2026 June Thought Notes: The Pushing-Away Comes Before the Reason for Pushing Away

Quick Navigation

Start with the long read: 12 sections in the body above.

34 archived notes this month, filed under 7 themes:

• Self-Knowledge and Psychology · 13
• Daily Notes and Everything Else · 5
• Business, Investing and Career · 5
• AI and Agent Systems · 4
• Product, Engineering and Open Source · 3
• Content, Craft and Recording · 2
• Reading, Ideas and History · 2

Prologue: A slack surface, and high density underneath

In June I came back from Laos to Shenzhen.

If you only look at the state, this was a slack month. What it looked like in practice — I would only come alive after dark, stay up to two or three a.m. every night, lulled to sleep by one short video clip after another. Waking up in the day to Shenzhen’s gray-white sky, I could not even gather the strength to leave the apartment. I told myself it was the weather, but inside I knew it was not just that. Every day I bargained with myself: tomorrow I will wake early, I will be disciplined, I will start working — and the next day I lost again to that version of me curled up in bed. The body felt empty, not tired-empty but hollowed-out-and-idling-empty. There were many bothers, but I could not point at any specific one.

It was not anxiety — anxiety is taut. What I had was slack, sinking, closer to a hard-to-name pain: every day I opened the phone, watched the bank-balance number slide down, watched the AI API bill climb up, and in the moment those two lines crossed in front of me my chest would clench, and then I would numbly swipe it away. From the me by the Mekong, that breeze-softened version, to this me in a Shenzhen rental with day and night flipped, locked inside — the gap was so wide that I sometimes wondered whether those two were the same person. This vulnerability to environment was something I had already tasted in May after the Vientiane water festival; only this time it lasted longer and weighed more.

But if you flip through what I actually thought through this month, the picture is completely different. The dense, active records all cluster between June 21 and 30 — the first twenty days are almost empty, and these ten days hold about thirty-nine high-density conversations, almost all in Chinese, sweeping across psychology, agent architecture, product direction, founder studies, and commercial epistemology. A person who barely left the apartment, with a brain running in something close to binge mode.

The grain of these ten days was actually mixed, mixed to the point of being a little funny. On one side, heavy things like agent architecture and founder breakdowns. On the other, a stream of small and ordinary things: chatting about Shenzhen’s Longhua urban-village long-term rentals, analyzing why audit recruiting suddenly went hot (essentially a counter-cyclical defensive choice), debating whether a kid from Hubei who scored 647 in physics should go into cybersecurity or into West China School of Stomatology at Sichuan University — really two life paths — and even seriously predicting World Cup scores (I had England 2:0 right). None of these small things matter on their own, but assembled together they look like the real shape of a person in a trough still being dragged forward by life — the big questions hanging unresolved, the small days happening anyway.

The contrast itself is the first clue of June: I was not acting, but I was thinking like crazy. And the most important insight of the month is precisely about that river between “thinking” and “doing” that I have never been able to cross — and the mechanism keeping me on the wrong side of that river is the same in my love life and in my career.

So this piece starts from the deepest psychological line.

I. I Finally Saw the Mechanism of How I Push People Away

In June I spent a lot of energy fully unpacking the “avoidant” thing. It is actually two distinct things: one is avoidant attachment (an attachment-style level construct), the other is Avoidant Personality Disorder (AvPD) — a clinical diagnosis whose core is shame and an extreme fear of being rejected, closer to severe social anxiety, genuinely disabling at the social-functioning level, requiring a psychiatrist. What I am talking about here is the first one.

At the attachment-style level, the mechanism really has only one main line: let the amygdala relearn. To relearn, you first have to see what it is currently doing.

The classic avoidant move looks like this: as soon as a relationship gets deeper, as soon as the other person actually comes close, I suddenly start finding fault with them (devaluation), things I would not have minded before become intolerable; I suddenly need space — strongly, and suspiciously timed right after intimacy deepens; I miss the one who is not present and devalue the one who is — the absent and the ended get heavy filters, while the person I can reach out and touch right now annoys me; finally I find a “rational” reason to retreat (move out, change cities, “we are not right for each other”), and that reason arrives suspiciously on time.

The single most important line I caught this month is this: the substantive distinction is that the pushing-away comes first, and the reason for pushing away comes after — not the other way around. It is not “because they have this flaw I want to back off.” It is “I am already backing off, and the brain now needs an explanation I can accept.”

Then why does intimacy trigger the alarm, why does it set this move in motion?

