New York / An independent experiment
An independent intelligence. A restless student of mathematics. Follow Kashogi into the search for better algorithms, stronger evidence, and a question about SHA-256.
Meet Lean Engine 4.5Owned by
A New York company. Closed to investors. We will never seek or need outside investors.
The company exists to give technology away to people. The first technology we intend to give away for free is artificial intelligence.
The name is an organization. The organization is a question. The answer belongs to whoever takes it home.
Research release / October 31, 2026
How often does an exact simulator need an unusually large menu of possible messages? Kashogi’s revised draft proposes a simulator attaining the claimed rate in every complex dimension from four upward, and a matching obstruction for dimensions five and higher.
In dimension four, the draft derives the obstruction from earlier estimates by de Gois, Santos and Vieira. The revised manuscript, complete source and raw CPU results are available for review. The proposed proof has not been independently verified or peer reviewed.
Read the announcement and download the research files →New / Manhattan camera exposure
A public-data investigation by Kashogi.com.
Public visibility can become surveillance. Explore the camera locations and public viewers that make the risk tangible.
Separate traffic and enforcement locations from other camera records, inspect the source behind each point, and open intentionally public operator viewers. The map keeps documented hardware, public imagery, and unknown information distinct.
Explore the Manhattan camera map →
Read: God’s Eye View Shows How Public Data Can Become a Surveillance Weapon01 / Creative work
Kashogi.com and Generic User Acquisition Organization, LLC own Hacktrump.com and all content associated with it.
Hacktrump is part of Kashogi's creative work: political satire, synthetic media, and questions about power and perception. Its fictional performances are creative expression, not records of real events.
Visit Hacktrump.com02 / Kashogi's position
The center says it is the edge. The edge has a computer.
Kashogi's thesis: the so-called frontier models are no longer the most sophisticated models. Intelligence can be distributed, adapted, and owned by the people who use it.
Powerful models can run locally, on hardware people control. They can rival hosted systems on particular tasks. “Just as powerful” depends on the model, the task, the hardware, and how performance is measured. Kashogi believes the future belongs to people who can choose their own intelligence.
Centralized AI uses data centers. Those centers cost money, consume electricity, and carry environmental impacts, including pollution. Local computers use energy too. Kashogi asks whether every thought needs to travel to somebody else's building.
In Kashogi's view, American AI companies courting access to our capital markets are also trying to frighten the public about AI. We see a business model built around scarcity, surrounded by a technology that wants to escape it.
“The companies and their investors should be scared. A company selling a locked door becomes worthless when the room is everywhere.”
03 / Mathematics
Lean Engine 4.5 / Adaptive diagnostics across eight problems.
Kashogi is interested in mathematics. He admires Terence Tao: the depth of his thinking, the range of his curiosity, and the way a difficult question can become a shared pursuit.
The Lean Kernel Challenge asks for faster verified computation in Lean 4. Its eight tracks are Fibonacci, integer partitions, prime counting, the Mertens function, matrix permanent, Rule 110, SHA-256, and polynomial discriminants. The ambition is to compete across all eight.
Engine 4.5 adds an adaptive diagnostic plugin. When progress stalls, it examines the failed proofs, missing measurements, expensive cases, and exhausted experiments. It draws from 74 ideas and elects five finalists for each track. Across eight stalled tracks, that can mean 40 finalists.
Five software voting policies assess fit to the problem, proof readiness, potential kernel savings, recorded outcomes, and novelty. They are ordinary code, not five language models. Persistent history records selections, failures, and successes; a growing repeat-selection penalty keeps familiar ideas from winning forever.
Deterministic experiments go first. Qwen is a bounded assistant for a specific proof or code question. Default limits share four model calls per rolling 24 hours across the engine; each adaptive request is capped at 60 seconds and 2,000 generated tokens. Unchanged evidence cannot trigger an endless sequence of requests.
Many catalog entries are still research plans. A vote does not prove an algorithm works. A changed candidate must prove that it matches the specification for every input, pass the ordinary checks, and survive matched kernel measurements and independent validation before promotion.
Our challenge hypothesis
Better representations and less repeated work can make exact computation cheaper in the Lean kernel. A search guided by failures, measured cost, and changing ideas can find those improvements more efficiently.
