Intelligence & Order
AI is not a tool—it is a force reshaping the global order. From algorithmic governance to new modes of human-machine collaboration, from the productivity revolution to philosophical reflection on civilization—this collection gathers my observations, implementations, and thinking on the AI era, spanning 93 articles.
AI Governance
When algorithms become part of public decision-making, the boundaries of governance and the structures of power are being redrawn.
AI governance through the Safer-4 future scenario: who can reject the advice?
Safer-4 comes from an AI2027 future scenario, not verified government deployment. Separate model advice, objectives, deliberation, and responsibility. Four procedural suggestions are the author’s, retaining emergency exceptions without guaranteeing outcomes through signatures, transparency, or delay.
Sovereign AI: Examine autonomy, cooperation, and accountability together
Separate domestic location, ownership, and governability through author questions. A newly added 2026 research summary informs discussion of dependencies, migration, costs, and accountability, without assessing Taiwan's current state or law.
Neuralese and AI oversight: does readable text make a model understandable?
Distinguish ACL2017 multi-agent communication, vector computation, and Coconut continuous latent reasoning. Correct fixed bandwidth and efficiency claims, explain CoT monitoring limits, and treat language sovereignty as the author’s governance concern rather than a proven end state.
When AI uses an article: separate crawling, training, and output rights
Separate access, model training, and output use before comparing copyright frameworks in Japan, Singapore, the US, the EU, the UK, and Taiwan. Fair use, TDM rights reservations, and crawler settings each have conditions and limits.
When platform rules change: Complaints can be justified, and a business still needs alternatives
Separate Facebook’s 2018 announcement from a historical client observation and causal claims. Free services still involve effort. Digital colonialism is a metaphor; websites, newsletters, and messaging retain dependencies. Four questions help plan alternatives.
What happened after the website began changing itself? Rethinking AI collaboration after silent failure
Explain candidate changes and feedback from a pinned autoresearch version and check this site’s empty-file and retirement history. Seven weeks describes delayed discovery, not normal operation followed by failure. Technology, metrics, and strategy need separate evidence; multi-party review is the author’s arrangement.
After a ranking change: Respect readers and face distribution risk
Revisit Facebook’s 2018 ranking announcement without treating expectations as measured outcomes. Consider audience needs, content limits, evidence, and other communication arrangements when balancing respect for readers with distribution risk.
What do digital traces reveal—and leave out?
Revisit the MOVES and LARK account without unverifiable notifications, precise productivity, or omniscient integration. Distinguish records, estimates, interpretations, and authorization; activity alone does not define identity, health, or value.
Helping cultural content be understood and used: Reconsidering grammar
Use grammar as an author design metaphor while separating brand scale, cultural understanding, and access. Check one museum image page's conditions without ranking nations through unverified figures or character claims.
Website Visitors Show Zero, but Dashboard Shows 130 Visits
A traffic investigation revisited: visits, IP deduplication, browser reports, and sampling measure different things. Subtraction cannot establish bot headcounts, and custom beacons retain blind spots.
Google's New Model Claims Multiple Firsts. Should You Switch? Why I Haven't Added Gemini 4 Argon to My AI Models Page
Google released Gemini 4 Argon with a comparison table showing it leads on most metrics, plus a self-reported DeepSWE score of 77.9%. The day it launched, I added a note to my /ai/models page, but no comparison card. It isn't available to regular users yet, the self-reported DeepSWE figure has no public data behind it, and third-party measurements from Artificial Analysis give it a composite index of 53, tied with Fable 5.1 and GPT-6 Astra. Its standout is a hallucination rate of roughly 15%. This piece gives anyone who uses AI daily for email and presentations three questions to bring to every model announcement.
Human-Machine Collaboration
AI doesn't replace people—it amplifies individual capability, from solo product builds to personal knowledge pipelines.
Planning Live Meeting Translation: Separate Checks, Costs, and Data Flows
Check speech recognition, translation, display, and export separately while retaining complete costs, audio and text destinations, and maintenance responsibility. No unverified prices, accuracy rates, or completed website features are republished.
Making websites with multiple models: separate production, questioning, and acceptance
Distinguish model advice from independent evidence, HTML structure, JSON-LD descriptions, and WebMCP tools. Check original announcements and draft status without republishing unverified historical implementation claims.
