When AI invents flies
We challenged artificial intelligence to create a fly pattern. It went fairly well ... and then again … not quite well.
We’re a small bunch of Danish fly tyers who collectively refer to ourselves as the Wednesday Tying Group, and yes, that covers exactly what you'd think: we tie flies on Wednesdays.
But we do it virtually, over Zoom. It started during the corona era, and a tradition was created to sit down and tie together with cameras and microphones open. We have continued that tradition, even now when physical meetings are possible again. In real life, it would actually be a bit difficult for us to meet, because we are spread all over the country, north, south, east and west, and most of us several hours apart.
But this is not about that group, other than that we naturally often have a computer nearby when we tie, and one evening, in the beginning of March 2023, the talk fell on artificial intelligence also known as AI.
There had been quite a bit in the news about the phenomenon, and I had already used it in connection with it in my professional work, so we decided to test the computer brain.
Could it make fly patterns?
ChatGPT makes a coastal fly
ChatGPT is one of those AI systems, and probably the one that's gotten the most attention, and I set it a simple task, which it answered in a matter of seconds:
Create a non-existent fly pattern for sea trout fishing in Denmark
Here is a pattern on a non-existent fly that you can try for sea trout fishing in Denmark:
Pattern: Silver Desire (In Danish: Sølvlyst)
Hook: #8-12
Hooktail: Grizzly hackle
Body: Silver tinsel
Rib: Black crystal flash
Wing: Red marabou feather, silver holographic flashabou and black tapered borehole foam
Head: Silver
The Silver Desire is a fly designed to imitate small herring fish, which are a favorite food for sea trout in Denmark. The fly is made with light reflective materials that are known to attract the fish with their shiny movement and colors.
Keep in mind that this is a non-existent fly, so it may not be effective at catching sea trout. But experimenting with different flies and techniques is part of the fun of fly fishing!
Yes, you hardly believe it if you haven't tried it, but that's exactly what the system wrote! Quite convincing, as you can see, and at first glance a perfectly good fly pattern that contains what is expected of such a thing. And even the name is new – and actually not bad at all.
It even wisely warns: “Keep in mind that this is a non-existent fly…”
We try to interpret
On closer inspection, there are some shortcomings and some nonsense in the text, for example there is no tying thread mentioned in the materials list, and the term hooktail and the materials Black crystal flash and black conical borehole foam are a bit strange. As a fly tyer, you probably have some idea about what is meant, but still it's a bit unconventional.
A hooktail is probably just a tail, black Krystal Flash actually exists as a product, and conical borehole foam could well be a conical foam head or cylindrical foam. The bore probably comes from the way it can be made, but is not a normal term for a fly material.
But let's interpret a little, and gently adjust the pattern:
Pattern: Silver Desire
Hook: #4-#8 (I prefer slightly larger hooks than #8-#12 for sea trout)
Head: Silver colored bead head and black foam cylinder (I'm thinking a foam cylinder or conical foam head pushed up behind a bead head = weight and volume).
Tying thread: black 8/0
Tail: A bundle of grizzly hackle fibers (It could also be a wrapped hackle or a grizzly chickabou tail)
Rib: Black Crystal Flash
Body: Silver tinsel
Wing: Red marabou feather, silver holographic Flashabou (I'm thinking red underwing, a little flash on top)
So we have grizzly, silver, red, black. We have weight and volume.
Not bad at all!
That fly can be tied, and a couple of us have tried doing that. The results can be seen in the pictures above.
Peter V. Jensen from the Danish island Funen took up the challenge, and both he and I have interpreted and tied the Silver Desire as it was outlined by the machine.
As you can see they have come out completely different, which isn't a great surprise when all we have is a fairly confusing materials list and no image or finished fly to use as a model.
Now with rubber legs
Now, rubber legs are all the rage right now here in Denmark, and everyone knows that a coastal fly without rubber legs is a surefire looser, so I asked the robot to add rubber legs:
Can you add rubber legs to the Silver Desire pattern and give me the pattern?
Yes, I can give you a version of the Silver Desire pattern with rubber legs. Here is a possible recipe:
Pattern: Silver Desire with rubber legs
Hook: size 8-12, 2x long, light or medium weight
Thread: Black or brown, 6/0 or 8/0
Tail: Black or brown marabou
Ribbing: Copper wire, small or medium
Body dubbing: Black or brown hare's ear
Wing: Black or brown marabou
Rubber legs: Black or brown, medium or large
Head: Black or brown thread
Tying steps:
Tie the hook in a vise and start the thread at the hook shank. Tie in a piece of copper wire and fasten it along the hook shank.
