Skip to main content

AI

This is where I write about AI and Data Science. As a data scientist/machine learning engineer, I read many books on machine learning and AI and try to develop my skills and knowledge on the topic. I tried to dangle a bit into AI Strategy and discuss what I have learned and what I thought was a good way of approaching AI as a strategic asset and opportunity.

Understanding what AI is

I think a big problem in the discourse about AI is anthropomorphism. We tend to assosiate AI as sometihng of a co-worker, an entity that will remove jobs or talk to us like a partner. It is a common pitfall. I catch my self saying thank you to the LLM after getting a good answer. It is a tool, not a person or a pet. It takes an input and gives an output with some degree of entropy.

We need to shift the discourse from the AI takes over the world to AI is a tool that can help us tremendously, but also be used to nefarious purposes. Any talk of artificial intelligence as something of a world altering threat is grounded in a baseless urge to make something into us. We should be careful to direct our policies and efforts towards combating a dragon that in reality is a windmill. If you hack at the windmill long and hard enough, it might fall down and not mill any grain no more. And as Don Quixote, the regulators will ride glorisly into the sunset as the farmers realize a critical tool that made their life so much easier has been destroyed by a self-grandazing hero.

Data Science

I think the notes on the Data Science Project Start-Up Phase are cool; I have worked a lot on the data science aspect of projects and have learned some hard-learned lessons, and I think many of those will be interesting. I thought the Confident Learning topic was interesting, and I recommend going there to get more knowledge about that topic.

NLP or LLM

I wrote a lot about LLMs in 2023, some of which may be outdated, but it is still a good starting point for LLMs.

Links

Thoughts

  • Even though machine learning LLMs such as GPT-3, GPT-4, and LLama are non-deterministic. There are a few parameters that determine the output; there are few outputs that determine the inconsistency.
    • Seed - a random number that the model uses to start calculations.
    • System Fingerprint - describes the state of the engine.
    • temperature / top_p - how the model samples based on log probability