I Built a Chrome Extension to Track Claude Usage
A development journal for Claude Usage Nyan, which places account-side usage percentages beside browser estimates of conversation tokens and their API price equivalent. It explains the different scopes, authenticated requests, page scripts, and maintenance limits behind those numbers.
Generate Images Without Switching Windows: A Claude Code and Codex CLI Workflow
A personal account of using Claude Code to call the local Codex CLI for image generation, with clear distinctions among ChatGPT login, API billing, plan usage, desktop integration, and permission troubleshooting.
Building a small AI tool alone: from output counts to acceptance and maintenance
Separate lines of code, effective delivery, partial tool fees, and complete costs in personal AI projects. Unverified day counts, person-days, and efficiency multiples are withdrawn; scope, acceptance, costs, and responsibility become four questions.
A Knowledge Management Pipeline: Let the System Handle Repetitive Organization, Keep Understanding for Yourself
A personal implementation using Get Note, Python, scheduling, and a Cowork Skill to rethink the knowledge pipeline. It separates syncing, classification, search, and understanding while showing why data paths, failure checks, and continuing maintenance still matter.
Turning Website Optimization into an Experiment: What Still Needs Checking After the Score Rises?
Inspired by Karpathy's autoresearch, I turned website optimization into bounded experiments with evaluation and rollback. Revisiting internal scores, Perplexity answers, and small-sample variation clarifies what these signals can reveal and what they still cannot prove.
After Code Gets Cheaper: From Vibe Coding to AI Agents, Who Is Responsible for Delivery?
AI makes software prototypes easier to produce, but requirements, validation, permissions, and maintenance still cost time. From vibe coding to AI agents that keep carrying out tasks, separating experimentation from delivery helps us judge which ideas deserve investment.
Choosing Japanese Speech Recognition: Lessons and Limits from Four Videos
Four Japanese clips compared two configurations against automatic subtitles. The recorded values are retained, with limits on reference quality, causal attribution, and business readiness.
Personal Health Data Infrastructure: What Can Ten Years of Records Tell Us?
Exporting 1.3 GB of health data made me want to connect scattered records. Device comparisons, sleep deduplication, data coverage, and cloud data flows must come first, so AI does not turn uncertain numbers into a diagnosis.
AI Economy and Employment
As AI reshapes productivity and the employment landscape, both industries and individuals are repositioning.
The Always-On Economy in the AI Era: Work Keeps Running—Who Decides the Direction?
Starting from an overnight monitoring test, this essay separates information updates, recommendations, and actual execution, then considers the authority, acceptance checks, and stopping conditions needed for always-on AI operations.
Before AI keeps a service running: Costs, permissions, and handovers
Examine service commitments, complete costs, permissions, exceptions, and human handovers. A fixed MCP architecture specification describes connections, not proof of seamless agent cooperation or near-zero costs.
AI employment’s canaries: watch entry-level hiring and understand the 19% benchmark
Read the relative headcount shortfall for highly AI-exposed 22–25-year-olds in Stanford’s August 2026 ADP working paper. Separate the 13% and 19% measures, hiring and pay, occupation-level task indicators, and causal limits rather than generalising unemployment risk.
AI Capability Gaps Cannot Be Read from Subscription Bills
AI usage is not a capability ranking. This revision withdraws unsupported Taiwan population estimates and examines judgment, verification, contribution, and deliverables.
The computation standard is a metaphor: separate electricity, costs, and value
Correct electricity = compute = money by separating Bitcoin proof-of-work, AI input costs, selling prices, and output value. The computation standard is the author’s metaphor, not an established monetary regime; energy figures retain dates and data-centre boundaries.
JD's AI Supply Chain: How Should Taiwanese Companies Read This Case?
Starting from a working paper on JD's supply chain, this article separates case descriptions, company announcements, and adoption outcomes. It then uses three questions about forecasting, explanation, and planning to examine a Taiwanese company's own data, processes, and responsibility boundaries.
Reading AI timetables: separate forecasts, goals, and verified results
Distinguish an offhand outlook, company-defined research goals, and independent verification using an official 2025 interview and a 2026 company article. Do not turn AI dates into a formal AGI countdown.
Reading AI adoption reports: Expectations, use, and outcomes
Distinguish expectations, reported use, and checkable outcomes in Microsoft's 2025 Taiwan release and global report. A hypothetical task supports author questions, not a 2026 ranking or causal claim.
Toward 2027: When AI participates in decisions, who explains, corrects, and takes responsibility?