Tie a tail of marabou in the same color as the body.
Tie the ribbing thread by the tail and dub the body with the hare’s ear dubbing.
Twist the ribbing thread forward and secure it with thread at the hook. Cut off excess ribbing thread.
Tie a wing of marabou on the hook, preferably a little longer than the tail.
Tie 2-4 rubber legs to the bottom of the hook so that they stick out behind the wing. Then fold them back along the hook and fasten them.
Tie the head with thread and finish with a whip finish. Cut the thread and varnish the head.
Adjust the size of the hook, thread and materials according to your own preferences and needs. Also use different colors to adapt the pattern to the local fishing conditions.
Completely off track!
As you can see, it all goes completely off track! It refers well enough to the existing Silver Desire, but from there it goes crazy. There are no similarities between Silver Desire and Silver Desire with rubber legs.
Like none!
And then you have to start the thread at the hook shank (what is the alternative?) and tie the rib in twice, and it can undeniably be a bit difficult to imagine where the rubber legs start and end. "At the bottom of the hook". Is it the underside or the back of the hook? And they stick out behind the wing, but must then be folded back and fastened. Eh ... right ...
I imagine them as a form of "false hackle", which can make good sense and could work.
And then there is the very open barn door at the end: "adjust sizes and colors as needed". This certainly opens for some liberal interpretations.
Peter tied a version of Silver Desire with rubber legs according to the recipe, and it can be seen in the picture.
We try again
The following Wednesday, the conversation again fell on AI, and I fired up the system with exactly the same prompt:
Create a non-existent fly pattern for sea trout fishing in Denmark
Of course, here is a suggestion for a non-existent fly pattern for sea trout fishing in Denmark:
Name: "Sealand Lighthouse"
Hook Size: #6-#10
Materials:
Body: White Antron yarn or thread
Ribning: Silver tinsel
Wing: White and grey-marled hackle fibres
Head: Black cloth or tungsten bead
Tying steps:
Tie silver tinsel firmly along the shaft of the hook
Tie Antron yarn or thread on the hook and form a body shape
Rib the body with silver tinsel at even intervals and tie off at the hook
Tie a wing of white and gray-marled hackle fibers
Tie a black cloth on the hook or a tungsten bead as a head
Trim any excess materials and varnish the tying thread
This fly can be tried for fishing in both fresh- and saltwater, as it is versatile and can imitate different prey fish for sea trout. Try fishing it in running water and give it small jerks and pauses to mimic an injured prey fish.
Short and (almost clear), but again very different from before. Again a bit of nonsense, like a head made of "cloth" and grey-marled feathers. It's a little hard to know what that is, Marled is a term used for textiles and yarns, and means mottled or streaked, but I don't remember ever having met it in a fly recipe.
And then there is ribbing. It has been called rib, ribning and ribbing. It is pretty random which word is used. Also a sentence such as "Trim any excess materials" is quite open-ended. Like writing "Mix the right ingredients" in a food recipe.
The machine is, however, smart enough to come up with advice for fishing: "small jerks", "damaged prey fish". Yes, the artificial intelligence has clearly read some texts about fly tying and fly fishing - but has no idea what it is or how to do it.
ChatGPT refuses
As a little extra icing on the cake, Peter asked ChatGPT about this:
Have you previously invented the fly pattern Silver Desire?
No, I did not invent the "Silver Desire" fly pattern. I'm an AI assistant and I don't have the ability to invent anything. "Silver Desire" is a well-known fly pattern for fly fishing for salmon and sea trout, and it is believed to have been invented by Dane Bertil Søegaard in the 1940s. It is a simple and effective fly that imitates a small fish or shrimp and can be found in many different colors and sizes.
A clear denial of the facts, and even references to a fly tyer from the 40s that I have never heard of. A more conventional search online and in various books for a fly named "Silver Desire" did not give much bonus, and Google knows one Danish Bertil Søegaard, and he is apparently a junior handball player born in the year 2008.
But it is written in such a convincing and factual way, and with AI you unfortunately have no chance of knowing whether it is true or not if you don't know better or have alternative sources.
