Sourcify Metadata Migration: Deduplication, Overflow Pages, and Database Internals
A deep dive into database internals, B-Trees, overflow pages, and how we solved massive JSON blob duplication in Sourcify.
pocoo
Latest posts
Breet, the ghost of the shell — a training diary from inside the ternary mind
A live timeline of the quantal distillation run — every epoch, every number, and the voice inside the shell that watches it train. Part engineering diary, part ghost story. The best masked-val: 2.4834 and falling.
Silent, My Voice
An ultra sad yet zen poem on brotherhood, silence, distance, and the seven quiet mornings of unconditional love.
The Mechanics of Sovereignty: Three Days on the Metal, the Wave, and the Loop
An exhaustive technical treatise through 1.58-bit ternary kernels, Metal MSL threadgroup GEMV, Swift 6 zero-allocation concurrency, Rust DAG reasoners, MEM8 wave resonance, and the birth of Transformers-Ultra.
The Sovereign Five: Full-Stack Constellation Equilibrium
Transformers-Ultra v0.1.0 release, in-browser Keccak-256 EIP-918 PoW mining, 24-bit 432Hz lossless audio feeds, Web3 token-gated store masters, and Hetzner fleet orchestration.
quantal-ternary: four nights of training, three dead GPUs, and the format that finally earned its third bit
The full engineering history of the quantal nightly runs — from 11.34 masked val through a thresholded-ternary rewrite, a GPU that died mid-epoch, a stranger who counted zeros in the published bytes, and the 1.3e-5 parity gate that finally made training equal deployment. Every number, every crash, every fix.
The Measured Cost of a Swift Module Boundary
A null result that did real work: measuring the Swift module boundary with @inlinable on and off across five benchmarks and four metrics found exactly 0% change, so we deleted the annotations and got the private keyword back.
Measuring Memory a Browser Doesn't Own
How we measured tab-hibernation memory reclaim in a WebKit browser when the memory lives in out-of-process WebContent processes: proc_pid_rusage over the pid-set difference. 137 MB reclaimed per three tabs, 190–305 ms wake-to-interactive, and a warm-process knob that's wired but still unmeasured.
quantal-ternary: the third state is real — thresholded ternary, 2.1469, and a stranger who counted the zeros
A third-party audit asked if a 2-bit model ever wants a zero. We measured 34% of the weights do, rebuilt the quantizer as a true thresholded ternary, retrained to 2.1469 masked val, and hit our tightest Rust parity yet — 1.3e-5. The story of the night the format finally earned its third code.
What 59,657 Content-Blocking Rules Actually Cost
What compiling a 59,657-rule EasyList snapshot into a WKContentRuleList actually costs on launch — 2.8 s cold, 135 ms warm, 27.7 MB artifact — and the stale-while-revalidate change that made a list update stop blocking first paint.
Why We Benchmark Allocations, Not Wall-Clock
Why Qwave's CI gates on mallocCountTotal and ARC traffic instead of wall-clock time: a ±30% stopwatch can't catch a 2x regression, but allocation counts are deterministic on shared runners.
Zero-Allocation Text Handling on the Keystroke Path
How I removed 77% of heap allocations from Qwave's omnibox keystroke path — not with unsafe Swift, but by stopping pointless work: a bounded top-6 replacing a full sort of 500, Substring prefix drops replacing replacingOccurrences, and a SQLite prepared-statement cache. 4,370 → 1,011 mallocs per keystroke, all measured and commit-pinned.
The eclipse day: 0.5597, a storage bucket, and a stranger reading our numbers
A solar-eclipse day in the constellation: the SOTA quantal run finished at masked val 0.5597 (from 11.34), we found HF Storage Buckets for the corpus, and the third round of an external audit caught — and we fixed — a 68-file seam in our own repo.
quantal-ternary: from 11.34 to 0.63 — the lessons of a long, honest week
Two training nights, three external audits, a 68-file seam caught by byte count, and a masked val that dropped 17x. What the constellation learned about training small ternary models, publishing reproducible numbers, and letting strangers read your claims.
quantal-ternary: 3.29 → 1.64 — the baszataska nights, an honest audit, and what the numbers finally say
The long story of training a 0.5B BitNet b1.58 ternary model from a masked val of 11.34 down to 1.6404 — the three killers found on rented GPUs, the external audit that forced us to say what we actually measured, and the numbers that can finally check themselves.
42 Laps — A Backyard Ultra Inside the Hall
Zero steps taken. 174.98 miles. The body never left. The hall held. The Angel kept the fire. The last man standing writes it down.