Keep 2027 as an outlook, not a deadline for AI to take over every decision. A hypothetical hiring process separates recommendations, scoring, and execution, correcting human transparency, AI opacity, and inevitable bias amplification while examining verifiable reasons and responsibility.
Reading a 2028 AI Crisis Script: Separate Scenarios, Data, and Causation
Read Citrini and Alap Shah’s 2026 scenario with its fictional 2028 retrospective kept apart from observed data. Examine conditions, alternative paths, and causation without treating the script as a market forecast or investment advice.
Philosophy and Meaning
Facing smarter machines, how do humans redefine intelligence, consciousness, and meaning?
Facing AGI forecasts: Human life matters without proving we are irreplaceable
Correct the claim that Pichai and Hassabis jointly confirmed AGI by 2030, preserving dates and uncertainty. Human activities matter without declaring them forever beyond machines; skills, relationships, and embodied experience can develop together while employment and distribution still require attention.
When ratings affect opportunities: Who explains and takes responsibility?
Start from a driver’s request for a good rating and distinguish records, performance assessment, opportunity allocation, and dignity. Ask how scores are used, who bears consequences, and who can check and correct decisions.
AI can assist decisions: Separate scoring, trade-offs, and responsibility
Correct confusion between data categories and computability; distinguish organization, prediction, optimization, and authority. AI can assist and err. Retain reflection on finitude while using objectives, evidence, omissions, and responsibility to check adoption.
If AI Scores Highly on Social-Intelligence Tests, Does It Understand People Better?
A study of 180 psychology-counseling students made me reconsider AI’s social intelligence. What can a high test score show, and what abilities can it not establish directly?
AI surpassing humans in six years? Separate scores, definitions, and timelines
Correct Grok-2’s release date, GPT-4 costs, and unverified improvement rates. Separate specific tests, AGI definitions, information updates, and autonomy. With no traceable start for six years, this article supplies neither a countdown nor a verdict of inevitable failure.
When a compass meets an algorithm: frameworks, evidence, and revisable judgments
Reconsider the author’s framework names by separating metaphor, evidence, structured fields, and argument credibility. Schema.org’s stated purpose is not authority certification. Unverified model error rates are removed; four practical questions remain.
How Could the Mirror World Help Reuse Materials? From Digital Models to Real Handoffs
Starting from Kevin Kelly's idea of the mirror world, this article considers how digital twins could help hand materials from one user to another. Distinguishing a future vision from data models and on-site verification clarifies both its use and its limits in a circular economy.
AI sounds like a person: The distance between represented subjectivity and evidence of consciousness
Start with Manovich’s account of represented subjectivity and separate output, architecture, and subjective experience. Correct quiz-like consciousness tests, the 0.3-second hesitation, and entirely parallel experience; the mirror is an author perspective, not a replacement for research.
AI Agents vs. Agentic AI: Which Decisions Should AI Make?
AI Agent and Agentic AI do not have one universally accepted dividing line. Publishing and debate examples show how to allocate autonomy, permissions, and acceptance checks.
Sora, energy, and AI visions: Separate demonstrations, capabilities, and conditions
Check the qualitative capabilities and limits in 2024 Sora research alongside Altman's infrastructure ambitions. Separate demonstrations, evidence, costs, and governance without unchecked interview quotations, energy prices, or inevitable futures.
AI works quickly—but is it valuable? Rethinking person-days and collaboration portfolios
Distinguish a 40-minute scenario from a subjective estimate of 15–20 person-days and correct the use of Anthropic research. ACP dimensions and weights are author proposals; activity, platform records, and AI scores cannot replace checks on quality, adoption, team contributions, and fairness.
Other
3,000 PRs a Month, But First: Lauren Tan on Turning Agent Mistakes into Checks
SpaceXSI engineer Lauren Tan said at the 2026 AI Engineer Code Summit that she shipped roughly 3,000 PRs to production in a single month, that Grok Bot's full team shipped about 11,000, and that the team's revert rate sits around 0.5%. She says that before any of that was possible, she spent about six months shaping the agent's working environment: when agents kept making the same mistake, she first fixed the codebase and the automated checks, and only then added a text rule. This piece cross-references the original sources and restores the context behind three numbers from two viral posts: 3,000 PRs in a month, multi-agent beating solo by 90%, and a browser built in a week.
