My new job is AI sommelier and I detect the bouquet of progress

12.06, Wednesday 22 Mar 2023

I made an AI clock for my bookshelves! It composes a new poem every minute using ChatGPT and mysteriously has an enthusiastic vibe which I am totally into. Kinda. Maybe. Well, see below.

Here are photos on Twitter. (Check the thread for more pics.)

The e-ink screen shows:

Eleven-thirty eight, don’t hesitate /

Time to savor life, don’t be late.

And it totally blew up. (rn: 5,987 likes, 802 retweets; 1,165 reactions on LinkedIn.) So I need to do something with all that interest. It is super gratifying!

This post is not about that.

BUT having the clock does mean that I’ve been glimpsing AI-generated poetry pretty regularly for the past few days, plus the time I was making it, so now I have opinions.

Opinions about flavours of large language model, of all things.

11:51, time to be bold, /

Minutes tick by, stories untold.

So… the clock displays a rhyming couplet based on the time. The prompt also feeds it a concise description of the room from its pov, so it sometimes refers to books or the rug, or its own self as a small screen.

I say it’s generated with ChatGPT but technically what I mean is that it’s a completion using a model from OpenAI called gpt-3.5-turbo.

A model is a giant matrix of numbers that represents how likely it is that one word comes after a previous sequence of words. Models take months and $x00 million in capex to train. (GPT-5 is currently being trained on an estimated $225 million of NVIDIA GPUs – graphics cards.) There are now several large language models in the world though OpenAI’s models are the most available for use.

And they vary! In character and behaviour!

So whereas gpt-3.5-turbo is the model behind ChatGPT, I could have used OpenAI’s previous major model, text-davinci-003, commonly called GPT-3 (I don’t have access to GPT-4 yet).

I spent a morning looking at poetry composed by both models.

If I were an AI sommelier I would say that gpt-3.5-turbo is smooth and agreeable with a long finish, though perhaps lacking depth. text-davinci-003 is spicy and tight, sophisticated even.

(Perhaps I AM an AI sommelier. I just made that up. Perhaps I can put that on my LinkedIn.)

So gpt-3.5-turbo kinda leans towards the vapid with its prose. It’s more readable, but there’s less variety, and - like an excitable puppy - it frequently runs on. It’s hard to get it to stick to 2 lines for the poem. It will fib about the time if that means it can get a rhyme.

text-davinci-003 is more likely to pick ten-dollar words. It hits those 2 lines reliably and in sometimes surprising ways. With my AI literary critic hat on (another new career) I way prefer it.

BUT: gpt-3.5-turbo is 10% of the price.

That AI clock costs me $1.80 per day to run. It’s composing a new poem every minute so that’s 3,600 completions a day.

I prefer davinci’s words… but do I like them enough to pay $18/day? Punchy. No.

Though if anyone were to commission me for lobby art, davinci is what I’d reach for. (You know how to reach me…)

The clock strikes one-thirty-eight, /

Afternoon sun shines bright with fate.

Another difference between models is instruction tuning.

Once you’ve hoovered up all the text on the internet and trained your model, you can do something called “fine-tuning” which is to bias it to respond in a certain kind of way. You can feed it, say, a corpus of scientific papers, and then the model will respond in a way that sounds more like that.

The model behind ChatGPT, gpt-3.5-turbo, has been fine-tuned based on actual user interactions, as ranked by OpenAI:

To make our models safer, more helpful, and more aligned, we use an existing technique called reinforcement learning from human feedback (RLHF). On prompts submitted by our customers to the API, our labelers provide demonstrations of the desired model behavior, and rank several outputs from our models. We then use this data to fine-tune GPT-3.

In particular, the actual user interactions consist of questions, chat, requests, and so on, not just a partial sentence to be completed.

Which explains ChatGPT’s agreeableness and, well, chattiness.

The best way to notice instruction tuning is to play with a model that hasn’t been instruction tuned.

Facebook’s large language model, LLaMA, is available to researchers and has also recently leaked. Simon Willison details how to run LLaMA on your own laptop if you’re so inclined. And if you can get your hands on the model (ahem cough).

