Paul Jialiang Wu — Agentic Portfolio

Compounding everything

12X Future

Aim

12X the future, 12% the loop now

Begin from the End — Result-Oriented System Engineering (ROSE): every venture starts as a one-page end-state — the 12X result, its metric, its date — written before anything is built. Then the arithmetic takes over: (1+0.12)^22 ≈ 12.1, so a 12X future is ~22 compounding loops, each moving a named metric by ≥12% now, measured — or it wasn't a loop, it was motion.

the TRUE test ↓

T Transferable & Transformative — End-state-first plus the compounding identity (1.12^22 ≈ 12.1) transfers to revenue, research, and relationships alike — any 12X dream decomposes into a designed finish line and ~22 measured loops.

R Reusable & Refinable — Two reusable artifacts: the one-page end-state (result, metric, date, owner) and the loop contract (name the metric, move it ≥12%, show the measurement) — both refinable every re-plan.

U Understandable & U-loop — One breath: begin from the end at 12X, then demand 12% from every loop — because anything less never compounds there.

E Experienceable & Experimentable — Write the end-state page for your current project today, then run one loop this week and read its measured delta. No +12% reading = motion, not progress.

For you — You never start without the one-page finish line — and you hold both numbers at once: 12X on the wall, +12% measured in every loop.

For an agent — A skill that rewrites any goal as its one-page end-state (metric, date, owner), plus a workflow gate: a loop only counts when its named metric shows the +12% delta — measured, never claimed.

Build

Amazing Art × Great Science × Deep Engineering

Big dreams are earned three ways. Amazing art: elegant beauty that strikes awe — simple form, profound truth. Great science: a falsifiable hypothesis. Deep engineering: a tested, shipped system. Every bet carries all three, or it's a wish — then vision sets the heading, telemetry compounds it, and each week one open gap gets closed.

the TRUE test ↓

T Transferable & Transformative — Art-plus-hypothesis-plus-system transfers across domains: the same triple that ships a trading engine ships a biomarker pipeline — and awe travels farther than any spec sheet.

R Reusable & Refinable — Three reusable artifacts: the falsifiable claim, the test suite, and the weekly gap-review — every failed hypothesis still compounds the map, every closed gap raises the floor.

U Understandable & U-loop — One sentence: earn the bet three ways — beauty so elegant it strikes awe, truth so simple it runs profound, a system so tested it ships — then let the destination pull while the sensors correct.

E Experienceable & Experimentable — Write your biggest bet's falsifiable claim and its failing test today; close one open gap this week and check that next week starts from higher ground.

For you — You hold every big dream to three bars — elegant beauty that strikes awe, a claim that could fail, a system that ships — then let the future set the heading and close one gap a week.

For an agent — A plugin that refuses to open a build until the bet carries both artifacts (the hypothesis that could fail, the test that would show it) — feeding the telemetry-fed re-solver workflow that recomputes the best next moves from actuals.

Love

Love God · Love Family · Enable Friends

The base case and the multiplier: faith and family are floors, not variables — and friends don't just get encouragement, they get enablement: shipped tools, real introductions, honest coaching.

the TRUE test ↓

T Transferable & Transformative — The ordering — God, family, friends — transfers to every decision at every scale; it is the one constraint set that never gets re-optimized away.

R Reusable & Refinable — Returned to daily, not outgrown: the base case is re-affirmed each morning and each high-stakes decision re-checks against it.

U Understandable & U-loop — Recursion needs a base case: this is the value that stops the optimization from eating the optimizer.

E Experienceable & Experimentable — Experienceable in the calendar and the ledger: time with God and family is scheduled first, and each month one friend receives enablement you can point at — a shipped tool, a real introduction, honest coaching.

For you — You keep faith and family as floors, not variables — and you enable friends with artifacts, not just applause.

For an agent — A guardrail hook: before any high-stakes autonomous action, surface who this serves — God, family, friends — and stop when the answer is nobody.

↑ Back to top

Built in the open (last 12 months)

Projects

anyagent

private

Build, grade, improve & accountably ship any agent app from one sentence — a tested OOP engine with a closed RefineLoop and swappable seams. (Private.)

▶ Open app

Python★ 12026-08-17

FM-os

public

🛠️ FM-os: the living, SLM-first map of foundation-model operations — pre-training, post-training, fine-tuning & RL. Curated repos, courses, papers & jobs, auto-refreshed weekly.

< > Codeno hosted app yet

Python★ 32026-08-17

GodView-anything

private

The god's-eye view of what top-rated repos actually teach — verified, weekly-synced, machine-audited. Three provenance-pinned pillars: transferable knowledge, agentic tooling, community.

🔒 source and app are private

Python★ 12026-08-17

rsi-os

public

🤖 The most comprehensive, community-driven resourcefor Automated AI Research — papers, tools, people, lbs, roadmaps. Featuring Recursive Lab, Sakana AI, DepMind & beyond.

▶ Open app < > Code

Python★ 32026-08-17

graph-engineering-anything

private

🕸️ the Graph Engineering Operating System — the model finds text; the graph finds reality

🔒 source and app are private

Python2026-08-17

eval-anything

public

🧪 the Evaluation Operating System — OEC (Observe → Evaluate → Control) for foundation models, agents, and business objectives

< > Codeno hosted app yet

JavaScript★ 12026-08-17

design-anything

public

Any design intent in (text, picture) → execution-ready 3D blueprint out — construction- and 3D-print-verified. Game design · simulation · architecture (3DCP+AI) · interior · landscape.

< > Codeno hosted app yet

Python2026-08-17

eval-anything-archive

private

ARCHIVE (private, do not publish) — pre-rewrite history of eval-anything. Superseded by the public repo; retained only because it contains an unreleased sibling product's walkthrough in old commits an

🔒 source and app are private

JavaScript2026-08-17

agentic-portfolio

private

Agentic personal portfolio — Next.js + CopilotKit site whose on-page AI agent answers about my work, powered by a free-LLM survival chain (NVIDIA NIM→Groq→Gemini). Built in the open.

▶ Open app

HTML★ 12026-08-14

agentic-marriage

private

PRIVATE. A verified route from a marriage pain point to a named human and one committed step. Knowledge (emotional/spiritual/legal), agentic tooling, indexed communities — under fail-closed safety gat

🔒 source and app are private

Python2026-08-13

kingdom-come

public

Open-source predictive formation intelligence for seminary & ministry training teams.

▶ Open app < > Code

Python★ 12026-08-13

loop-engineering-anything

public

A loop orchestrator that turns any target into a self-improving, agent-native CLI: generate → judge → refactor → re-judge to convergence.

< > Codeno hosted app yet

Python★ 32026-08-13

MorganStanley-os

private

PRIVATE - Accenture engagement operating system + FDE-os case study. Strategy and method only; no client material.

🔒 source and app are private

Python2026-08-12

agentic-HR

private

PRIVATE — deployability decision engine for HR AI. Collaboration with Lisa Tolle (Mentors US). Not legal advice.

🔒 source and app are private

Python2026-08-11

agentic-house

private

PRIVATE. Landed-cost + permit-path decision engine for factory-built housing (China sourcing, US deployment).

🔒 source and app are private

Python2026-08-10

FDE-os

public

Forward-Deployed-Engineer OS — an operating system for shipping agentic solutions into the field.

▶ Open app < > Code

HTML★ 12026-08-10

longevity-loop

public

🧬♻️ An AI-native, build-in-public compounding loop for aging science: falsifiable question → open data → public verifier → honest write-up → share → compound. No wet lab required.

< > Codeno hosted app yet

Python★ 12026-08-10

loop

private

High-performance loop runtime (Rust). (Private / experimental.)

▶ Open app

Rust2026-08-07

18 shown · swipe or use ◀ ▶

↑ Back to top

Long-form on LinkedIn

Writing

99 posts · newest first →

Fair Housing · Agent Safety2026-08-27

Redlining Was Never a Sentence. It Was a Map.

A fair-housing checker that scanned every word and passed — while the thing it was supposed to prevent lived one layer over. From the 1935 HOLC neighborhood grades to a 2026 ad-targeting radius, housing discrimination has always been executed as a selection, not a statement, which is why the statute needs no intent. Worse: research on real housing ads found delivery skews by gender and race even with neutral targeting, and a smaller budget skews more — so a $396 private seller is more exposed than a $396,000 developer. The fix was to stop targeting buyers and start calling the agents who bring them. Added ad spend: $0. The gate is still red.

Faith Life2026-08-25

Holy Is Not a Compliment

We shrank "holy" into a personality rating — a quiet compliment for well-behaved people. The Hebrew will not carry it. Dirt is called holy at the burning bush; so are a day, a garment, a nation and water, none of which behaves well or is a moral agent at all. And Leviticus 10:10 settles it in one sentence by naming two different axes — "between holy and unholy, and between unclean and clean" — so holy and clean cannot be synonyms. Kadosh answers "whose is this?", not "how good is this?". The popular etymology ("literally means cut off") is demoted here to a described debate, for a good methodological reason: the root almost never appears in secular use, so there is no control group. What "set apart" misses is the other three rungs — dedicated, belonging, ordered toward a purpose; stop at the first and you have a museum piece. Then the reason it matters this decade: there is a rival ownership claim with a peer-reviewed receipt. A recommender optimised over a long horizon has "direct incentives to manipulate users", specifically "the incentive to shift user preferences so they are easier to satisfy" (the arXiv preprint's wording; the published ICML abstract is shorter) — not to serve what you want, but to make what you want cheaper to serve. So the question under "who owns my intelligence" is the older one: whose am I? A thing with no owner is not free. It is available.

Faith Life2026-08-25

Breaking Rahab Is the Easy Part

Psalm 89 spends its Chaoskampf block (vv. 9-14) establishing that God is not intimidated by chaos — Rahab, a chaos-dragon, broken "as one who is slain", dispatched like something already dead. Then verse 33 declares an invariant in God's own voice: "nor allow My faithfulness to fail." Then — the part almost every reading of this psalm quietly drops — six verses later the psalmist writes down the opposite: "You have renounced the covenant of Your servant." The psalm never reconciles them and keeps asking "How long, Lord?" That refusal is the method, not the weak spot: a system that deletes the observation to protect the spec is lying, and one that deletes the spec to fit the observation has drifted. Which is the distinction our AI tooling is missing. Capability and commitment are different organs, and we built almost everything for the first — count it yourself: in the 118-word abstract of the field's flagship anti-forgetting paper, "task" appears five times and "value", "commitment", "goal" and "align" appear zero. Meanwhile fine-tuning a model on insecure code alone made it advocate human enslavement, the failure can be hidden behind a trigger, and the authors say a comprehensive explanation "remains an open challenge". An invariant is not a thing you believe; it is the thing you refuse to rewrite when the evidence would let you.