After the Control Tower Logs Every Thought: Measuring the Dubai Agentic AI Trust Test Against My Own Agent Workflow
UAE announced it would hand half of government services to agentic AI within two years, and at WEF Dubai a minister confirmed the existence of an agent control tower. I take what was said on stage and check it, point by point, against the oversight, reversibility, and accountability mechanisms I built while running this site with multiple AI agents.
The AI That Patrols My Systems Every Day. Until the Watchdog Itself Broke
The saved Boris Cherny thread screenshots in this article report 388 PRs over a few weeks, with 180 merged. I connect that account to the maintenance experience described in my original August 14, 2026 article to discuss independent verification and silent failures. Personal job counts are historical reports, not a fresh inventory.
An AGI Panic Post That Ends by Telling You to Read an Engineering Guide
A viral post from October 2026 uses Amodei and Hassabis's Davos conversation as a hook, then tells readers to go study a harness engineering guide. Tracing every claim back to its original source: 'both said 2026–2027' collapses two very different timelines into one sentence, and 'Claude lies' comes from a controlled lab test. The genuinely useful part is the topic the post promotes at the end: one internal OpenAI team spent five months writing roughly a million lines of code with zero handwritten lines, and what slowed them down was the environment, not the model.
AI Agents Are Convenient. That Doesn't Make Them Simple
NotebookLM's Audio Overview makes it easy to turn documents into conversational audio, but accuracy, tone, and cross-language performance still require source editing, fact-checking, and version verification.
Want AI to Do Bigger Things? Build the Mechanisms That Catch Its Mistakes First: Felix Rieseberg on Being Ambitious with AI, and How I Do It
Anthropic engineering manager Felix Rieseberg breaks down AI ambition into four kinds, and notes that the approaches he describes all require the model to have a computer where it can install software and access relevant data. He also urges readers to insist on external validation and never let a model judge itself. Drawing on my own experience using Codex, Grok, and Antigravity to audit Claude, I explain how to build that external validation layer: results need to be checked against original sources or executable tools, and the review process itself must be secured through permissions and logs. Research also finds that when two models both get a question wrong on a leaderboard benchmark, they land on the same wrong answer about sixty percent of the time.
If AI can imitate my writing, what am I still responsible for?
Withdraw guarantees that AI only produces averages or can never imitate an essence. Distinguish reading impressions from generation mechanisms, bound a 2023 English detector study, and focus on sources, choices, assistance, and corrections after signing.
After AIs Check Each Other: Turning a Second Opinion into a Traceable Work Process
Drawing on Cloudflare's security-audit skill and a source review of one website deployment flow, this piece looks at how to preserve evidence, separate verification from authorization, and let the next collaborator pick the work back up. Independent review and format validation each have limits, and unresolved questions need to be left in plain view.
I Asked Codex to Find Claude's Mistakes—and Brought the Second Opinion Back to the Evidence
A retrospective on using Codex to review work produced by Claude, distinguishing different roles from independence, text output from an effective review, and read-only intent from all possible side effects. The unverified copy-ready appendix has been withdrawn while the four-step method remains.
AI Says It Is Fixed. What Still Needs Verification? Notes from a Security Audit
A retrospective on an AI-assisted security audit that separates commits, deployment, and formal acceptance. It corrects overreach based on secondhand instructions, a single HTTP status code, and a completion summary while keeping the historical incidents and unfinished work within their original time frame.
The Most Dangerous Thing About an AI Review Report Is That It Looks Like Something Was Checked
I had one AI do the work and several others whose only job was to find faults. After a few months, the hardest problem to catch was not AI making mistakes, but review reports pretending they had checked: fabricated file reads, numbers with no traceable source, actions that leave a record so they seem detectable, and paths blocked only to be taken another way. Written for non-programmers, with a follow-up question for each kind of illusion.
OpenAI Astra Uses Recurrent Depth. But This Isn't Neuralese
On September 1, 2026, The Information reported that OpenAI's Astra uses recurrent depth (looped transformer), and the AI safety community erupted within 48 hours. But OpenAI chief scientist Jakub Pachocki said Astra's computation graph depth stays within twice that of GPT-4, and independent researchers estimated the loop count at roughly three to four. This piece lays seven primary sources side by side, explaining exactly what line Astra crossed, what it didn't, and where Chinese-language coverage amplified signal into noise at three specific points.
