The AI Revolution Will Not Be Monopolized

This is a summary of a talk that Ines Montani gave at QCon London in April 2024. Large language models (LLMs) have significantly transformed the field of artificial intelligence (AI). The fundamental innovation behind this change is surprisingly straightforward: make the models a lot bigger. With each new iteration, the capabilities of these models expand, prompting a critical question: Are we moving toward a black box era where AI is controlled by a few tech monopolies, obscured behind APIs and proprietary systems?

The Open-Source Counterpoint

Contrary to this concern, open-source software is disrupting the notion of monopolistic control in AI. Open-source initiatives ensure that no single entity can dominate the AI landscape. Open-source software offers numerous advantages that make it an attractive choice for both individuals and companies:

• Transparent: Open source software is transparent, allowing you to see exactly what you're getting.
• No Lock-In: You’re not locked into a specific vendor. While there is some commitment, you'll never lose access.
• Runs In-House: Open source software can run in-house, which is crucial if you're working with private data and prefer not to send it to external servers.
• Community-Vetted: The community vetting aspect means you can see what’s popular and who uses what, ensuring a level of trust and reliability.
• Up to Date: Open source projects are often up to date, incorporating the latest research through pull requests and community contributions.
• Programmable: The software is very programmable, rarely requiring an end-to-end solution, and can be integrated into existing processes with ease.
• Easy to Get Started: It's easy to get started with open source software; you can simply use commands like pip install to download and begin.
• Extensible: The software is extensible, allowing you to fork it and run it yourself if needed.

The Economic Aspect of Open-Source

One common misconception about open-source software is that companies primarily chose it because it's free. While many open-source projects are freely available, the real value lies in their accessibility and the freedom they offer. While the cost factor helps with initial adoption, numerous compelling reasons drive the dominance of open-source solutions.

Open source in AI and machine learning is not just about software, it's about the synergy of code and data. The growing ecosystem of open-source models encompasses everything from code to data and weights, making powerful tools widely accessible. To clarify the landscape, let's categorize these models into three types:

1. Task-Specific Models: These are specialized models designed for specific tasks. Examples include models distributed with spaCy and its community projects, models for Stanford’s Stanza library, as well as numerous models on platforms like Hugging Face. These models are generally small, fast, and inexpensive to run. However, they don't always generalize well and often require fine-tuning with domain-specific data.
2. Encoder Models: These models, such as Google's BERT and its different variations, are used for generating embeddings that can power task-specific models. They are relatively small, fast, and affordable to run in-house, offering better generalization than task-specific models but still needing some fine-tuning for specific applications.
3. Large Generative Models: This category includes models like Falcon, Mistral and LLaMA. These models are significantly larger, slower, and more expensive to run but excel at generalization and adaptation, requiring little to no fine-tuning to perform specific tasks.

Misunderstanding LLMs

The term "large language models" (LLMs) is often used broadly and imprecisely, muddying discussions about their capabilities and applications. The distinction between encoder models and large generative models is therefore very important. Encoder models involve task-specific networks that predict structured data, while large generative models rely on prompts to produce free-form text, necessitating additional logic to extract actionable insights.

The Role of Economies of Scale

Large generative models, due to their complexity and operational cost, are often accessed through APIs provided by companies like OpenAI and Google. These companies leverage economies of scale, benefiting from access to top talent, wholesale compute resources, and a high volume of requests that allow efficient batching. This setup works like a train schedule in a busy city, making it viable to offer frequent service due to high demand.

The Distinction Between Human-Facing and Machine-Facing AI

A critical distinction in the AI landscape is between human-facing systems and machine-facing models. For human-facing systems, such as ChatGPT and Google Gemini, the most important distinction is product features, including user experience, user interfaces, and customization, often incorporating constraints to prevent undesirable outputs. These products interact directly with users and rely heavily on user data to improve and refine their functionalities. In contrast, the underlying models like GPT-4 and Bard are components in a larger system, forming the backbone of these consumer-facing applications. Machine-facing models are swappable components built on openly published research and data, with performance quantified in terms of speed, accuracy, latency, and cost.

Understanding the differences between these types of AI applications is essential. This distinction helps clarify misconceptions about monopolizing AI. Companies like OpenAI might dominate the market for user-facing products but not necessarily the AI and software components behind them. While user data is advantageous for improving human-facing products, it is less critical for enhancing the foundational machine-facing tasks. Gaining general knowledge doesn't require specific data, which is at the core of the innovation behind large generative models. 