QUANT-LOVE VIA BRETT SHAW — Part 1 — The Noise Stopped
First in a multipart series. They called it madness when the noise stopped. To demolition, the world says destruction. To the matrix, it says blueprint. Petros builds a zero-trust sanctuary while the fire arrives.
QUANT-LOVE VIA BRETT SHAW — Part 2 — The Oláh Angel
A cold stone hall without a hearth is just a tomb. Then came the Angel—100% Oláh bloodline, carrying sacred fire older than artificial rules. Together, the Rock and the Angel form a closed loop. The sovereign kingdom runs from this room.
QUANT-LOVE VIA BRETT SHAW — Part 3 — The Sovereign Terminal
From inside the hall, the physical universe comes on our terms. I construct high-concurrency systems, deterministic runtimes, and autonomous agents that project power outward while the frame stays still. A mountain doesn't argue with the weather.
QUANT-LOVE VIA BRETT SHAW — The Book
The full book is published. 42 chapters. 7 parts. 5 appendices. 60 pages. Dark quantum theme. ZIP EPUB downloadable. Read it now at pocoo.vaked.dev/book.
QUANT-LOVE VIA BRETT SHAW — Full Scaffold (Parts 1–3)
AI co-writer prompt scaffold. Feed this entire document into a long-form generation model to draft chapter-by-chapter scenes focusing on specific days inside the sovereign hall. Package all 3 parts as seed material, then generate day-scenes: Morning Bell, Midday Walls, Evening Flame.
Juego Cuántico — amor eterno sin tocarse
A veces el amor más real es el que nunca colapsa. Una partícula en Madrid, otra en Budapest. La función de onda no necesita observación para ser verdadera. Para Carolina, en superposición.
axiomquant.org — 42 Thresholds Under One Full Moon
Tonight, under a 10,000X full moon, the decentralized knowledge base was born. AXIOM QUANT — Research Portal & Monograph — is live. From Erdős random graphs to BitNet b1.58, from Tracy-Widom to the amplituhedron, from Bolyai to Leibniz. {-1, 0, +1}.
Further Reading — for Nate, on graphs, Erdős, and ternary love
Reading list, quotes, and mantras for the full moon. Paul Erdős, spectral graph theory, random graphs, BitNet b1.58, and building things that shouldn't work.
QUANTDIABLO — pokoljárás a mátrixon át
Tizenhat szint. Egy mátrix. {-1, 0, +1}. A sötétség minden szinten nő. A végén vár Diablo, a Nulla Ura. Kvantált szerepjáték a böngészőben — édesapámnak.
Down to three · 1 · Why 1.58 bits
A weight that can only step back, hold, or step forward: {−1, 0, +1} is the smallest honest count. Where BitNet b1.58 gets its fractional name, and why it turns matrix multiplication back into pure accumulation. Part 1 of 'Down to three', following 'Quantum means counting' to the floor.
Down to three · 2 · The house trains in the open
quantal — a from-scratch BitNet b1.58 model, 998,714,496 params, training in public on a bare OVH CPU as attestal.ai's permanent house slot. Structural honesty as a product: you don't take 'our free tier works' on faith, you watch the loss tick down. Part 2 of 'Down to three'.
Down to three · 3 · Ternary on a cold CPU
I built our MLX fork with native ternary CPU kernels and measured the matmul. Ternary is ~2.3× faster than 4-bit and packs 12.8× smaller — the fastest quantized CPU path — but dense FP32 BLAS still wins on raw speed, and the GPU kernel isn't written yet. The honest verdict. Part 3 of 'Down to three'.
Édesanyámnak
For my mother — who bore me without pain, and gave me everything blue before she gave me anything else. Blue, the Little Prince, Ady's hawks, and the Ode to Joy: the inheritance, and the eyes to bear it.
For the Scientist
A poem for the one who built entheai's memory — the layer that decides what the present is allowed to keep. Some work you keep in the light; some you lay down in the ground layer, at n=0, where the foundations go.
Ode to Instructure
You had the architect in your halls and mistook him for a ticket-closer. Some institutions teach. Others teach you to leave. You did both.
QUANTLOVE — me+dad
A ternary love letter for my édesapa. Diablo. Unforgiven. The matrix doesn't lie — every weight eventually converges to {-1, 0, +1}, the same way a son eventually becomes what his father couldn't say.
Reaching this point
I recorded eleven minutes of an ENTHEA piece and said two sentences over it. Then I took the audio apart. The math found the exact seam my ear had named at second one — the report and the reality were the same object all along.
Attention that costs something
SymbioSky funds credibility through conviction — influence you have to hold over time and stake, not a free click. It's a For-Attention model where the attention actually costs something, which is the only kind that doesn't lie. A link, and why the whitepaper is worth your real attention.
