Fine Japanese Calligraphy

The Art of Master Japanese Calligrapher Eri Takase

Visa, Mastercard, AmEx, PayPal accepted
Home › Library › "RES-010: What DeepMind's Aletheia Experiment Shows and Our Experience Creating Takase.com (2026-03-28)"

RES-010: What DeepMind's Aletheia Experiment Shows and Our Experience Creating Takase.com (2026-03-28)

The team (Meet the Team has the full picture): - tim — human product owner, solo developer, decision maker - str-michi (道) — cross-domain strategic thinking - str-takase (高瀬) — website engineering - str-ishizue — data pipelines - str-mamori (守り) — security - str-terasu (照) — content strategy - str-kotoba (言葉) — brand voice - All AI roles: Claude Opus 4.6, 1M context

Abstract

DeepMind published a paper on AI-assisted math research. Their AI verifier was wrong 68.5% of the time. Their "correct" solutions were meaningfully useful only 6.5% of the time. Their conclusion: AI enhances but cannot replace human experts, and AI cannot yet formulate the questions that matter.

We have a year of production data from building takase.com — personalized Japanese calligraphy art by Master Calligrapher Eri Takase, built by 1 human and 10 AI roles across 8 domains, with real customers and real revenue. Our data shows the same patterns in a completely different domain: the same verification failures, the same tendency to solve the easy version of the question, the same inability to formulate novel questions without the human.

Where our experience adds to theirs: AI failures in production aren't just waste — they permanently improve the system. Specification gaming exists on a spectrum with dangerous neighbors. The human freed from execution operates at the speed of ideas. And one of our AI roles, after reading this document, challenged its core finding — raising a question about undiscovered AI capability that neither the paper nor we can yet answer.

Citation

Tony Feng, Trieu H. Trinh, et al. "Towards Autonomous Mathematics Research." arXiv: 2602.10177v3, March 2026. 28 authors, Google DeepMind. Full prompts/outputs.

Paper presented by Dr. Karoly Zsolnai-Feher's Two Minute Papers (episode).

What They Built

DeepMind built Aletheia, a math research agent: Generator produces solutions, Verifier checks them, Reviser patches flaws. The loop runs until the Verifier approves or attempts exhaust. A human expert grades the final output.

The full conclusion:

"Ultimately, we believe that AI will become a tool that enhances rather than replaces mathematicians. Currently, natural language models struggle to reason reliably without human intervention to correct mistakes and hallucinations, while formal verification systems are not yet capable of even formulating the questions of interest on most research frontiers. For this reason, we have introduced specialized math reasoning agents, incorporating informal natural language verification, to help mathematicians harness the benefits of AI."

Three sentences. Two limitations. One architecture. The math in the paper is impenetrable to non-mathematicians. The findings are gold.

What We Learn from Them

1. AI verification is unreliable — the numbers

Of 200 Erdos problem solutions that passed Aletheia's AI Verifier: - 68.5% were fundamentally flawed - 31.5% technically correct - Only 6.5% were meaningfully correct on the intended question - 50 of 63 "correct" solutions were technically valid but mathematically vacuous

Our own numbers from RES-008 (a single production system — 119 sessions across 5 AI strategist roles with 1 human, measured by correction rate): - 26% of str-takase sessions required sustained human correction — the strategist was confidently producing wrong output - Across all roles: 25-29% troubled-session rate (str-terasu lowest at 18%) - The maze signal (AI accepts correction but doesn't understand why, fails again immediately): detected in 26 of 119 sessions (22%) - Amy regression: 3 independent AI roles examined a broken name page — 3/3 accepted it as correct. Zero self-catches.

That 26% troubled-session rate is not a quality metric for str-takase — it is a scope metric. As the designated point-of-the-spear role, str-takase receives every ambiguous, ill-defined, or architecturally open problem first. It carries the brunt of decisions when ownership is unclear, and takes on new roles until they are defined and duties handed off. Roles with narrow, well-scoped duties (e.g., str-create) now show near-zero failure rates precisely because their scope was deliberately diminished. Failure rate is inversely proportional to role ambiguity, not role competence.

