Canada’s AI Blind Spot: Why the Application Layer Matters

Unlocking Canada’s AI Potential: Why the Application Layer is the Real Game Changer

Scott Stevenson, Co-founder and CEO of Spellbook, highlighted Canada’s AI blind spot, emphasizing an overemphasis on infrastructure and research over the application layer. He argued that government, investors, and founders are excessively preoccupied with GPUs, data centers, and deep R&D, neglecting practical applications atop existing large language models like OpenAI, Anthropic, and Cohere.

Mr. Stevenson introduced the “scrub” mindset from David Sirlin’s “Playing to Win.” A “scrub” is a player handicapped by self-imposed rules, playing not to win but to adhere to a “romantic” game version. They label effective tactics “cheap,” believing victory comes from doing things the “proper” or “hard” way.

Applying this analogy to AI, Mr. Stevenson likened “scrubs” to those fixated on training models and complex evaluations. Successful application layer companies, conversely, simply build on existing models. This distinction is critical for Canada’s AI development, often prioritizing perceived difficulty over pragmatic commercialization.

Spellbook, launched in 2022, pioneered the first generative AI product for the legal sector. Serving over 4500 customers across 80 countries, it specializes in contract drafting and reviews, recognized globally as the most used AI tool. Despite its success and a Rocket Ship Award, Spellbook frequently faces rejection from Canadian AI funding programs.

Canada has allocated over $2 billion in AI subsidies, yet Spellbook often qualifies for none. Mr. Stevenson attributes this to a “scrub mindset” in funding, favoring “hard” or “defensible” projects like data centers, model training, and deep R&D. This overlooks pragmatic applications reaching a wide customer base.

Mr. Stevenson argued that training custom AI models is often inefficient. Bloomberg GPT, costing $10 million, was quickly surpassed by GPT-4. Harvey abandoned its custom legal model for general-purpose ones. Fine-tuning, while appealing to investors, encourages hallucination, isn’t adaptable to diverse preferences, lacks real-time updates, and poses privacy risks.

Spellbook’s success demonstrates an alternative: the company never purchased a single GPU, engaged with data centers, or employed an “AI researcher.” Its 200-person team comprises “AI builders” and “AI hackers” focused on developing legal domain features. Their approach prioritizes Retrieval Augmented Generation (RAG) to embed vertical knowledge and cater to client preferences.

Mr. Stevenson questioned if Canada’s culture and incentives are steering companies wrongly, prioritizing “hard” over “pragmatic.” He emphasized that billion-dollar AI companies emerge from the application layer, not from training models or owning infrastructure. He urged greater focus on building simple, pragmatic applications atop existing models, rather than romanticizing data centers, GPUs, and deep R&D.