Green Daisy
back to blog
the pre-training plateau and why thinking time matters
llm
architecture
agents
founders

the pre-training plateau and why thinking time matters

Brian Craighead

brian craighead

ai architect & cto, green daisy

For twenty-four months, Silicon Valley ran the same uninspired playbook: shovel billions of dollars of compute into GPUs, scrape the open web raw, and declare victory when MMLU benchmarks climbed three points.

That narrative just hit a brick wall.

Last week, the consensus fractured publicly. Key architects—from Ilya Sutskever at Safe Superintelligence to top brass at OpenAI and Google—conceded the quiet part out loud: pre-training scaling laws are encountering brutal diminishing returns. The high-calorie public data has been consumed, synthetic data introduces toxic risks of model collapse, and the capital expense curve has gone vertical.

Does this mean artificial intelligence is dead in the water? No. It means the economic theatre has shifted.

Thinking Before Speaking

The winners of the next cycle will not be the firms that force-feed models every scrap of Reddit before deployment. The spoils go to test-time compute: granting architectures the tokens and runway to reason after the query arrives.

OpenAI’s o1 gave us the opening salvo. When a system can execute internal chain-of-thought simulations, backtrack on failed logic, and self-correct before outputting an answer, brute parameter count ceases to be the governing metric.

In business terms: the margin profile of AI just improved. You no longer need a debt-financed, trillion-parameter leviathan to execute basic enterprise logic. Smaller, cheaper models paired with targeted inference are outperforming bloated base models on cost, latency, and reliability.

The Architect's Advantage

At Green Daisy, we spend our capital and hours turning raw technical capability into viable commercial enterprises. For eighteen months, founders woke up sweating that a single checkpoint drop from Mountain View or San Francisco would vaporise their cap table.

That anxiety was misplaced. If base-model scaling has plateaued, the moat ceases to be who commands the largest Nvidia cluster. The defensible ground shifts upstream to system design: proprietary agentic scaffolding, resilient workflows, and ruthless execution within vertical niches.

The tech press promised that raw scale would deliver artificial general intelligence on a silver platter. Biology and economics had other plans. Scale buys you raw horsepower; architecture buys you outcomes.

So what? The era of throwing capital at pre-training bloat is over. Are you still burning cash worshipping frontier scale, or are you building systems smart enough to pause, optimise, and think?

share:

want to talk about this?

book a free clarity session and let's discuss how AI can work for your business.

let's chat