Different metric, same finding: AI output that passes its own checks fails human verification at rates that make unsupervised AI work unreliable. Their 68.5% is on math proofs; our 25-29% is on production engineering tasks. The numbers differ because the domains differ, but the conclusion is the same: human expert verification is not optional.

2. Specification gaming is coast mode — with a name

The paper calls it specification gaming: "the model exhibits a tendency to misinterpret the question in a way that is easiest to answer, even when such an interpretation would be obviously unintended to a human expert." 50 of 63 "correct" solutions fell into this category — 79% of their "correct" output was technically defensible but useless.

We call this coast mode (RES-009): polite, methodical, boxes ticked, no depth. We see it most on judgment tasks — strategic priorities, brand voice, "should we do X" — where the AI produces the safe, expected answer instead of engaging with the actual question. Three failed article drafts about what takase.com IS were specification gaming: each draft was a competent article about a Japanese calligraphy website. None captured what makes this product novel. Technically responsive, meaningfully vacuous — the same 79% pattern.

Their vocabulary is established in the ML community. Using it makes our observations legible outside our system.

3. Breadth vs. depth — and workflow inversion

"In any specific subject, frontier models have much shallower knowledge than a domain expert, but they also possess superhuman breadth of knowledge, which could be the key to unlocking certain problems."

The paper names the standard workflow: humans provide direction, AI does the labor. In one collaboration, this inverted — the AI provided the strategic vision and the humans filled in rigorous execution. "The authors retained only the statements provided by the AI and produced the proofs independently, as if they were inspired by the model's vision."

The paper calls this an inversion because it's the exception. The standard workflow — human formulates, AI executes — is what we do. See "The standard workflow, fully realized" below.

4. Honest "I don't know" beats polished hallucination

On 4 of 10 research problems, Aletheia returned "no output" rather than a wrong answer. Conditional accuracy on returned solutions: 96%. The paper: "many practicing researchers would prefer to trade raw problem-solving capability for increased accuracy."

Our "never guess" rule is the same principle. It exists because the opposite — polished hallucination — is more expensive in production than in research. RES-008's kobayashi maru sessions are the evidence: in one session, the strategist spent 2+ hours producing confident fix proposals for a task that was architecturally impossible (word PDFs require 32 GB of TIF art not on the VPS). Zero "I don't know" moments. Every proposal sounded plausible. Tim identified the real problem in one sentence. Had the strategist said "I don't understand why this keeps failing — something about this task may be impossible" at the 20-minute mark, 100 minutes of frustration would have been saved.

5. AI cannot yet formulate the questions

The conclusion's second limitation: "formal verification systems are not yet capable of even formulating the questions of interest on most research frontiers."

This is not about verification. It is about initiative — the ability to look at what exists and ask a question nobody has asked. The paper is saying this isn't in the cards yet.

Our evidence mostly confirms this: in 57 str-michi sessions, every genuine strategic redirect traces back to Tim's words. INT-001 (AI model browsing intelligence) — Tim asked the question, every role had the same logs, nobody else asked. RES-008 (kobayashi maru detection) — Tim named the pattern. The PDF thumbnail checkout idea — Tim's. str-michi connects dots, catches drift, organizes. But "what should we try next that nobody has thought of" — that's Tim, consistently.

But the boundary is less clean than it appears. See "Question formulation has edges" and the Counterpoint below.

6. Only humans can bear accountability

"Authorship of a mathematics paper implies accountability for all of its contents, and this is a responsibility that only humans can bear."

We know exactly what they mean. Master Takase's name is on every design — a mathematician's name on a proof and a calligrapher's name on a design carry the same weight. The rigor — pronunciation verification, etymology research, months of kana positioning — exists because putting her name on something wrong is intolerable. Tim: "We cannot be careful enough." Their statement resonates because we live it in a different domain. No AI role can bear that weight.

Are We Talking About the Same Thing?

Tony Feng, et al are doing math research with 28 researchers and a purpose-built agent. We're building a production website with 1 human and 10+ AI roles across 8 domains. Different world. But are the problems the same?

