Why companies should consider controlling their own AI

We don’t have to tell you that AI has been thé hot topic for about a year since the release of ChatGPT. This technology promises to improve efficiency and productivity across the board for many industries.

Big Tech is undoubtedly investing heavily in this technology in the form of SaaS products. The list is extensive and keeps growing monthly:

• Azure OpenAI with GPT-3.5-Turbo and GPT-4
• Anthropic with Claude (3)
• Mistral with Mistral-Medium and Mistral-Large
• Google with Gemini
• Cohere with Command

All these SaaS products have one thing in common: they are deceptively cheap to use. This should raise a red flag. Why is this?

Investors prioritize growth and revenue over profit. This is the classic tech disruption model

1. Use venture capitalist money to grow at all costs
2. Get control over the market by offering the lowest prices to get a monopoly of the market
3. Raise prices.

It’s a strategy that could leave your business vulnerable in the long run.

Github Copilot, which is part of Microsoft and uses Azure OpenAI for code predictions. Which is now running at an average loss of $20 per month per user.

In the past, we’ve witnessed this phenomenon with Uber. In 2015, I could hail an Uber in Brussels for just 2(!) euros, effectively benefiting from investor subsidies. Once Uber established a stronghold in the market, it evolved into a duopoly with Lyft, leading to a universal increase in prices.

Currently, we find ourselves in the initial phase of the ‘tech disruption’ cycle with generative AI, where investor subsidies partly cover every transaction to capture the market. This cycle is inevitable, making it essential for businesses to adapt and brace themselves for the accompanying changes. The question is not if but when these changes will impact your business.

Once the winner is clear and companies or even governments have deeply integrated these services into their workflow, transitioning becomes nearly impossible, especially if these services become too costly. This is because integrations have been tailored to the model’s highly specific generative predictions. Altering the API or the model might necessitate rewriting most integrations.

The choices you make now will significantly influence how vulnerable your business is to future changes concerning AI.

So what can you do? You can’t stay behind; everybody is experimenting with AI!

Luckily, we don’t have to fall for this tactic; there is an alternative. Many AI models are released for free and can be used for commercial use. A prime example is Mixtral-8x7b-Instruct, trained by a French company Mistral and released under the Apache 2.0 license.

This model is slightly better in quality than Azure OpenAI GPT 3.5-Turbo, and it’s completely free! In combination with Hugginface Text Generation Inference or Nvidia Triton Inference Server it becomes very easy to implement your own Generative AI API. The only cost you have is the on-premise hardware or cloud servers you have to rent.

Instead of outsourcing your AI, you now possess the capability to manage the AI models you deploy, offering several advantages:

• By retaining control over your data, you avoid transmitting it to a third party, which could be vital for complying with regulations such as GDPR.
• You maintain authority over the AI system, enabling you to dictate when updates should or should not occur. In contrast, outsourcing your AI to a third party grants them the power to decide when updates take place, potentially altering the predictions you depend on.
• You possess a clear understanding of the costs in the ensuing months and years.

The most significant drawbacks associated with this approach are the initial costs, particularly in the case of on-premise AI, and the technical expertise required to deploy and maintain the system independently.

Controlling your own AI does not look as complex as this generated image :)

What is the best course of action?

If you’re thinking about experimenting with generative AI, our recommendation is to utilize these SaaS products for the initial validation of your business case. This approach is not only cost-effective but also relatively simple to set up. We should maximise the subsidies from the investors for now :-)

However, once you’ve confirmed the business value and are prepared to transition from PoC to production integration, it’s crucial to consider the following questions:

• Do you want to outsource your AI requirements?
• Are you comfortable sending your data to a third party?
• Are you willing to incorporate AI into your business without a clear understanding of the potential costs in the future?
• What are the risks if the third party modifies the AI model? Will this significantly impact your business?

While all questions may not apply to your specific use case, those that do should prompt serious consideration about taking control of your AI. This could involve adopting open-source AI models and potentially operating them on-premise.

At ReBatch, we bring our expertise to help you establish your first PoC, guide you in addressing these questions, and transform the learnings from the PoC into a full-fledged production system. If an on-premise system is your preference, we will not only assist and advise on selecting the most suitable AI model but also implement it on the most efficient hardware to effectively resolve your business case.

Find me on LinkedIn if you have more questions!