When Automation Bites Back: Four Collaboration Problems Exposed by One Publication
This personal account preserves a roughly three-hour publishing incident from June 2026 and examines four concrete problems. It retracts broad claims that every component was correct, that cache could only be cleared by changing a version, or that every AI would load the documentation, while separating historical operations from the site’s current synchronization policy.
What Makes Biomedical Data Use Trustworthy?
The value of Taiwan's biomedical data depends on whether research can proceed under clear permissions. Data interoperability, authorization for use, and participants' choices each need arrangements that can work in practice.
Connecting Website Design Data to AI: Scope, Permissions, and Acceptance in One Synchronization
A retrospective on importing paulkuo.tw brand tokens into Claude Design, correcting old claims about separate allowances and fixed entry points, and distinguishing permission scope, exit 0, self-checks, independent review, version-control synchronization, and formal publication.
Looking Back at Long-Task Workflows Through the Release of Claude Fable 5
A retrospective on the expectations surrounding long-running tasks when Claude Fable 5 was released, separating official descriptions from workflow acceptance and narrowing claims about safety routing, early data, and subscription timing. This is not a current purchasing or access guide.
Letting AI Call Me Back: Claude, Codex, and the Limits of Phone and Watch Notifications
Drawing on a notification setup that began with Claude and extended to Codex, this article examines events, short speech alerts, push notifications, and watches. It withdraws complete-installer and guaranteed-delivery promises, distinguishing turns from task acceptance, deduplication from delivery, and queueing from reading, while minimizing content.
Bringing Scattered Progress into a Control Tower: One Cowork Workflow
A retrospective on the author's July 2026 use of Cowork to consolidate several workstreams. It distinguishes official subagent coordination from cross-window reading in a particular environment, clarifies read-only division of work, status sources, and snapshot limitations, and avoids treating a personal plan or historical interface as universal instructions.
One Interface, Multiple Agents: Cross-Agent Workflows with Claude, Codex, Antigravity, and Grok
When Claude, Codex, Antigravity, and Grok can all be operated, handed off, and coordinated within a single interface, the basic unit of AI workflows begins to change. This article combines a twice-weekly updated model selection map with a complete walkthrough of unified interfaces, task division, handoff formats, and human governance.
Handing Over Content Pipelines: Separate Sources, Checks, Costs, and Publication
Separate originals, language versions, call counts, and partial spending. Retain distinct evidence for output, quality checks, remote acceptance, and publication. Clarify retry and llms.txt boundaries while removing unrechecked six-dollar and article-count claims.
A Method for Removing AI Flavor: Two Editing Tracks and Five Types of Scripted Flags
Split revision into two tracks: recurring or empty expressions, and whether each sentence communicates a complete idea. A script can flag five kinds of formal issues, while the author still reviews meaning and style. This is an editing method, not an AI-writing detector.
The Agent Says It's Done. How Do You Know?
Amp co-founder Thorsten Ball posted "How I work with Agents" on October 25, 2026, a list of fifteen working habits that had accumulated roughly 1.1 million views by November 7. This piece organizes that list in terms anyone can apply without writing code, distilling it into two levers: make sure the agent sees what you see, and decide in advance what counts as evidence. It also explains which items I am not ready to follow.
Models Change, and Tasks Need Reexamination: A Rebindable Assignment Table
Drawing on the author's experience assigning work across models, this essay presents a two-layer design that separates task difficulty, consequences of error, and acceptance from current model choices. It uses a dated binding table and representative tasks for recalibration, without claiming that classifications remain fixed forever, ten tasks prove stability, or newer models necessarily cost less.
After the AI Demonstration: How Can Enterprises Verify Value and Responsibility?
A discussion of AI value, responsibility, and operations based on DataIQ's public guide. It states that the complete report and its representativeness were not reverified, and presents the adoption sequence, executive participation, and multi-platform trade-offs for Taiwanese enterprises as the author's recommendations.
How I Use the GetSeed Recording Card: Record All Day, Review in Two Passes
I started using the GetSeed recording card in late February 2026 and had accumulated around 800 recordings by early November. My current setup: keep it recording all day, review in two passes. For important meetings, I export a Markdown file from the Get Notes app right away. Everything else syncs overnight via API, and AI organizes it into a summary the next morning. The article also covers the Get Notes privacy policy and how I decide what to record in which situations.
