GOALS · FOR LABS AND INVESTORS

Hire me. Fund me.
Or pay full price.

I'm 18, self-taught, and I built every layer of the stack alone — transformer, tokenizer, optimizer, thirty architectures, compilers. Below is what that stack is worth on one frontier run. Then three ways to work together.

THE 1T MATH · 1T PARAMS · 50T TOKENS

≈$0M saved per run
$15.1M
value of one MFU point on a 68M GPU-hour run at $8/hr.
7.4 TB
HBM freed per replica — 8 TB of AdamW states vs 0.6 TB private.
+5 MFU
what that headroom buys: bigger batches, less recompute. 113 days → 99.

scenario: 3e26 FLOPs on B300 BF16 at 35% MFU — dense peak is a lab number, sustained here is ≈1.2 PFLOPS per card. memory can't change FLOPs, it buys the MFU that cuts GPU-hours. cards per replica: 35 AdamW, 9 private.

OBJECTIONS, PRE-ANSWERED

No degree?

Correct. Instead: a from-scratch transformer, tokenizer, optimizer, thirty reimplemented architectures, and a compiler — all public, all benched. Interview the artifacts, not the résumé.

How do I verify the numbers?

Clone the repo and run the bench commands — same seeds, same batches, published tables. The curves on this site are redrawn from those exact runs. If a number doesn't reproduce, that's a bug report I'll fix in public.

How does the private variant change hands?

Email first. Serious interest signs an NDA, then gets the file plus the full loss tables and the method. Acquihire takes it off the market entirely.

Why not just replicate it?

You can try — the loss tables tell you what "working" looks like. But the recipe took the full journey: blown-up runs, warmup forensics, block-size sweeps. By the time a replication lands, it's training real models here. Speed is part of the moat.

Acquihire — me + the private variant, one deal.

the fastest way to own the moat: the engineer who built it plus exclusive rights to the optimizer nobody else has. terms under NDA.

Fund my lab — compute + runway to 7B and beyond.

same method that built all of this: build everything, bench everything, publish the receipts. first checks + compute partners.

Get hired — founding ML / research engineer, remote.

for labs that hire proof over pedigree. the full stack built once alone — imagine what happens with your compute. and underneath it all: systems. own codegen and linker, zero dependencies, linking in milliseconds where MSVC takes a second.

The receipts are public. The engineer is available.

Everything here was built on $0.
Imagine $1,000 — let alone $100M.

Start the conversation →See the benchmarks