A boon for developers: AI models are getting smaller and cheaper

It’s been another busy week for AI practitioners and enthusiasts. OpenAI has released a mini version of their GPT-4o model. The company says the GPT-4o mini will cost 15 cents per million input tokens and 60 cents per million output tokens, making it over 60% cheaper than the GPT-3.5 Turbo.

Meanwhile, almost simultaneously, Meta released the latest version of their family of open-source AI models. The free-to-use model, called Llama 3.1, boasts benchmark performance on par with proprietary models from OpenAI, Google, and Anthropic.

What makes Llama 3.1 truly impressive is the algorithmic efficiencies that Meta has managed to squeeze out of the model; Llama 3.1, with 405 billion parameters, is comparable to OpenAI’s GPT 4, which is rumored to have over a trillion parameters. This means that Llama 3.1 is less computationally intensive to run, and is also open source.

Not to be outdone by the Americans, French Mistral also released an updated version of their model, called Mistral Large 2. It is also cost-effective to use, but not fully open source like Llama (developers must obtain a license to use Mistral’s model for commercial purposes).

According to OpenAI, they envision a future where models are seamlessly integrated into every app and website. But for this future to become a reality, models need to become smaller, more efficient, less power-hungry, and more affordable so they can be deployed at scale.

Mark Zuckerberg recently argued in a blog post that open-source AI models will overtake proprietary models. “I believe the release of Llama 3.1 will be a turning point in the industry, with most developers moving primarily to open source,” he said.


Easier to experiment
This push towards more affordable and open-source models bodes well for Indian developers, say AI experts. “The lower costs will allow developers to experiment and build real-world applications without the previous financial constraints. Since these models offer capabilities similar to frontier models, developers can integrate advanced AI features into their applications. This democratization of access to advanced AI models can accelerate the development of innovative solutions across industries,” says Tanusree De, executive director & responsible AI leader for technology consulting at EY Global Delivery Services.

There are also less obvious ways developers can benefit from these models.

Debdoot Mukherjee, chief data scientist and head of AI & demand engineering at Meesho, says one of the key things about a large AI model being open-sourced is that you can now generate a lot of training data at scale, which you can use to refine smaller models. “Previously, if you wanted to create hundreds of millions of training examples, it was practically impossible to use GPT 4 or something like that at that scale because it was very expensive. But now you have an open-source alternative.”

Rahul Lodhe, global head of engineering for SAP Business AI Copilot Joule, says that to benefit from models like GPT-4o mini and Llama 3.1, Indian developers should focus on gaining proficiency in programming languages ​​like Python and working with AI frameworks and tools. “A good understanding of software engineering principles, along with awareness and application of AI ethics and responsible AI considerations, is essential. Additionally, it is crucial to participate in open-source projects, engage with AI communities, and continuously update your knowledge with the latest AI research and tools. By doing so, developers can not only improve their skills but also contribute to the global AI ecosystem, making a significant impact both locally and globally.”