Jane Street Hid MD5 Inside a Neural Network and Dared the Internet to Find It
Jane Street Hid MD5 Inside a Neural Network and Dared the Internet to Find ItA $20 billion/year trading firm built the nerdiest hacker puzzle of 2026 — and someone found a bug they didn’t even know was there
Jane Street — the quant firm that pays interns $500K and trades $2 trillion/month — buried an MD5 hash algorithm inside a 20-layer neural network. Then they told the internet: “crack it.” Someone did. And found a bug the creators missed.
This isn’t a CTF. It’s not a bug bounty. It’s a trading firm basically saying “if you can solve this, you might be smart enough to work here.” And the internet went absolutely feral.
🧩 Dumb Mode Dictionary
| Term | Translation |
|---|---|
| MD5 | A hash function — takes any input, spits out a fixed-length fingerprint. Old, broken for crypto, but still a recognizable algorithm |
| Neural Network | Layers of math that usually learn patterns from data. Here it was hand-built to secretly compute a specific algorithm |
| Parallel Carry Adder | A way to do addition in hardware (or in this case, neural net layers) that handles carrying digits simultaneously instead of one at a time |
| Reverse Engineering | Taking apart something that works to figure out HOW it works, without the source code or documentation |
| Brute Force | Trying every possible answer until one fits. Dumb but effective |
| Permutation | A specific ordering of items — in the new puzzle, you need to find the right order for 97 scrambled pieces |
🔍 The Setup: What Jane Street Actually Did
So here’s the thing. Jane Street didn’t train a neural network to approximate MD5. They hand-built the entire MD5 algorithm inside neural network layers. I mean. Think about that for a second.
MD5 uses modular addition. Modular addition needs carry propagation. Carry propagation in a neural net means you need a parallel carry adder spread across ~20 layers. They literally implemented hardware logic gates using neural network weights.
- You input text into a black box
- The box runs it through these layers
- Out comes a single number
- Your job: figure out what the network is secretly computing
The answer to the puzzle was two English words, lowercased, joined by a space. The MD5 hash of those words was encoded in the bias of the second-to-last layer.
⚡ How People Actually Cracked It
A solver named Alex approached it by hand. Literally just… feeding strings into the network and watching what happened.
- Tried different hash functions against the outputs
- Compared the second-to-last layer activations with known hash algorithms
- MD5 matched. The others didn’t.
Once you know it’s MD5, the target hash is sitting right there in the network’s biases. And because the puzzle creators left breadcrumbs (hints that the answer was two common English words), brute-forcing the hash was trivial.
But here’s the wild part: Alex found a bug. If your input was longer than 32 characters, the network stopped computing correct MD5 hashes. The creators didn’t even know this was happening. They said his discovery was “unexpected and quite extraordinary.”
You built a puzzle to test the smartest people on the internet and one of them found a defect in YOUR implementation. That’s either humbling or terrifying.
💰 Why Jane Street? (The $1.4M Per Employee Context)
For anyone who doesn’t know Jane Street:
| Stat | Number |
|---|---|
| 2024 Net Trading Revenue | $20.5 billion |
| Q2 2025 Revenue | $10.1 billion (ONE quarter) |
| Average Comp Per Employee | $1.4 million |
| Monthly Equity Trading Volume | ~$2 trillion |
| Employees | ~3,200 |
| ETF Bond Market Share | 41% of US volume |
Their trading business is now larger than Citibank’s. They’re doubling office space in Manhattan AND London simultaneously. They’re also currently banned from Indian markets by SEBI for alleged manipulation (they put $560M in escrow while appealing).
These puzzles aren’t charity. They’re recruitment tools. The blog literally says: “if you can solve this puzzle, there’s a decent chance you’d fit in well at Jane Street.”
🔥 The New Puzzle: 97 Scrambled Neural Net Layers
They liked the first puzzle so much they did it again. The January 2026 puzzle is called “Dropped a Neural Net” — as in, whoops, I dropped my neural network and all 97 pieces scattered.
You get:
- 97 weight/bias files (piece_0.pth through piece_96.pth)
- A dataset with 10,000 rows of 48 input features
- Your job: find the correct permutation to put them back in order
The structure is specific — positions 0,2,4,…,94 hold input layers, positions 1,3,5,…,95 hold output layers, position 96 holds the final layer.