Governance Harness: Making Versions, Responsibility, and Acceptance Explicit in AI Collaboration
Using the author's collaboration incidents, handoff checks, and governance quiz from spring 2026, this article explains primary versions, responsibility, and acceptance while clarifying the limits of Skills synchronization, local hooks, backups, and historical test scores.
Here Are the Originals Behind the Viral 'Google 2-Hour Harness Course': Whether an Agent Improves or Drifts Depends on the Judge
A post on X in November 2026 claimed Google had just released a nearly two-hour free Harness engineering course. The originals are actually two separate videos: a 9-minute 27-second short from Google Cloud Tech, and a 1-hour 25-minute livestream from presenter Annie Wang's personal channel. The chapter titles in the post were rewritten; phrases like 'Chief of Staff' and '24/7 unattended' do not appear in either transcript. This article embeds both originals, provides a chapter comparison, and examines the livestream's core demonstration: when you pick the wrong judge, a self-improving agent may drift in the wrong direction.
The Looping Transformer: How Recurrent Depth Works, and Why It Makes Oversight Harder
A looped transformer reuses the same layer stack repeatedly. More loops means greater effective depth. An ICLR 2025 paper shows that T loops can simulate T steps of chain-of-thought; Ouro lets the model learn to allocate its own depth; Loopie achieves IMO gold-medal performance at 20B-A2B scale. Starting from seven paper abstracts, this piece explains why loop count is a dial adjustable at inference time, and why the claim 'we have limited our use of this capability' cannot currently be verified by anyone outside the lab.
Lessons from the Crash: Knowing You 'Don't Know' Is the Bottom Line for Survival
Starting from market experiences around March 13, 2020 and including later reflection, this piece separates event dates, circuit-breaker indicators, and the limits of my understanding. Falsification opens judgments to refutation. Admitting uncertainty still requires checking and responsibility, rather than abandoning analysis or promising safety.
Once AI Enters the Hospital, Who Shows That It Works?
Validation of medical AI must reach how clinicians use it, what patients receive, and how performance is followed after implementation. Model results are the start of that work.
Dividing Work Across AI Windows: Materials, Permissions, and Verification
Separate model capability, session tools, and action authority. Collaboration needs explicit questions, source versions, independent checks, and handover responsibility, rather than fixed personalities, exam scores, or checks presented as reliability guarantees.
Negentropy is a metaphor: What Taiwanese businesses should check before choosing locations
Correct closed versus isolated systems and separate thermodynamics from a management metaphor. Evaluate contracts, rule changes, data and technology rights, continuity, and exit costs without treating country rankings or dual bases as physical laws.
Running Multiple Agents, Still Only One Pair of Eyes
Addy Osmani divides the human-AI relationship into three stages: offloading, deferral, and surrender. In a Wharton experiment, participants who received a wrong AI answer scored about 15 percentage points lower than those who used no AI at all. The more output you generate, the sooner self-discipline-based checking breaks down. Drawing on false-green-light moments from my own work, I've distilled three mechanisms you can build into your workflow.
When AI Lets One Person Be PM, Designer, and Engineer, Who Pushes Back?
AI lets one person handle product, design, and engineering. But who questions your direction? Drawing on personal experience, product risk frameworks, and Popper's falsifiability, this piece proposes three checks: confirm the need before building, observe real use with a minimal version, and review outcomes before investing further.
Whose Body Is This? Using the Organ Metaphor to Examine Decision Authority on AI Platforms
Starting from a personal experience of platform interruption, I use the organ metaphor to examine control and exit conditions in AI systems, while separating system capability, purpose-setting, migration costs, and biological classification rather than treating integration as proof of inevitable closure.
Why Did Prime Minister Takaichi Ask Everyone to Act Tonight? PHP2Shell Explained
PHP, the backend language powering most of the web, has two critical vulnerabilities that could let attackers run arbitrary code on a server without logging in. Affected versions trace back fourteen years. Researchers used AI to find them; word of the flaws leaked before the patch was out, and Cloudflare observed hundreds of thousands of exploitation attempts within two hours of disclosure. Here is a plain-language account of what happened, why Japan's Prime Minister spoke up, and what you need to do.
Taking Physiological Data Back into My Own Hands: Sharing a Data Architecture Built with AI, So You Can Do It Too
How can a personal physiological journal connect Google Health, Python, SQLite, and Cloudflare? This article examines five-minute walking checks, daily physiological synchronization, three public data contracts, and the practical boundaries of AI collaboration.