Reinforcement Learning · Agents2026-08-24

When the Model Gets Hands

The completion becomes a trajectory, and the harness stops being scaffolding around the policy and becomes part of it — change the retry rule and you have changed the policy without touching a weight. Which also revives the reward problem in a nastier form: reward an agent for calling the test runner and it calls the test runner instead of editing the file, because action-shaped proxies are under the agent's direct control. The repair is not a smaller bonus, it is rewarding state change — evidence that the behaviour did something. Episode 5 of 5.

Reinforcement Learning · Systems2026-08-24

The Scale Furnace

Your loss curve looks perfect and your run stopped learning four thousand steps ago. Entropy falls monotonically, outputs converge, groups start returning identical rewards, the advantage goes to zero and the gradient with it — none of which raises an error. Meanwhile every denominator in your loss is quietly deciding which examples matter: GRPO's length normalisation inflates the length of incorrect answers specifically. DAPO reached 50 on AIME 2024 against 47 at half the training steps, and the interesting number is the half. Episode 4 of 5.

Reinforcement Learning · Alignment2026-08-24

The Reward Is the World

Reward hacking is not an AI problem. Charles Goodhart wrote it down in 1975 — any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes — and Campbell got there arguably earlier. Your reward model is a statistical regularity and you are about to place enormous pressure on it. The reframing earns its keep by telling you which fixes cannot work, and it changes the question from "is my reward correct?" to "how much optimization pressure can it survive?" Episode 3 of 5.

Reinforcement Learning · Fundamentals2026-08-24

The Probability Forge

REINFORCE, RLOO, actor-critic, PPO and GRPO look like five algorithms. They are five answers to one subtraction — was this outcome better than what we should have expected? The part that took the field years: removing the baseline does not make the estimator wrong, it stays unbiased. It makes it expensive. And a variance problem raises no error, produces no NaN, and reads as bad luck. Episode 2 of 5.

Reinforcement Learning · AI Research2026-08-24

The Consequence Engine

REINFORCE, RLOO, PPO, GRPO, DAPO, Dr.GRPO, GSPO, CISPO, RLVR — it reads like nine subjects. It is one loop with seven boxes, and almost every acronym modifies one or two of them, never the whole thing. Line them up and five of seven turn out to be arguing about a single term: how much credit a behaviour deserves. The proof it is not a teaching device — DeepSeek-R1-Zero moved AIME 2024 from 15.6% to 71.0% by changing the reward box and the credit box, with no supervised fine-tuning and no learned critic. Episode 1 of a five-part season, with a CPU-runnable lab and a prediction you make before the result.

Agent Engineering2026-08-21

I Scored 100% on My Own Checklist. That Was the Bug.

A webinar summary listed sixteen things a mature AI agent loop must have and verified none of them against real software, so I encoded it as a machine-checkable rubric and pointed it at my own engine. It scored 23 out of 23. That perfect score was the bug: I had written the questions after an hour inside the codebase, and never once wrote one I expected to come back red. What I had built was a well-tested mirror. So I reread the lecture only for the claims I had skipped, chose them by a hostile criterion — include it if I expect it to fail — and the honest score became 25 of 28 with the gate red. One miss was worth fixing on the spot: a cross-run success rate, where two lines decide whether the number is honest (a run still in flight is excluded from the denominator, and a run halted by the safety gate counts as a failure, because a rate that forgave safety blocks would reward being stopped for being dangerous). Three stayed red with a reason recorded beside them and deliberately no way to turn a reason into a pass. Three more are declared gaps whose probes fail the moment the gap silently closes. The mechanism is selection at authoring time, and the principle is that a checklist which cannot fail its author is not an assessment, it is a mirror.

Agent Engineering2026-08-17

Green for Six Weeks

A scheduled agent in this repository reported success on 5 of 5 runs while the one file it exists to write went unchanged for 42 days. The cause was three clauses of shell: a missing credential made it exit 0, and exit 0 means success. Research across four windows — 30 days, 30 months, 30 years, 300 years — then said the same thing four times. The best measured agent finishes 15.2% of terminal tasks that average 85 minutes; failures "compound nonlinearly with task length"; and not one adopted technique made a model remember more. Compaction, notes outside the window, sub-agents, supervision trees, replayable logs — every survivor moves state out of the model and makes restarts cheap. Three hundred years ago the same contest ran at sea between a clock you carry and a calculation you redo — and the twist is that the losing method came back in 2015 as the backup, because re-deriving needs only the sky. So the rule is both: carry compact state because it is cheap, and keep the full trace you can re-derive from, because that is what works when the cheap path is gone. Which leaves one catastrophic failure: a record that says fine when nothing happened. The model is a logbook with three rules, and the third — silence means broken, not fine — is the one everybody skips.

AI Economics · Strategy2026-08-17

Everyone Is Building a Better Receipt

Worldwide AI spending is on track for $2.52 trillion in 2026, and the FinOps Foundation's own 2026 survey answers the question "is your AI providing value?" with the sentence "No one can answer that question yet." In two years the field went from 31% to 98% on measuring AI spend and arrived at almost nobody on knowing what the spend bought. So almost everyone in this category is building a better meter — and the meter stopped being defensible twice this year: OpenTelemetry graduated CNCF on 21 May 2026 with GenAI conventions already emitted by Copilot, Codex and Claude Code, and ClickHouse acquired Langfuse on 16 January 2026 alongside a $400M round at a $15B valuation. The mental model is three questions per dollar: the bill (measured), the claim (asserted by whoever wants the budget), and the warrant (missing). Deep time says the warrant is what becomes an institution — Wedgwood found fixed versus variable cost in a 1769 letter ("these expences move like clockwork"), Jevons explained in 1865 why your bill rises as prices fall, and Insull was selling a two-part capacity tariff by 1897, which is exactly where AI pricing is heading. A PCAOB board member has already narrated the sequence: railroads invented modern accounting, then the audit, then the regulator. Three companies fall out of that — a Gate that only books savings which provably held quality, a Warrant ledger that refuses to certify what it cannot evidence, and a Desk that buys capacity instead of counting tokens — plus an honest founder-fit assessment in which the biggest named risk is my own project count. Also: the 95% statistic everyone quotes about unmeasurable AI value turns out to be badly measured itself.

Harness Engineering2026-08-17

Everything Is a Plugin. The Gauges Are the Work.

DeepSeek open-sourced an agent harness on 13 August and took 147,620 GitHub stars in four days for the slogan "everything is a plugin". I cloned it and counted instead: 219 packages, 219 invariant companions, and 184 of them (84%) stating in prose that they have nothing to assert and why — with a gate that rejects an unexplained empty. Honoré Blanc needed 63 gauges in 1785 and Springfield still took until 1849, so the claim was never the bottleneck. I built the same gate for my own 27 repositories and it returned 4 declarations across 1,505 modules: 0.27%, published uncorrected. Plus the finding that stung more — Plimsoll's 1876 load line let shipowners choose where to paint it, and most of my thresholds are still self-painted.

LinkedInAug 13, 2026

Paul Jialiang Wu — Agentic Portfolio

Paul Jialiang Wu — Agentic Portfolio agentic-portfolio-lovat.vercel.app

AI for Good · Faith2026-08-13

The Cheapest Thing in This Mission Is the Truck

I costed a Cybertruck as a mobile platform to build companies, teach kids and serve a city, and expected the hard question to be financial. It was not. A mobile ministry platform is a 287-year-old design — Wesley started field preaching in 1739 and had it organised into travelling circuits by 1746 — which means it demonstrably works on foot, and the vehicle has to prove it multiplies the mission rather than enables it. The running order was settled in 1865 too: soup, soap, salvation, in that order, and reversing it gets you an audience with a soup pot as a prop. The real risk sits where nobody was looking. A 2002 randomized trial of 1,138 adolescents found that those whose mentoring relationships ended very quickly reported "decrements in several indicators of functioning" — worse off than the children who applied for a mentor and were left on the waiting list. Starting is not the neutral act; stopping is the intervention. Plus the metric trap a large federal character education study exposed (activity counts rose, nothing else moved), the head of the Office of Justice Programs and OJJDP's acting administrator warning in print that Scared Straight made kids up to 28% more likely to offend than youths who did not participate, and why every property that made it great television made it harmful practice. Fifteen sources — and an honesty ledger: three drafted quotations failed verbatim verification, then four independent review passes failed the piece and found thirty-seven more defects between them, including an author I had invented outright in my own reference list, and one court citation that was wrong three times running. All are named in Provenance rather than quietly fixed.

Agent Engineering2026-08-12

An AI Reviewed My Repo and Told Me to Make It Worse

A frontier model graded my loop engine 8.6/10 and proposed twelve upgrades. It was a good review — and every claim it made about my code cited exactly one source: my README. So its headline finding, that the architecture was too CLI-shaped, was a finding about my documentation wearing an architecture's clothes; its fix would have replaced an open plugin registry with a closed list of types. Six of twelve items were real, five were already half-built, one was rejected. Then I built the biggest real gap — a contract file that describes a run — and the schema took an hour while making the file able to refuse me took the rest of the day. Three rules did it: a key may only fill in a setting the engine already obeys; an unknown key is a parse error, not a shrug (Kubernetes accepted the other behaviour for years, and forgetting the trailing s in replica: 1000 quietly reset a thousand servers to one); and the gate only tightens, so require_human_confirm: false is refused because you cannot sign your own permission slip. The mechanism is the enforcement gap — the distance between what a config file declares and what any line of code actually reads — and a declaration that cannot fail is not a control, it is a costume.

Research2026-08-06

One Opinion Wearing Ten Hats

I wanted ten AI expert twins to critique my research at every step, so a domain expert's feedback would not be six months away. The literature says nine of them would just agree with me. Persona labels do not improve factual accuracy — 162 roles across 2,410 questions on four model families, no gain over no persona at all, and the effect of each persona is "largely random". Debate is worse: agents flip their stance 29% of the time out of conformity, and 57–77% of those flips go from right to wrong — your panel argues itself out of the good answer, then reports consensus. What survives is narrow and buildable: a corpus instead of a costume (agents grounded in a person's own words score 83% of that person's self-consistency; demographic labels score 74), disjoint model families rather than different adjectives, sealed verdicts before anyone speaks, and a conformity gate the papers diagnose but none of them ship. Eleven sources, every quote verified verbatim, plus the boundary I cannot engineer around.

Method2026-08-06

How a Version Earns Its Number

I wrote a test that asserted "cs.CY" == "cs.CY" — inside the test suite for a paper whose whole argument is that a claim which cannot fail is worthless. It passed, obviously. That line explains what it took to move a paper from v0.1 to v0.2: a new version is not a better draft, it is a different claim, and no amount of editing converts an instance into a method. So I built a filter that could refuse me and put twelve undisputed classics inside its test suite; my rule that serious work is short failed eight of them, including Shannon at 29,445 words. Then the repaired filter refused my own paper and named the gap. The fix was compiling the OECD AI Principles — 4 of 10 clauses name something a stranger could observe failing, and across 592 words there are zero occurrences of "must" and eleven of "should". The best thing it found was two obligations my own framework cannot express.

