Reasonable AI — the Golden Age of AI Programming

[short url: https://chief.sc/reasonable-ai]

When LLMs just appeared, I was explaining them with a joke. It went like this:

LLMs are like goat sacrifices. You’d sacrifice a black goat and ask for the rain. If the rain comes, this was a great prompt. If not, you try a white goat, a black sheep, and so on until you get the result you want.

Fortunately the causal relationship between prompts and ChatGPT responses seems stronger than that between the rain and the sacrifices.

I first asked Anthropic folks how they are going about their safety agenda (the Constitutional AI) last year at SciPy. It was fascinating to hear that it consists of a series of dialogues with the models, where they are asked about their confidence in the results, their safety, etc.

On April 21–22, I’ve attended the first Full Stack Deep Learning workshop on LLMs, the LLM Bootcamp. Over two days we went through the basics of LLMs, the workflows around them, and the best practices of queuing them.

The emerging techniques are all around the way you construct the prompts and also chain them. Effectively, we’re plotting dialogues.

I call it the Age of Socratic AI, or Reasonable AI. We are engaging in conversations with AI that elicit meaning. We make the most basic assumption that it has the information we need and can provide it in the form we need, e.g. as an explanation or a how-to plan of action. We consider it an imperfect oracle that has to be assuaged, and asked questions in very specific ways to get the reply we need.

The history of myth and sci-fi is full of tropes of such an oracle. E.g. there’s an oracle at the end of the universe that can give the answer to the meaning of life. But if it says 42, you need some follow up questions to interpret that. You might be also required to use a specific tone of voice, and open with certain incantation to get the attention of the oracle.

So long as the oracle can talk, аnd maintain a reasonable discussion, it becomes a human-like partner that we can bring our whole human history and culture to bear on. Especially now that it encompasses all of our culture!

A lot of heuristics reinvent Socratic method, Aristotelian logic, and scholastics. With ChatGPT, we progress a thousand years — from the primordial goat sacrifices to mediaeval universities with trivium, where we are learning how to reason with logic, syllogisms, and are conducting debates in a rigorous manner. Eventually we’ll graduate to a Hegelian dialectics, covering to the truth on an ever expanding spiral of knowledge evolution.

What does it mean for programming? We are entering a new Golden Age. As my friend Anna Nachesa, a googler, has tweeted, saying that you don’t need programmers in the age of AI is like saying you don’t need magicians in the age of magic. The beauty of ChatGPT API is it’s all strings and dialogues. It’s an API! If you are a functional programmer, your Scala or Haskell skills were not easily applicable to TensorFlow or SGD. Training on A100s is harnessed by a whole bunch of frameworks tailored to Python consumers.

But get what? The LLMs are trained! And now we query them. We get back strings and we use embeddings, with vector stores. We build distributed system with asynchronous glue. If you are a Scala programmer, this is your time! Use Akka, Futures, and even the new Direct Style to conduct those dialogues, shove those strings into data stores, pump them through Spark, run ETL, send them to Kafka, iterate, reiterate, backtest. Maintain dialogue stores, evaluate and monitor model quality, build and deploy your own validation and filtering frameworks. Rejoice in being able to elegantly compose all of these systems using the powerful tools you know and love!

On the scientific front, we need a new kind of a university, a Socratic AI University, where we’ll learn to talk to each other and to the FMs in a rigorous natural language of scientists. Here’s to all of us who discuss professional topics with precision, who write and read papers conveying the meaning of science. All of it will come to life, and the abilities to reason will bear fruit. Long Live Meaning (LLM), and welcome to the age of Reasonable AI.