It is amazing to run this on your own machine!

But - and this is the best way I can communicate this - it’s like talking to someone who is asleep, or hypnotised.

This AI sommelier says that LLaMA mumbles. It’s like talking to an ancient wizard who repeats the last half of your sentence and then rambles on a bit and then starts talking in circles.

Whereas ChatGPT – we’re interacting! It’s awake!

Now I know (or at least assume) that neither of these models are sentient, but the difference is night and day.

Instruction tuning!

In cozy shelves, I do reside, /

It’s nearly noon, the clock confides.

AND THEN there’s ChatGPT’s quiet obsession with progress.

My prompt for the AI clock tells the model that it’s a rhyming clock, and about its embodied situation, and also gives it some pointers about how to respond. In particular I whispered this line in its ear:

Be imaginative and profound. Sometimes refer to your physical situation.

Now, I experimented with a variety of prompts.

I tried out adding the word “playful”. I tried “motivational,” and “poetic.”

All of those modified the vibe.

I also tried a prompt which asked the clock to sometimes refer to the future. I wanted solarpunk epigrams!

I included a little detail - about humanity and progress and the galaxy - just as I provide detail about the physical situation: the room and the rug, the bookshelves, the books and the Lego that the small screen is next to.

But I found that, when I included the idea of the “future,” this dominated every response from the model.

Like, every poem included a reference to the future; variety collapsed.

And the only way I can read that is that there is something in the instruction tuning of gpt-3.5-turbo which includes a belief in progress? A bias towards the future, rather than pastoral conservatism, which can manifest in an almost pushy fashion from time to time?

And so a mere mention of the future reinforces this bias and brings it to the surface.

I don’t know how explicit this is - maybe it’s encoded unknowingly in the bias of the rankings from the OpenAI team - but, with my AI sommelier hat back on, I sense that progressivism is in the bouquet, somehow.

Because even without the gravity of the “future” concept in the prompt, this motivational hustle still comes through from the AI clock – and I didn’t put it there.

Tick tock, don’t mock, it’s 9:29 on the dot. /

Time to rise and shine, leave the bed and get in line.

Look: an analogy.

Back in 2014, Facebook conducted a vast experiment in which it manipulated information posted on 689,000 users’ home pages and found it could make people feel more positive or negative through a process of ‘emotional contagion’.

In a study with academics from Cornell and the University of California, Facebook filtered users’ news feeds - the flow of comments, videos, pictures and web links posted by other people in their social network. One test reduced users’ exposure to their friends’ “positive emotional content”, resulting in fewer positive posts of their own. Another test reduced exposure to “negative emotional content” and the opposite happened.

The study concluded: “Emotions expressed by friends, via online social networks, influence our own moods, constituting, to our knowledge, the first experimental evidence for massive-scale emotional contagion via social networks.

This was unethical.

But at least it was being studied! It was performed knowingly!

OpenAI, to its credit, does deep work into the unintended capabilities and societal impact of GPT. There’s the GPT-4 “Safety Card,” as previously discussed, including a “red team” (a group which tries to do the bad thing, to see what happens, so we can anticipate and prepare) which investigated the possibility for automated propaganda via large language models:

Based on GPT-4’s performance at related language tasks, we expect it to be better than GPT-3 at these sorts of tasks, which increases the risk that bad actors could use GPT-4 to create misleading content and that society’s future epistemic views could be partially shaped by persuasive LLMs.

I would suggest, for the next Safety Card, the red team investigates what happens when consulting ChatGPT is widespread by students, in business, and by politicians – and when there is a gentle systemic bias of this kind towards (hand waves) “the future.”

(And, if so, could there be a instruction-tuning hack on the societal psyche of another nation… a la state-sponsored fashion hacks…)

Maybe it’s fine! Maybe we’ll all get in our rockets and inhabit the galaxy without really understanding why! Maybe we don’t need to sit and contemplate!

Maybe I’m imagining things and none of it means anything at all!

Most likely scenario tbh. But still.

More posts tagged:
Follow-up posts:

If you enjoyed this post, please consider sharing it by email or on social media. Here’s the link. Thanks, —Matt.