Press "Enter" to skip to content
Credit: benzoix on Freepik

Scaling LLMs Without More Data: Why Reasoning is the New Inference Paradigm

For decades, the prevailing wisdom in artificial intelligence has been that more data is the key to better performance.

Consider a simple example: if you were trying to build a model to predict the probability of a coin toss, observing 100 tosses might give you some insight. Observing 1,000 or 100,000 tosses would bring you closer to the true 50/50 probability. Similarly, large language models (LLMs) — like OpenAI’s GPT family, Anthropic’s Claude and Meta’s LLaMA — have historically relied on ever-growing datasets to improve their capabilities.

But what happens when you’ve exhausted all the human written data available? There’s speculation that this may already be the case. The AI field has begun experimenting with synthetic data — training models on outputs from other models. While this approach shows short-term promise, research suggests it might not lead to sustained improvements in capability over time.

This challenge has prompted a paradigm shift. The latest innovation isn’t about training on more data, but rather, making better use of existing models during inference — the process where a model generates an output in response to a user query. OpenAI’s introduction of the o1 ‘reasoning’ models exemplifies this breakthrough.

The innovation behind reasoning model

Instead of relying solely on pre-trained knowledge, these reasoning models take additional computational steps during inference to ‘think’ through problems. This includes generating intermediate reasoning steps before producing a final answer.

If you’ve ever asked ChatGPT a tricky question and followed up with, “Are you sure?” you may have noticed it corrects itself. This self-reflection capability can be harnessed systematically. By prompting the model to double-check and refine its answers before delivering them, performance improves significantly without the need for additional training data.

The implications are profound. Traditionally, scaling LLMs involved acquiring vast amounts of data and using significant computational resources to train larger models. Reasoning models, on the other hand, achieve better performance by using more compute during inference — a much cheaper and more accessible alternative.

Opportunities for businesses

This shift presents a game changing opportunity for businesses. Reasoning models like OpenAI’s o1 can unlock new frontiers in AI adoption and automation:

1. Immediate capability boosts: These models provide out-of-the-box improvements in performance, reducing the need for costly retraining or data acquisition.

2. New use cases: With enhanced reasoning and reliability, reasoning models can tackle complex problems previously thought unfeasible, such as automating intricate manual processes or creating AI agents that handle diverse and arbitrary inputs.

3. Cost efficiency: While these models require more compute during inference, this is vastly less expensive than training larger models. Over time, firms can expect the costs of these capabilities to drop even further.

4. Customization potential: Combining these models with proprietary data in a privacy-aware manner allows businesses to develop tailored solutions that stand apart from competitors.

Current challenges and future potential

The reasoning model paradigm is not without its limitations. These models are currently slower — sometimes taking up to 60 seconds per query — and more expensive for token-intensive processes. Additionally, they don’t yet support multimodal inputs like images or audio, which limits their versatility.

However, these challenges are temporary. As research progresses, we can expect faster inference times and expanded multimodal capabilities. The investment in reasoning models today will likely yield significant dividends as the technology matures.

The time to act is now

The rise of reasoning models marks a pivotal moment for AI innovation. Businesses that embrace this paradigm will be positioned to lead in their industries, unlocking capabilities that go far beyond chatbots or static automation tools. By combining the enhanced reasoning power of these models with proprietary data, firms can develop transformative solutions tailored to their unique challenges.

For those concerned about the current costs and speed, it’s important to recognize that the trajectory of technology adoption is clear: Prices will come down and performance will improve. Waiting on the sidelines could mean missing the opportunity to lead in this next wave of AI advancement.

As the paradigm shifts from ‘bigger data’ to ‘smarter inference,’ businesses have an unprecedented opportunity to redefine what’s possible with AI.

Author

  • Joseph Lo, image

    Joseph Lo is the head of enterprise platform at Broadridge, which provides critical infrastructure powering corporate governance, capital markets and wealth and investment management in the financial services industry.

    View all posts

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *