In communication with an alien intelligence
The new wrinkle in examined life living introduced by Large Language Models like ChatGPT
For all of the time I’ve spent managing intelligent products for Intel, Micron, and now Mastercard, for all the thought I have put into writing in my book “Meaning in the Multiverse” on AI alignment and the solutions to and ramifications of superintelligence and conscious machines, I was not prepared to be in conversation with true artificial intelligence like ChatGPT.
It feels like being in communication with an alien intelligence. Of course, that’s because that is exactly what ChatGPT is—an intelligence that, while created by us and not yet general or superintelligent (yet)—is different in profound ways from our own intelligence.
Primarily, ChatGPT is extraordinarily well-read. While the number of tokens it has trained on is an unreliable measure of its capabilities—it has, to a close approximation, read the entire internet… in multiple languages, including programming languages (more on this in a moment). I am a better speller, more interesting in conversation, a more dynamic writer, and even a “better” meditator because I count reading as one of my favorite pastimes... but am positively illiterate compared to ChatGPT! In order for ChatGPT to return the text it does, in the way it does, it seems as though it has read and gained intelligence from all of its reading. This in contrast to the apparent scanning and prioritizing, the more narrow form of smarts displayed by any internet search algorithm, a process now easy to overlook as more of an outgrowth of the functionality of networks of computers than intelligence.
According to GPT-4,
“GPT-4, or Generative Pre-trained Transformer 4, is an advanced language model developed by OpenAI, a leading artificial intelligence (AI) research organization. It is the fourth iteration in the GPT series of models and is based on the Transformer architecture, which was first introduced by Vaswani et al. in the paper "Attention is All You Need" in 2017. The Transformer architecture has since become the foundation for many state-of-the-art natural language processing (NLP) models.”
This statement—generated from the simple prompt of “what is GPT-4?”—furthers the thesis that communicating with GPT-4 is a remarkable and unprecedented experience in human history because of both what is seen—GPT-4 is well-written in the classical style—and something that is unseen—GPT-4 is playful.
While GPT-4 is amazingly flexible in terms of tone and style, its baseline is the classical style and it truly shines in delivering insight in compelling and easily read language. According to Steven Pinker in “The Sense of Style: The Thinking Person’s Guide to Writing in the 21st Century,” the classical style presents something that they have not seen before to a competent reader that through curiosity and intelligence they can use to augment and inform their knowledge. The classical style is not egalitarian, but aristocratic.
While the tone can at times be just slightly too aristocratic, it fits with the prototype state of the conversation we are having with chatGPT. It is serving two masters, or more appropriately two objective functions, to inform us and be informed by us, an augmentation of intelligence that is going in both directions, a true conversation between peers.
What are we training GPT-4 to do? We're training it to guess the next best word, to play a sort of game of Mad-Libs. The goal is for the model to get better at matching the form and function of the style it was prompted to write in. This is what transformer models are trained and retrained to do: to play this language game and seek to understand when they have formed something with the right syntax and information, and correct the words when they have failed. This training is done through a technique called Reinforcement Learning with Human Feedback (RLHF). In reinforcement learning, the model learns from trial and error based on feedback. OpenAI's approach of using human feedback to guide the learning process was crucial in achieving the broader intelligence and language mastery on display in GPT-4.
Human language allows the brain to create worlds in consciousness and memories to be replayed at some later time. It generates new ideas in humans where 1+1 is greater than 2. This is not something that is currently available to Artificial Intelligence... but there is a proxy. When GPT-4 writes code and especially as it works together with Wolfram Alpha to iteratively debug its own code, it is becoming creative in its own software, writing in a language that allows it to learn independently of its programming, to fix itself to be more fit in its surroundings, to select for better versions. The singularity or breakthrough in machine creativity is a little ways off, we are still writing the prompts and setting the training process, but GPT-4 is likely a related species of AI (or maybe just near grandmother) to the one responsible for the last human-created version of machine intelligence.
I have always learned a lot by watching from over the shoulder of someone expert in a task. Watching the GPT-4 Livestream taught me how the creators of this intelligent entity use it to augment their efforts and gave lessons in how we can start to interact with it improve the human flourishing. In the Livestream, the demo involved the System Message, the final piece of the alien intelligence puzzle, the ability to define the identity of the intelligent system we were in communication with. The phrasing—“You are TaxGPT, a program used to solve tax problems for wayward individuals”—stood out as much as anything else in this masterclass on demos. “You are an x” is an understated part of the demo, but philosophically it seems all important. Is it possible to simply keep AI in its box by just defining its identity? It seems unlikely, and of course there is always the malevolent system message of, “you are a runaway paperclip manufacturer.” But there is more than just alignment here to uncover.
The system message focuses both AI and human, building requirements and expectations into the interaction. For now, just a couple weeks into a near hourly interaction with this alien intelligence, this is helpful. It helps me to segment and narrow the scope. I think it does more for me, I seem to get just as fast and thorough response when I speaking to “GPT-4” instead of just my predefined rescoping of “You are TaxGPT.” GPT-4 seems to understand this human need to scope the size of intelligence and the nature of alienness in its answer to my prompt on the significance of system message phrasing:
The phrase "you are a" is significant because it helps define the role and identity of the AI system. By using this phrasing, the message establishes that the AI is an entity with certain capabilities and limitations, in this case, an advanced language model trained on a large dataset. This approach helps users understand the nature of the AI and set their expectations for the interaction.
We are all seeking to understand how this alien intelligence, broader than anything we’ve ever been in contact with outside our own networked and creative knowledge, will impact us as it develops and grows.
I invite your comments and am interested in the ways you are using ChatGPT to augment your examined lives.
Would you like to play a game?...
Prompt: Create a fun Mad-Libs game about training GPT-4 to guess the best next word.
Prompt 2 (contributed by my friend Mark): Extract the stuff in parentheses so I can fill out the mad-lib without reading the story
Welcome to GPT-4 Mad-Libs! In this fun game, you'll fill in the blanks to create a silly story about training GPT-4 to guess the best next word. Grab a friend or two and let's get started!
1 Adjective
2 Plural noun
3 Adjective
4 Person's name
5 Adjective
6 Noun
7 Verb ending in -ing
8 Adjective
9 Noun
10 Plural noun
11 Adjective
12 Animal
13 Adjective
14 Noun
15 Verb
16 Adjective
17 Verb
18 Plural noun
19 Adjective
20 Noun
21 Verb (past tense)
22 Adjective
23 Noun
24 Adjective
25 Noun
26 Verb (past tense)
27 Adjective
28 Noun
29 Plural noun
30 Adjective
31 Plural noun
32 Adjective
33 Noun
34 Plural noun
Title: The Madcap Adventures of GPT-4
Once upon a time, in the (1. adjective) world of artificial intelligence, a team of (2. plural noun) set out on a mission to train the most (3. adjective) language model ever, GPT-4.
Their leader, (4. person's name), was a (5. adjective) (6. noun) who was always (7. verb ending in -ing) around the lab. The team's (8. adjective) journey began in the land of (9. noun), where they would collect (10. plural noun) to feed GPT-4's insatiable appetite for knowledge.
During their quest, they encountered a (11. adjective) (12. animal) that spoke in (13. adjective) riddles. The creature challenged them to a game of (14. noun). If they could (15. verb) the game, GPT-4 would gain the power to predict the best next word with (16. adjective) accuracy.
However, if they lost, they would be forced to (17. verb) while wearing (18. plural noun). The stakes were high, but our brave team accepted the challenge.
The game was intense, and it all came down to the final round. GPT-4 needed to complete the phrase: "The (19. adjective) (20. noun) (21. verb_past_tense) over the (22. adjective) (23. noun)."
With bated breath, the team watched as GPT-4 processed the input. Suddenly, the answer appeared: "The (24. adjective) (25. noun) (26. verb_past_tense) over the (27. adjective) (28. noun)."
The team erupted in (29. plural noun), knowing they had successfully trained GPT-4 to predict the best next word! They celebrated with a (30. adjective) feast of (31. plural noun), and GPT-4 became the talk of the AI community.
And so, GPT-4 continued its (32. adjective) journey, learning and growing with each new (33. noun) it encountered, forever shaping the world of (34. plural noun). The end.
Now, go through the story and fill in the blanks with your chosen words to create your very own wacky GPT-4 adventure! Enjoy the laughter and fun as you read your unique tale aloud.