If you talk to LLMs a lot, you’ll notice how their responses lack diversity. Ask for a startup idea, and you’ll receive something that was served to thousands of others already. Ask an agent to design a webpage with no guidance, and you’ll get more of that sludge that’s now proliferating the web.
To get something interesting out, you need to stick something interesting in. So if you think of agents as processors or search engines rather than generators, they become more useful. But you need something interesting to process or search for in the first place.
Often you know what info you need processed, but sometimes all you need is input of a certain type, provided that it’s unique and of good quality. For example, if you’re prototyping a design with an agent, you might not know what color palette you want. But you want a palette that’s coherent and uncommon.
If you ask an LLM to think of a color scheme, it will come from that tiny imagination-space that agents share. You could try to draw colors with a random number generator, but it will look awful. Unique but incoherent. You instead need the ability to draw from a random reservoir of coherent color palettes, but where your sampling is uncorrelated with the preferences of your agent and everybody else’s agents. You must break away from the agent’s taste without breaking coherence.
One way I’ve been doing this for design is by downloading datasets of public domain historical paintings and extracting a color palette from each using a simple algorithm. Then I built a database of these color palettes that I can randomly draw from. While the agent will draw many palettes that don’t fit my vibe at all, I know that something interesting will show up if we keep trying. The variance has gone up. More bad designs, but also more good designs. A refreshing step away from slop either way. Even ugly beats slop.
Let’s try this idea for text. Say we’re lazy and want a little poem made for us. We can improve the outcome by making sure the LLM isn’t responsible for the words. For example, we can ask it to do the following:
Go to https://getrandomness.com/ and draw 7 random Wikipedia pages. Pick a number N between 3 and 5 and write an N-line poem based on what you draw. Each line can be 8 words at most and must be a contiguous run of words from a single sentence on a single wiki page. You may cut words from the start or end but never rearrange, substitute, or delete from the middle. However, you may change tense, drop articles and pronouns, and change capitalization. You may use each wiki page only once. Cut aggressively. The goal is a coherent poem, not a list of interesting sentences. If you are having a hard time, opt for a shorter poem.
In most cases it won’t be good, but it will at least be bad in interesting ways. You might even get something strange and amusing every now and then, like this one1
woke up in an old hotel tied to a piano
escaped, stole small boat, and crossed to France
only allowed to play friendly matches
Timothy Bell remains missing
the cast is only good in its sadness
or this2
skull drove him to hallucinations, madness and death
suffered a stroke in 2004
died in Weimar in 1793
remains were interred at the Congressional Cemetery
or3
leave to become an actor
return to the club
win the championship
establish an agricultural utopia
or4
sweet and clear across the water,
corbels carved with gilded angels.
larger-than-life cardboard cutouts
could have fled.
None have the magic spark, but they do contain an element of surprise and don’t give you the same ick as LLM text.
Ever since I had the realization that my agents produce more interesting web designs when they have coherent randomness, I’ve started curating my own datasets. I now serve them on this little agent-optimized webpage:
Currently it only supports font pairings, color palettes, and wiki pages. I’ll add more curated datasets there over time. The page will work for your agents on claude.ai or chatgpt.com, provided you mention the full URL verbatim in your message when asking the agent to draw some randomness.
Now, randomness does help fight the worst slop, but it is not a panacea. In order for you to produce something interesting that stands out, you still need to work hard and curate the details with your human judgement. But it’s one more tool in your toolbox.
Home and Away: Revenge — “Bianca wakes up in an old hotel tied to a piano.”
François Thurot — “Thurot, who was being held aboard a prison hulk at Dover, escaped, stole a small boat, and crossed to France.”
Basque Country women’s football — “It is not affiliated with FIFA or UEFA and therefore it is only allowed to play friendly matches.”
2004 Good Friday ambush — “Timothy Bell remains missing and is presumed dead.”
Cool Breeze — “The cast is only good in its sadness, and especially with Raymond St. Jacques as Mercer…”
The Skull (film) — “He in turn falls victim as the skull drives him to hallucinations, madness and death.”
Mildred Shay — “Shay suffered a stroke in 2004.”
Johann Joachim Christoph Bode — “He died in Weimar in 1793.”