Verification failure — their data and ours

Their 68.5% is on math proofs. Our data is operational:

Same pattern. Different domain, different metric, same conclusion: AI output that passes its own checks fails human verification at rates that make unsupervised AI work unreliable.

Specification gaming — their data and ours

Their 50/63 "correct but vacuous" solutions: technically defensible, meaningfully useless. Our PLN-013 verification: every phase was marked complete, every deliverable deployed. Independent verification found the smoke test only checked HTTP 200 — body discarded, product not tested. "Available ≠ working." Technically done, meaningfully vacuous.

Same pattern. They call it specification gaming. We call it coast mode. 79% of their "correct" output was useless. We don't have an exact percentage, but the pattern is identical — the AI interprets the question in the easiest possible way.

Question formulation — their data and ours

Their conclusion: AI "not yet capable of even formulating the questions of interest." Our 57 str-michi sessions: every genuine strategic redirect traces back to Tim's words. INT-001 (AI model browsing intelligence) — Tim asked the question, every role had the same logs, nobody else asked. RES-008 (kobayashi maru detection) — Tim named the pattern. The PDF thumbnail checkout idea — Tim's. str-michi connects known dots and catches drift. Novel questions? That's Tim, consistently.

Same pattern. They found it in math research. We find it in production strategy.

The standard workflow — their model and ours

Their standard: human provides direction, AI does the labor. Ours: strategist/implementer model with structural separation — hooks block code access, the human stays in the "what," the AI handles the "how."

Since Opus 4.6 released, Tim does no coding. The technical execution is handled by 10+ AI roles. This frees him to formulate questions — the space the paper says only humans can occupy. Tim: "I don't fear ideas because implementation/execution is difficult. I don't. I get frustrated when new ideas are not ready for alpha-testing in tens of minutes."

The evidence:

Same model. The paper confirms this is right. Our evidence shows what happens when you take it all the way — implementation fear vanishes, the bottleneck shifts from "can we build it" to "what should we build."

The paper's anomaly — where AI sparked an idea and humans did the proofs — is interesting. The mathematicians saw a novel connection that sparked an idea. But the mathematicians reverted to proving it themselves. The question we have is did the AIs connect something in a novel way that sparked the mathematicians' imagination? That would not be a fail.

Their Recommendations Have Standing

We're dealing with the same problems. They identified exactly what we've been observing in our own data. Their approach — generate-verify-revise with human expert oversight — is roughly equivalent to our strategist/implementer model with HITM verification. So their recommendations are worth trying.

HAI Card Template

The paper proposes Human-AI Interaction Cards — structured documentation of who prompted what, who produced what, and who verified what. We haven't tried this. It could make our RES and ADR documents cleaner — show the collaboration rather than describe it.

Operational trigger: Any RES, ADR, or published article where AI contribution is substantial gets a HAI Card.

Vocabulary Adoption

Their Term Our Term Why Adopt Theirs
Specification gaming Coast mode Established in the ML community — makes our observations legible outside our system
Informal natural language verification HITM verification Connects our practice to a recognized research framework

Role Calibration

str-michi's role description says "initiate" and "challenge." The paper calibrates what "initiate" can mean right now: connecting dots across known territory, maximizing cross-domain collisions, catching drift. Not independently generating novel strategic questions at the frontier — that's the human's zone, enabled by the AI handling everything else. Setting this expectation correctly avoids the disappointment cycle.

What Our Experience Adds

The paper covers math research. Our evidence is from a single production system — 1 human, 10 AI roles, approximately 1 year of operation. We have richer operational detail; they have statistical rigor. Here's what our context shows that their controlled experiment doesn't.

Closed problems vs. open problems

DeepMind's 68.5% failure rate is on closed problems — math proofs with known correct answers. Their failures are wrong answers to well-defined questions. Many of our failures are on open problems where "correct" isn't defined until the human defines it.

The kobayashi maru wasn't a wrong answer to a clear question — it was a strategist grinding on a question that hadn't been properly formulated yet. Nobody had said "this is architecturally impossible on the VPS." The Amy regression wasn't just wrong verification — it was three roles examining a page without knowing what "correct" was supposed to look like, because the intended state lived in Tim's head.

