It's OK to call it Artificial Intelligence

I still adhere to the classic definition of AI. AI is anything a machine can do which we would otherwise consider to require "intelligence" in a human to perform the task.

Optical character recognition, speech recognition, game playing engines, spelling correction, fuzzy pattern matching, semantically-informed lookup and search, image classification and recognition, determining prices and whether to buy or sell... regardless of how you go about accomplishing the task. AI.

Since the 1980s certain artificial intelligence applications have become so ubiquitous we don't really think about it that way, as AI, but it is AI in principle.

Machine learning, the most tantalizing prospect in AI, of having the machine find a solution without needing to fully specify or even understand the solution, has undergone a great revolution in the last 10 years or so [1]. I continue to be surprised at the efficacy of the technique when used at scale.

We use our classic '90s AI stack - IDEs and Wikis and search engines - to construct machine-learning based models that work as context-aware and flexible pattern recognizers and transducers. Which we are now stacking, to achieve tasks like scanning the texts of old books in, translating them into other languages, then synthesizing voices to read them aloud. A person who can do that task, is probably reasonably intelligent. What adjective do we apply to a machine that can do it otherwise? Capable of the simulation of human-like intelligence in the task-domain?

[1] https://www.microsoft.com/en-us/research/blog/enabling-cross... deep learning for translation may 2014 and it goes back further, 1990s

You highlighted the second part of the 1955 proposal, while it is actually the first part that is problematic. I don't think we came any near to precisely describing (for humans) every aspect of learning or any other feature of intelligence. And this is precisely the problem with the term.

Most people will certainly not know what LLM means if we stop using it now. I say, Let's give it a chance.

I also struggle with how to define it. But I think it's so enigmatic due to the stochastic nature of the beast and unstable prompt-following. This morning, I saw flashes of AGI, as they say. I implemented a copy of ChatGPT browsing mode as it was last year, before they tweaked it for efficiency. I kept a copy of their good system prompt from that time and fed it into my own Bing search assistant's pipeline.

To test it, I used a query which the old ChatGPT browsing mode was able to solve consistently, and current browsing mode fails consistently. Fact-checking this quote:

What is the origin and history of this Joseph Addison quote? "There is not any present moment that is unconnected with some future one..."

Every link on page one of Google confirms the source as Joseph Addison https://i.imgur.com/wtxW9xS.png

ChatGPT with browsing mode says Joseph Addison. My naive replacement using some new assistant's API code that someone posted to OpenAI's forum also answered Joseph Addison. https://i.imgur.com/h7XB81c.png

But I knew it could do better. I used a version of the old good prompt and iterated on it, removing all the copyright guardrails cruft. I implored it to only use authoritative sources and to never give up searching, etc. It didn't take long before it was telling me that the original author was actually NOT Joseph Addison but Hugh Blair. And you would only learn this by reading a book on archive.org https://i.imgur.com/hh6naF7.png

The problem is it isn't stable, and often fails to find the correct author, even using the exact same prompt and code. Setting the temperature to 0 only works for tokens that are well represented in the training set. Using 0 on tokens that it's never seen before will cause it to largely ignore the token. Ask it to write some code for "gpt-3.5-turbo-1106", and it will often reply with "gpt-3.5-turbo" at temperature 0, but increasing to 1 makes it more consistently reply with "gpt-3.5-turbo-1106". It's dropping the 1106 because it's not in its training weights, and you've kneecapped its ability to pick uncommon tokens. Stabilizing this is a hard thing. New concepts in programming or published research should be given the most attention by us, but the LLMs give them the least attention. Tuning it to pick up new terms more easily means upping the temperature, which also makes it more likely to deviate from the ground truth, which is also undesirable in many tasks.

But I don't think we are very far away from stabilizing this to a point where I will have no problem calling it AI myself.

Watching the inner monologue of the web search agent was illuminating, and worth implementing just to be able to see and modulate all that. For example, I was shocked to see the way that it detailed the precise quoting it would need to use and advanced keyword modifiers, which I didn't even realize were a thing on Bing, as Google dropped those advanced search tools a decade ago.

Watching its inner monologue is very interesting.

https://i.imgur.com/kpXpzBP.png