Making Sense of LLMs - Questions, answers, and alpacas.

This is the seventh installment in a series about LLMs. You can find the previous article here:?Making Sense of LLMs - GPT4

"Jack and Judy were lying on the floor dead. There was a puddle of water and broken glass on the floor. How did they die?"?This sentence is an example of a?yes or no riddle, which I remember fondly from childhood. My dad would tell us the end of an obscure and mystifying story, like the untimely demise of Jack and Judy, and we kids needed to uncover the plot. The twist was that we could only ask yes or no questions. It felt immensely gratifying to slowly peel away uncertainty and get to the bottom of the story.

I like to think that my dad wanted to encourage us to think creatively and develop a taste for cunning questions. However, he was more likely suing for peace between cantankerous siblings on particularly long drives in the car.

This week a team from Stanford released a new model called,?Alpaca. So far, so uneventful for a week that sported big reveals every single day of the week. What is extraordinary about Alpaca is not its release but how it was created. The researchers came up with a clever, new trick on how to cheaply train their LLM.

I will skip the technical details and focus on the idea. The researchers started with the smallest version of Meta's newest open-source model,?LLaMA?(one must appreciate the camelid pun). They then paired it up with a more sophisticated and powerful model from?OpenAI?and instructed the better model to train its weaker sibling. The technique is called?self-instruct?and essentially plays out like a complicated version of the yes or no riddle game. The two models engage in a back-and-forth of thousands and thousands of questions and answers until the weaker model has learned all it can from its counterpart. Instead of training models by themselves, the researchers found a way to make LLMs learn from each other.??

To add insult to injury, the Stanford team managed to fine-tune Alpaca with a budget of around $600, compared to the?eye-watering, multi-million dollar bills, companies usually have to put up with to?train a foundational model.

Another implication of Alpaca's success is much graver and should give everyone in the industry pause. Proprietary data seemed to be a source of real competitive advantage. Without high-quality data, no high-quality model. Or, as the experts say, "garbage in, garbage out."???

But by using automated, self-instruct learning techniques, the data can stay perfectly safe, yet all the knowledge can still escape. Worse, once knowledge is out in the world as an LLM, you can't put it back into a bottle. Companies like OpenAI will undoubtedly put up a legal wall around this scenario, making it against the terms of use to train other models with their API. But first of all, it is impossible to prove what queries are used for, and second, one of the most powerful, state-of-the-art models recently?leaked to the public. It won't be the last.

There is a joke among philosophers that students enroll in their classes because they crave truths and yearn for answers. But studying philosophy to get answers is like becoming a catholic priest to date women.?

We all want answers, but questions can be powerful too. People tend to think carefully about what they might reveal with any given answer. Less so with questions. Most people only ask what they actually want to know. This way, just like with the yes or no riddles, questions can be more revealing and truthful than answers. Recent developments in LLM technology remind us of this overlooked fact.?

Last but not least. If you are curious about Jack and Judy's mystery. They were goldfish and died because their bowl broke after falling down during an earthquake.

You can find the next article here:?Making Sense of LLMs - Artificial emotions and real product management.