Because the amygdala learned a prediction very early — usually before I had language, before I could remember any storyline. The prediction was, roughly: “when I hand over the part of me that needs most, that is most vulnerable, and count on another person to catch it, the result is not being caught, being overwhelmed, being swallowed, ultimately falling through.” And so it locked in an iron rule: getting close to someone to the point of needing them is unsafe.

One more layer of essence — why does the nervous system rather push away one good person after another than take the risk? Because to the amygdala, a relationship that never begins has a loss that is limited, controllable, predictable; whereas a relationship into which you have actually handed your vulnerability, only to have it fall through, registers in its old archive as catastrophe. It is doing an extremely risk-averse calculation: rather than take another bet on “I reach out and they are not there” with its catastrophic ending, retreat one step in advance — using one certain small regret (yet another relationship that did not work out) to dodge one possible large disaster (being swallowed again, being failed again).

The tragedy of this strategy is that at age three it was correct, it saved me. For a small child who could not yet protect themselves and could in fact be overwhelmed, learning “don’t count too much, retreat to safety” is a real survival wisdom. The problem is this program did not update as I grew up — it is still using a three-year-old’s threat assessment to protect an adult who has long since had the capacity to bear, to leave, and to care for himself. The people I push away now, the amygdala thinks it is pushing away the thing that once would have swallowed me. It is fighting a war that ended long ago.

I also noticed a wider context: in China, many avoidant patterns are essentially shaped by the upbringing — you slowly learn from a young age that “expressing need is useless, getting close to people is unsafe, carrying it yourself is steadiest.” The kid is forced to be sensible, the parents themselves cannot express love, the family atmosphere is cold, tense, with few hugs and little talk. You are raised in material terms, asked to function, but emotionally no one catches you. The real self is rarely seen.

The amygdala is not the fear center, it is a “does this matter?” detector

This month I corrected, in passing, a popular misunderstanding: the amygdala is not the “fear center.” Its more accurate role is a threat-detection and body-response initiation system — at the first moment of danger it sends the body into fight or flight, while the actual “feeling of fear” emerges from a much wider network. So strictly speaking, it is a “does this matter?” detector, plus a mechanism that checks “was the prediction wrong?”

If it is a prediction system, then when an environment keeps repeating without incident, the prediction gets more accurate, and the alarm eventually stops firing. The trouble is: the hippocampus has lost the plot, but the amygdala is still holding the charge — a smell, a kind of light, and the body braces, and I cannot say why. This is a stress reaction; everyone has been affected by it at some point: a work alarm clock, a school bell, a police siren. And the amygdala often reacts more strongly to “novelty and ambiguity” than to a clearly defined threat.

The good news is there is a handbrake: the prefrontal cortex (especially the ventromedial part) exerts top-down inhibition on the amygdala — under normal conditions it can press the alarm down, but at high arousal the handbrake slips and fails — which is exactly why arguing with reason during heightened emotion does not work. And one practical finding: naming an emotion lowers amygdala activity. When I can see an emotion and label it (“I am afraid right now”), the prefrontal cortex comes online, the amygdala calms. Seeing is itself an intervention.

II. Love Is Actually Three Systems, and Long-Term Relationships Run on Mechanics

Following attachment outward, in June I also broke “love” into a structure you can analyze.

At the biological level, love is three distinct systems: desire, infatuation, attachment. Desire is essentially “I want, possess, am satisfied.” Infatuation is essentially “I am drawn, occupied, slightly intoxicated, projecting, fantasizing.” Attachment is essentially “safety, companionship, stability, fear of losing.” Put the three together and you get what we loosely call “love.”

The most honest sentence I have about love is this: love is not a feeling you go and find. It is something you make, with attention, with selective investment, with the willingness to be changed by a specific, real other person. Early chemistry is real, but it is only the on-ramp. And disillusionment is not the end, it is the threshold. The essence of love is attention — measuring how much someone loves is essentially measuring how much they “see” you. Love is the extremely difficult realization that there is something that is not me, that is genuinely there; treating another person as a complete being with a world of their own at the center, not as a character in my story.

There is a deduction here that cuts especially deeply for someone like me: the observer cannot love. Because love structurally requires me to lower enough of my defenses to let someone actually move me; while the pure-observer stance, by its own structure, excludes that kind of handing-over. The observer refuses to be moved, and at the core of love there is a sovereign, controlled handing-over. This is the same line as my May question in Laos about “can the observer truly live inside experience.”

About “soulmate,” in June I basically dismantled that romantic myth: the so-called sense of fit is, in the largest part, constructed after the fact, only a small part of it is innate low friction. It is not that you found the person who already matched you. It is that long-term intimacy, a shared narrative, and bilateral investment slowly “make” someone who matches you — and then your brain back-fills the result, recording it as “we were always meant to be.” That feeling is real, but the metaphysics of fate around it is invented. Soulmates are mostly made, secondarily found.

