Model Citizens, Why AI Value Is The Next Business Yardstick

The effectiveness of a company's AI model is now a key performance indicator. Since the AI boom, generative AI has been used for everything from composing birthday invitations to writing software code. This has led to a proliferation of language models (large and small) and their associated applications.

Recent years have witnessed AI leaders pushing model boundaries, boasting ever-increasing parameter counts—Llama's latest models, for instance, are trained on 70 billion parameters. However, this trend necessitates a reassessment of AI model development strategies.

Diminishing Returns on Model Size

Shane McAllister, lead developer advocate at MongoDB, highlights a crucial juncture: "The methods that previously yielded significant intelligence gains (increasing compute power and parameters) are now showing diminishing returns." The sheer volume of internet data accessible to even the largest models presents a limit to what can be learned effectively. While models are becoming more intelligent, excessive power is often unnecessary for typical business applications.

The focus has shifted. Enterprises are finding the most value in models providing accurate, domain-specific expertise—an area where general-purpose LLMs often fall short due to outdated or inaccurate data, resulting in unreliable responses.

Chris Mahl, CEO of Pryon, emphasizes this point: "The debate over model size misses the mark. Companies achieve impressive results by combining the reasoning capabilities of large models with specialized knowledge using techniques like RAG and fine-tuning. It's not an 'either/or' choice; the real innovation lies in integrating both approaches to solve specific business problems."

Both experts agree that most AI tasks are of low-to-medium complexity (summarizing documents, creating emails, basic data analytics). Employing massive models for these tasks is akin to using a supercomputer to send a text message—overkill and inefficient.

The Importance of Rightsizing

McAllister stresses the importance of "rightsizing" AI projects. Careful selection of appropriate language models, considering their scale, is crucial. The tendency to default to ChatGPT, he argues, is akin to the past practice of always choosing IBM—a knee-jerk reaction that ignores optimal solutions. Rightsizing should be a core component of AI governance, as not every task requires the power of GPT-4.

This rush to implement AI may stem from a desire to quickly capitalize on the technology's potential. However, the computational costs of large models only become fully apparent over time.

McAllister notes that in regulated industries, smaller models often outperform larger ones because they can leverage highly specialized data crucial for accurate responses in those sectors—data not fully captured in the training of general-purpose LLMs. Furthermore, enterprises can utilize multiple SLMs through intelligent model routing or reasoning engines, selecting the optimal model for each task dynamically.

The Cost of Application Emissions

Running large LLMs in production is expensive, consuming significant computing power and electricity. Smaller models offer lower costs and reduced energy consumption, leading to lower "application emissions"—a concept ripe for formalization in the tech industry. McAllister also points out that SLMs offer increased deployment flexibility, particularly beneficial in resource-constrained environments or those with strict data security requirements.

John Nay, CEO and co-founder of Norm AI, counters that while smaller models enhance governance and data sovereignty, the broader regulatory concern for increasingly autonomous AI centers on assessing its output against relevant laws and regulations, a challenge not solely addressed by model size.

Smaller models do bolster governance and data sovereignty, becoming increasingly vital as regulations evolve. Developers are prioritizing minimizing reliance on large, centralized models to ensure data compliance and residency requirements. Ultimately, true AI value lies in suitability, not scale.

Siqi Chen, CEO and founder of Runway, emphasizes the importance of considering the value and volume of work: "Businesses should assess whether the cost of a sufficiently capable model is a significant portion of the human labor cost for the same task. In most cases, the AI model cost is a small percentage. Then, consider if demand is elastic with cost. For some tasks, lower cost increases demand; for others, it doesn't. In the latter case, using the most capable model makes sense." He acknowledges that while SLMs excel in vertical use cases, their economic value is often limited, and generalized models can outperform specialized ones.

Model Implementation: Not a Weekend Project

Julian LaNeve, CTO of Astronomer, cautions that enterprises seeking SLMs may need to self-host, a significant undertaking. Hosting, scaling, and maintaining LLM infrastructure requires substantial investment, particularly in personnel. However, he believes models will continuously improve in speed, efficiency, and cost, making currently marginal use cases more viable in the future. He cites Astronomer's experience fine-tuning an SLM for summarizing data pipeline failures, only to later replace it with a superior, more cost-effective frontier model.

The Future of AI Value

The AI industry is recognizing the need for precision over brute force. The optimal approach isn't simply avoiding overkill, but selecting the right tool for the job. The ideal AI model is rationalized, rightsized, and robust—not excessively large, too small, or inappropriately sized.

The above is the detailed content of Model Citizens, Why AI Value Is The Next Business Yardstick. For more information, please follow other related articles on the PHP Chinese website!