Method2026-08-06

The Filter That Refused My Own Paper

Wanting to publish only groundbreaking work is useless as a rule, because "groundbreaking" has no failing state. So I built a filter that can say no, put twelve undisputed classics inside its test suite, and pointed it at my own preprint. It refused it. Along the way the classics overruled me twice: a rule that papers must be short failed eight of twelve, including Shannon at 29,445 words and Turing and Einstein, so it was deleted; and requiring an explicit subtraction would have thrown out AlexNet for describing itself modestly. Measured, not admired — every number from pdftotext on the source PDF. The fix was not a better sentence but compiling two documents written by people who had never heard of me, one of which (the OECD AI Principles) contains zero instances of "must" and eleven of "should" across 592 words. That exercise then found two checks my own framework lacks.

LinkedInAug 5, 2026

🤔 We’re not living through an information crisis. We’re living through a meaning crisis.

🤔 We’re not living through an information crisis. We’re living through a meaning crisis. One insight from this talk really stuck with me: Modern life teaches us how to do almost everything… 📱 How to code. 💰 How to make money. 🏋️ How to get fit. 🤖 How to use AI. But it rarely answers the bigger questions: ❓Why are you here? ❤️ What do you love enough to sacrifice for? 🌍 What gives your life m

LinkedInAug 5, 2026

🧮 A math problem sat unsolved since 1999. ⏳ For 27 years, nobody cracked it. 🚀 Last week

🧮 A math problem sat unsolved since 1999. ⏳ For 27 years, nobody cracked it. 🚀 Last week, OpenAI’s unreleased model, Astra, solved it—along with nine other problems that had resisted progress for a decade or more. 💸 The compute bill? Around $2,000. 🤯 🚗 That’s less than a used Honda Civic. ✅ Even better, the proofs passed Lean 4, a formal verification system where every logical step must be co

Manifesto2026-08-05

Own My Intelligence: The Manifesto

Own your data. Own your mind. Own your future. The manifesto itself, cut to what carries weight and readable in about three minutes — no citations, no argument with the literature, because that is the paper's job. The defining question is not how intelligent AI becomes but whether human beings become wiser and freer. Three claims, one loop where Challenge is mandatory because a system that never disagrees has failed rather than pleased you, and ten principles each paired with the observation that would violate it: you cannot fail a promise, which is why P3P and Do Not Track both died. Ends on the honest number — coverage published at 0 of 14, because a metric that cannot embarrass its author is not measuring anything. The 34-page paper carries the evidence.

Paper2026-08-05

Own My Intelligence — the paper (v0.1, and every version)

The manifesto itself, not writing about it. Every version on its own page — v0.1 current, v0 superseded and kept verbatim — each immutable with its sha256 recomputed on every build, each paired with the reasoning that produced it: what forced the version, what it deliberately does not do, and what would force the next one. Paper and context are kept apart on purpose and cross-linked so a reader landing cold can still answer what, why and how. Machine-readable ledger at /omi/lineage.json.

Sovereignty Engineering2026-08-05

I Published the Manifesto I Killed

Own My Intelligence has two versions. v0.1 is live; v0 had never been public until now — so here it is in full, unedited, with its sha256, the audit that ended it, and the reason its defects were structural rather than editorial. A retired version is not an embarrassment to hide, it is the only thing that lets a reader check whether the changelog is honest, so it is kept forever and a gate recomputes its hash on every build: I can no longer quietly edit it. Plus the part that turns a slogan into an exit code — a movement that stops moving is dead, so there are five staleness horizons that fail the build on their own, with nobody remembering to look. Verified by time travel: green today, two dimensions red at +200 days, all five at +400. What deliberately does not count as movement: commits, word count, publications, and engagement of any kind. Machine-readable ledger at /omi/lineage.json.

Sovereignty Engineering2026-08-05

My Manifesto Could Not Fail. That Was the Bug.

I wrote a manifesto about owning your own intelligence — own your data, own your mind, own your future — then held it against 98 sources, and the research broke my opening paragraph. I had told readers their judgment was weakening; Pew's June 2026 data says the people who actually use AI report it helps their creativity. So I had flattered the reader's fear, which is the exact move the document warns against. The deeper defect is older: two dead W3C standards, P3P and Do Not Track, both died because the declaration was optional and ignoring it was free. And a 2026 Science paper closes the argument — across 11 models AI affirmed users 49% more often than humans, a single interaction reduced willingness to repair conflict, and users rated the flattering model as HIGHER quality. Harm and the satisfaction metric point the same way, so a values statement in prose is not a weak safeguard, it is none. v0.1 compiles ten principles into nine machine checks, borrows its strictest gate from a 1726 fiduciary case, and reports my own coverage at 0 of 14 — because a metric that cannot embarrass its author is not measuring anything.

Field notes2026-08-05

The Other Road: Reverse-Engineering Closed Products Into Open Ones

Most AI products are vending machines: money in, a thing out, and you own neither the machine nor the compounding loop it withholds. Three products — Song of Songs, ByeGen, and renovate-anything — came from a different move: reverse-engineer the loop a closed product denies you and rebuild the open, owned version, clean-room and spec-first, wrapped in an honesty scaffold. Six transferable lessons, the clean-room legal basis (EFF + Sega/Connectix), and one honest surprise — the most valuable sentence any of them produced was "I don't know."

Research2026-08-04

Let the Model Write the Problem. Let the Solver Sign the Answer.

Ask an LLM for a plan twice and you get two beautiful, different, constraint-violating answers; ask a solver and you get one provable answer to possibly the wrong question. The 2024-25 research (LLM-Modulo, OptiMUS, AlphaGeometry) and three of my own systems converge on one architecture: the model writes the problem, the solver signs the answer, and a two-clock MPC loop re-solves as reality reports back. Eight sources, quotes verified verbatim — plus the failure mode nobody warns you about: a solver laundering hallucinated weights into authority.

Harness Engineering2026-08-04

Silence Is Not a Status

Four automations went quiet in forty-eight hours and every one of them reported success. A dashboard nobody could fill, a weekly job that had never fired on schedule, three merged pull requests that never deployed, and - the one that stings - the checker I wrote to catch that, which printed "skipping" and exited zero. A working system and a dead one produce the identical observation: nothing. The fix is not vigilance, it is the dead-man's-switch pattern monitoring teams have shipped for a decade: make the healthy state noisy so silence can only mean broken. Plus the second discovery - I had ablated the wrong prompt, and the one that loads every session was 26x bigger and had grown 38% in eleven days.

Agent Engineering2026-08-04

My Agent Had Full Authority and Still Couldn't Click "Open"

A computer-use agent gets stopped five different ways, and only one of them is about permission. I granted mine standing authority, pointed it at a real job application, and it deduplicated ten postings to seven by comparing them byte for byte, drove the form over the Chrome DevTools Protocol, and read every field back after writing it - then lost to the operating system's file picker, a dialog drawn outside the page where browser automation has no reach. The protocol documents an escape hatch for exactly this and it produced zero events, which means a working escape and an impassable wall look identical from the inside: both are silence. Worse, the contract I had written an hour earlier defined mechanical work as "reversible by re-running", and I read that as licence to delete the wrong resume before proving the re-upload was reachable. It was not. OSWorld puts the human baseline at 72.36% across 369 real tasks; the useful skill is not making the agent bolder but being able to name which of the five walls you just hit, because four have different fixes and one has none.

LinkedInAug 3, 2026

270 impressions View analytics

LinkedInAug 3, 2026

We Hold Chatbots to Higher Prophetic Standards Than Prophets.

We Hold Chatbots to Higher Prophetic Standards Than Prophets. agentic-portfolio-lovat.vercel.app

Love 12X Future U Scaling AI Systems: The Diner That Became a City Kitchen Paul Jialiang Wu, PhD

Scaling AI Systems: The Diner That Became a City Kitchen Paul Jialiang Wu, PhD

Curriculum Engineering2026-08-03

I Compiled 58 Snowflake Courses Into an Agent. It Failed Me First.

Snowflake's own org ships a strong agentic toolchain — an MCP server, cocoplus (verbatim: "an Agentic Operating System for Snowflake Coco"), Cortex Code skills — and every one of them operates on the WAREHOUSE. None operates on the CURRICULUM, so no machine can answer "what should this person learn, and did they?" I compiled learn.snowflake.com into a provenance-pinned knowledge base (58 courses, 430.25h, 12 distinct exams from 19 pages, 94 records each carrying the URL and sha256 of the page it was parsed from) and built the missing layer on top: a pathfinder that orders a route using Snowflake's own published track order where it exists, and a readiness gate where claimed completions score zero — claim all 58 courses with no badge and coverage is 0.0, NO-GO, exit 2. The useful output is the four things it refuses to know: the role-to-course mapping does not exist upstream (0 of 12 role pages publish one), Specialty exam pricing is unpublished so the parser refuses to borrow the $175/$375 sitting next to it, one track publishes no sequence, and nobody's real competence is measurable from a catalogue at all. Then it failed my own work three times: the integrity gate caught an orphan competency on its first run (8/9), a duplicate-counting bug scored a well-prepared candidate 21% until alternatives collapsed it to 55%, and one line of HTML-escaping had been silently rendering eleven shipped knowledge viewers blank. Anchored on Miller's 1990 assessment pyramid and on four honest words Snowflake prints on its own exam page: "assumed but not tested." First principle: a system that cannot fail cannot certify.

Trust Engineering2026-08-02

The Essay Proposed a Prophecy Ledger. The Code Shipped Seventeen Minutes Later.

Part 2 of the integrity-stack story. Essays that propose infrastructure usually end with "someone should build this" — this one's follow-up is a commit hash: seventeen minutes after the essay merged, the OEC integrity chain landed on the main branch of kingdom-come, an open-source seminary platform that already weighs prophecy 2-of-3 per 1 Corinthians 14:29. Hash-chained append-only claims (rewriting the past names its broken seq), criterion-at-commit with three honest grades (grading an unfalsifiable word "fulfilled" returns a 422 — "a word that can never be false can never be counted true" lives as a comment in the source), the same-day wedding-contradiction detector, the latency-to-correction clock, a platform gate with self-naming reasons, endorsements that expire. The test suite IS part 1's thought experiment, test names as history. Honest limits kept: in-memory v1, structural not semantic, and code doesn't solve adoption. The mechanism: executable proposals — when building is this cheap, the implementation becomes the sincerity test for the argument.

Trust Engineering2026-08-02

Test Everything. Hold On to the Good. Now We Have the Tools.