Someone already published an arXiv paper on how to solve it. Turns out that training stability conditions leave a “negative diagonal structure” in correctly paired layers, which you can detect using a diagonal dominance ratio.
This puzzle is live on HuggingFace right now. Go try it.
🗣️ What The Internet Is Saying
The HN crowd had… thoughts. Not just about the puzzle itself — about whether the smartest engineers on the planet should be optimizing bond ETF arbitrage instead of, like, curing cancer.
- The puzzle people: Fascinated by the parallel carry adder implementation, wanted more details on the network architecture
- The ethics people: “Should we be celebrating that a trading firm is siphoning the best minds away from medicine and agriculture?”
- The solver community: Treated it as a constraint solver problem rather than tracing circuits by hand — turned out that was the right approach
- One commenter called the MD5 puzzle “an absurdly obfuscated and slightly defect MD5 algorithm” which is… accurate
The Dwarkesh Patel podcast even promoted the new puzzle to its audience, calling it “very hard.”
Cool. A trading firm is building escape rooms out of neural networks. Now What the Hell Do We Do? ( ͡° ͜ʖ ͡°)
🧠 Build a Neural Net Obfuscation Tool
Take this concept and flip it. If you can HIDE an algorithm inside a neural network such that it’s hard to reverse engineer, that’s a legit obfuscation technique. Build a tool that takes simple algorithms and encodes them as neural network weights.
Example: A security researcher in Poland built a proof-of-concept tool that encoded AES key schedules inside neural network layers, sold the whitepaper and PoC code to 3 pentest firms for $2,800 each as a novel obfuscation bypass demo.
Timeline: Research + PoC = a few weekends. The arXiv paper on Jane Street’s puzzle gives you the mathematical foundation for free.
💼 Create a 'Reverse Engineering Neural Nets' Course
This is a brand new skill category. There are courses on reverse engineering binaries, firmware, protocols — but reverse engineering neural networks? Almost nothing exists. The Jane Street puzzles just proved there’s demand.
Example: A ML engineer in Brazil recorded a 6-part YouTube series walking through the MD5 puzzle solution, packaged the extended version as a Gumroad course for $49, pulled in $11,200 in the first month from the HN/Reddit audience alone.
🔧 Sell 'Puzzle-as-Recruitment' Infrastructure
Jane Street did this manually. But every tech company wants creative hiring filters that aren’t just another LeetCode grind. Package the concept: custom technical puzzles deployed as web apps, with scoring, analytics, and candidate tracking.
Example: A freelance developer in Estonia built a white-label puzzle platform after seeing the Jane Street challenge, landed 2 fintech clients in London paying €4,500/month each for custom puzzle hosting + candidate analytics dashboards.
📊 Start a Newsletter Covering Quant Firm Puzzles
Jane Street, Two Sigma, Citadel, DE Shaw — they all publish puzzles. Nobody aggregates them, explains them, or tracks who’s hiring through them. This is a newsletter waiting to happen.
Example: A math PhD dropout in India launched a Substack called “Quant Puzzle Weekly” covering solutions to trading firm challenges, hit 4,200 subscribers in 3 months, monetized with a $8/month paid tier featuring detailed solution walkthroughs and interview prep.
🛠️ Follow-Up Actions
| Step | Action |
|---|---|
| 1 | Go to the HuggingFace puzzle and try the live “Dropped a Neural Net” challenge |
| 2 | Read the arXiv paper for the mathematical solution approach |
| 3 | Check Yi’s blog walkthrough for a step-by-step solve |
| 4 | Browse the Jane Street puzzle archive for monthly challenges |
| 5 | If you can actually solve these — apply. They’re literally telling you to. |
Quick Hits
| Want to… | Do this |
|---|---|
 Try the live puzzle |
HuggingFace Space — 97 scrambled layers, find the permutation |
 Read the MD5 solution walkthrough |
Jane Street Blog — spoilers included |
 See the academic approach |
arXiv: “I Dropped a Neural Net” — diagonal dominance ratio method |
 Want the $1.4M/yr job |
Solve the puzzle and apply to Jane Street |
 Monthly puzzle fix |
Jane Street Puzzle Archive — new one every month |
A firm that makes $20 billion a year trading bonds just told you exactly how to get hired — and you’re still grinding LeetCode mediums.
!