One of the reasons I prefer to write code for computer programs rather than interacting with people is because computer programs have deterministic outputs. A person can say "yes" to something you need them to do but then do something completely different because they believe that is the right thing to do. Or they may fabricate some facts to convince themselves that is the right thing to do. You need to pick up on subtle signals to determine if they are going to do what you expect or if they are going to do something different. With computer programs, the output is always binary. Either it works or it fails. When it works, I understand why it worked, and when it fails, I can figure out why and fix it. However, this logic does not apply when it comes to pretraining and fine-tuning large language models (LLMs).
We saw this week with Google's Gemini. LLMs are first trained using the available public and private datasets from the internet. The first step is to use raw, unfiltered datasets that go into these models. More data is better than good data. If you equate pretraining LLMs to manufacturing a large piece of equipment, the raw materials often go through multiple layers of preprocessing before the material gets refined. However, imagine if the manufacturer decided it was not worth refining the raw materials, so they fed all the crap and junk into the manufacturing process. While the pretrained output from an LLM is a reflection of the data fed into it during pretraining, if you don't like the output, you can't really blame the model.
For example, in the case of Gemini, Google most likely did not anticipate or intend the historical-figures-who-should-reasonably-be-white result. So they injected their own version to mellow it down. Now the model took that input literally and started to apply it on its own.
The important thing is how one of the largest and most capable AI organizations in the world tried to instruct its LLM to do something, and got a totally bonkers result they couldn't anticipate. This event is significant because it is a major demonstration of someone giving a LLM a set of instructions, and the results being totally not at all what they predicted.
**As Emad Mostaque from Stability AI puts it, "Not your models, not your mind."**
While, Google and OpenAI are advocating politicians to make some regulations to stop the "bad" actors, they very well know it is to protect themselves than anything else.
I would clearly stay away from using Google Gemini or any closed source large language models from big companies and start to look at open source models, set up your own infrastructure for fine-tuning and inference. It is better to put efforts around building our own predictable open source Large languages model (LLMs) as a community and contribute to building our society rather than trusting these idiots.
What are your thoughts? Happy Friday and welcome to another crazy week in AI...!
