Temporada 5 · Descolonizar la Singularidad · JVV-T5-16

Dataset Absinthe: distilling decadence in an intoxicated GAN

I The green fairy enters the GPU

Absinthe — cloudy liquor, 72 % ABV, chlorophyll macerated with wormwood — inspired the symbolists to saw through the ribs of reality.
In 2026 a Paris collective scans 18,000 fin-de-siècle lithographs, bottle labels, cabaret photographs and cursed poems; it turns them into DATASET ABSINTHE v1 [7.3 GB].
Train a GAN on those green-saturated images and the model begins to hallucinate arabesques dripping wormwood: melted pixels, verses scrubbed of morality, dancers with no body, only tulle and turquoise smoke.

“Absinthe dissolves the tongue; the GAN dissolves the author's signature. Both leave a poisonous glow you drink with your eyes.”

II Ethyl-algorithmic recipe

Ingredient	Weight	Toxic function	GAN domain
Wormwood [thujone 35 mg/L]	1 L	Visionary, convulses GABA	128-d latent noise 💀
Burnt sugar	250 g	Sweetens the bitterness	Loss smoothing [Label-Noise 0 .05]
Iced water	1 L	Releases the louche [opalescence]	Style-Mixing layers
Toulouse-Lautrec lithograph	3,000 img	Degenerative contour	High-freq patch 32×32
Decadent poems	6,000 lines	Narcotic language	Text-embedding condition

Still: CUDA RTX 4090, batch = 24, ADA = on.

III Formula for latent louche

\Lambda = \frac{\sigma_{G_\ell} \; \cdot \; \text{Thuj}_{\%}}{\mu_{D_\text{real}} + \delta}

\sigma_{G_\ell} = style deviation ℓ
Thuj% = % synthetic thujone introduced as noise
\mu_{D_\text{real}} = mean real probability from the discriminator
\delta = 0.01 avoids dividing by zero

If \Lambda ≥ 1,6 the image “louches”: misty green colour, distilled edges.
Optimum value found ≈ 1.9 after 14k iterations.

IV Results: drunken postcards

Conditional prompt	GAN output	Hallucinogenic effect
`courtesan+halo`	Liquefied face, aquamarine halo	Mild VR dizziness
`poem: spleen`	Drop-by-drop calligraphy, withered background	Smell reports an aniseed aroma [synaptic]
`louise+bridge`	The Seine bridge dissolved into sugar cubes	User feels vertigo for 3 s

Clinical observation: 18 % of testers said they heard a 14 kHz hum during inference — a “symbolist mosquito”.

V Home ritualsafe

01

Pour 30 mL of absinthe at 45 % ABV [or strong green tea].
02

Drip iced water -- 3× at a 1:3 ratio -- and watch the louche.
03

While it clouds, generate an image with the prompt “verlaine+pixie”.
04

Contemplate the picture for 60 s, take a sip.
05

If you notice edges melting in real life, close the visor; the liturgy is complete.

[Drink responsibly.]

VI Risks & antidotes

Risk	Symptom	Countermeasure
Decadent network — the model degrades, generates pure noise	Loss oscillates > 4.8	Reduce the thujone noise, reintroduce 10 % “healthy” photos
Market fetish — a brand sells “AbsintheGAN™ prints”	Appropriation without credit	CC-BY-SA licence + 15 % to the Bibliothèque Nationale archive
Spleen overfitting — nostalgic outputs depress people	Users feel a “green heaviness”	Inject a magenta colour module, a line of Rimbaud to oxygenate

VII Coda

“Between 72 % alcohol and 18 GFlops there is no border, only filaments of binary chlorophyll.”

Dataset Absinthe does not try to replicate the Belle Époque; it tries to grind it in a CUDA mortar and turn it into a psychovisual tonic reminding the user that every vintage filter hides a cultural poison.
Drink in micro-doses, exhale unstable verses and archive your melancholy as a transparent PNG: tomorrow another model will distil your pixels until the palette turns toxic neon.