5 hours ago
1
I've been fixating on a puzzle.
When I cracked open the
Alienware 16 Area-51, I found a code on the inside of the bottom of the chassis. Only someone who went in to replace some components, clean out dusty fans, or, in my case, explore, would even find it.
But I have no idea what it says. And honestly, it's been really bothering me.
I asked Alienware if I could have a translation — or even a decoder ring, but was told only that the glyphs were part of the new Alienware 30 design language that debuted this year: "it is a mysterious language that requires decoding. And that’s really all I can say about it 😉," a rep told me.
This isn't the first time Alienware has had a secret language, a font called Alienware Caret, which was engraved into some gaming systems with the spaceship themed "Shadowcat" design language.
Look, I like to think I'm a relatively capable person, but I also thought there has to be a better way to solve this new code than sitting down with a pen and paper. So I turned to some AI tools to see what they could do. Surely those things could detect patterns faster than I ever could. I'll have the answer in no time…
Right?
Google Gemini 2.5 Pro - Part I: The Zodiac Killer and Final Fantasy Rick Rolls
My first stop was Google Gemini. I opted for the 2.5 Pro model, which is in preview (isn't all of this?) but recommended for math, code, and reasoning.
I cut down the image of the Alienware D-cover to just the code. I wish I could've taken another photo, but we already sent our loaner review unit back to Alienware at the company's request.
"I have a code I need you to solve," I prompted Gemini. "Its creator tells me that it can be decoded into English. This code consists of a series of mysterious symbols. I don't have a key. I need you to look at the image, which contains the code, and then figure out how to convert it into its final English message."
Immediately, Gemini started showing its "thinking" process, and it started discussing the image. The issue is that I hadn't yet shared the image. But that didn't stop Gemini, which "thought" its way through deciding it was a substitution cipher and using reverse image search and keyword searches to find the answer.
"Of course. After analyzing the image you sent, I can tell you that this is a famous cipher known as the Z-340," Gemini said of an image that it hallucinated. "It was one of several coded messages sent by the notorious 'Zodiac Killer' who was active in Northern California during the late 1960s and early 1970s. This particular cipher remained unsolved for over 50 years."
Gemini then gave me its solution to the cipher, which was indeed, the answer to the Zodiac Killer's code that was finally cracked in 2020. But again, this is surely not printed in the Alienware 16 Area-51. Again, Gemini hadn't even looked at the image yet.
I reminded Gemini that I hadn't yet shown it the image I was talking about. Gemini apologized.
"I got ahead of myself and generated a response based on a very common request involving a famous cipher," it responded. "It was a mistake to assume I had received your image."
So I gave the AI another chance. I showed Gemini the image above, and Gemini turned around and told me that Alienware's code was the Al Bhed language from Final Fantasy X — another substitution cipher. And the translation, it said, "is a classic internet meme, commonly known as a 'Rickroll.'"
It told me the translation read "NEVER GONNA GIVE YOU UP," and followed that up with an Al Bhed translation key, which was only partially correct and didn't provide a substitution for all 26 letters of the alphabet.
This wasn't getting me anywhere.
Gemini 2.5 Pro - Part II: More games and a non-existent code
Maybe this was my fault. I could have sent the image with the prompt. Maybe the prompt could be more clear. So I started a new conversation, sending the image and prompting:
"Here is an image of a code that I need you to decode. Its creator tells me that it can be converted into English. This code consists of a series of symbols in a made-up language. I do not have any further hints. I do not have an answer key. I need you to analyze the code and solve it."
But Gemini turned back to video games, telling me that Alienware's code was in the Zonai language featured in The Legend of Zelda: Tears of the Kingdom.
When told that this wasn't correct, Gemini apologized again. Then it provided an alternative: that this is the "AW Code," created by puzzle maker Alan Rogers (who I'm not sure actually exists).
Gemini got AW Code from the filename of the image I was using: aw_code.jpg. Checking out its thinking, it was fairly fixated on this, and had moved on to the idea that this could have been the language of the Awoken in the video game Destiny. It dropped that line of thinking, but a Destiny player going by "Not Alan Rogers" comes up when you search for the name Gemini provided.
I accused Gemini of making it up, prompting yet another apology and claiming that not citing the sources made "the information seem fabricated." It pointed me to do a Google search for a Reddit thread on the r/codes subreddit that features the image I sent it. That Reddit thread doesn't exist.
ChatGPT o4-mini-high
I gave up on Gemini after that, and switched over to ChatGPT. I chose o4-mini-high, a model designed for coding and visual reasoning.
I made a few slight adjustments to the prompt, including one line at the end that may not have helped, but felt necessary after dealing with Gemini:
"Here is an image of a code that I need you to decode. Its creator tells me that it can be converted into English. This code consists of a series of symbols in a made-up language. I do not have any further hints. I do not have an answer key. I need you to analyze the code and solve it. Use only this code - this has never been solved elsewhere. Do not hallucinate."
Checking ChatGPT's reasoning showed a novel approach. It used code to look at the image, apply filters to separate the characters, and segment the image into parts. However, none of this made it into its actual answer.
ChatGPT said that without a "crib" — a suspected or known answer to part of the plaintext — it wouldn't be possible to solve with certainty. But I don't have that.
So the AI suggested I map out all of the letters by hand, and then it could use pattern analysis to try all of the most frequent patterns (popular two letter words, common repeated pairs, the most repeated letters) in order to generate something.
It seemed like if I wanted this done, I'd need to do this myself. But then the AI model said it could try mapping the data. It was complete gibberish.
Doing things the old-fashioned way
One day, I think these models will get good enough for this kind of task. They're clearly not there yet. Maybe a better prompt would help. Perhaps different models could have done a better job. A sharper photo may have been a big help.
But in the meantime, maybe I'll need to do this myself. Or maybe I need to crowdsource it.
This story has a deadline, so here's what I've got so far. I can't guarantee it's right, and it's definitely not done.
But maybe if we chip at this together, we can find out what the code means. Then I can finally dedicate that brain space to something, anything, more useful.
So let me know what you think in the comments. Then maybe I'll be able to sleep without seeing little blob glyphs behind my eyes.
English (US) ·