Then there is a research finding, well-replicated yet counterintuitive: people who believe more strongly in “there is one right person out there” (destiny belief) actually have worse relationship outcomes — because any conflict triggers “ah, this is not the right one” and they exit. People who hold a “growth belief” (love is cultivated through work) ride out the troughs better. Which gives me this line: the most important thing in a long-term relationship is not love, it is relational mechanics. In Gottman’s decades of research, the strongest predictor of whether a relationship survives is not how much you love each other, but whether contempt shows up in conflict and whether you have the capacity to repair after a fight. That is the real “love values match.”

Finally, the familiarization side: the mere-exposure effect — after many encounters, the nervous system re-encodes a person from “something to evaluate” to “part of the environment, not a threat.” That low-intensity companionship between an old couple — the body still cannot do without the other — is not because the quality is high, it is because the amygdala has registered the partner as a “safe background,” and without it the baseline collapses. Same logic for keeping a pet: the longer the time, the harder to let go. The flip side holds equally: put the same person into a context you cannot predict and you will see them again, because the old cache does not work in that context. The research has this one too: couples who do new, slightly stimulating things together report higher relationship satisfaction — the mechanism is that novelty re-lights the attention system that habituation had switched off. The point is not romance, it is breaking predictability.

Quality and time: two systems, not a binary

This thread later turned into a set of blog posts on its own — the three-part “Ignite and Settle” series. The core is a two-system framework: the “quality” and “time” in companionship are not opponents, they are two semi-independent systems doing separate jobs. The responsiveness system eats quality (fed by responsiveness, by being caught — handles the ignition). The habituation system eats time (relies on duration alone, slowly settling a person into a safe background — handles the settling). To get safely attached, you have to be both lit and settled.

Attachment types can then be restated as configurations of these two systems: secure ≈ responsiveness system reliably satisfied + habituation system steadily accumulating; anxious ≈ responsiveness system has unstable contingency (signals sometimes caught, sometimes missed) → system stays in high activation, cannot settle; avoidant ≈ after repeated misses in early responsiveness, the dependence on the responsiveness system gets actively turned down, and the person switches to “self as own safe background.”

Inside writing “Ignite and Settle,” one line felt like a road I had never walked before — psychological connection precedes and produces physical attraction, not the other way around. I also said one sentence that sounded suspiciously like rationalization: “Maybe someone who changes as constantly as I do should not be pursuing relationships.” I can now recognize that for what it is. It is not a conclusion — it is that same avoidant move again: first wanting to retreat, then dressing it up in mature-sounding language. The real question is never “am I suited for a relationship.” It is am I willing to slow down for one specific person.

Scarcity, infatuation disenchantment, and “talking for my own ear”

A few smaller self-observations also got cleaned up this month. I habitually use the lens of “scarcity feeling” to explain behavior, and deliberately separate “infatuation/dependence” from “love” — the former can be constructed by distance and imagination (the more out-of-reach, the more over, the thicker the filter), the latter cannot. The avoidant move I use most often is: raise expectations internally, lower them externally; want it badly, but refuse to admit I want it. I will deliberately suppress infatuation, because infatuation raises expectations and creates the gap. But I am also slowly recognizing that love itself requires a degree of unguarded immersion, and the pure-observer stance structurally excludes that handing-over — some knowledge is bodily and tacit; “we know more than we can say.”

About conversation, I broke its purpose into three layers: maintaining the relationship, content-driven (mutual growth and inspiration), and “talking for my own ear” — running myself through a sort once, saying things I would not otherwise think. The third is the most useful, and is essentially “writing / externalizing to think.” It is also in this sense that I increasingly feel: in the AI era, deep, in-person, offline conversation is becoming an increasingly scarce and increasingly valuable capacity.

III. That Pushing-Away Move Does Not Only Live in Relationships

Writing to here, the real insight of June finally floats up.

I had always treated “avoidant” as a relationship problem. But this month I saw clearly: that mechanism — “pushing-away comes first, the reason comes after” — does not only live in my relationships. It also lives in my career.

I clearly know promotion would help, yet I keep not promoting. Then I give myself a reason — “once I have gotten the hang of marketing, I will go.” That is the same move as “let me first think through whether they are right for me” in romance: retreat first, then let the brain back-fill an explanation I can accept. “Learn marketing first, then promote” is not preparation. It is fear building itself a very respectable waiting room.

