Two reputable news organizations — Reuters and The Information — recently reported sources claiming that recent drama around OpenAI’s leadership was based in part on a massive technological breakthrough at the company.
That breakthrough is something called Q* (pronounced cue-star), which is claimed to be able to do grade-school-level math, and integrate that mathematical reasoning to improve the choosing of responses.
Here’s everything you need to know about Q*, and why it’s nothing to freak out about.
The problem: AI can’t think
The LLM-based generative AI (genAI) revolution we’ve all been obsessing over this year is based on what is essentially a word- or number-prediction algorithm. It’s basically Gmail’s “Smart Compose” feature on steroids.
When you interact with a genAI chatbot, such as ChatGPT, it takes your input and responds based on prediction. It predicts the first word will be X, then the second word will by Y and the third word will be Z, all based on its training on massive amounts of data. But these chatbots don’t know what the words mean, or what the concepts are. It just predicts next words, within the confines of human-generated parameters.
That’s why artificial intelligence can be artificially stupid.
In May, a lawyer named Steven A. Schwartz used ChatGPT to write a legal brief for a case in Federal District Court. The brief cited cases that never existed. ChatGPT just made them up because LLMs don’t know or care about reality, only likely word order.
In September, the Microsoft-owned news site MSN published an LLM-written obituary for former NBA player Brandon Hunter. The headline read: “Brandon Hunter useless at 42.” The article claimed Hunter had “handed away at the age of 42” and that during his two-season career, he played “67 video games.”
GenAI can’t reason. It can know that it’s possible to replace “dead” with “useless,” “passed” with “handed” and “games” with “video games.” But it’s too dumb to know that these alternatives are nonsensical in a basketball player’s obit.
The Q* solution: AI that can think
Although no actual facts are publicly known about Q*, the emerging consensus in AI circles is that the technology is being developed by a team led by OpenAI’s chief scientist, Ilya Sutskever, and that it combines the AI techniques Q-learning and A* search (hence the name Q*).
(Q-learning is an AI-training tool that rewards the AI tool for making the correct “decision” in the process of formulating a response. A* is an algorithm for checking nodes in a graph and looking for pathways between nodes. Neither of these techniques is new or unique to OpenAI.)
The idea is that it could enhance ChatGPT by the application of something like reason or mathematical logic — i.e., “thinking” — to arrive at better results. And, the hype goes, a ChatGPT that can think approaches artificial general intelligence (AGI).
The AGI goal,…
2023-12-07 10:41:03
Original from www.computerworld.com rnrn