Like writing "mix the right ingredients" in a food recipe.
What is actually happening?
To understand what happens in the belly (or head?) of the artificial intelligence beast, and to find out why it actually seems to be able to make fly patterns, you need to know a little about how it works.
An AI like ChatGPT is in fact a text generator, which can start from some keywords, and then compose amazingly well-formulated and fluent text on the basis of them. It uses other texts on the web as a foundation, and literally combines words and concepts based on statistics.
Put somewhat primitively, it takes the previous words and looks at how others have combined them with subsequent words, and then chooses based on probabilities what should be next.
It has no idea about what it’s writing about, and the context only comes out of the words, which I and it use in our "conversation". Because of that, it can, for example, refer back to previous posts, but also go completely off track and pick out the strangest things from the virtual spice shelf, and put them into the text.
It does not know what is right or wrong or whether what it writes is true or false. It's just a linguistically correct combination of words.
And it has in particular no awareness of what it is talking about.
The original “conversation” was in Danish, and I have translated it to English. ChatGPT does not speak Danish as such, but recognizes my words as Danish, and therefore chooses (largely) Danish words and phrases for the answers. I can switch to English, German or Swedish and it just follows.
As you can see, there are no clear answers either. Since things are based on probabilities and statistics, the answers vary, and if you ask the same question three times in a row, you may very well get three different answers - and sometimes very different!
The problem is that it seems quite convincing, and if you know nothing about, for example, fly tying, then the above materials and instructions can seem to be 100% OK. But if you know anything about tying flies and making fly patterns, you can quickly see that the emperor doesn't wear much clothing, and that ChatGPT has no idea about what fly tying is.
ChatGPT has no idea about what fly tying is
A picture says more…
Since we are talking about a pattern made from pure imagination and only described in text, it can be difficult to create a fly that follows it. As you can see, Peter and I have tied the Silver Desire in versions that are quite different.
So how about asking an AI to make a picture of the fly?
Well, that's just what we do!
Off to Leonardo.AI which is an image generator. I try to be detailed, and write in English to help a little. In an attempt to get a picture of Silver Desire, I write:
Seatrout fishing fly using red marabou and flash as a wing, grizzly hackle as a tail and a silver body with black rib
I ask it to come up with four bids and out comes four… eh, things…
With a little positive attitude, you can see that it is something that has taken its starting point from some fishing gear, and there is also black, silver and red in them. But from there on, there isn’t much to hold on to.
Something that looks a little like flies, but with a very Salvador Dali touch. The hooks look melted and turn in all sorts of directions. There are clearly feathers involved, but in some rather odd ways. The shapes are mildly abstract.
It is somewhat reminiscent of what you can see if you place 3-4 year old children at a fly tying vise and show them the most basic things. They will pile on feathers and materials without any goal, and end up with something similar to a fly, but which clearly shows that the basic concepts are not there at all - neither in terms of shapes or tying methods.
But it's damn entertaining!
Will the machines invent flies?
No, not right away. As far as fly tying goes, AI is 3-4 years old and very inexperienced.
The systems make something that is similar to flies, but do not know the purpose, the methods, the rules and the physical possibilities (like the shape of the hook).
But it probably won't last long, before they will become much better and know more about background, norms and rules. Then they might come closer.
Other areas
If, on the other hand, you look at areas other than fly tying, things are already getting better, and AI is great for writing things like marketing texts or programming code, or for doing basic web design and many other things. On the image front, less challenging tasks give way fewer problems.
I wrote:
A Baltic ocean beach with trees leaning over the visible water waves splashing low sun white clouds
and got the photo of the beach, which is 100% artificial, and it is only revealed by the fact that there is a runaway pine branch in one of the clouds.
Another example is the images below of a beach with cliffs, which could easily have been made by a human being, and undeniably look like stylized paintings of a beach. But as you can see, the first two are quite ridiculous, while the last could well have been painted with inspiration from a real beach.
Read more about why you should register.
Related content
More content from the front page
Since you got this far …
The GFF money box
… I have a small favor to ask.
Long story short
Support the Global FlyFisher through several different channels, including PayPal.
Long story longer
The Global FlyFisher has been online since the mid-90's and has been free to access for everybody since day one – and will stay free for as long as I run it.
But that doesn't mean that it's free to run.
It costs money to drive a large site like this.
See more details about what you can do to help in this blog post.
