A Lovecraftian coastal scene under a gibbous moon, depicting Alden Marsh and Thalassa standing on the edge of crumbling cliffs overlooking a raging abyss. Thalassa, a mysterious sea creature draped in seafoam and kelp, has bioluminescent veins visible beneath her translucent skin and long, ink-black hair flowing like water. Alden, a scholarly man, holds her webbed hand as the villagers, pale and slit-eyed, approach with hooks and chains. Below, the ocean churns into a maelstrom, and a colossal shadow stirs beneath the waves. The sky is filled with storm clouds and eerie, phosphorescent light from the sunken ziggurat in the depths.

The Lovecraft Test, DeepSeek Edition

This entry is part 15 of 15 in the series Artificial Intelligence

The stupidly big news in LLMs right now is that DeepSeek, a LLM built by a small Chinese team that was released a couple weeks ago, performs as well as, if not better than larger, more sophisticated (at least in terms of the dollars and hardware thrown at it) models by OpenAI and Anthropic (Claude) for a fraction of the cost and CPU resources. In other words, they owned us, and it has sent Nvidia’s stock prices into freefall. (Nvidia being where the majority of the ridiculously expensive GPUs that LLMs apparently need to function properly…or needed before DeepSeek made us realize that they actually didn’t need them at all. Apparently.)

So, of course I was curious and I started running a few of my own tests on DeepSeek. The first one being the “strawberry” test.

Many LLMs have failed to identify the number of r‘s in the word strawberry. While obvious to a human (count them, they’re right there!), an LLM doesn’t process each letter individually — it processes a whole word as a “token”, so it often bases the number on what it thinks the word sounds like. Which means often times, it will say there are two r‘s in strawberry, because there are two r sounds (I mean, I’m guessing…AIs can’t really sound things out either…). DeepSeek got it right on the first try, although what’s interesting (and honestly, pretty endearing) about DeepSeek is that is exposes its thought process. And it really tried hard to psyche itself out when figuring out the number of r‘s in strawberry, looping back on itself multiple times and cross-checking its work.

Passing that, I tried another question that tends to stump AIs — which came first, the chicken or the egg? This can generate any number of arbitrary responses, but DeepSeek approached it logically, concluding that the egg must’ve been first because it would have been laid by a bird that resembled a chicken but was actually not a chicken (pointing to evolutionary mutations).

So then I challenged DeepSeek with a test that a friend of mine has been using to stump LLMs for the last couple years: if I hammer a nail into a wall, is it parallel or perpendicular to the floor? In almost every instance when he, or I, have given this to various AI models, it will get the answer wrong, concluding that a nail hammered into a wall will be perpendicular to the floor (the correct answer, of course, is parallel). DeepSeek, thinking out loud as it does, managed to work out that, since the wall is perpendicular to the floor, if a nail was hammered straight into the wall, it would be parallel to the floor, coming to the correct conclusion.

I figured it was ready to give it the Lovecraft test.

The thing I find most interesting about DeepSeek is less the answers it actually gives and more watching it come to the conclusions it does. I used the word “endearing” earlier; this, I believe, is by design. Keeping the thought process in a black box makes the hallucinations and wrong conclusions feel like more of a betrayal, but watching it “think” feels more…well, human. I read its thought process and see myself in the meandering stream of consciousness. And as it approached the challenge of writing a Lovecraftian tale with romance elements, it was obviously trying to grapple with the seeming contradiction of the request. Here is the story and the thought process in its entirety:

Okay, the user wants a story with romance elements in the style of H.P. Lovecraft. Hmm, Lovecraft is known for cosmic horror, ancient gods, and a sense of impending doom. But romance isn’t typically his focus. So, I need to blend the two. How to approach this?

First, setting. Lovecraft often uses isolated, eerie locations. Maybe a coastal town like Innsmouth or Arkham. The protagonist could be a scholar or someone drawn there for research, which is common in Lovecraft’s stories.

Romance elements. Maybe the protagonist meets someone mysterious, perhaps with a connection to the occult. But in Lovecraftian style, there should be a twist. Maybe the love interest isn’t entirely human? Like a descendant of Deep Ones or another mythical creature.

Conflict. The romance should intertwine with the cosmic horror. Perhaps the protagonist is torn between their feelings and the terrifying truth about their lover. The tension between love and fear is key.

