1 minute read

AI is the technology everybody talks about right now. It is a revolution in many sectors. Here are a few of my thoughts about this technology:

  • Is anyone getting replaced by AI?

Some people think AI will replace humans, while others claim it will not happen. The AIs of today give predictions powered by statistical models trained on large datasets. A part of the value an human brings is in capabilities AI haven’t been able to demonstrate so far, judgment, critical thinking, creativity for example. While there are cases of LLMs demonstrating some capabilities, their increased capabilities come with risks. I think AI must remain a tool.

  • What about society?

In the current internet era, people look for information on social networks without verifying sources. AI is getting better at content generation. It will eventually be very hard to tell the difference between fake and real content. Without traceability on the generated content, a malicious actor could use AI as a mean to spread fake news, for example. To me, this form of content is internet pollution. It has no real value for anyone, it only makes noise until it gets disproved.

  • More data is better

I heard this statement too often and couldn’t help but think this contradicts textbook theory. This article published on MIT technology review website confirms that academic theory. An example of an open-source model trained on fewer parameters is given. When compared with bigger, proprietary models, it features comparable performance. The main criteria when forming training datasets is quality, not quantity.

  • Your data is now my data

Because this myth is widely spread, AI-driven Businesses, rather than buying the data necessary to their use case, now find it interesting to use their customers data. Why? Because it’s much cheaper to modify a data policy than purchasing data representing millions of users. In some industries, I think this could be problematic. Once my data has trained a model, there’s no way to remove it from there, and the model will be in operation until there’s a better model available. My competition might even benefit from that model, which is something I do not wish to happen. I think opting in/out is necessary, and it’s not widely implemented yet.

References

  1. https://arxiv.org/pdf/2412.04984; Frontier Models are Capable of In-context Scheming; Alexander Meinke, Bronson Schoen,Jérémy Scheurer,Mikita Balesni, Rusheb Shah, Marius Hobbhahn