Some interesting examples of GPT-4's abilities.
In the first, it uses a Linux terminal to iteratively problem-solve locating and intruding into a poorly-secured machine on a local network.
In the second, it navigates through a text-based environment, after which it accurately draws a map of the rooms it walked through.
From Bubeck et al (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4: https://arxiv.org/abs/2303.12712
Sparks of Artificial General Intelligence: Early experiments with GPT-4
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition. The latest model developed by OpenAI, GPT-4, was trained using an unprecedented scale of compute and data. In this paper, we report on our investigation of an early version of GPT-4, when it was still in active development by OpenAI. We contend that (this early version of) GPT-4 is part of a new cohort of LLMs (along with ChatGPT and Google's PaLM for example) that exhibit more general intelligence than previous AI models. We discuss the rising capabilities and implications of these models. We demonstrate that, beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting. Moreover, in all of these tasks, GPT-4's performance is strikingly close to human-level performance, and often vastly surpasses prior models such as ChatGPT. Given the breadth and depth of GPT-4's capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system. In our exploration of GPT-4, we put special emphasis on discovering its limitations, and we discuss the challenges ahead for advancing towards deeper and more comprehensive versions of AGI, including the possible need for pursuing a new paradigm that moves beyond next-word prediction. We conclude with reflections on societal influences of the recent technological leap and future research directions.
We do not and cannot know every contingent fact of evolutionary history and every biochemical detail of every cellular process. But evolution gives us an in-principle explanation of how all of that came to be, makes copious well-tested predictions, and offers a framework for discovering specific details in specific cases. No other alternative framework exists which can claim to do so. The same seems to hold here. 4/4
it does credit assignment of reward signals and learns how to achieve goals. I’m not aware of another paradigm that can do all that, theoretically.
To be clear, I’m of course not claiming that we know (or could hope to know) every detail of how we do every cognitive task (and I’m also leaving the hard philosophical problems of mind such as qualia aside in this discussion). The situation is analogous to that with the theory of evolution: 3/n
It’s the fact that, so far as I’ve ever come across, the connectionist school of cognitive science which has birthed this paradigm is the only game in town theoretically, in the sense that it has a fairly complete, convincing, mechanistic story for how our own minds work: it explains how world models form, how systematicity and combinatorial reasoning arise from grounded, pattern-based thought, it explains the immense plasticity of the brain, it explains how 2/n
The main reason I expect roughly the current paradigm (multimodal large language models with tools and inner monologue, trained via self-supervised learning and reinforcement learning) to go all the way isn’t just the fact that it’s currently achieving tasks that, only a few years ago, basically everyone familiar with the field would have called “AI-complete” (i.e. if you can do them, you’ve basically solved AI). 1/n
The empty dishwasher is clirty.
There’s a lot I love about the rationality community: it’s full of intelligent, open-minded, kind people who really want to make the world better.
One thing I don’t love is some high-profile people routinely bullying others, often even without good arguments backing them up.
A few minutes ago in a parking lot, a car very slowly backed up towards the car I was sitting in, while my driver saw it coming and blared his horn.
Each second the probability of impact increased, until the impact occurred.
This is basic stuff.
Martingale ≠ random walk.
Here’s a fun game: identify all the problems with this graphic.
In which Bing asks appropriate clarifying questions to solve a complex puzzle, then works through to the (nearly) correct answer step by step.
(It ignored my admonition not to search the web, but I don't think this puzzle exists on the web.
The puzzle was invented by Alexander Cohen.)