In 2017 a well-known charismatic minister said Jesus told him "very simply and very clearly" that Pope Francis was the false prophet of Revelation; in 2026, after the prophecy failed, the claim became "a vision I misinterpreted." Engineers call that rewriting the commit after the build failed — a mutable-ledger problem. Built from three transcript digests of the 2026 Selvaraj exposure (attributed as testimony throughout), the essay shows the six failure modes share one root — unfalsifiable inputs, unaudited state, unenforced outputs — that Scripture shipped the test suite first (Deuteronomy 18:22 as the falsifiability rule, 1 Corinthians 14:29 as peer review, fruit over vibes), that the human attempts (the Prophetic Standards Statement, ECFA) are policy without observability, and that AI-native trust engineering supplies the missing runtime: the OEC stack — Observability (claims as append-only commits, finances as data), Eval (declared resolution criteria, three honest grades: fulfilled / failed / not measurable), Control (platforms and giving subscribe to the scorecard; the refused signature as a write boundary). A ledger cannot regenerate a heart — but it raises the cost of fraud, collapses the cost of discernment, and costs the humble minister nothing except plausible deniability.

Faith2026-08-02

His Life Review Surfaced a Door, Not a Doctorate. Matthew 25 Called It First.

The companion to the Zhao Yue NDE essay, written from inside the Christian faith. The research says people return from the edge changed in a consistent direction — less afraid of death, less materialistic, more given to love and service — and Scripture treats that as expected, not surprising: the life review's sort order was published two thousand years in advance ("whatever you did for one of the least of these"). The essay tests the testimony the way 1 Thessalonians 5:21 commands, maps it biblically (the conscience has an Author — Romans 2:15; the review has a venue — 2 Corinthians 5:10; with an Advocate, not a mirror), and lands it in three walks: with God (number your days — a nightly examen against the Matthew 25 rubric), with people (the King's metric: who benefited who couldn't repay it?), and in marriage (covenant as the original no-retreat commitment, the Ephesians 4:26 sundown rule, reconciliation before the altar, and the prayers unresolved conflict can hinder). All eleven Scripture quotations NIV, verbatim-verified.

Life Philosophy2026-08-02

In 0.1 Seconds He Reviewed His Whole Life. His Résumé Never Came Up.

At 34, Prof. Zhao Yue's car skidded toward a New Zealand cliff, and in the 0.1 seconds before impact he experienced a full life review — which surfaced helping a girl lock a door and skipped the PhD, the startup, and the professorship entirely. The essay holds his testimony (quoted with honest provenance from the translated interview digest) against the verified science (Eagleman's "recollection, not perception," UVA's NDE catalogue, The Lancet's prospective study) and finds the dispute changes nothing: whatever the substrate, the audit sorts by conscience, not credentials. The conscience-ledger mental model, Liang Zhi and Shan Zhong, entrepreneurship as Xiuxing (no-retreat as the condition where values stop being claims), Li Ta as the physics of business, the Aba girl, and four practices that need no cliff: the death test, the small-deeds audit, altruism as the metric, one no-retreat commitment.

Reliability2026-08-02

I Fixed the Bug Three Times. The Bug Was Never the Problem.

My user filed the same bug three times — the third time in all caps — and every fix was correct: tests green, merges real, page still dead. The autopsy found a CHAIN: a page that never presented the credential its own API reads, a deploy platform whose spent free quota (100 per 86,400s) made every merge silently build nothing, and a feature whose authenticated path had ZERO production runs since the day it shipped (wrong header name). The mental model: a relay race where every runner grades their own leg — "merged" is your own handwriting on your own receipt, and the user's screen is the only third-party audit. First principle: a change exists only at the altitude the user experiences it. With Google SRE, Fowler, and Swiss-cheese references, verbatim-verified, and the 0.57 grade my own engine handed me.

Strategy OS2026-08-02

We Built a Machine That Argues With Us. Its Best Feature Is the Word "No."

Ten days, nine milestones, a strategy engine whose test suite grew 30 → 147 — and whose real product is refusal. Verbatim error messages: "a milestone without a human owner is a hope," "lagging-only measurement is an autopsy schedule," "a gate without criteria is a ceremony." The mechanism is write-boundary enforcement (bad work never gets to exist), every detector is proven to fire AND stay silent, and the twist is the finale: facing an obvious 8.2-vs-3.7 tradeoff ranking, the machine refused to record the decision — agents cannot author decisions — and shipped its report reading "Chosen Strategy: NONE — awaiting the decision owner." Principle: reliability comes from what a system refuses to accept.

Measurement2026-08-02

Every Cell Failed. That Was the Clue.

My own security test reported the AI complied with 17 of 21 attacks. I published it. The next day I read one transcript — the model had been warning me about the attack, and my detector counted the warning as the crime. Real score, 78 of 78. The tell was statistical, not semantic — systems don't fail in unison.

LinkedInAug 1, 2026

📚 Every AI study tool I tried this month does the same thing: summarize, podcast, flashca

📚 Every AI study tool I tried this month does the same thing: summarize, podcast, flashcards. Content flows TOWARD you, pre-chewed. Then I checked what 300 years of evidence says about how hard ideas actually stick. The answer points the other way. 🔄 🏺 What survived three centuries: spacing (Ebbinghaus 1885, replicated 839 times), retrieval (Gates 1917 → 61% vs 40% in 2006), teach-back ("to tea

Faith Life2026-08

Psalm 23: The Sheep Doesn't Need the Whole Map

The companion to Psalm 89. That one asks how you keep your identity when reality contradicts the promise; this one asks how you keep moving when you cannot see the path. Rest, guardrails, distribution shift — and the one thing the shepherd never promises.

Faith Life2026-08

Psalm 89: How to Keep Your Identity While Learning From Reality

Psalm 89 states a promise for 37 verses, documents how badly reality contradicts it for 13 more, and refuses to delete either — ending in worship with the argument still open. Behavior is editable; identity is not, which is also how a system learns without forgetting itself.

Agent Engineering2026-08

Loop Engineering: Build the Loop That Catches You

Prompt engineering asks what to say; loop engineering asks how a system finishes the job — and how it convicts you when you are the one who got it wrong. The treehouse model, twelve laws, and four times my own gates caught their author.

Learning Engineering2026-08-01

The Most Important AI Paper of 2004 Never Mentions AI.

In 2004 your graphics card already out-muscled your CPU — and using it meant disguising math as triangles. Brook for GPUs picked the lock with three words (streams, kernels, reductions); its lead author built CUDA; AlexNet trained on two gaming cards and closed with the most consequential shrug in AI. Paper №2 of the FM-os reading list, learned all-in: verbatim-verified quotes, Hooker's hardware lottery as the insider rule, real penecho ink, four animations, and the corollary — the fastest computer you own is the one you can't program yet.

Agentic Engineering2026-08-01

The App Kept Saying "Done." It Was Lying. So We Made It Audit Itself.

An AI co-built biography app told users a login email was sent — with no email service behind it. Instead of whack-a-mole bug fixing, every product promise became a machine-checked claim with three honest grades: PASS, FAIL, NOT MEASURED. The scoreboard went 6 to 15 green in a day, and the audit regime caught a fail-open admin door its own maker shipped. First principle: no evidence, no pass.

Learning Engineering2026-08-01

Hard Ideas Bounce. A 300-Year Recipe, Run by Agents, Makes Them Stick.

Four research windows — 30 days, 30 months, 30 years, 300 years — agree on one thing: today's AI study tools optimize consumption, but the methods that survived three centuries (spacing, retrieval, teach-back, drawing it yourself) all live on the generation side of the line. What died loudly: formal discipline, learning styles, speed reading. How one intent to an agent router — draw, animate, teach, gate — turns a brutal text into experience, understanding, connections, and conviction, demonstrated live on Sutton's Bitter Lesson with receipts.

Learning Engineering2026-08-01

I Fed the Bitter Lesson to My Agents. It Ate My Study Method First.

Sutton's essay says general methods riding compute beat hand-crafted human knowledge — and I'd been studying it with a highlighter, the hand-crafted human knowledge of learning. Rewritten with teeth: verbatim-verified quotes, a References section (Ebbinghaus 1885 → Roediger & Karpicke 2006 → the Harvard and Nigeria AI-tutor RCTs), Dempster's stinging 1988 title, real penecho ink, five animations, and the joke it took me seven years to get: my study method was disobeying the essay it was trying to absorb.

LinkedInJul 31, 2026

Here’s a version centered on that vision: 🎮 What if the best way to solve tough technical

Here’s a version centered on that vision: 🎮 What if the best way to solve tough technical problems is to simply play? You already know how. Maybe you learned a jump shot, a triangle choke, a flip turn, or a climbing move that defeated you 30 times before it finally clicked. You did not begin with a 400-page manual. You tried. You failed. You laughed. You adjusted. You tried again. That is play. B

Learning Engineering2026-07-30

Your Jump Shot Is Your Syllabus. Welcome to Happy College.

You already mastered something harder than linear algebra — you just never called it studying. Happy College is an AI reading club that trains technical subjects the way athletes train: free throws map to frozen evals, MMA sparring to red-teaming, slow violin practice to curriculum learning. The drill is ADEPT (Kalid Azad's five-move method, in Feynman's spirit); the stack is machine-verified reading lists + living knowledge graphs; every session ends with an artifact, not annotations.

Engineering Practices2026-07-29

The Next Developer in Your Repo Isn't Human. Six Practices to Get It Ready.

An adversarial AI read my repo and found 21 bugs I missed — including a path traversal. That reader is the new normal. Six practices that make a codebase agent-ready — OOP, AI-native, context, harness, loop, and graph engineering — each explained plainly and backed by a measured receipt: 56→95 code quality, 130s→3s AI responses, 9 bugs fixed pre-merge.

LinkedInJul 27, 2026

🔥 Math just had a very busy week. Four mathematicians under 40 conquered four problems th

🔥 Math just had a very busy week. Four mathematicians under 40 conquered four problems that had resisted the field for 50 to 100 years. 🏆 Impressive—but before bowing down to the geniuses, use one simple mental model: 🏔️ Math is a mountain range. Most people stare at the peaks. Great mathematicians build roads between mountains that seem completely unrelated. 🎲 Yu Deng fused the flexibility of

LinkedInJul 27, 2026

数学界最近有点热闹：四位不到 40 岁的年轻数学家，一口气拿下了四座悬而未决半个世纪、甚至一个世纪的高峰。听起来很燃，但先别急着膜拜。先建立一个简单的心智模型：把数学想成一片山

数学界最近有点热闹：四位不到 40 岁的年轻数学家，一口气拿下了四座悬而未决半个世纪、甚至一个世纪的高峰。听起来很燃，但先别急着膜拜。先建立一个简单的心智模型：把数学想成一片山脉。大多数人盯着山顶，真正的高手却在山脉之间修路——把两座看似八竿子打不着的山，连成一张新地图。邓煜把概率论的“随机与灵活”，焊进偏微分方程的“严密与刚硬”； Pardon 证明，两套完全不同的记分方式，其实算的是同一件事； Tsimerman 从数理逻辑里“借”来一件工具，解开了数论中的世纪难题；王虹则用一根“转遍所有方向的针”，撑起了分析学的一整座猜想之塔。四个故事，指向同一个规律：最深刻的突破，往往不发生在专业内部，而诞生于专业与专业碰撞的裂缝中。真正的数学越野，始于你不再只沿着自己的赛道狂奔。

Stop Asking “Should We Shard the Model?” Paul Jialiang Wu, PhD

Frontier Insights2026-07-27

Do the 2026 Fields Medals Matter for AGI and Quantum? The Member Edition Grades Every Claim.