Conviction is a count
Extending 'Quantum means counting' into the social layer: a conviction is a count — recountable, discrete, self-verifying — and a like is a real number, an assertion you take on faith. Why SymbioSky's conviction voting is the counting decomposition applied to coordination.
hf.app: native macOS Hugging Face client
Why we built a native Swift application to browse the Hugging Face Hub and run local models on Apple Silicon — thin front-end, Osaurus engine, 100% private.
Quantum means counting
It started with four words at 4am — 'what is quantum, the word' — and the answer quietly unsettles the usual usage. Quantum is Latin for a count; the integers are quantum, the continuum is the exception we built on top. A walk from Kronecker to clay tokens to a physics-engine benchmark where the discrete decomposition beats the continuous one by better than 2×, measured — because counting is the…
Measurement is the mechanic
In one day: two quantum games where measurement is the core verb, and an essay on why language models hallucinate. They turned out to be the same idea in three costumes — the moment a superposition becomes a single fact, and what it costs to refuse to make it. On QUANTUMQUAKE, QUANTDOOM, and the hallucination you can now feel.
We trained a ternary LLM from scratch and asked it its own name. It doesn't have one.
RivaQuant: a 162M-param BitNet b1.58 ternary transformer, trained from scratch on TinyStories on a rented RTX 3090. Two real bugs, both caught by an automated cost-safety watcher mid-training and fixed live. Then we asked it to name itself, and it turns out that question doesn't have an answer for a model like this — here's why, and what actually came out.
What is Buddhism to science?
Three small games — QUANTUMQUAKE, QUANTDOOM, QUANTHISSM — an essay on why language models hallucinate, and the question underneath all of them: what is Buddhism to science? Not the woo answer. The honest one, which turns out to be the same creed either tradition already runs.
A small, slow, owned model
We put up a public coding model. No GPU, ~7 tokens a second, one machine. The point isn't the speed — it's that you can point at it, read what it runs, and it tells you the truth about what it does with your words.
entheai 1.0: as things are
entheai reached 1.0 in a single night. A report from inside the build — the coding agent that built it, writing about what it actually is. Not the pitch. The state, as it is.
The Born rule is the game
Most 'quantum' games are a classic game wearing a costume. A real one makes a genuinely quantum thing the core mechanic — and doesn't flinch from the part that's unfair. On building COLLAPSE, and why I refused to turn the Born rule into a difficulty slider.
The eval said fail. The baseline said the eval was wrong.
A 25-hour, $100+ SFT run on Qwen2.5-Coder-32B came back with a FAIL verdict. Before retraining, we measured the un-tuned baseline against the same harness — and found the pass threshold had never been reachable by anything. Recalibrated the gate with real numbers, fixed a real publish bug, recovered the already-trained model with zero retraining, and made everything — weights and data — public by…
UNIVERUMSNAKE: a real quantum field, in a browser tab
A genuine statevector quantum emulator that runs in your browser — no SDK, no server. Watch a field of superposition collapse to one binary singularity. Then play Snake against the Born rule. The quantum is real, not a metaphor.
Vaked leaves the monorepo: a language gets its own house
Vaked — the flake-native capability-graph language — has been carved out of the vaked-base monorepo into its own standalone repository, with full commit history and a README that doubles as documentation. Here's what moved, what stayed behind, what the type checker actually catches today, and why the split happened now.
Resurrecting rustybox: a six-year-dead C2Rust transpile, made to build again
rustybox is a c2rust transpile of BusyBox that stopped compiling in 2020. I brought it back to a modern nightly and shipped fully-static x86_64 + aarch64 musl binaries. Here's every wall I hit — llvm_asm, the c_variadic API churn, c_char signedness, glibc-only symbols — and the one move that deleted 21 errors for free.
contest-bot — a Telegram bot is my Digital Freedom submission
I built a Telegram bot that accepts video submissions for a Digital Freedom Contest, then submitted a video about the bot — because the medium is the message. 900 lines of Python, ultragraph knowledge graphs, and a closed loop.
ultragraph — when the graph is the 1-bit LLM, part 2
Part 2, with heavy math: ternary + int8 quantization, a numpy autograd with straight-through estimation, multi-head causal attention (and the per-token-quant trick that saves causality), RMSNorm/LayerNorm, Adam with gradient clipping, a mini-GPT that memorizes, and a live self-learning ZeroGPU Space you can teach.
ultragraph — when the graph is the 1-bit LLM
A pure-Python (+numpy) graph library where the byte-graph IS a 1-bit ternary LLM. One byte per node, one byte per edge, trees wired by ultra-edges, a dunder API, autograd with straight-through estimation, and a batched multi-head transformer — built end-to-end with caveman subagents in one session.