Introduction

I have wondered for some time whether the gold rush to develop large language models (LLMs) might be heading down a blind alley in terms of development. And last week, the release of Deepseek gave some indication I'm not entirely wrong.

Reliability Issues

Let me explain:

There are well-known issues with using them to provide reliable answers. They often hallucinate or make things up, and they're not always dependable for facts. Even when used as a translator between English and coding languages, the AI can fabricate information. I once asked a friend about an API they were using for a job, and the LLM returned plausible-looking results that were tailored specifically for the task—results that weren't actually available in the original API. The LLM had imagined them. The other options for the API were done correctly, but since they didn't include those specifics, the response was misleading.

Call Center Example

In another instance, there's a service (probably more than one, as far as I know) aiming to replace human call centers, which are staffed by presumably bored individuals, with climate-destroying LLMs [0]. The setup allows users to write scripts for these LLMs, similar to how you'd write scripts for a person.

But the service also used the LLM to implement the branching logic, though where AI isn't very strong. Traditional boolean logic is more of a mainstay for ordinary computers and ordinary computer programs, the CPU, and even the GPU [1] itself can evaluate logic quickly.

Alternative Approaches

I haven't had the opportunity to build this myself, but a general knowledge Q&A system might be better implemented by asking an LLM what search query would yield answers to a specific question. Then, using that query and having the LLM summarize the results could be effective.

This approach could also work for more complex inquiries.

Local Models

Now, here's something you may not have considered: if your PC has a reasonable modern graphics card (like the GeForce 1050 Ti I've been testing with, a card which dates back to 2016), you can download much smaller language models that run on your PC.

These are slower and less powerful but still sufficient for capturing language—though they might not capture a vast amount of actual facts. This approach could use less power and yield better answers.

Of course, the big players in AI might already be doing something like this, but they're not sharing details. We only know they're experimenting with increasingly larger models, and it's unclear if scaling leads to better outcomes.

DeepSeek's Approach

But here's a key point: DeepSeek has developed many of these smaller models and linked them together to achieve better results than the larger AI players expected, on a small model. Plus, you can download some of there models onto your own PC. In fact I did just that and used a deepseek related model on my own PC to proof-read this article, for grammar and spelling.

Conclusion

If we can run effective models on reasonable hardware locally we can have the use of good AI tools with less cost and not just climate or financial cost, but the risk to privacy and confidentiality is better managed when all the data the model ingests stays on your organisation's PCs.