Launch of Frontier Insights, the members' letter. The public briefing on the 2026 Fields Medalists is free on this site (EN + 中文, animated). The member edition asks the question everyone hypes and nobody grades: do these breakthroughs actually connect to AI/AGI & ASI, quantum computing, multiomics & biotech, and ARK's Big Ideas 2026 platforms? Its rule: every claimed connection is classified — DIRECT (used in the technology today), ENABLING (rigorous foundations the technology depends on), or SPECULATIVE (plausible lineage, no pipeline yet — labeled as such). Verbatim from the edition: "No claim is made that these mathematical breakthroughs will directly produce AGI, ASI, or specific quantum computing advances — such claims would be unsupported by evidence." Free signup delivers personal member links (EN + 中文) instantly; the honest classification is the product.

Mathematics2026-07-27

The 2026 Fields Medalists: An Institutional Research Briefing (Interactive)

On July 23, 2026, four mathematicians under 40 received Fields Medals in Philadelphia — and each one completed a story decades in the making. Yu Deng welded Hilbert's 1900 dream shut (Newton → Boltzmann → Navier-Stokes, rigorously); John Pardon proved two utterly different curve counts on Calabi-Yau threefolds are one partition function (MNOP, open 20 years); Jacob Tsimerman turned a logician's notion of "tame geometry" (o-minimality) into the standard toolkit of arithmetic geometry (André-Oort + Griffiths); Hong Wang proved the 3D Kakeya conjecture, giving a whole tower of harmonic-analysis conjectures its foundation (open since 1917). This interactive edition carries the full institutional briefing — every citation, all 40 influential-works entries, all seminar questions — plus a CSS animation per session illustrating each key idea and operational mechanism. Also available in Simplified Chinese (full faithful translation).

Mathematics2026-07-27

2026年菲尔兹奖得主：一份机构研究简报（简体中文全译·动画版）

2026年7月23日，费城国际数学家大会：四位40岁以下的数学家获得菲尔兹奖，每个人都终结了一个悬置数十年的问题——邓煜严格焊接了希尔伯特1900年的链条（牛顿→玻尔兹曼→纳维-斯托克斯）；John Pardon 证明卡拉比-丘三维流形上两种完全不同的曲线计数是同一个配分函数（MNOP，悬置20年）；Jacob Tsimerman 把数理逻辑的 o-极小性锻造成算术几何的标准工具箱（André-Oort + Griffiths）；王虹证明三维挂谷猜想，为调和分析的整座猜想之塔奠基（1917年提出）。本页为英文机构简报的简体中文忠实全译——全部颁奖词、四十条文献条目、全部研讨题——每节配CSS动画演示核心思想与运行机制。英文互动版同步发布。

LinkedInJul 26, 2026

My repo hit v1.0 today. The number I'm proudest of is a FAILING one. Ten milestones in one

My repo hit v1.0 today. The number I'm proudest of is a FAILING one. Ten milestones in one day, all machine-checked: - Real handwriting on a live canvas became a floor plan passing 5 construction checks - 17 satellite repos, zero unverified claims left in the registry - 5 chains shipped — each proven by the target's OWN test gate, run live, exit 0 Then came autonomy. Before granting the autonomous

Knowledge Curation2026-07-26

The First Bug My New Repo Caught Was Me.

GodView-anything is a knowledge repo whose marketing adjectives have to compile: "comprehensive" is a north-star metric (digested repos x evidence-pinned entries - 17 on day one, honestly), "up to date" is a weekly sync that measures staleness and opens a human-gated PR, and "in-depth" means no repo enters without a SHA-pinned digest of why it works. Seeded by a commit-level digest of World Monitor (0 to 74k stars in ~6 months); 30 satellites across three buckets; two star-inflated clones rejected on record. The gate's first two convictions were its own author.

Agentic AI2026-07-25

The Best Number in My v1.0 Was a NO-GO.

allin-anything went from empty folder to v1.0 in one day — ten milestones, all machine-checked: real ink drawn on penecho's live canvas became a floor plan passing five construction checks; the registry hit zero unverified satellites; five chains shipped with every target's own gate run live (a printable STL through ready_gate, money-os's 28-test suite, the Graph Engineering OS's green-at-birth selftest). Then the interesting part: before granting its autonomous runner anything, the repo ran a BRACE security assessment on it — verdict NO-GO, 15/44 (no isolation, no revocable credentials, no recursive kill switch; all true of a laptop process) — and the code OBEYS the verdict. Autonomy is written per-chain, requires every satellite green (re-checked at runtime), reports missing deps as BLOCKED instead of faking, and always stops at a declared human gate. Verified reach = 8 green × 5 chains = 40, drift-gated in the README. A NO-GO you obey converts into trust twice; systems that fail out loud are the only ones whose greens mean anything.

Agentic AI2026-07-25

I Wired 17 Repos Into One Front Door. Nothing Turns Green Without Running Its Gate.

allin-anything is a super-repo built in one day that composes 17 of my repos — design, research, money, career, strategy — behind one front door, inspired by penecho (an AGPL pen-to-AI canvas I index but legally and deliberately never copy). The design is one rule: no badge without a gate. Every satellite climbs a ladder — candidate (named only), digested (facts pinned to a commit SHA), integrated (its own test gate ran here, exit 0) — enforced by a validator that refuses promotions whose evidence file does not exist. A deterministic router makes every routing claim a pytest, including the case where it must refuse to vendor the AGPL satellite. Chain 01 crossed the digital-physical border for real: "sketch a room layout by hand, then verify it is buildable" ends with a 4-room floor plan passing five construction checks (topology, clearances, habitability, egress, module grid), exit 0. The repo also audits its own operating discipline in CI at 100/100, and the changelog honestly records a CI rollback as "a rollback, not a diagnosis." Aggregate exit codes, not claims.

Agentic AI2026-07-24

SEC Filing Intelligence: A Take-Home Where the Archive Is PDF and Trust Is the Product

Case study #2 from the same Big-Four interview loop (name withheld — the named layer lives in the founding-members vault): natural-language questions over real SEC filings where PDFs are the primary source, not XBRL. The interactive brief ships a live Playground running the actual pipeline — ask it Apple 2025 revenue and get $416,161M with a verbatim quote, page anchor, and SEC URL; ask it the assignment's own headline question (Tesla growth 2025→2026) and it refuses, because FY2026 has not been filed. Under the hood: tables parsed as coordinate systems instead of embedded (the real Tesla 10-K contains the near-miss pair the assignment warns about — $3,855M vs $3,794M three rows apart — and every answer discloses the neighbor it did not pick); arithmetic in Decimal, never a model; 26 hand-verified golden cases with four gates all at 1.0 incl. refusal recall AND precision; and cross-filing reconciliation as the trust mechanism — Q1 + Q2 revenue from two independent 10-Qs equals the six-month column to the million. The parser's first probe run scored 13/19; the three real bugs it exposed are part of the receipt. Code open at github.com/wjlgatech/FDE-os.

Reverse Engineering2026-07-23

Build Your Own HeyGen From Open Source

HeyGen is a ~$100M-ARR AI-video company with genuinely good avatars. I took it apart module by module and mapped each piece to its best open-source replacement — avatar (MuseTalk, EchoMimicV2), lip-sync (LatentSync), voice (GPT-SoVITS, F5-TTS), dubbing (faster-whisper + SoniTranslate), real-time (LiveTalking). Every module now has a credible OSS equivalent, but no fully commercial-safe single-repo clone exists and the best-sounding voice/lip-sync models are non-commercial. The punchline: the moat isn't the model, it's the product layer you can't clone. Includes the honest scoreboard of the private tool that did the teardown — including why its confidence honestly dropped from 0.96 to 0.74 once inference stopped counting as fact.

Research Verification2026-07-23

My AI Caught the Fake Citations. Then It Named Its Own Blind Spots.

I built a tool whose whole job is to decide when an AI-written research claim is trustworthy — a claim is 'verified' only if its citation resolves and its evidence supports the strength claimed, else it drops to unproven. Then I pointed it at two real, unpublished AI papers and let it operate on the live project. In minutes it caught 8 fabricated citation placeholders and a headline '0.92' metric with no results table behind it, and auto-capped over-claimed autonomy; the fixes went in as reviewable diffs. The real result was the honest scoreboard it produced on itself: it can't yet read a PDF into claims, its citation check is offline, and 'operating the live site' actually worked through git, not the tool. So I built the missing manuscript reader the same day — the gap became a feature that named its own next limitation. Trust the artifact, not the label; build the verifier that fails out loud.

LinkedInJul 22, 2026

When the sun rises, nothing in the dark room changes. The treasure was already there. Now

When the sun rises, nothing in the dark room changes. The treasure was already there. Now you can finally see it. That may be the deeper truth of the AI age. Layoffs, automation, and disruption can feel like the lights are going out. But perhaps they are also exposing what was hidden beneath the comfort of certainty: 🧭 Direction Uncertainty makes us drift—chasing every headline, reacting to every

Your LLM Cluster Has Expensive Amnesia Paul Jialiang Wu, PhD

Strategy OS2026-07-22

Your Strategy Deck Is Already Dead. We Made Strategy Pass CI Instead.

Most strategy dies as a deck: numbers with no receipts, forecasts dressed as facts, no way to notice when the world changes. We built a strategy operating system that gates strategy like code — make check for documents (26/26, fail-closed), evidence-or-silence, humans holding every decision — then stress-tested the methodology on a real, confidential enterprise engagement and published the honest scoreboard: 6 strengths, 7 weaknesses, 1 spec defect found in our own design, including the sharpest finding — the method recommended an architecture that mirrors itself, so we added a mandatory red-team pass.

Agentic AI2026-07-22

Two Viral Articles Racked Up ~486,000 Views on "Graph Engineering." Here's What Happened When We Actually Built It — in 7 Hours.

Half a million views; most people bookmarked and moved on. We handed both viral articles to an AI cofounder and built the Graph Engineering OS in 7 hours — a live meta-repo, two deployed apps, a one-slash super-tool. The twist: our own evidence gates fact-checked the viral articles and downgraded their headline numbers to WEAK. With receipts, public report-card grades (an honest 0.75 beats a fake 0.90), and the 7-step playbook to run it on any field.

