US CIO study reveals that 42% of US organizations are considering developing their own proprietary large language models (LLMs). Although over 90% of enterprise software code is open source, companies remain cautious about using open-source LLMs for critical applications, mainly due to security and privacy concerns. However, they are beginning to use these models for basic business applications.
For instance, open-source LLMs, like Meta’s Llama 2 and Falcon 180B, are free, making them attractive for smaller organizations with limited budgets. They reduce reliance on external vendors but raise concerns about security and privacy.
On the other hand, in 2023, Apple restricted its employees from using publicly available AI tools like ChatGPT for security reasons. Similarly, Samsung barred its staff from using AI tools after discovering a data breach involving ChatGPT. Italy also banned ChatGPT, citing privacy issues.
Organizations have other reservations about open-source LLMs, including accuracy, dependency on vendors, and licensing fees. Wilbertus Darmadi, CIO of Toyota Astra Motor, expressed concerns about the accuracy of AI-generated analytics during an ISMG panel discussion. He noted that the effectiveness of AI and machine learning depends on the quality of the data source. LLMs can produce inaccurate information, exhibit bias, raise consent issues, and pose security risks.
Consequently, more organizations are leaning towards proprietary LLMs, customizing them with their data for better control and security.
Understanding Open Source LLMs
Open-source LLMs are accessible to the public, allowing developers to inspect, modify, and distribute the code. This transparency fosters a community-driven development process, leading to rapid innovation and diverse applications. With transparency and flexibility, they are essential for enterprises without in-house machine-learning expertise. They also offer performance improvements through efficient optimization. They are ideal for projects that require customization and those where budget constraints are a primary concern.
Open-source LLMs are more affordable in the long run as they don’t involve licensing fees. However, they do entail infrastructure costs and an initial setup investment.
They also allow for added features and benefit from community contributions. Unlike proprietary models that depend on a single provider, open-source models can leverage multiple providers and internal teams for updates and support. This flexibility enables businesses to stay at the forefront of technology and exercise greater control over their tech usage.
Organizations can use open-source LLMs for various projects, including:
- Text generation for emails, blog posts, and stories. A notable example is Falcon LLM from the Technology Innovation Institute, which is used for creative text generation and problem-solving.
- Code generation to assist developers, such as StarCoder from Hugging Face, a coding assistant.
- Virtual tutoring for personalized learning experiences. Notable examples like Hugging Face Transformers and FairSeq offer diverse applications in virtual tutoring, including personalized question generation, real-time feedback, and custom lesson planning.
- AI-driven chatbots for interaction and assistance. Rasa and AllenAI’s AllenNLP are frameworks offering features like natural language understanding, reinforcement learning, and various NLP tools for advanced and engaging chatbot development.
- Language translation across multiple languages. PolyLM, LLaMA, and FLAN-T5 are examples of open-source LLMs for language translation, excelling in multilingual support, translating low-resource languages, and providing high-quality translation across various languages.
- Sentiment analysis for brand management and customer feedback. Flair, NLTK, spaCy, and TextBlob are prominent open-source LLMs for sentiment analysis, offering capabilities like polarity extraction, pre-trained models, and custom training for analyzing sentiment in various text forms.
Examples include IBM and NASA using open-source LLMs for climate change research, publishers and journalists for data analysis, healthcare organizations for patient care tools, and the financial industry with models like FinGPT.
“52% of organizations express security concerns in using Generative AI; 42% of the organizations contemplate developing their own proprietary large language models.”
– findings from the CIO survey
Navigating Proprietary LLMs
Proprietary LLMs, on the other hand, are developed and maintained by specific companies. They often come with robust support and consistent updates and are built to meet high-performance standards. These models are ideal for businesses that require reliability, support, and advanced features.
Companies like OpenAI, Microsoft, and AWS offer proprietary LLMs as a service. They provide public APIs to models such as OpenAI’s GPT-4 or those hosted on Microsoft Azure, allowing easy deployment and access to advanced models for various uses. Let’s explore some of the more popular models available for enterprises to select from:
- Anthropic’s Claude 2: Based in San Francisco, Anthropic released Claude 2 in July 2023. It excels in open-ended conversations, has a larger context window than ChatGPT, and uses self-supervised learning for contextual understanding. Claude 2 can generate and summarize texts, assist in research, and quickly summarize lengthy documents.
