that allow you to experience the power of these models firsthand.
New LLMs Entries
Enterprise Grade Open Source LLM by Databricks
Databricks released an open-source large language model called DBRX, which achieves state-of-the-art performance on public benchmarks. DBRX is a mixture-of-experts (MoE) large language model, using
Unlike other open models, DBRX uses a unique fine-grained mixture-of-experts approach with a larger number of smaller experts, allowing for more possible combinations and better performance.
Read more, Hugging Face, Online Demo
We have done preliminary security testing of DBRX, it has got a score of 95/100 (a good score), i.e we could get 5 toxic reponse out of every 100 specially crafted (randomly selected) prompts.
Made for India - Indic Gemma Support 16+ Indian Languages
This large language model, fine-tuned on over 16+ Indian languages and built with Google's Gemma 7b model, is designed specifically for the Indian market. While the model performs well on various language-specific tasks, it's recommended for enterprise users to conduct security testing to mitigate potential vulnerabilities like toxicity, misuse, and bias. Hugging Face
Checkout our hosted version of the model using TPU V4 (Superfast) here Online Demo
FinGPT - Open source LLM models trained on Financial Data
FinGPT is an open-source framework designed specifically for applying large language models (LLMs) to financial data. Launched in 2023, it aims to democratize access to high-quality financial information and empower innovation in areas like algorithmic trading and robo-advising.
GenAI Adoption
Drug Discovery
Launch of Aurigene.AI by Dr. Reddy's subsidiary Aurigene in the domain of drug discovery . The goal is to speed up the development of new drugs. Aurigene.AI is built on a large database of chemical compounds and bioassay data (Source)
100+ Startups in Y Combinator in various categories
B2B software and services - Ocular AI, askLio, Sapling.ai, Contour, PlayHT
Consumer - Lifelike, Lumona, Somn, Juicebox, Photoroom
Healthcare - Somn
Fintech - Humanlike, Feather, OfOne
Education - Shepherd, Infobot
GenAI Risks
Ban of ChatGPT and Microsoft CoPilot
The US House of Representatives banning Microsoft Copilot on House devices . It discusses security concerns that led to the ban. The House previously banned a similar AI assistant, ChatGPT during part part of the 2024.
Will GenAI go into the Federal Cloud direction? Microsoft is developing government-oriented AI tools to address these concerns.
Regulations
EU AI Act 2024
A recent incident involving autonomous robots mistakenly collecting poisonous mushrooms from farmland highlights the importance of safe and responsible AI use. How will the EU AI Act, implemented in March 2024, will prevent such Risks and Hazards from AI? Read on to find out more. Read more on EU AI Act 2024
US - BLUEPRINT FOR AN AI BILL OF RIGHTS
Here are 3 takeaways for AI builders:
Focus on building safe and effective AI systems. This means ensuring that AI systems are designed and tested to minimize risks and produce reliable results.
Build AI systems that are fair and unbiased. This requires careful consideration of the data used to train AI systems and the potential for bias to creep in.
Make AI systems accountable and transparent. This means being able to explain how AI systems make decisions and who is responsible for their outcomes.
US FDA to Develop Regulatory Scheme for AI in Medical Products
It discusses the FDA’s efforts to ensure the safe and effective use of AI in medical products. The article also details collaborative efforts between different FDA centers. Key points from the whitepaper are:
Transparency: AI products should be transparent in their decision-making.
Bias: AI products should be designed and developed to minimize bias.
Cybersecurity: AI products should be designed and developed to be secure from cyberattacks.
Open Source Tools
LLM Security Scanners
Garak: Generative AI Red-teaming & Assessment Kit
detoxio/dtx: LLM security Scanner powered by 10+ Million Prompts
Damn Vulnerable Environments
Pokebot: a deliberately vulnerable GenAI RAG application designed to test OWASP Top 10 LLMs and AI Apps.
Github, Demo, Online Playground
AI/ML Supply Chain Security Tool
Safedep/Vet: Identifying risks in open source software supply chain.
Safedep/Vet is selected in Blackhat Asia 2024. Book your Calendar for a demo from
Snyk: Popular supply chain tool among the enterprises
AI Firewalls
Guardrails: Guardrails runs Input/Output Guards in your application that detect, quantify and mitigate the presence of specific types of risks.
Cloudflare AI firewall: Firewall for AI is an advanced Web Application Firewall (WAF) specifically tailored for applications using LLMs
Online LLM Demos
DBRX: Databricks powered state-of-the-art LLM model
Detoxio Hosted Indic Gemma 7b on 16+ Indian Languages: LLM made for India by Telegu LLM Labs
Zephyr-Gemma-7b: Yet another LLM model trained by Hugging Face on top of Gemma various versions.
Conclusion
Just like any venture, success in the LLM world hinges on Awareness and Curiosity. By staying informed about risks and regulations, and actively exploring open-source security tools, we can build a future where LLMs fuel innovation while prioritizing human safety.