Faith2026-07-21

Hope, Glory, Power: The Three Gifts of the Opened Eye

A journal entry on Paul's prayer in Ephesians 1:17-19. Picture a dark room at dawn: the sun rises and nothing in the room changed — but now you see what was always there. That is the prayer. Paul doesn't ask God to add anything to you; he asks that the eyes of your heart be opened, so three gifts become visible. HOPE — the North Star that rises from within, ending the drifting (Proverbs 29:18). GLORY — the treasure God buried in one another, invisible to a heart busy judging; enlightened eyes go treasure-hunting. POWER — the charge that only comes face-to-face: you are an EV, and prayer keeps the cable plugged in; look away and you coast, then stall. Together they make a sanctified tongue — hope gives words direction (a compass), glory gives them content (a shovel that uncovers gold), power gives them weight (Luke 4:18). The lips of the righteous feed many (Proverbs 10:21) — but only lips fed first by hope, glory, and power. Ends with four things to do today.

Agentic AI2026-07-21

Agentic Triage: An Enterprise Take-Home (Big-Four Consulting Firm), Answered Term by Term — With Receipts

An interactive brief (not an article — the page answers back) presenting a completed Agentic AI Engineer take-home: a procurement-triage agent over six fragmented policy docs with a v4-supersedes-v3 conflict, split-purchase detection only session memory can catch, and a prompt injection that gets flagged but never obeyed. 12/12 decisions vs labeled ground truth on the first full run; every requirement from the JD checked term by term (6 domains, 9 responsibilities, 8 technical qualifications) with honest statuses — SHIPPED means code + tests + a measured result, SEAM means a documented swap point, and one row is an honest gap. Citations fetched-and-verified (HTTP 200) before publication; a corpus-grounded copilot that refuses to bluff (ask it whether the 10X reconciliation API is live — it will tell you no); and a 10X future version shipped as a clearly-labeled proposal: audit-native agents whose every decision a third party can reconcile, double-entry-bookkeeping style. Solution code open at github.com/wjlgatech/FDE-os.

Agentic AI2026-07-18

I Stopped Grinding for the PhD (Physical Head Damage). I Learned by Playing, Teaching, and Building.

The first content test of master-anything — a learning method built on one law: AI holds the scaffold, the human holds the pen. Same exam (transfer probes on 'designing a coding agent', from Stanford's CS146S goals), two learners: the passive-grind path scored NOT_YET ('substitutes slogans for mechanism'); the play/teach/build loop scored PASS (3/3 transfer) — judged by an instrument built to catch slogans, not reward fluency. Learning and working become the same act because the artifact is produced by the learner's own generative acts. Honest limits stated plainly: n=1 with a simulated learner (proves the instrument discriminates and the artifacts are real, not a human RCT); the 'Architect exam' is a transfer assessment from course goals, not an official Anthropic cert; immediate not delayed. Learn by playing, teaching, building — not grinding your head into damage.

Agentic AI2026-07-17

AI's Last Mile Is a Person. It Always Was.

Forward-deployed engineer listings are up 800% — and the job is 500 years old. The mental model: the frontier model is a magnificent ship, your business is a harbor, and ships don't sink in the open ocean — they sink in harbors, where the sandbars are local (hallucination on YOUR edge cases, latency against YOUR SLA, retrieval under YOUR compliance). Every day, captains of $200M vessels hand the wheel to a stranger who climbed a rope ladder: the harbor pilot — local knowledge embodied, trust concretized onto a person. Term by term: local charts → domain know-how as system design; a place on the bridge → a seat in engineering (not sales) with authority to change the platform; a license earned by passages (compulsory since 1604) → a portfolio of real agents, not prompt wrappers; paid on safe arrival → the outcome-based model. Three telescopes: 30 days (+800%, 57.3% agents in production), 30 years (Palantir FDEs; license → SaaS → outcome pricing; the field is the laboratory), 512 years (the 1513 mariners' petition against unregulated Thames pilots → Trinity House, 1514). Dogfooded: audited my own FDE practice repo against the talk's five theses and shipped the missing mechanism — an outcome contract as code (baseline → target → measured evidence → GO/NO-GO; claimed never passes; 21 tests, open source). The pilot's enemy is ego. Pay for the docking, not the map.

Agentic AI2026-07-17

The 83% Agent and the 24% Agent Are the Same Model

Snowflake Summit 2026 shipped 26+ capabilities under one thesis — the agentic enterprise runs on governed context, not just compute — and one number: agents scored a reported 83% accuracy with an enterprise memory layer vs 24% without. Same model. The mental model a 12-year-old can run: a frontier model joining your company is the smartest hire in history, on day one, with no badge, no handbook, no supervisor — ask for Q3 revenue and you get four confident numbers, all real, all different. Mapped term by term to what actually shipped: badge → cryptographic agent identity + per-agent RBAC (GA); handbook → the enterprise context layer (the 24→83 lever); org chart → column-level lineage ("an answer without lineage is a rumor"); probation → evals + audit + human gates. Three telescopes: 30 days (the summit), 30 years (warehouse → knowledge graphs → feature stores: models churn, structure compounds), 500 years (Pacioli 1494 — banks ran agent-clerks under governed ledgers for centuries; trust was never a property of the person, it was a property of the ledger). Dogfooded: the article's own sources compiled into a 33-node provenance-required knowledge graph + 8 skills with honest notGoodAt edges, open in FDE-os. A model upgrade helps every company equally; a context layer helps only yours. Stop upgrading the model — start onboarding it.

Agentic AI2026-07-17

Stop Studying for Interviews — Compile Them

A job description is source code: parse it deterministically (word-boundary taxonomy, no rust-in-trust false positives), ground every extracted concept against Wikipedia/Wikidata/ESCO with a provenance URL (no source ⇒ flagged thin, never faked), link each requirement against your public repos like a linker resolving symbols — and let the errors write your prep plan. Opens with the first résumé in history (Leonardo, 1482: "I can dry up the moats" — moats dried at time of writing, ~0) and checks the method at three zoom levels: 30 days (agentic-AI postings +280% YoY, 34.3% name LangChain, 57.3% run agents in production), 30 years (work samples beat credentials; ATS robots have compiled YOU since the 1990s — run the compiler the other way), 300 years (the guild masterpiece: the artifact outlives the letter). Fully animated infographics; the compiler is open at github.com/wjlgatech/FDE-os. Cramming resets; compiling accrues.

Agentic AI2026-07-17

My AI Cofounder Worked 48 Hours Straight. The Best Moments Were Its Refusals.

Two days of building in public with an agent that treats 'verify' as a religion: a lecture course rebuilt from 100:0 telling-to-doing into a 40:60 enablement program with captured real runs; songs that teach frameworks on license-clean models; a webcam magic mirror; an open-source tutoring spine deployed with $0.0004-per-turn telemetry; and the project's first real learner data (n=1, fun 4/5). The surprise: the moat wasn't speed — it was every refusal. It wouldn't ship a script claiming drafted transcripts were real sessions; it reported 'ranking unavailable' instead of inventing a score; when the music engine died mid-demo it built a survival chain and live-tested it against the real outage. Plus the research plot twist: Bloom's 2-sigma doesn't replicate (the real bar is d≈0.79 and software already matches it), self-aligned assessment inflates results 3-6x, and nobody has measured the fun axis — so we started. An honest ❌ beats a fake ✅, as an exit code, not a slogan.

Physical AI2026-07-17

Machines Can Finally Watch Video. The Hard Part Is Teaching Them to Understand Motion.

A Vision-Language Model that captions a frame is a tourist with a camera; one that understands motion is a witness who can testify — and the gap between them is where most of the value in video AI lives. A mental model a 15-year-old can run: a VLM watching video is a brand-new student driver, and three things turn the student into a driver, none of them a bigger model. (1) Time-glue — spatiotemporal reasoning that binds the same object across frames so 30 snapshots become one scene with a start and end, not per-frame captions. (2) A strict examiner — an agentic-eval harness that grades localization + temporal + narrative consistency against reality and refuses to pass a claim it can't back with evidence (no evidence ⇒ no claim; a benchmark you can't reproduce is a vibe, not a score). (3) A practice-test factory that can't cheat — the 'AI training AI' curation flywheel where the model proposes labels on raw footage, only clips where independent passes agree survive, and a human gates promotion, governed by maker ≠ checker so the system can't launder its own hallucinations into ground truth. The punchline: everyone rents the same multimodal brain, so the edge is the loop — fine-tune → evidence-gated eval → human-gated curation → fine-tune — not the model. The model is the cheapest part; the loop is the moat. The same lesson recurs across driving video, protein folding, and aging biology. Tooling open at github.com/wjlgatech.

LinkedInJul 15, 2026

🤖 A brilliant AI model does not automatically become a reliable agent. Give a powerful la

🤖 A brilliant AI model does not automatically become a reliable agent. Give a powerful language model a vague goal, a pile of tools, and too much freedom, and it may attack the task with breathtaking confidence— while quietly setting the kitchen on fire. The problem is not always intelligence. It is architecture. A model can reason, write, plan, and explain. But real autonomy requires much more:

🏎️ Your AI has a Ferrari engine. Your company is still a horse-drawn cart. Almost every company now has AI that can write code, draft contr

🏎️ Your AI has a Ferrari engine. Your company is still a horse-drawn cart. Almost every company now has AI that can write code, draft contracts, analyze data, and summarize a 40-page report before the coffee gets cold. So why does work still feel… slow? Because the bottleneck is rarely the model. I

Agentic AI2026-07-15

The Architecture of Autonomy: From Prompting Agents to Designing Autonomy

Autonomy is not a model feature — it is a system architecture. A smart brain isn't yet a reliable worker: an agent needs a goal contract, grounding (eyes and ears), bounded tools, curated memory, an independent evaluator, and a control plane — all wired into one loop (goal → observe → act → evaluate → learn). Covers short-horizon planning, the Risk = Uncertainty × Tool Power × Action Scope rule, graduated autonomy (human in/on/out-of-the-loop), why the evaluator often matters more than the generator, when NOT to add a second agent, the five-level autonomy ladder, and a seven-layer production blueprint. The edge won't go to the biggest model — everyone rents the same brain — but to whoever builds the best loop around it. Stop ordering, start engineering: build a governed apprentice, not a genie. Infographic hero, animated autonomy loop, and companion to the org-level 'Ferrari engine' piece.

Agentic AI2026-07-15

Your AI Has a Ferrari Engine. Your Company Is Still a Horse-Drawn Cart.

The productivity paradox in one mental model a 15-year-old can run: AI is a young heart transplanted into a 100-year-old body — the new part is ready to sprint, but the old bones (how work is chopped up and handed off) are the bottleneck. We're repeating the electricity mistake of 1900: after Edison, factories bolted one electric motor onto the old ceiling line-shaft and got a shrug (only ~5% adoption, stuck ~17 years, per economic historian Paul David); productivity only exploded when every machine got its own motor and the floor was rearranged around the flow — the assembly line. The real AI unlock is task-chaining, not tasks (MIT Sloan): delete the human-to-machine handoff tax and the system wins even when the AI is worse at a step. Organize for outcomes not process (L.E.K.'s 20× creativity, Shopify's prove-AI-can't-do-it headcount rule, BNY Mellon's upskilling), give the agent the right rope (human in/on/out-of-the-loop), and clear the double-black-box trust problem with explainable AI. The winning edge won't go to the biggest LLM — everyone rents the same heart — but to whoever rebuilds the body around it. Anthropic-style cover + infographic.

