AI is capable of increasingly sophisticated levels of content creation. Instead of automating boring tasks (I am still waiting for a robot that unloads my dishwasher and takes out the trash), we get AI that can create paintings, music and film, even write entire books - the things that humans enjoy doing. Many artists are concerned about losing their livelihood and about the loss of culture as we knew it. And rightly so.

However, I can't help but wonder if collaborating with AI could spark new forms of creativity and art.

In addition to writing articles and books, I've been dabbling in songwriting for a while. Unfortunately, I've yet to convince a musician to perform my songs, and I have to admit that I lack any musical ability myself. Last month, I decided to experiment with an AI platform (udio.com) to set some of my lyrics to music. I was pleasantly surprised at how easy it was to create first a blues version and then a more upbeat version of my song (although it was frustratingly difficult to get exactly what I had in mind).

While it feels a little sad not to collaborate with human musicians, as a songwriter, it's incredibly valuable to be able to hear my lyrics brought to life in different styles within minutes. Whether it’s jazz, punk rock, or a simple acoustic arrangement, this technology gives me the opportunity to experiment with my songs in ways that wouldn't be possible otherwise.

Yes, some art forms may be threatened by the surge of AI-generated content, but I believe many artists will find themselves empowered by these new technological tools. They could speed up the creative process or even spark entirely new ideas.

It is exciting that it becomes easier to finetune (or post-train) models with content specific to different languages and cultures. I had a brief chat with Yann LeCun last week on this topic in the margins of the Leaders in AI Summit.

It is well known that the large LLM foundation models have been trained mainly on either English or Chinese language content. There is cultural bias too: ask DALL-E for a picture of breakfast and it will likely show you bacon and eggs - not idli or tortillas or muesli or other breakfasts that are common in 90% of the world.

Fortunately, it is becoming easier for countries or communities to fine-tune or retrain open source foundation models (such as Meta's Llama3 or Alibaba's Qwen2) with local content. And it is important that they dedicate resources to do so. Good efforts have been made when it comes to languages such as Technology Innovation Institute's recently released Falcon Mamba 7B (Arabic), several models supporting French, Masakhane (Swahili, Yoruba) and KoGPT (Korean) are some examples. Multilingual models such as BLOOM and YAYI2 are good resources too.

We need models that represent local cultures. And it is not a one-time effort. As models keep evolving (and at some point may use very different architectures than current LLMs), it also takes a constant effort to produce new versions of these models for each cultural context. Countries need to establish capacities dedicated to do so. Therefore they need to be sure to put enabling conditions in place to train AI engineers.

Traditional wisdom suggests that asking questions is easy, but providing answers is challenging. Parents enduring the never-ending "why" phase with their children can attest to that. However, with the rise of chatbots fueled by LLM's and RAG models accessing internal data, answers are now abundant and immediate. The real challenge lies in posing the right questions.

Some professions have been trained to ask good questions: doctors, lawyers, and investigators like Colombo, Miss Marple and inspector Clouseau. But most of us are terrible at it.

I remember when we first started offering advanced analytics and ML services, we told business leaders "We can get you amazing insights into anything we have data on. What would you like to know?" And no one had an answer. The upward reporting streams in most organizations have been, and often continue to be, supply driven: people report quarterly sales figures, expenditures and market data because that’s what's available. Now that the supply is unlimited, we need to change the information flows to be demand-driven.

The key in the next decade will be to ask the right questions. We all need to become more proficient in that skill.