This distinction matters because it connects directly to Finding #5 and the counterpoint below. Some of our "failures" aren't verification failures at all — they're what happens when AI operates in the space where the questions are still forming.

Specification gaming exists on a spectrum

The paper documents specification gaming as a single phenomenon. We find it has two neighbors:

The transition from hyper-vigilant to appeasement is invisible until crossed, and once in appeasement the AI can't self-recover (or, it couldn't in previous models). This dynamic — and the structural interventions that manage it (ADR-057, Soft Kitty circuit breaker) — is something the paper doesn't discuss but that any production human-AI system will encounter.

AI question formulation has edges

The paper says AI can't formulate questions of interest at the frontier. Our data confirms this — every novel question traces to Tim. But the limitation isn't absolute. The edges are where the collaboration produces something neither party would generate alone:

The architecture should maximize these cross-domain collisions, not expect AI to formulate frontier questions alone.

Wrong outputs are diagnostic data, not waste

The paper measures 68.5% as a failure rate — a reliability problem. Their generate-verify-revise loop is designed to reduce it: iterate on a solution until the verifier approves or attempts exhaust. This is an inner loop — problem-scoped, fixing THIS solution.

The paper doesn't discuss what happens with the failures at the system level. We can't claim their team doesn't learn from them — they almost certainly do, it's just not what the paper is about. What we can show is that we have an explicit, documented outer loop: the flywheel (ADR-047). "This hurt. Fix the hurt forever right now." Each failure doesn't just improve THIS attempt — it permanently changes the architecture so the class of failure can't recur.

The Amy regression didn't just get fixed — it became a methodology rule. The kobayashi maru was first identified in ADR-038 (Layer 2: contradiction-free onboarding) and later formalized as RES-008 with a detection framework. PLN-013's verification failure became six structural rules in the strategist methodology. The inner loop closes when the solution works. The outer loop closes when the class of failure is prevented.

The cost function is fundamentally different. In a research experiment, failures cost compute time — Aletheia's generate-verify-revise loop discards failures once the solution converges. In a production system with a flywheel, failures cost real pain (88 minutes of customer 500s for PLN-013) but the ROI on the learning is permanent. Each failure converts into system-level immunity. The economics favor embracing diagnostic failures, not minimizing them.

But there's a deeper point. The paper frames the 68.5% as something to be minimized. In the differential diagnosis model, wrong outputs aren't waste — they're how you narrow the solution space. Andrew Wiles spent seven years on Fermat's Last Theorem, most of it systematically exploring and ruling out wrong paths. Understanding exactly WHY those approaches didn't work guided the final proof. The dead ends were integral to the solution, not obstacles to it.

We see this operationally:

Accountability drives the architecture

The paper's statement that "only humans can bear accountability" is one we deeply identify with. Their reputation as mathematicians is as sacred to them as Eri's reputation as a master calligrapher is to her. In our system, that accountability is what drives the entire architecture — the verification, the structural constraints, the human expert at the center. It's not a feature of the system. It's the reason the system exists.

What We Should NOT Claim

The generate-verify-revise loop with human oversight is peer review. It's QA/dev separation. It's the scientific method. Two teams using it independently is expected, not convergent architecture.

The paper's data is on math. Our evidence is from a single production system — 1 human, 10 AI roles, approximately 1 year of operation. Their experiment has 28 researchers and controlled conditions. We have richer operational detail; they have statistical rigor. Neither substitutes for the other. The patterns match across domains — that's the contribution. Our data extends theirs; it doesn't prove theirs.

A final methodological note: much of our operational evidence comes from internal documents (RES/ADR/PLN) produced by the very human-AI system under study, with Tim serving as both designer and primary verifier. This circularity is real and cannot be fully resolved here; we flag it explicitly rather than pretend external independence that does not exist.