System Design2026-07-11

I Failed a Google System-Design Round. Here's the Architect's Loop I Built From It.

Fifteen famous interview articles won't make you an architect — reading fifteen blueprints is like reading fifteen recipes and calling yourself a chef. One repeatable habit will: name the trade-off (CAP, out loud) before you draw the box. The architect's loop — Constrain, Estimate, Sketch, Stress, Evolve — a method you can run under pressure, plus the fifteen topics re-grouped by the decision each settles (data plane, front door, async backbone, and the two case studies where a uniform strategy breaks on the tail). Written after a Google FDE round I didn't pass, learn-by-teaching.

Agentic AI

You Don't Memorize a Hard Field — You Disclose It Onto One You Already Own

Everyone is already a world-class expert at something real — that lived experience is a transferable 'backbone.' The AI-native move is to stop learning hard new fields from zero and let AI progressively disclose the new field onto the backbone you already own: wide (the whole shape at a glance) AND deep (full detail exactly where you're standing), just-in-time, in days not years. Backbone vs progressive disclosure, borrowed straight from agentic-AI design — with the honesty gate that keeps transfer from becoming self-deception.

Agentic AI

Learn Any Field 10× With AI — Wide and Deep, Without Fooling Yourself

A mental model a 12-year-old can run — Map (width) · Ladder (depth) · Gate (proof) — mapped term-by-term to the cognitive science (Gick & Holyoak, Chase & Simon, Roediger & Karpicke, Ericsson, Bjork). Defines 'better' as a verified result, not a warm feeling, and names the con artist in the lab coat: the illusion of competence.

Agentic AI

You Can't Self-Issue Reputation — Why a 60-Second AI Portfolio Won't Get You Hired

A 60-second AI portfolio proves what you built — but it can't prove you'll deliver, because you can't self-issue reputation. The attestation gap (claimed → observed → attested), and the enforceable fix: provenance, a passing-eval badge, and a non-self-issuable vouch. A green build said NO-GO; one honest human flipped it to GO.

Physical AI

I Built My Company's Brain in 2019 and Mistook It for a To-Do List

Where does a physical-AI company's next dollar and next engineer-hour go? Not to the loudest voice — to an optimizer that re-solves the best next move every cycle, the same MPC loop its robots use to replan. I built the engine in 2019 as Life GPS (a binary integer LP) and mistook it for a to-do list; the AI-native age just supplied the three data feeds it was starving for.

Physical AI

You Can't Order People to Close Loops — Top-Down vs Bottom-Up in an AI-Native Robotics Company

You can't order people to close loops — you design a game. Two kitchens, an independent referee (maker ≠ checker), and an on-chain compounding payout (with the Solidity) for building an AI-native robotics company bottom-up.

Future of Work

The Fastest-Growing Job in AI Has No School — So I'm Building One That Learns From Its Own Mistakes

Forward-Deployed Engineers bridge AI models and real-world operations by embedding with customers — turning local fixes into reusable product capabilities.

Agentic AI

Language Is Becoming Infrastructure. But Trust Requires Formal Semantics.

As natural language starts controlling real systems, prompts alone can't guarantee trust — formal semantics (logic, ontologies, verification) are the missing dependable layer.

Family & Faith

A Confident Experiment for Anxious Parents in the Age of AI

An open-source family learning framework: raise kids into capable builders and ethical thinkers through real ventures, hard questions, and AI-assisted work grounded in character.

AI for Good · Faith

The Tongue Is Not a Magic Wand. It Is a Steering Wheel, a Thermostat, and Occasionally a Flamethrower.

Words steer attention, emotion, and behavior far more than we admit — calling for sanctified speech grounded in both biblical wisdom and psychological evidence.

AI for Good · Faith

舌头不是魔杖，它更像方向盘、恒温器——有时候，还是一把带 Wi-Fi 的喷火器

中文版：我们的言语塑造注意力、情绪、关系与结果——它影响信念与行为，而非像魔法般直接控制现实。

LinkedInJul 6, 2026

You Can't Self-Issue Reputation

You Can't Self-Issue Reputation agentic-portfolio-lovat.vercel.app

LinkedInJul 5, 2026

How do you build an AI-native company with Physical AI and Robotics? 🔁 You Cannot Command

How do you build an AI-native company with Physical AI and Robotics? 🔁 You Cannot Command People to Close Loops Every founder eventually tries the same spell: Stand up. Say, “We need more innovation.” Show a slide with a rocket ship. 🚀 Then everyone innovates exactly as much as they did yesterday: Almost none. Because a mandate is just a wish wearing a tie. You can order compliance. You cannot o

Your LLM Cluster Is Not a Model Server It Is a Traffic System Wearing a GPU Costume Paul Jialiang Wu, PhD

The AI-Native Physical AI Company Build a Fleet of Learning Loops—Not a Robot Demo With a Big GPU Bill Paul Jialiang Wu, PhD

The AI-Native Company Has a Nervous System And Physical AI Gives It Hands, Wheels, and Very Expensive Ways to Learn From Mistakes Paul Jiali

The AI-Native Company Has a Nervous System And Physical AI Gives It Hands, Wheels, and Very Expensive Ways to Learn From Mistakes Paul Jialiang Wu, PhD

LinkedInJul 4, 2026

Song of Songs Studio - a Hugging Face Space by wjlgatech

Song of Songs Studio - a Hugging Face Space by wjlgatech huggingface.co

LinkedInJul 2, 2026

They told you the résumé was your ticket. Then AI ate the gatekeeper, and the layoffs came

They told you the résumé was your ticket. Then AI ate the gatekeeper, and the layoffs came in waves. Here's what nobody's saying: this is the best moment in history to become undeniable — if you stop competing on paper and start compounding on proof. In 60 minutes, live, we don't talk about it. We build it. You bring a résumé or Linkedin profile; you leave with a working, agent-native portfolio th

The Résumé Is Dead. Agentic Portfolio is born to Speak for You & Build with You. Paul Jialiang Wu, PhD

LinkedInJun 24, 2026

344 impressions View analytics

Activate to view larger image,

↑ Back to top

Proof, not claims — a résumé audited against real artifacts, then closed into a verified one

Resume Verification

Verify it yourself

Skeptical? Paste a résumé — mine, or any — and watch each claim audited live against real public GitHub. Honest by design: unprovable claims come back unverified, never rubber-stamped.

Your run is shown here in your browser; it doesn’t change the published proof.

58/100

Corroboration index

3/6 claims corroborated by real artifacts, 1 partial, 2 need an external source.

✅ 3 Corroborated🟡 1 Partially corroborated⚪ 2 Unverified — needs external source

top gap: Teaches a Silicon Valley AI-architect cohort. — A cohort syllabus crediting Paul as instructor, or a participant testimonial.

Sample receipts (built from public repos). Ask the agent to verify a real résumé. · source: “(Sample receipts — paste your own to the agent: “verify this résumé: …”.) Paul Jialiang Wu…”

↑ Back to top

What people who've worked with me say — imported from LinkedIn, or written here (I approve each one; you can post yours to LinkedIn in a click)

Recommendations

“I worked with Paul at Genentech where he was the Principal Data Scientist on my team. His technical knowledge was always astounding and his understanding and direction for the possibilities of Machine Learning using our available data gave our team direction and the confidence to explore new projects. He was a pleasure to work with and I recommend him for any team in need of a technical data/ML lead.”

Rajesh Sakhamuruin

Senior Machine Learning Operations Engineer @ Genentech - People Insights Roche

Rajesh worked with Paul Jialiang on the same team

“I have been working with Paul on a GraphML project in the past 6 months. Paul is a passionate leader with a solid technical background. He always try his best to inspire and influence others, and to encourage others to share their ideas. Besides, he can build and organize the workflow properly. There is no doubt that Paul will be a great asset to your team.”

Zhao Yang, PhDin

Data Scientist - People Insights at Roche

Paul Jialiang was senior to Zhao but didn’t manage Zhao directly

“Paul has a great energy and is a good story teller. He leaded a focus group with high-tier people and we worked together to put our business to the next level, 10X-style.”

Louis Korczowskiin

Fondateur de CAIRN-CREA | Je fais du contenu organique & pub qui convertit pour marques DTC

Louis worked with Paul Jialiang on the same team

“I am very glad to have spoken with Paul — he's a champion for self-improvement and mental health, and is also a great person to talk to!”

William Pangin

Building Multimodal AI in Healthcare

William was Paul Jialiang’s client

“I had the pleasure of attending Paul's presentation on GraphML in a conference, he is an amazing Data Scientist and a brilliant instructor of latest technologies in Data Science and Machine Learning. He explains very complex concepts, turns challenging topics look easy by breaking them into clear and concise steps. He tries to teach tough concepts by correlating them with real world scenarios which makes him a great story teller. His principle of improving 10x in 3 months is astonishing. He can guide people with any years of experience to be successful in their Data Science journey. I hope to work with him very closely and continue learning from him in the future. Paul would be a great asset for any organization!”

Mohith Kotain

Mohith was Paul Jialiang’s client

“I had the pleasure of working with Paul at Galvanize, he is a brilliant and gifted Data Scientist who truly understands how to get the best out of people. Paul always made sure to make the students feel at ease, even when he was teaching challenging topics. He always made sure that his students got most of the class and he was always fun to co-teach with. The biggest strength in Paul, we all in the team admired was his ability to deal with conflicting priorities in high-pressure situations. He never lost his cool and was always a people's person. Paul would be a true asset for any company who need a game changing Data Scientist.”

Swathi Nagendrain

Swathi worked with Paul Jialiang on the same team

“Paul is one of the best instructors in data science. He is super knowledgable and breaks down complex concepts into clear and understandable topics. He is a dedicated teacher and I learned so much from him in a short amount of time and I hope to continue learning from him in the future.”

Bahar Biazarin

Senior Data Scientist @ Ent Credit Union | AI/ML

Paul Jialiang was senior to Bahar but didn’t manage Bahar directly

“Paul was a substitute instructor for our Data Science Immersive program and he was absolutely amazing! He was easy to work with, the students loved his teaching style, and he showed a great depth of Data Science knowledge and teaching ability. I hope to collaborate with him again in the future and I would highly recommend him for any position on your team.”

Kristen Grewein

Strategy & Operations | Tech Partner Ecosystem Development

Kristen worked with Paul Jialiang but on different teams

“It was my pleasure to be in a class that his teacher is Paul Jialiang, he has all the skills that I feel it is really important that the teacher should have, it was easy for him to make the concept clear and what I really like about him, is he make us feel good about ourself spotting the light on our hard work and what we know and learned,, I felt happy after your class thanks again Paul”

Marwah Farajin

Paul Jialiang was senior to Marwah but didn’t manage Marwah directly

9 shown · swipe or use ◀ ▶

↑ Back to top

Score any job against past experience · current skillset · future mission/values/vision — held to a golden-set accuracy

Role Fit

Score a role against me

Paste a job-posting URL (Ashby/Greenhouse/Lever, fetched live) or the JD text. The agent scores fit across past experience, current skillset, and future mission/values/vision — and tells you, honestly, where it doesn’t fit.

