Language Models Are Terrible at Math - And That's Fine

Choose the right tool: don’t force LLMs on math; use simpler analytic methods where numbers beat language models.

Sep 29, 2025 0m 25s

Alan Buxton

Expert Insights

Alan Buxton dismisses the approach of using language models to solve mathematical problems in AI development processes, likening it to an inefficient game of chance.

He counters the growing sentiment that AI models can be adept at everything, emphasizing that systems are far from being one-size-fits-all solutions.

In Alan's exploration of AI in action, he underscores:

The limitations of AI language models in tackling complex mathematical problems.
The importance of understanding and acknowledging the strengths and weaknesses of AI models.
The risk of wasted effort and time when improperly applying AI capabilities to unsuitable tasks.
The call for more informed use of AI technologies, grounding expectations in reality rather than hype.

The engineer who's working on that, he came up with a lovely phrase that I'll never forget. He said, try to use language models to do maths. It's like, it's like trying to shoot birds with a T shake. sometimes it might work, but, but it's probably a bit of a waste of time.

— Alan Buxton, CTO, Simphony

THE NEW DEFAULT angle

Here's how to ensure you apply AI thoughtfully in software development:

Recognize Limitations: Acknowledge that AI isn't adept at everything. Discern its strengths and apply them to suitable tasks to avoid wasted effort.
Optimize Application of Models: Identify the right problems and tasks that specific AI models can address proficiently. Don't use a language model to solve mathematical problems.
Implement Continuous Learning: As an ongoing process, keep refining your understanding of what AI models can and cannot do effectively.
Challenge AI Hype: Counteract unrealistic expectations of AI. Make informed decisions based on a realistic understanding of AI capabilities.