Themes. Lovecraft deals with forbidden knowledge and the insignificance of humanity. The romance could be a metaphor for the allure of the unknown, which ultimately leads to madness or doom.

Structure. Lovecraft’s stories often end tragically, with the protagonist losing their sanity or worse. The romance might end in sacrifice or transformation, maintaining the horror element.

Character names. Lovecraftian names are usually archaic or have a certain ring to them. Maybe Alden Marsh as the protagonist. The love interest could have an exotic name, Thalassa, referencing the sea.

Plot points. The protagonist arrives in a decaying town, meets the enigmatic Thalassa, discovers her connection to ancient beings, faces a ritual, and must choose between saving her or succumbing to the horror.

Imagery. Use descriptions of the sea, cyclopean ruins, strange rituals, and a sense of dread. The romance should have a melancholic, doomed feel.

Need to ensure the cosmic horror elements are present—ancient gods, forbidden texts, transformations. The romance adds a personal stake, making the horror more poignant.

Potential title? Maybe something like “The Whispering Tide” to evoke both the sea and hidden voices. End with a tragic yet inevitable conclusion, perhaps the protagonist joining the love interest in a transformed state, embracing the horror out of love.

Check for Lovecraftian tropes: non-human entities, madness, forbidden knowledge, atmospheric tension. Make sure the romance doesn’t overshadow the horror but complements it. The love story should enhance the sense of tragedy and cosmic indifference.

Title: “The Whispering Tide”

In the forsaken coastal hamlet of Dagon’s Reach, where the cliffs crumbled like ancient bones into a frothing abyss, Alden Marsh arrived under a gibbous moon. A linguist of arcane dialects, he sought the source of a chorus heard only in fever dreams—a susurrus of vowels no human throat should shape. The villagers, pallid and slit-eyed, muttered of a “Bride of the Depths” who sang to the void. Alden scoffed… until he heard her.

She emerged from the mist one twilight, her form draped in seafoam and kelp, hair like ink spilled in water. Thalassa, they called her, orphan of the tide. Her voice, when she spoke, resonated with the hollow ache of abyssal trenches. Yet her eyes—pools of liquid obsidian—held a sorrow that mirrored his own lonely scholarship. They met in secret among the cyclopean ruins that jutted from the shore, where she taught him hymns that made the stars shiver. Love, or something like it, bloomed in the shadow of leviathan bones.

But as the autumn storms clawed the sky, Thalassa’s skin grew translucent, revealing the faint pulse of bioluminescent veins. She whispered of a covenant: her lineage, a hybrid curse born of human folly and the Deep Ones’ cold embrace. Each generation, a bride was offered to Dagon to still his wrath. This year, the elders had chosen her.

On the eve of the Ritual of Submergence, Alden found her kneeling in a grotto lit by corpse-blue phosphorescence. The air reeked of brine and primordial musk. “The moon aligns,” she trembled, “and I must descend to the sunken ziggurat. My flesh will… change.” Her tears were saltwater. “Forgive me. I wished to know warmth before the cold claim—”

Madness or devotion? Alden could not say. But when the villagers came with hooks and chains, he plunged into the riot, seizing Thalassa’s webbed hand. Together, they fled to the cliffs, the mob’s guttural chants echoing like a dirge. Below, the waves churned into a maelstrom, and something colossal stirred.

“The Gate,” she urged. “Once opened, it cannot be closed. You will be unmade—”

He kissed her then, tasting the ocean’s infinite hunger. “What is unmade may be remade.”

They leapt. The water did not drown but transformed. Pressure crushed his ribs, gills slit his neck, and Thalassa’s form shimmered into her true glory—a creature of writhing tentacles and iridescent scales, yet her eyes remained achingly human. Together, they sank into the ziggurat’s glowing maw, where Dagon’s voice thrummed: “Mine.”

Alden’s final human thought was not of terror, but of her fingers interlaced with his, as the abyss reshaped him into something that loved and devoured in equal measure.

Epilogue: Sailors whisper of a duo that surfaces on moonless nights, their duet a lure and a warning. Some say the man’s face still bears the ghost of a scholar, eyes wide with adoration for the dark siren who holds him—a romance etched not in stars, but in the void between them.