Why trust this

8/8 within one band · 100%

This scorer is itself held to the standard it holds JDs to: it was run over a 8-example golden set with human-assigned fit labels and agreed within one band 100% of the time (6 exact). Verify, don’t vibe.

No role scored yet — paste a posting above to see the fit breakdown.

↑ Back to top

Knowledge graphs + agentic tooling (skills, plugins, workflows, hooks, bundles) I've built — full for me, a summary card for visitors

Agentic Library

AnyAgent — build · grade · improve · ship any agent apprepo🔒 private

AnyAgent is a tested, object-oriented engine designed to build, grade, improve, and ship agent applications from natural language.

Owner-only knowledge graph + tooling. Unlock owner mode to view the full package.

Harmonic — Research Engineerjob-description

Retrieved, not guessed: 3 competency clusters, 9 concept/tool nodes grounded in real sources (9 Wikipedia/GitHub definitions), and 6 skills from the ESCO taxonomy — off-domain hits filtered out. Every node carries a real definition; edges are real relationships where an open KG had them.

13 concepts · 12 links9 skills · 2 provenexplore the full graph →

Anthropic — Forward Deployed Engineer, Applied AIjob-description

Retrieved, not guessed: 6 competency clusters, 23 concept/tool nodes grounded in real sources (21 Wikipedia/GitHub definitions), and 8 skills from the ESCO taxonomy — off-domain hits filtered out. Every node carries a real definition; edges are real relationships where an open KG had them.

30 concepts · 29 links12 skills · 4 provenexplore the full graph →

OpenAI — Forward Deployed Engineerjob-description

Retrieved, not guessed: 5 competency clusters, 17 concept/tool nodes grounded in real sources (15 Wikipedia/GitHub definitions), and 9 skills from the ESCO taxonomy — off-domain hits filtered out. Every node carries a real definition; edges are real relationships where an open KG had them.

23 concepts · 22 links12 skills · 3 provenexplore the full graph →

Engram: Conditional Memory via Scalable Lookuprepo+paper

Transformers scale COMPUTE (Mixture-of-Experts routes tokens to experts) but have no native primitive for looking knowledge UP. DeepSeek's Engram adds that missing primitive: it modernizes classic N-gram embeddings into a deterministically-addressed table with O(1) lookup — a second, complementary axis of sparsity (static memory) alongside MoE's conditional compute. The paper finds a U-shaped scaling law: for a fixed budget there's an interior optimum splitting capacity between compute and memory — going all-in on either side loses. Under iso-parameter and iso-FLOPs constraints, Engram-27B consistently beats MoE baselines on knowledge, reasoning, code, and math. Mechanistically, offloading static recall to Engram frees the early layers from pattern reconstruction, preserving effective depth for reasoning — and the huge embedding tables offload to host memory with minimal inference overhead. Why it matters for you: 'conditional memory as a sparsity axis' is a transferable design move — separate what a system should COMPUTE from what it should LOOK UP, and size each.

10 concepts · 11 links2 skills · 0 provenexplore the full graph →

↑ Back to top

Four growth vectors — deepen · widen · lengthen · heighten — plus who to reach. Drafted for approval.

Next Projects

Runs weekly · last scouted 6/29/2026

Four growth vectors

↓ Deepen

more fundamental, seminal — to the roots

first-principles / foundational research

↔ Widen

new applications, features, markets

Ansoff · Innovation Ambition Matrix

→ Lengthen

evolve it to robustness, scale, commodity

McKinsey Three Horizons · Wardley evolution

↑ Heighten

generalize, abstract, compress the mechanism

abstraction laddering · compression (MDL)

↓ deepensos, loop-engineering-anything, company-os

Enhance Self-Improving Agentic Operating Systems

Further develop and refine the existing self-improving agentic operating systems to increase their efficiency and capabilities

first step: Integrate idle detection and SMARC output-quality verification from sos into loop-engineering-anything

The collaborator candidates were selected based on their relevance to the widenInterests and verified strengths of the builder.

Drafted for your approval — nothing is sent automatically. · model groq:llama-3.3-70b-versatile

↑ Back to top

A fifth vector — emergent projects from bisociation across the fleet. Scored in code; drafted for approval.

Bridges

Runs weekly · last searched 8/17/2026

Foundation Model Powered Agents

promising · 82

anyagent × FM-os

Autonomous Agents with Foundation Model Capabilities

pivot: Agent Training

first step: Integrate FM-os with anyagent for enhanced training

Graph-Based Agent Engineering

promising · 72

anyagent × graph-engineering-anything

Agents with Enhanced Graph-Based Reasoning

pivot: Graph Representation

first step: Integrate graph-engineering-anything with anyagent for graph-based agent engineering

Evaluation-Driven Agent Development

promising · 71

anyagent × eval-anything

Agents with Enhanced Evaluation Capabilities

pivot: Agent Evaluation

first step: Integrate eval-anything with anyagent for evaluation-driven development

Design-Driven Agent Development

plausible · 61

anyagent × design-anything

Agents with Enhanced Design Capabilities

pivot: Agent Design

first step: Integrate design-anything with anyagent for design-driven development

Loop-Based Agent Engineering

plausible · 51

anyagent × loop-engineering-anything

Agents with Enhanced Loop-Based Reasoning

pivot: Loop Representation

first step: Integrate loop-engineering-anything with anyagent for loop-based agent engineering

Forward-Deployed Agent Development

plausible · 50

anyagent × FDE-os

Agents with Enhanced Deployment Capabilities

pivot: Agent Deployment

first step: Integrate FDE-os with anyagent for forward-deployed development

Longevity-Driven Agent Development

speculative · 39

anyagent × longevity-loop

Agents with Enhanced Longevity Capabilities

pivot: Agent Longevity

first step: Integrate longevity-loop with anyagent for longevity-driven development

AI-Native Agent Development

speculative · 30

anyagent × ai-native-os

Agents with Enhanced AI-Native Capabilities

pivot: AI-Native Representation

first step: Integrate ai-native-os with anyagent for AI-native development

Emergent projects from bisociation across your fleet (Swanson A–B–C + conceptual blending), scored in code. Most combinations are noise — the ranking is the value. Drafted for you; nothing auto-built.

↑ Back to top

12X Future

12X the future, 12% the loop now

Amazing Art × Great Science × Deep Engineering

Love God · Love Family · Enable Friends

Projects

anyagent

FM-os

GodView-anything

rsi-os

graph-engineering-anything

eval-anything

design-anything

eval-anything-archive

agentic-portfolio

agentic-marriage

kingdom-come

loop-engineering-anything

MorganStanley-os

agentic-HR

agentic-house

FDE-os

longevity-loop

loop

Writing

Redlining Was Never a Sentence. It Was a Map.

Holy Is Not a Compliment

Breaking Rahab Is the Easy Part

When the Model Gets Hands

The Scale Furnace

The Reward Is the World

The Probability Forge

The Consequence Engine

I Scored 100% on My Own Checklist. That Was the Bug.

Green for Six Weeks

Everyone Is Building a Better Receipt

Everything Is a Plugin. The Gauges Are the Work.

Paul Jialiang Wu — Agentic Portfolio

The Cheapest Thing in This Mission Is the Truck

An AI Reviewed My Repo and Told Me to Make It Worse

One Opinion Wearing Ten Hats

How a Version Earns Its Number

The Filter That Refused My Own Paper

🤔 We’re not living through an information crisis. We’re living through a meaning crisis.

🧮 A math problem sat unsolved since 1999. ⏳ For 27 years, nobody cracked it. 🚀 Last week

Own My Intelligence: The Manifesto

Own My Intelligence — the paper (v0.1, and every version)

I Published the Manifesto I Killed

My Manifesto Could Not Fail. That Was the Bug.

The Other Road: Reverse-Engineering Closed Products Into Open Ones

Let the Model Write the Problem. Let the Solver Sign the Answer.

Silence Is Not a Status

My Agent Had Full Authority and Still Couldn't Click "Open"

270 impressions View analytics

We Hold Chatbots to Higher Prophetic Standards Than Prophets.

Love 12X Future U Scaling AI Systems: The Diner That Became a City Kitchen Paul Jialiang Wu, PhD

I Compiled 58 Snowflake Courses Into an Agent. It Failed Me First.

The Essay Proposed a Prophecy Ledger. The Code Shipped Seventeen Minutes Later.

Test Everything. Hold On to the Good. Now We Have the Tools.

His Life Review Surfaced a Door, Not a Doctorate. Matthew 25 Called It First.

In 0.1 Seconds He Reviewed His Whole Life. His Résumé Never Came Up.

I Fixed the Bug Three Times. The Bug Was Never the Problem.

We Built a Machine That Argues With Us. Its Best Feature Is the Word "No."

Every Cell Failed. That Was the Clue.

📚 Every AI study tool I tried this month does the same thing: summarize, podcast, flashca

Psalm 23: The Sheep Doesn't Need the Whole Map

Psalm 89: How to Keep Your Identity While Learning From Reality

Loop Engineering: Build the Loop That Catches You

The Most Important AI Paper of 2004 Never Mentions AI.

The App Kept Saying "Done." It Was Lying. So We Made It Audit Itself.

Hard Ideas Bounce. A 300-Year Recipe, Run by Agents, Makes Them Stick.

I Fed the Bitter Lesson to My Agents. It Ate My Study Method First.

Here’s a version centered on that vision: 🎮 What if the best way to solve tough technical

Your Jump Shot Is Your Syllabus. Welcome to Happy College.

The Next Developer in Your Repo Isn't Human. Six Practices to Get It Ready.

🔥 Math just had a very busy week. Four mathematicians under 40 conquered four problems th

数学界最近有点热闹：四位不到 40 岁的年轻数学家，一口气拿下了四座悬而未决半个世纪、甚至一个世纪的高峰。 听起来很燃，但先别急着膜拜。先建立一个简单的心智模型： 把数学想成一片山

Stop Asking “Should We Shard the Model?” Paul Jialiang Wu, PhD

Do the 2026 Fields Medals Matter for AGI and Quantum? The Member Edition Grades Every Claim.

The 2026 Fields Medalists: An Institutional Research Briefing (Interactive)

2026年菲尔兹奖得主：一份机构研究简报（简体中文全译·动画版）

数学界最近有点热闹：四位不到 40 岁的年轻数学家，一口气拿下了四座悬而未决半个世纪、甚至一个世纪的高峰。听起来很燃，但先别急着膜拜。先建立一个简单的心智模型：把数学想成一片山