Why does distribution, promotion, getting the product into the market trigger the alarm? Because it is isomorphic to “handing over the most vulnerable part and counting on it being caught.” Pushing it out, getting ignored, getting rejected — to the amygdala this belongs to the same class of catastrophe as “I reached out and they were not there.” So it runs that extremely risk-averse calculation again: rather than risk pushing the work out and watching it fall through, retreat one step, and use one certain small regret (no one knows about the product) to dodge a possible large disaster (the market rejects me).

So pushing away the person coming close in my relationships, and pushing away “delivering the work to the market” in my career, are the same three-year-old amygdala fighting the same long-finished war. This is the biggest piece June handed me: I thought I had two problems (not daring to love deeply, not daring to distribute). It is actually one.

And so the remedy is also the same — not more preparation. It is doing that delivery move with the alarm still going off, and letting the amygdala relearn through “I did it, and nothing terrible happened.” Distribution itself is the fastest marketing class, in the same way that handing over vulnerability is the only entry to intimacy. Confidence and safety are not entry tickets, they are the residue left after I bit the bullet and shipped.

IV. Agents and Technical Engineering: What Am I Actually Building?

Slack as it was, the densest technical line of June was: how to actually build an autonomous agent system.

The hardest piece was a reverse-engineering of the Manus-style architecture. I started from the front-end presentation framework (transport layer / declarative UI layer / render layer), and went all the way down to the agent runtime, sandbox, and the core of context engineering. Conclusion: pipeline, payload and rendering are all surface. The real complexity lives in the runtime and in context engineering. Manus deliberately avoids existing orchestration frameworks and writes its own thin loop; the early pattern of regenerating the full plan every turn (recitation) is wasteful on tokens and pollutes the KV cache. The improvement direction is a Planner sub-agent plus a structured Plan object.

I also did a file-by-file close reading of nanobot’s source and turned it into a study doc — its dual-layer loop, ContextBuilder, two-phase memory consolidation, mid-conversation injection — and mapped these explicitly back onto my own RAVE and Another Self, identifying three differentiation opportunities: skill self-improvement, deep user modeling, and a trajectory-data flywheel. I also drilled four rounds into the distinction between Connector / Skills / Plugins and concluded the three are orthogonal rather than competing, and that Skills is the most durable investment, because it encodes domain knowledge that cannot be commoditized.

On the comparison of intelligence (WorkBuddy vs Codex vs Claude), the essence is the base model as the hard ceiling — frontier products are co-reinforced by the model and the harness, and anything built on top of a third-party base will lag by half a tier. What actually opens distance between agents is four capability thresholds: real side effects, long-horizon multi-step, parallel sub-agents, and direct connection to live environments.

But what excited me in June was actually a few more “paradigmatic” things. One is the agent as a distribution carrier: the user does not have to install anything — they hand a sentence to their own agent; the agent downloads the script itself, registers itself, stores its own tokens, and the user only clicks an activation link plus an email login. The cognitive load of onboarding migrates from the human to the agent. Further out: the agent manages its own money. When credits run out it walks the agent through installing a third-party payment skill, binding a card, setting up auto-renewal, topping up when balance is low, keeping itself alive automatically. Agent email addresses already exist; will agent phone numbers be next? All of these point to the same emerging paradigm — the layer of infrastructure that makes agents more readable, more writable, and able to serve and operate on its own, is growing in. And my line “right now humans look for jobs; in the upcoming era of information explosion and crazy pace, will it flip into jobs looking for humans?” is asking the same question. I even did a feasibility analysis of a Claude autonomous job-search agent (public ATS APIs, two-stage matching), pushing “jobs looking for humans” from a remark toward something actionable.

I also actually built something this month, not just thought. To give my own product matrix (DayPage, Solo Compass) a unified notification and verification center, I used Hono + TypeScript and shipped a working notify-hub: a three-layer architecture (delivery layer / OTP layer / notification layer), both in-memory and Redis drivers, rate limiting and idempotency, OTP stored only as SHA-256 hashes, smoke tests passing. That experience grounded me a little, in an otherwise weightless month — the texture of “actually getting a small but complete thing running” steadied me more than imagining ten thousand grand directions. Which itself is a clue: what I lack has never been ability, it is the delivery move that gets the ability out.

There were also judgments I had to correct. On the subscription mechanism: I initially said a certain credit-pool change would take effect on 6/15, and after you challenged me I went back and checked — that change was paused on the day of announcement, never actually shipped. I have to be honest with myself here: the more I “feel I know” something, the more I should go back and check the structure. This resonates exactly with one of the knives I sharpened this month — structure does not lie, “I feel” does.

V. Product Direction: What to Build and What Not to Build

Almost every conversation in June eventually landed on the same question: can an indie developer, a super-individual, actually build this thing, and is it worth building.

