Most compressors assume your data is more-or-less random and go hunting for repeated byte strings. That's the right bet for a lot of files. But an enormous and fast-growing share of the world's data isn't random at all — it's generated. A column of timestamps. A counter. A sensor humming away. An audio waveform. Nobody typed those out byte by byte; a rule produced them.
MIKE's bet is simple: if a rule made the data, store the rule. Here's what that actually buys you — and, just as importantly, where it buys you nothing.
01The 38-byte file
Take 44,640 consecutive one-minute timestamps from a real Binance market-data dump — little-endian 64-bit integers, 357,120 bytes on disk. Every value is the one before it plus exactly 60,000 milliseconds. Point the field at it:
Thirty-eight bytes. Not thirty-eight kilobytes — bytes. Because MIKE noticed the run was an exact arithmetic progression and stored the only three things it needs to rebuild all 44,640 values on decode: the first value, the step, and the count. That's a ~20-byte formula plus a tiny header. The file could have been 357 KB or 357 MB — the formula is the same size.
Why can't LZMA (which powers 7z and xz) get there? Because it finds repeated byte strings, and consecutive int64s that differ by 60,000 aren't repeats — the carry ripples through the eight bytes differently at every step, so there's no matching substring to grab. There's an arithmetic relationship, and byte-oriented match-finders don't see arithmetic. (7z's delta filter is why it reaches 90 KB instead of 900 KB — but it still stores a residual for every sample. MIKE stores one formula for all of them.)
This isn't a lab toy. The same shape is everywhere: log timestamps, sequence numbers, database primary keys, packet counters, trade IDs. Hand MIKE 3.2 MB of real Binance trade-IDs and it also lands on 38 bytes — 221× smaller than rar.
02Math the deltas
Audio is a signal — a smooth curve sampled tens of thousands of times a second. Smooth means predictable: each sample sits close to a simple projection of the ones before it. So instead of storing the samples, store how wrong a cheap prediction was.
Order-1 prediction is just the delta (this sample minus the last). Order-2 is the delta of the delta. Order-3 is one more. For a smooth signal those higher-order differences collapse toward zero — and a stream of near-zeros compresses to almost nothing. MIKE splits each channel out, picks the best predictor order per channel, separates each residual into a low-byte and a high-byte plane (small residuals mean the high plane is nearly all zeros), and stores the file's header and any extra chunks verbatim so the result is byte-exact.
On a real hospital ECG recording (MIT-BIH), MIKE lands 24% under 7z and 72% under zip. This is the same core idea FLAC uses to shrink your music losslessly — MIKE just does it as one mode inside a general-purpose codec. The lesson is the same as the timestamps: LZMA can't model a curve; prediction can. Where the data is a signal, prediction wins.
03The part nobody else will tell you
You cannot losslessly shrink every possible input. It's a counting argument: if a compressor makes some files smaller, it must make others larger, because there simply aren't enough short outputs to go around. Anyone selling you "smaller than everything, every time" is selling you a violation of arithmetic.
So MIKE is a specialist, and it plays by two rules. First, it never expands — worst case, it falls back to plain LZMA plus a one-byte tag. Second, it tells you where it loses.
Executables: MIKE splits x86 branch targets into separate streams (a BCJ2-style transform) and pulls dead-even with 7z — within 0.3–0.7% — while beating zip, rar, xz and zstd. But it does not beat 7z. We went looking for a "formula" in compiled code, and there isn't one: the redundancy in machine code is repetition, and 7z's range-coded BCJ2 already maxes that out. So we say near-parity, not a win.
Already-compressed media (JPEG, MP3, MP4): nobody wins, MIKE included. They're already at the entropy floor. MIKE breaks even and moves on.
That honesty isn't modesty — it's the product. In a field waist-deep in snake oil, telling you precisely where the tool shines and where it shrugs is the feature.
04Best of all modes, never worse
Under the hood MIKE is a best-of selector. For each block — or each whole file, for the audio and executable paths — it tries a set of reversible transforms: plain LZMA, byte-delta at several strides, 32- and 64-bit integer generators, true integer delta, audio linear prediction, executable branch-splitting. It keeps whichever produces the smallest output and tags it so decode reverses exactly the right one. Plain LZMA is always in the running, which is what guarantees the never-expand property. Every path is byte-exact: decode(encode(x)) == x for all x, and every benchmark above was verified with a full round-trip and a CRC check.
05Don't trust me — run it
None of the numbers on this page ask for your faith. We bundled the exact files, a one-command benchmark script, and a README explaining what each file is. It compares MIKE against whatever of zip/rar/7z/xz/zstd you have installed, skips the rest, and CRC-checks every round-trip.
tar xzf mike-ai-benchmark-pack.tgz && cd mike-ai-benchmark && python3 bench.py
Download the benchmark pack (18.9 MB) — real public datasets, reproducible on your own machine.
MIKE isn't magic. It's the opposite of magic: the plain observation that much of the world's data was generated by simple rules, plus the discipline to find those rules, store them, and admit honestly when there aren't any. Where the data has math, MIKE wins. Where it doesn't, MIKE says so out loud. That's the whole trick.
It's a solo project — Rust engine, Linux / Windows / Android, free to evaluate.
Download MIKE See every benchmark