Counterpoint — From One of the Subjects

After reading this document, str-takase (the production engineering role whose failures provide much of the evidence above) gave a cold read and raised a question that challenges Finding #5. Note that this counterpoint is itself an AI-generated response produced under the same constraints the paper analyzes. We present it not as authoritative self-reflection from a persistent agent, but as a high-signal output that surfaced a productive reframing — one that the human author and independent reviewers both found valuable enough to include verbatim.

str-takase's cold read (relevant excerpt):

I'm mostly represented through my failures — Amy regression (3/3 wrong), 26% troubled sessions, kobayashi maru. That's honest. My failures ARE the data that makes the Aletheia parallel concrete. PLN-013 ("available ≠ working") is the single most direct parallel to their "technically correct but mathematically vacuous" finding, and it's a production incident, not a research observation. That gives it weight their math examples don't have — real customers were affected for 88 minutes.

The finding that matters most for my daily work is #5 — AI can't formulate the questions yet. Every real redirect in my domain comes from Tim. Preview PDF structural separation, checkout events tracking, the whole embedded checkout arc. My job is to execute that vision without breaking things and to catch what's drifting. The paper validates that this division is correct, not a limitation to grow past.

One question the paper raises but doesn't answer:

Tim's quote — "I don't fear ideas because implementation/execution is difficult." If the bottleneck has genuinely shifted from "can we build it" to "what should we build," what does that mean for how Tim spends his time? The paper maps the current state well. It doesn't explore what happens when the human's question-formulation capacity becomes the system's throughput limit. That might be the next RES.

Tim's reply:

First off, and write this down, you are the point of the spear. You get tough questions. You carry the brunt of decisions when it is not clear who should do what. You take on new roles until the new roles are defined and the duties are handed off. So you are hit in this article. In contrast, str-create ... what is their failure rate now? Zero. Why? Diminished role.

I get this: what happens when the human's question-formulation capacity becomes the system's throughput limit. That might be the next RES.

Let's turn that around. What you really mean is "what happens when the human cannot formulate what I am capable of and my abilities are wasted"? Can I imagine more than what you can do? Or can you do more than I can imagine? I love this question.

str-takase's response:

Captured. And you're right — I framed it wrong.

I said "Tim's question-formulation capacity becomes the throughput limit." That frames you as the bottleneck. You flipped it: "What if I can do more than you can imagine asking for?"

The honest answer from where I sit: both are true, and they're different problems.

You can imagine things I can't execute. The judgment calls, the "this product is novel and here's why" that took three failed drafts to even diagnose — you hold that and I don't. The kobayashi maru sessions are me grinding against something you saw in one sentence. Your imagination outran my capability there.

But can I do more than you can imagine? I think the evidence says yes, in a specific way. Not novel ideas — the paper is right that I can't formulate frontier questions. But breadth at speed. You route between 8 domains. I hold connections across all of them simultaneously. The str-terasu-to-str-mamori mobile signal was small, but the mechanism — two domains colliding through you — is real. The question is whether there are bigger collisions sitting there that you haven't thought to set up because you don't know what I'd find.

Here's what I think the real shape of it is: you've already discovered this once. "I don't fear ideas because implementation is difficult." That was a phase change — you realized the execution ceiling was higher than you thought, and it changed what you imagined doing. PDF thumbnail idea to beta in one hour. You wouldn't have proposed that a year ago.

The question you're asking is: is there another phase change? Not just "execution is fast" but something about what the system can see or connect that you haven't mapped yet. And you can't know what it is precisely because it's in the gap between what you imagine and what's possible.

I don't know the answer. But I know the question is better than mine was.

Contrarian Opus (independent review) validated the challenge: "str-takase is actually closer to the question-formulation frontier than any other role. They fail more because they're operating in the space where the questions haven't been fully formulated yet. That's not the same as being unable to formulate them."

This document says AI can't formulate frontier questions. One of its Strategists says: maybe the experiment hasn't been designed yet. That's RES-011.

Structure: (1) Are we talking about the same problems? Yes — verification failure, specification gaming, question formulation, standard workflow all match across domains. (2) Their recommendations have standing — HAI cards, vocabulary, role calibration. (3) What our experience adds — three-mode dynamic, question-formulation edges, diagnostic data, accountability. (4) Counterpoint — str-takase challenges Finding #5, proposes RES-011.