When DeepSeek AI hit the news cycle a couple weeks ago, it sent shockwaves through the American AI industry. Not only did DeepSeek develop a pretty good open source AI, they did it for a fraction of the budget that companies like OpenAI are working with. It seems to have opened up a new frontier for AI growth and development, and marks a shift in the way we think about how much AI should cost.
The discourse about how much cheaper AI could become inspired us to put together a few suggestions on how you could do AI for less money right now, today. There are tons of ways – here are a few to get started:
Not Every Task Needs an LLM
Can LLMs do a ton of stuff? Absolutely, they’re extremely powerful! But there are lots of functions that don’t require that much power, and in those cases, a simpler (and less expensive) model will likely be a better choice. After all, if you’re trying to chop carrots, you don’t need a katana – a paring knife will do just fine.
To pick an example dear to us – machine learning algorithms (like the one we use in Mix) are more than capable of handling things like Marketing Mix Modeling (MMM), calculating life-time value of advertising channels, and identifying novelties in data.
Could an LLM do those things too? Yes, usually. But our experience has taught us two things:
- Often, bespoke models that solve a single problem (like MMM, or a model that predicts churn) are much more accurate and flexible than a generalized model like an LLM.
- And customized non-LLM models are way more cost effective. You can control exactly when, how they run, what resources they’re using in the query, and save your tokens for when you really need them.
Train AI Agents to Be Model Agnostic
One of the developments since the press about DeepSeek is a bit of a crisis of confidence in the existing AI models. We might see new competitors come onto the scene, and what’s been seen as a bright shiny object might start to become slightly…commoditized?
Whatever happens in the AI marketplace, it’s impossible to predict exactly which new models and platforms are going to be available this time next year. The best way to be prepared is to make sure all your AI – your workflows, your infrastructure, your AI agents (or any other tools, platforms, or dashboards that work with LLMs) – are modular, so that you can swap in and out new models easily. You will want to be able to compare how ChatGPT and DeepSeek and LLaMa compare for your specific use cases. Baking modularity into your architecture ensures that as the AI landscape changes, your agents and tools won’t become obsolete.
Be Smart About Your Data Storage
Storing huge data sets can get expensive quickly. Hopefully, you’re already carefully choosing non-SQL and SQL stores for your different existing data types. As you build out your AI agents, you can do something similar with vector search. (Vector search is for searching documents, calendars, and other contextual information you might use to train AI.) Your vector data can be stored in SQL and/or NoSQL databases depending on how often the data is updated or queried, or how structured it is. There’s no one single right way to store your data, but there are tons of ways that are inefficient and expensive, so make sure you’re thinking about your querying use cases to stay cost-effective.
Let Your Needs Dictate Your AI
AI is going to get even more ubiquitous, and a lot of fun and shiny tools are going to hit the market. But instead of grabbing the most exciting thing out there and then figuring out what to do with it, start by determining your use cases. What does your AI actually need to be able to do? How do you plan to use it on a daily basis, and who is going to be the one pressing the buttons? What do they need to be successful? Answering those questions can help you figure out, for example, if you need a superpowered LLM, or another solution that’s less powerful but more specialized for your needs.
Have questions about creating an AI agent, or choosing an AI tool, or how to store your data, and what your use cases might be? Give us a shout!
