Julia Neagu is the co-founder and CEO of Quotient AI.
getty
Generative AI is reshaping industries, and the debate over control of the models driving it is intensifying. Once limited to tech giants, large language models (LLMs) are now more accessible. This article will examine the growing competition between open-source and proprietary LLMs, exploring how open models are challenging the dominance of closed systems and shaping the future of AI.
The State Of Open Versus Proprietary LLMs
Open LLMs are publicly available models like Meta’s Llama 3.1. Their code, architecture and sometimes training data can be accessed, modified and used for commercial purposes. A comprehensive list of open LLMs can be found here.
The rise of foundational models has created distinct cohorts of providers. Cloud infrastructure giants such as Google and Amazon Web Services (AWS) are embedding models like Gemini and Titan directly into their ecosystems. Social media giants like Meta and X have released their own LLMs, with X’s Grok-2 exclusive to its platform. AI leaders like OpenAI, Anthropic and Cohere also offer enterprise models like GPT-4, Claude and Command R. Among these, most remain proprietary, with only Meta taking the open model approach.
AI developers continue to default to proprietary models, especially as their first choice. The appeal lies in simplicity and reliability—these models offer high performance right out of the box without requiring extensive customization. By eliminating the need to source inference providers or manage complex infrastructure, they provide a seamless, plug-and-play experience that allows developers to focus on building solutions rather than backend logistics.
But as open LLMs become more competitive, developers are turning to them as cost-effective, privacy-focused alternatives. With full control over hosting and data management, these models help offer better customization, security and compliance, ultimately making them an attractive choice for businesses seeking performance without vendor lock-in.
Cost Efficiency
The cost of utilizing state-of-the-art open models versus proprietary ones is a major consideration. Open models can be cheaper, particularly if deployed on cost-effective hardware. According to recent pricing comparisons, open models like Meta’s Llama 3.1 can offer significant savings compared to comparable proprietary models. The total cost of ownership depends on factors like deployment scale, hardware and the choice of cloud or on-prem infrastructure.
Benchmark Performance
Meta’s Llama 3.1 has shown competitive performance in public benchmarks compared to proprietary models. For certain tasks, open LLMs can even surpass proprietary models when tested on domain-specific evaluation benchmarks, as shown in Quotient AI’s research.
Additionally, innovative solutions like Martian have shown that optimized open model routing can achieve up to a 97% reduction in costs on specific tasks while matching GPT-4’s performance, highlighting the potential cost benefits of open models.
Flexibility And Customization
Open models become especially powerful when fine-tuned—a process that involves adapting a pre-trained model to a specific dataset or task. Fine-tuning allows companies to enhance the model’s performance by training it on their own data, improving its accuracy, relevance and alignment with specific business needs. For example, a company working in medical diagnostics might need a model fine-tuned on specialized medical terminology and datasets.
Unlike proprietary models, developers fine-tuning open models have full access to the model’s weights. These are hosted on platforms like Hugging Face, enabling them to make deeper customizations and deploy the model on their own infrastructure.
Companies like OpenPipe are simplifying and reducing the cost of fine-tuning by providing tools that make the process more accessible. For instance, OpenPipe’s implementation of Together AI’s "Mixture of Agents" approach combined multiple fine-tuned agents to outperform GPT-4 on specific tasks such as customer support query handling.
Scalability And Resource Utilization
Open Model Scalability
Organizations can deploy open LLMs across various hardware setups, from on-premise data centers to cloud infrastructure. This adaptability allows businesses to tailor their scalability strategies based on their specific requirements and budget constraints.
Companies like Together AI, Fireworks AI, and Groq are emerging players that focus on making it easier to deploy, scale and optimize open LLMs. Their tools and platforms streamline the deployment process, improving scalability and making it easier for businesses to leverage the power of open AI.
As more providers enter the market, inference hosting is becoming increasingly commoditized, with competition driving down prices for hosting these models. This shift is making high-performance AI more affordable and accessible, and it is reducing barriers for companies looking to integrate open LLMs into their workflows.
Proprietary Scalability
Proprietary models, especially cloud-based ones, offer built-in scalability. Companies like OpenAI have invested heavily in the ability to handle hosting, load balancing and performance optimization for a seamless user experience. This hands-off approach suits businesses seeking simplicity, but it can become costly for large-scale deployments. While enterprise deals offer custom pricing and SLAs, they may add layers of complexity and require negotiation.
Considerations For Adopting Open LLMs
For companies considering the adoption of open LLMs, several factors should be evaluated:
Infrastructure And Deployment
Companies must plan their deployment strategy carefully. This includes selecting the right ETL solutions (like Unstructured.io), embedding models, vector databases (like Pinecone or Qdrant) and orchestration tools (like LangChain or LlamaIndex).
Customization And Fine-Tuning
As the ability to fine-tune an open model on specialized data can be a significant advantage, companies should evaluate how well the model can be customized to meet their unique needs.
Domain-Specific Performance
Companies must evaluate open models on domain-specific datasets to ensure performance, such as testing an open LLM on financial reports in the finance industry. Setting benchmarks for metrics like precision is crucial to avoid missing key insights or mishandling industry-specific jargon.
Looking Ahead
The landscape of AI is rapidly evolving. The growing ecosystem around open tools, platforms and deployment strategies will likely fuel further innovation and drive the widespread adoption of these models.
Companies are increasingly evaluating the trade-offs between control, customization and convenience when choosing solutions. While some organizations may benefit from the adaptability and transparency offered by open AI models, others might prioritize the integrated support and streamlined deployment of proprietary systems. With the increasing accessibility of open models, future breakthroughs in AI could emerge from either domain, ushering in a new era of high-performance solutions.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
{Categories} _Category: Takes{/Categories}
{URL}https://www.forbes.com/councils/forbestechcouncil/2024/10/09/how-open-models-are-disrupting-the-ai-status-quo-and-shaping-the-future-of-innovation/{/URL}
{Author}Julia Neagu, Forbes Councils Member{/Author}
{Image}https://imageio.forbes.com/specials-images/imageserve/6705768b5d1b4588b9139734/0x0.jpg?format=jpg&height=600&width=1200&fit=bounds{/Image}
{Keywords}Innovation,/innovation,Innovation,/innovation,technology,standard{/Keywords}
{Source}POV{/Source}
{Thumb}{/Thumb}