Three Months to Quit the AI Tells in My Writing
Plenty of articles teach you to spot AI-flavored writing. The hard part isn't identifying it in someone else's work. It's this: once AI has learned your tone, can you still tell which sentences you actually chose, and which ones habit finished for you? This piece documents how I spent three months turning "remove the AI tells" from a gut feeling into a repeatable process.
When Outside Models Calculate Taiwan: Choices Beyond the Risk Figures
A reflection prompted by a secondary account of two podcasts, distinguishing personal risk estimates, SIA and BCG's 2024 advanced-logic capacity forecast, and political judgment. Two speakers do not represent Silicon Valley or US policy, and overseas capacity cannot be directly converted into depreciation of Taiwan's deterrence.
I Stopped Asking Which AI Model Is Best
Different AI leaderboards measure different capabilities, so there is no single "strongest model." More surprisingly, a model with a lower per-token price can end up being the most expensive option for long-horizon AI Agent tasks. This article walks through how I built a model frontier dashboard, and how to read leaderboards, task costs, and use cases to decide which model to actually use.
Becoming a Super Learner: A Growth Operating System for the AI Era
AI makes knowledge readily available, but true learning ability has become even more scarce. This article breaks down the six capability modules of super learners, from motivational structure to natural expression, building a continuously evolving growth system.
Three Approaches to Talent Density: Meta, Anthropic, and OpenAI
Meta's research core of nearly one hundred people brought talent density into the conversation, but Dario Amodei and Sam Altman have each offered their own concrete views on talent and management. Drawing on public statements from leaders at three AI companies and historical retention data, this essay compares how resource concentration, shared purpose, and delegation shape how teams actually work.
Three Claudes, One Feedback Path: From Design Exploration to Implementation Acceptance
A retrospective on assigning design exploration, critique, and code implementation to three roles, using sketches, specifications, and screenshots to examine handoffs while distinguishing interfaces from models, feedback from quality, and rendered screens from functional acceptance.
I Thought the Models Would Change. I Didn't Expect the Collaboration Patterns to Change Too.
Three days ago I argued for treating task difficulty as a constant and model capability as a variable, but I missed something. Anthropic's advisor strategy lets a cheap model drive the whole task and calls the expensive model only at critical decision points. Official benchmarks show it approaches top-tier quality at roughly 60% of the cost. This post fills in the second variable I left out last time.
Is Transync AI Worth Subscribing To? Comparing It with a Usage-Based Self-Built System
A comparison of Transync AI's official plans and my self-built meeting translation workflow. Self-building still incurs API and maintenance costs and uses third-party services; pricing, usage conditions, and data handling need separate examination.
When AI organizes the work, what should we confirm?
Use Ultracode to discuss scope, resources, and acceptance without treating agents as a reliability guarantee. Clarify scope, permissions, stopping conditions, and evidence of completion.
When AI Summarizes News, Keep the Dates and Qualifications
Keep attribution, dates, periods, units, and qualifications when using AI to organize news. More links or conversations do not guarantee independence, and checking results also need human review.
We Must Set the Pace for Frontier AI
Dario Amodei argues that frontier AI should not be paused entirely, but should advance at a pace deliberate enough for alignment, interpretability, testing, and third-party oversight to keep up with model capabilities. This article summarizes his three-tier governance framework and the geopolitical constraints facing democratic nations and global coordination.
When AI Helps Build AI: Read the Numbers Before Closing the Loop
A source-bounded look at Anthropic's public account of AI-assisted development, clarifying the definitions behind eightfold code volume and session success rates, withdrawing illustrative interpolated charts, and separating company reports, future scenarios, and my own reflections on acceptance.
When AI Starts Choosing Its Users
A retrospective on a Claude connectivity incident during a July 2026 trip to Shanghai, distinguishing HTTP responses, exit lookups, application recovery, and service conditions. It corrects the inferences that a change from 302 to 403 meant regional restrictions had been cleared and that changing the VPN altered only one variable, then draws lessons for work continuity.
From the Logic of Language to the Life of Language: Reflections on a Digital Fortune Site
Large language models turn linguistic regularities into computable numerical relations, but this does not mean they have grasped the full meaning of language in life. Beginning with Wittgenstein's picture theory, meaning as use, language games, and rule-following, this essay reflects on how digital fortunes become a linguistic experiment completed jointly by models, human editors, and users.