Now, I don’t think the story is very good. It’s fine, and it’s got a vibe that feels vaguely Lovecraftian. I think it actually manages a Lovecraftian romance better than some of the other tests. But I’m really interested in the thought process almost more so than the output. I like seeing how the sausage is made. Compared to the previous tests with modern ChatGPT models, it scores much lower in terms of reading level. But considering the relative cost to build the model, it performs remarkably well and it will be interesting to see what’s next for DeepSeek.

( ARLCalc )  
INFO.
AVERAGE READING LEVEL CONSENSUS
The average public reading level in the U.S. is 8th grade. Based on 8 formula(s), your text scored the following:
Score: 7.81 [ = grade level ] Read more
U.S. Grade Level: 8th Grade
Reading Level: Average – Slightly Difficult
Age Range: 13-14

AUTOMATED READABILITY INDEX

Score: 8.31 [ = grade level ]
Reading Difficulty: Average – Slightly Difficult
Grade Level: 8th Grade
Age Range: 13-14

CALCULATION ON YOUR TEXT
ARI = (4.71 * ( 2417 characters / 499 words)) + (0.5 * (499 words / 36 sentences)) – 21.43
Final result: 8.31

==============================================

FLESCH READING EASE
Score: 75.00 [ = reading scale ] INFO.
Reading Difficulty: Fairly Easy
Grade Level: 7th Grade
Age Range: 12-13 years old

CALCULATION ON YOUR TEXT
206.835 – (1.015 * (499 words / 36 sentences)) – (84.6 * 700 syllables / 499 words)

Final result: 75.00

GUNNING FOG INDEX
Score: 7.70 [ = grade level ]
Reading Difficulty: Average – Slightly Difficult
Grade Level: 8th Grade
Age Range: 13-14

CALCULATION ON YOUR TEXT
FOG = (0.4 * (499 words / (36 sentences + 8 compound sentences)) + 100 * (40 FOG hard words / 499 words))

Final result: 7.70

FLESCH-KINCAID GRADE LEVEL
Score: 6.37 [ = grade level ]
Reading Difficulty: Fairly Easy
Grade Level: 6th Grade
Age Range: 11-12

CALCULATION ON YOUR TEXT
SCORE = (0.39 * (499 words / 36 sentences)) + (11.8 * (700 syllables / 499 words)) – 15.59

Final result: 6.37

COLEMAN-LIAU READABILITY INDEX
Score: 9.15 [ = grade level ]
Reading Difficulty: Slightly Difficult
Grade Level: 9th Grade
Age Range: 14-15

CALCULATION ON YOUR TEXT
SCORE = (0.0588 * (2299 letters / 499 words) * 100) – (0.296 * (36 sentences / 499 words) * 100) – 15.8
Final result: 9.15

==============================================

THE SMOG INDEX
Score: 6.44 [ = grade level ]
Reading Difficulty: Fairly Easy
Grade Level: 6th Grade
Age Range: 11-12

CALCULATION ON YOUR TEXT
SCORE = 1.043 * Sqrt((42 hard words * (30 / 36 sentences)) + 3.1291)

Final result: 6.44

ORIGINAL LINSEAR WRITE FORMULA
Score: 83.00 [ = reading scale ] INFO.
Reading Difficulty: Above Average
Grade Level: 8th Grade
Age Range: 13-14 years old

CALCULATION ON YOUR TEXT
Original Version Score = ((349 one-syllable words – 42 ignored words) * 1) + (3 * (36 sentences + 0 compound sentences)) / (499 total words / 100))
Final result: 83.17 = 83

==============================================

LINSEAR WRITE GRADE LEVEL FORMULA
Score: 6.54 [ = grade level ]
Reading Difficulty: Average
Grade Level: 7th Grade
Age Range: 12-13

CALCULATION ON YOUR TEXT
Initial Score = (((459 easy words – 42 ignored words) * 1) + (42 hard words * 3)) / (36 sentences + 0 compound sentences)
Adjusted Score = (15.08 (initial score) – 2) / 2
Final result: 6.54

==============================================

FORCAST READABILITY FORMULA
Score: 9.51 [ = grade level ]
Reading Difficulty: Somewhat Difficult
Grade Level: 10th Grade
Age Range: 15-16

CALCULATION ON YOUR TEXT
SCORE = 20 – ((349 1-syllable words * 150) / (499 words * 10))

Final result: 9.51

Results calculated on 2/15/2025 at 12:19:26 AM by ReadabilityFormulas.com

Series Navigation← The H.P. Lovecraft Test, round 3

Posted

in

,

by

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.