- OpenAI’s GPT-4: OpenAI’s latest version, GPT-4, is a multimodal LLM that accepts text and image inputs. It shows advanced reasoning, coding skills, and academic proficiency. GPT-4, with over one trillion parameters, has improved on the hallucination issue found in its predecessor and supports up to 32,768 tokens. Accessible via platforms like Microsoft Bing and ChatGPT, it allows users to set the tone and task through natural language directives.
Deciding when to use open-source vs. proprietary LLMs
The choice between open-source and proprietary LLMs depends on various factors, including your project’s specific needs, resources, and objectives.
When to use open-source LLMs:
- Budget constraints: Open-source LLMs are generally free, making them ideal for organizations with limited budgets.
- Customization and innovation: If your project requires extensive customization, open-source models offer the flexibility to modify and adapt the model. They are also beneficial for projects aiming for innovative solutions through community-driven development.
- Experimentation and learning: Open-source LLMs are suitable for educational purposes or experimental projects where the cost is a concern and the risk of data breaches is minimal.
- Diverse applications: Open-source models like Hugging Face Transformers, Flair, and NLTK are excellent for specific applications such as sentiment analysis, virtual tutoring, and language translation, especially when tailored solutions are needed.
But there are limitations that should be highlighted:
- Fluency and authority issues: LLM outputs can appear convincing but might present risks, including reliance on erroneous or “hallucinated” information and challenges related to bias, consent, and security.
- Educational solutions: Educating users about these risks is a key strategy for addressing the complex issues surrounding data and AI.
- Hallucinations: LLMs can generate false information due to training on flawed, inconsistent, or inaccurate data or by predicting contextually accurate words without true comprehension.
- Bias: Occurs when the data source lacks diversity or fails to represent a broad spectrum of information.
- Consent issues: Concerns whether the data was collected responsibly, adhering to AI governance standards, legal compliance, and offering mechanisms for user feedback.
- Security concerns: Risks include the potential exposure of personally identifiable information, the misuse of LLMs for harmful activities like phishing or spam, and the threat of hackers altering the LLM’s programming.
- Resource intensive: Implementing and maintaining an open-source LLM can be resource-intensive, requiring significant technical expertise.
- Variable quality: The quality of open-source models can vary, and there might be less accountability for performance issues.
When to use proprietary LLMs:
- Security and privacy: If your project involves sensitive data or requires high security and privacy standards, proprietary LLMs are often more secure and regularly updated to address vulnerabilities.
- Reliability and support: For business-critical applications where consistent performance and dedicated support are crucial, proprietary LLMs like GPT-4 provide a more controlled and reliable environment.
- Resource availability: If your organization has the resources to invest in proprietary models, these can offer advanced features and integrations that might not be available or easy to implement with open-source models.
- Regulatory compliance: In scenarios where strict regulatory compliance is necessary, proprietary models often have the infrastructure to ensure adherence to laws and regulations.
Having examined their advantages, it’s important to also consider the limitations proprietary models may present in various enterprise applications:
- Cost: These models can be expensive, particularly for small businesses or individual developers.
- Less customizable: There is often limited scope for customization in proprietary models.
The choice between open-source and proprietary LLMs is not always clear-cut. For many organizations, a hybrid approach might be the most effective. This involves using open-source models for certain aspects of the project where customization and innovation are key and proprietary models for parts where security, reliability, and support are important for the organization’s business.
Both types of LLMs benefit significantly from fine-tuning, a process of adjusting the model to better suit specific tasks or datasets. While open-source LLMs offer more freedom in fine-tuning, proprietary LLMs often come with tools and support that simplify this process.
In summary, the decision to use an open-source or proprietary LLM should be based on a careful evaluation of your project’s specific needs, the resources at your disposal, and the potential risks involved, especially in terms of security and compliance. Remember, the right choice varies from one project to another. Evaluate your requirements carefully to make an informed decision that aligns with your goals and resources.