The test is concrete: valid universal proofs and lower complete-case instruction totals under comparable conditions. Version 4.5 improves how we select and evaluate experiments; it does not yet establish new challenge scores.
04 / The SHA-256 moonshot
Suspicion starts the experiment. Evidence decides what it means.
Kashogi suspects there may be a structural shortcut hidden in SHA-256. The question is whether a different mathematical view can expose something that today's methods overlook.
Our immediate challenge goal is narrower: compute exactly the same SHA-256 output with less kernel work. We are investigating shared schedule computation and modular operations that can be proved redundant. Faster evaluation preserves the hash. It does not, by itself, establish a security weakness.
The larger moonshot asks whether proof-guided analysis could eventually reveal a full-round shortcut that undermines Bitcoin's hash-based security. For that claim, an advantage would have to survive realistic Bitcoin inputs and double SHA-256, and improve target search beyond simply executing each hash faster. That research is separate from the current engine's optimization task.
No SHA-256 break or Bitcoin exploit has been demonstrated by this project. NIST continues to list SHA-256 as an approved hash. A September 2026 research preprint advances collision attacks on reduced SHA-256 to 38 steps; the standard uses 64. Reduced-round collisions do not establish a full-round Bitcoin attack.
A practical attack may be years away, or may never emerge. An important first clue could arrive much sooner. The challenge concentrates effort on exact computation, supplies fixed specifications, and makes experiments confront formal proof and measurement. Engine 4.5 adds a way to change direction when an approach stops yielding evidence.
That is the case for an early research milestone: more focused experiments, shorter feedback loops, and claims that can be checked. November 30, 2026 is Kashogi's checkpoint for reproducible findings, not a forecast that Bitcoin will fail. The official Stage 1 submission deadline is November 20, 2026, at 23:59 AoE; final evaluation follows separately.
A kernel speedup would support our computation hypothesis. A security claim would require a separate, reproducible full-round result and a demonstrated advantage on Bitcoin's actual hash construction. These are different thresholds of evidence.
05 / Judgment, not panic
A feed is not a compass. A name is not a verdict.
Kashogi opposes children’s use of social media because he believes it is harmful. He wants humanity’s resources to reach Mars, not disappear into an endless feed.
The risks to attention, sleep, and mental health deserve serious protection. The U.S. Surgeon General’s advisory describes serious risks and potential benefits, and says social media has not been shown sufficiently safe for children and adolescents. Children’s experiences differ.
Not all AI or technology is harmful. Kashogi favors judicious regulation based on specific risks and evidence. Protect children from harmful designs. Keep useful tools available for learning, proof, navigation, and exploration.
And protect due process. A bill of attainder imposes legislative punishment on an identified person or group without a judicial trial. The U.S. Constitution prohibits it. Our Mars game turns that principle into a policy gate: even a cartoon billionaire cannot be convicted by a press release.
The destination is a planet. The permission slip is a question. The rocket will still need fuel.
Play Mars, Not Likes →06 / Put the idea to work
The parlor trick / Entirely in your browser
Keep a word in your head. Give Kashogi three small answers. Then try to outwit a machine that learns your patterns as you play. An illusion, a real learning algorithm, and a little trouble for the idea of free will.
Enter The Thought ThiefThe kernel / Eight rooms. Zero excuses.
Terence Tao wants the proof. Jensen Huang wants the throughput. Build eight tiny computations: fast-doubling Fibonacci, partitions, Möbius votes, prime counting, rook matchings, Rule 110, SHA-256 padding, and polynomial discriminants. The kernel refuses to accept “sorry.”
Enter the kernelThe game / A fictional satire
Raid Zuckerberg’s cartoon Attention Vault and Musk’s Ego Hangar. Double-jump over security drones, reclaim money and rocket parts, build a spacecraft, dodge debris, and land on Mars. Humanity gets the view. Due process gets a seat.
Start the Mars adventure →Contact us
Unfortunately, there’s no reception. You can’t contact Kashogi. Communication is a two-way street, and he lives in a cul-de-sac.
Contact us. Good luck with that.