How much does AI cost

I asked DALL-E to generate a self-portrait of an AI for this post, and the result is the image that heads the article... a bit strange, to be honest, a mix of a thinker and a mannequin, but that's what it generated without much "prompt engineering", as they call it now. It's the only thing I asked an AI for this article, by the way.

It often happens to us, when talking to potential clients, that when they launch into a custom development project, a new platform, a solution to a problem of their clients, with technology, they inevitably bring up in the conversation the latest trends that exist at the moment and they want to integrate them into this new development they are proposing. As the great César Astudillo says in this conversation with Ion Cuervas-Mons for the Intercambio Iónico Podcast (the podcast also looks very good), innovating is not "adding Californian sauces." To understand the reference, it is almost better to listen to him:

Chat with César Astudillo for the Ionic Exchange Podcast

But it is normal for this to happen, in a certain way the media itself and each one of us, when a new technology emerges. disruptive, we are the loudspeaker of it, you only have to see what has happened since the end of last year with ChatGPT, it would be very rare to find someone who does not know what it is, and it is relatively recent. It is normal that companies, seeing this, have the feeling of needing ride the wave ; To stay current, you don't want to be the company that's still making carts when the company next door is already making cars.

Mind you, it is undeniable that Artificial Intelligence will mark a before and after at a technological level and, more than likely, for humanity in general, and Blockchain is also a technology that has facilitated solutions that were previously impossible for certain problems, but you have to understand very well the context and the details of what is being talked about, before considering using one of these technologies to solve a specific problem.

Excel is a marvel, eh, partly proof of this is how widespread it is and the number of different profiles that use it on a daily basis. Excel is a perfect tool as a spreadsheet, to filter and process specific data, but it is used for countless things for which it is not needed, version 4 revisions, DEF (final) version are still sent by email, they are copied to a Drive, they are used for task management, dashboards of information that should be shared between teams, silos are created, in short, islands of information, each one managing its Excel.

What I mean by this is that before we consider introducing the latest technologies, it may be easier and more productive to introduce the penultimate ones.

It is usually more a tangle of uncertainties, a lack of clear processes, or a review of the same, a lack of focus, design and analysis, rather than a technological problem. Then, obviously, having done all the above, we must rely on technology, which is what it is for, and that it supports the project that we have proposed, but the order should be that. The first line on our page, apart from Technology for Humans , he says You tell us your idea and we develop the technological architecture to support it, And that is precisely why we help you conceptualize that idea, carry out all the research, analysis and prior design, and then propose the technological architecture to support it.

Come on, 612 words later (there is a counter in the editor I use, it's not like I'm an ultra-geek when it comes to counting words), let's look at options if someone wants to include AI in their project no matter what, and what economic ranges we can be talking about.

This is actually going to be the most common and I would almost say the most advisable in most of the cases we talk to. Using some company's Artificial Intelligence system, AI as a service, for example, to integrate some intelligent chat service (chatbot), there are multiple alternatives here and traditional user support systems, such as Intercom, are already announcing their chatbots with Artificial Intelligence (END in the case of Intercom: https://www.intercom.com/end ). This, in general, will be relatively cheap, there are free chatbots for little use, but it grows exponentially with the number of interactions we have (for example https://snatchbot.me/pricing is $99 per month for 10,000 messages and then $0.0099 per additional message).

Within platforms such as Google Cloud, we have a multitude of options to integrate AI developed by them, specifically in our projects, such as DialogFlow for conversational systems, both text and voice, https://cloud.google.com/dialogflow, services to convert voice to text or text to voice (https://cloud.google.com/speech-to-text?hl=es and https://cloud.google.com/text-to-speech?hl=es), translation services via API, https://cloud.google.com/translate?hl=es, Product recommendations https://cloud.google.com/recommendations , for example, in Retail. Many of these systems have a free version at the beginning (translations of up to 500,000 characters per month are free, for example), but the cost then grows rapidly with use.

Along these lines, a very interesting option that is going to gain strength is to develop our own APIs for our information that we want to make available to certain Artificial Intelligence models such as ChatGPT, so that it is then accessible within the context of a natural language model. This is not yet open to everyone, ChatGPT has a waiting list for plugins: https://openai.com/blog/chatgpt-plugins and, in any case, integrating ChatGPT into our project has a cost that can be quite high ($0.03 for 750 input words and $0.06 for the same but output, as of today). Perhaps today it is not something viable for most people, but it will surely gain strength.

Integrating third-party stuff is all well and good, but sometimes we want something of our own, tailored, very specific, that doesn't fit with any of the existing solutions. Well, in this case the first thing to know is that AI is expensive, a lot. It is estimated that to train ChatGPT in its GPT-3 version, the people at OpenAI spent $5 million. Just looking at the numbers, 3.55 million GPU (graphics processor) hours to train it, and comparing it with the GPU hourly costs of $0.35 per GPU per hour in the cheapest case it offers, https://cloud.google.com/compute/gpus-pricing , we can understand that it is not cheap. There are other, better alternatives to train an AI model in Google that with GPUs directly in Compute Engine, they have their unified platform for training learning models, Vertex AI, https://cloud.google.com/vertex-ai?hl=es , but to get an idea of ​​the cost of training your own model, the above is useful.

Of course, most likely in our project we do not need a model like ChatGPT, our own, and a much smaller Machine Learning model will be enough. Estimating the costs of these models is not an easy task, there are studies https://epochai.org/blog/trends-in-the-dollar-training-cost-of-machine-learning-systems, complex calculation tools: https://github.com/aipaca-mlops/ML-training-cost-calculator, but all this focuses on the computational cost part, or number of machines and cost of these machines.

It must be understood that machine learning models are fundamentally based on having a huge amount of classified data (in the case of supervised learning), and just generating all that amount of information and classifying it has a high human cost. Generating a dataset of 100,000 pieces of information, classified, even using a service like https://scale.com/ to annotate it, and then validating it with a small in-house team, will cost us between $10,000 and $85,000, depending on the complexity of the annotations we need, and this is just the cost of generating the data. (More detail in this article: https://medium.com/cognifeed/the-cost-of-machine-learning-projects-7ca3aea03a5c ).

It might seem from this whole article that we don't like AI, but nothing could be further from the truth. AI will undoubtedly revolutionise future technological developments and the interaction we have with different projects, but sometimes marketing is tempted to put the smart label on a fan accessible via WiFi, or a robot vacuum cleaner that doesn't pass over the same point twice and is able to draw a map of the rooms and, well, we're not saying that they aren't important things, but AI is something else.

It's possible that the same thing will happen in your next project, idea, problem that you want to solve with technology. It's possible that the AI ​​you want to integrate could be a traditional decision tree, which in itself isn't bad, it's simply the best solution for the specific problem you have. Maybe not, maybe it really makes sense for your project to have a machine learning model or use third-party services in that line, but then you have to take this into account in the costs and have the appropriate financing.

That, in short, would be the conclusion we want to reach with the article, integrating real AI today is very expensive, especially if we want to make it custom-made, and in many of the problems we face on a daily basis it is either not necessary, or there are many previous aspects of processes, priorities or less expensive technologies that would give a much better return on investment.