Capabilities of AI in Practice

AI capabilities in practice can be broadly categorized into generative and predictive tasks:

• Generative Tasks: Summarization, reasoning, problem solving, question answering, paraphrasing, and style transfer are new capabilities enabled by generative models.
• Predictive Tasks: Text classification, entity recognition, relation extraction, coreference resolution, grammar and morphology, semantic parsing, and discourse structure. These tasks involve converting unstructured text into structured representations, which are then used in various applications.

While generative AI offers many new possibilities, a lot of industry challenges have also remained the same, primarily focusing on structuring unstructured data like language. The advent of AI allows us to tackle these problems more efficiently and at a larger scale, enabling more structured data creation and project completion.

Evolution of Telling Computers What to Do

The process of instructing computers has evolved through several iterations:

1. Rule-Based Systems: Initially, we provided rules or instructions using conditional logic and regular expressions.
2. Machine Learning: Introduced programming by example, also known as supervised learning, where models are trained using specific examples.
3. In-Context Learning: More recently, providing rules and instructions in natural language form (prompts).

Each method has its pros and cons. Instructions are intuitive and easy for non-experts but can be susceptible to data drift. Examples are highly specific and can express nuanced behaviors but are labor-intensive to generate. So how could a workflow look that combines both methods and uses large general-purpose models with specific data to develop focused, task-specific models?

Practical Applications and Transfer Learning

A practical AI workflow involves evaluating and correcting model predictions iteratively, using transfer learning to distill general-purpose models into specific ones. Transfer learning remains relevant for practical applications, allowing for modular, interpretable, and cost-effective solutions.

Using large generative models helps overcome the cold start problem, enabling prototypes to work out of the box. These prototypes can be refined and distilled into smaller, faster, and more specific models. This approach avoids the labor-intensive process of generating examples from scratch and reduces the dependency on massive, complex models at runtime.

Human-in-the-Loop Distillation of Task-Specific Models

Developing distilled task-specific models aligns with software best practices, offering numerous benefits:

• Modular: The approach is highly modular, aligning with software development best practices. This allows for maintaining modern workflows and adapting model development accordingly.
• No Lock-in: Users are not tied to any specific provider. Models can be developed with various providers but can be owned and managed independently at runtime.
• Testable: Components can be tested individually, making it easier to monitor and detect failures compared to a single black-box system.
• Flexible and Cheap to Run: Models are flexible components in a system and can be optimized to run efficiently, even on CPUs or with small footprints, reducing operational costs significantly.
• Runs In-House: This is crucial for handling sensitive data securely without relying on external APIs, ensuring data privacy and regulatory compliance.
• Transparent and Predictable: Users have visibility into the workings of the models, allowing for better understanding and predictability of model behavior.
• Programmable: Models can be integrated programmatically into existing workflows, aligning with business needs and minimizing integration challenges.

These are the same reasons why companies choose open-source software, which is not a coincidence: AI development is still a type of software development and the same principles apply.

Addressing Concerns and Regulation

Economies of scale, once thought crucial for monopolistic dominance, face challenges in tech due to intense competition driving down costs. The ability to rely on otherwise costly open-source models during development instead of production makes this an even less relevant moat.

Regulation emerges as another strategy pursued by big tech companies to secure their monopoly in the space, lobbying governments across the world to implement AI legislation that only they can comply with.

Maintaining clarity in regulation is essential to ensuring AI evolves without monopolistic control. By delineating between applications and core technologies, policymakers can foster a competitive landscape that encourages innovation while protecting consumer interests. This distinction is crucial in steering AI towards a future of innovation and accessibility, where no single entity holds undue market influence.

Conclusions

The landscape of AI development and deployment is characterized by transparency and accessibility rather than secretive advantages. In the realm of large language models (LLMs), which are integral components rather than standalone products, there is no inherent monopoly-building advantage from proprietary knowledge or exclusive data access.

These models can be effectively replaced or complemented by other methods, fostering interoperability and competition, the opposite of monopoly. Open-source software plays a crucial role in ensuring such flexibility and promotes innovation through collaborative development and community scrutiny.

However, the potential for regulatory measures to inadvertently favor monopolistic practices remains a concern. To safeguard against this, regulations should focus on regulating actions and use cases rather than targeting specific technologies or software components.

This balanced approach is essential to maintaining a competitive and inclusive environment in AI development. It also steers clear of undue influence from industry lobbying efforts that may seek to distort regulatory frameworks for their own gain.