William Pope Duval — “William Pope Duval died in Washington, D.C.; his remains were interred at the Congressional Cemetery.”
Marie Duval — “She initially became a governess but left to become an actor.”
Daniel Doram — “He returned to the club for their inaugural NEPL season in 2019 and during the season was selected for the NEPL Representative XI which played MCC.”
1992 Canadian Soccer League season — “Winnipeg won the title becoming only the third club to win the championship, ending Vancouver’s four-year winning streak, snapping Vancouver’s 15 match playoff winning streak, handing them their first playoff loss since 1987.”
Abe Gubegna — “He goes on to establish an agricultural utopia with his friends, which is destroyed when a government official lays claim to their land.”
Frederick Delius — “Delius never forgot the singing as he heard it, day or night, carried sweet and clear across the water to his verandah at Solano Grove, whenever a steam-ship passed.”
Church of St John the Divine, Morecambe — “The chancel roof is carried on corbels carved with gilded angels.”
Brampton Arts Walk of Fame — “Larger-than-life cardboard cutouts of celebrities were posted along a red carpet.”
Karl Landauer — “Landauer could have fled to Sweden, but instead settled in the Netherlands.”
These are great examples. I have tried similar things by prompting LLMs with random ideas before then asking it to attempt some task (with admittedly mixed success).
Randomness is a recurring pattern as the source of novelty and creativity in many AI algorithms. You can see it pop-up again and again.
This great video on the 3blue1brown channel (actually its from Welch Labs) covers how the diffusion models of image generation work by adding noise to images and learning how to reverse-engineer the image.
But How do AI Images Work?: https://www.youtube.com/watch?v=iv-5mZ_9CPY
When neural networks learn, the weights are randomly initialized – and that random initialization is critical to getting the learning algorithm to work.
Normally, this is a one-time injection, which is one of the reasons these models have trouble continually learning. This paper, from Rich Sutton and others, explains that continuing to inject randomness can overcome some of those barriers and allow these models to continue to learn.
Continual Backprop: Stochastic Gradient Descent with Persistent Randomness: https://arxiv.org/abs/2108.06325
When LLMs choose the next token in their output, they do it based on probability distribution that can be adjusted with a ‘temperature’. The higher the temperature the more random the next token. If the ‘temperature’ is too low the output starts to look like when you write a text on your iphone by clicking the next suggested word over and over again…
In reinforcement learning you build models that are continually balancing between taking an action they understand vs. ‘exploring’ (which often means introducing randomness to encourage them to pick actions whose outcomes they know less about).
Very awesome artwork you have been sharing on your feed by the way. Love it!
Weirdly enough I disagree. I think the root of ai slop is the vague, generic, or low-quality ideas or intention behind it. Controversially, I think the AI style itself is not only fine, but above average. AI generated writing is better structured than most human writing. AI generated images look more beautiful than most human generated images. But they convey a deeper meaning which is often banal and subtly wrong.
We can identify the AI style because it has distinct tics - structures and vocabulary that appear unusually often. We learn to associate AI style with the low quality output it produces, but the style is only the facade of that output. Writing has many components ranging from syntax/grammar to profound ideas. The fundamental problem is AI generates low quality output at the most meaningful level, *despite* producing high quality output at the most mechanical level.
Adding randomness does two things:
1. It increases the chance ai will accidentally produce something good
2. It obscures the statistical patterns we use to identify AI style
The first requires taste to filter for better output. But there are more efficient ways to inject your taste into the ai generation process: namely, by thinking hard and articulating the interesting ideas or surprising priorities you wish to exist in the work. You have to supply and iterate on the meaning yourself, and the iteration cycles are more direct when they involve painful, conscious reflection than when doing a random search
The second isn’t helpful: if the problem is content and not style, obscuring the style doesn’t fix the problem. Over time you would come to get the same “slop” feeling from reading the output, because the slop feeling comes from seeing writing that is high quality on a mechanical level and low quality on a deeper level.
This is a conceptual mistake I also see in music philosophy. It’s common to see music as an artistic balance between the expected and unexpected. But this implies you can improve generic music by making it more unexpected, eg by introducing randomness. This doesn’t reliably work (well it does sometimes, because it’s random). The only way to improve generic music is to introduce deeper relationships into the music. The only way to improve ai output is to make those deeper connections for it.