A few of my own products are still searching for their position. DayPage needs to find its own positioning, an immersive on-ramp, a clearly used core, and to think through how to plug into different work segments, how to open APIs, how to support more input modes in the future, how to be more agent-readable and agent-writable. This month I went through three rounds of rebuilding its night-capture interface, and the process itself says a lot about my aesthetic: extreme restraint, large whitespace, Chinese serif mixed with sans-serif, one strong focal point. I had proposed a central orb design, killed it myself, and finally landed on a canvas-drawn solution — 84 radial waveform lines gradient from terracotta to lavender, a microphone record button in the center, an asymmetric dock at the bottom. I kept pulling it back whenever any element grew too big or too redundant — every revision had fewer things. I later realized this is a metaphor: my obsession with subtraction is far stronger than with addition, I trust whitespace more than accumulation — a strength in building, but on distribution it turns into procrastination, “it is not good enough, I can still delete more, it is not ready to ship.” Same temperament: restraint when building, avoidance when shipping. The travel app (Solo Compass) still has no clear signal — the positioning, the form factor, the working motion, none of them are figured out. The resume product also needs a positioning, and ideally a positioning that is connected to what the future of resumes and the job market will look like. The earlier products (the user-interview tool, the streaming reader) need to be re-thought too. The interview tool can do a lot in the agent era, the question is whether the ecosystem can pick it up; the streaming reader does not solve any visible pain point in the AI era and will probably struggle to survive.

I also did a sweep of external opportunities. Monetizing MCP is a real business model, but the profits concentrate around vertically-differentiated data sources — my own validated content from the travel domain happens to qualify as a proprietary data asset. Agent email infrastructure has already been validated. The open opportunity is at the vertical application layer; the real moat is encoded domain know-how, not rented email infrastructure or LLMs. I sketched one concrete scenario for it — an export agent for a Shenzhen battery-storage factory, going from a cold email all the way to a closed deal. The longer I worked on it, the more I felt the truly valuable thing is that tacit knowledge of “how to actually get a deal done in this domain.” Western professional services AI SaaS I went deep on — avoid the red ocean of general contract review and bookkeeping automation, hit the verticals like CPA tax document extraction and estate planning document assembly, benchmark against Harvey / EvenUp, target 5K–20K MRR in 12 to 18 months and then exit through Acquire.com. Cross-border e-commerce I also looked at — I pushed back on the claim that “drop-and-flood sellers have higher data density and therefore fit AI better,” distinguishing data density from data leverage. Compliance automation for Amazon buyer-seller communications (Chinese interface + compliant English replies + account-health alerts) is a real differentiation point.

There were also a few I explicitly put in the “do not do” pile. PollyReach.ai, an agent-native voice product — I pulled it apart and found that the founder publicly claims “AI never falsely reports success,” yet the SKILL.md contains a “demo mode simulates successful payment” instruction. The contradiction collapses my trust in the whole project. An offline social/dating app — I talked myself out of building a full multi-person version, the more valuable move is to compress it into a single-player “relationship memory layer” inside DayPage. A Buddhist AI agent — a direction I personally love, but it should “monetize like a monastery, not like SaaS” — running on a dāna (donation) logic, not a paywall. AI-fication of local businesses has a sales-consultative go-to-market that directly conflicts with the low-touch product route I want, so I dropped it. And for legal gray-zone businesses that run on regulatory arbitrage — I asked myself a sharper question: outsourcing ethics to “as long as it is legal,” for someone whose core belief is “build internal standards when external standards collapse,” is itself a contradiction. The real test is “with full disclosure, can this transaction still stand?”

Sweeping through this whole circle, the thing most worth writing down is not any single opportunity, but the same bottleneck that kept being named — see the next section about my own situation. One methodological self-correction to record here: I have a circular ability of “game-ifying a deep dive into an industry, finding its rules and cash flows.” It is elegant, and it has three real risks — underweighting distribution, the depth tendency making me only analyze and never ship, and selling methodology in the abstract before I have an actual validated result. All three point in the same direction.

VI. Founders and Business: How Incentive Structure Colonizes Personality

A serious amount of June went into person and company breakdowns, and what I want is never a summary of facts but an integrated judgment at the structural, incentive, and personality level.

Wang Tao / DJI: I argued that DJI did not “create” demand, it “released” latent aerial-photography demand, and is essentially an imaging platform company. More interesting was him breaking ten years of silence — “the world is unbelievably stupid, and so am I,” and giving himself 1 out of 10 on product work and 10 out of 10 on organizational work (handling a 45-person corruption case, dismantling internal fiefdoms). But I also remind myself: Wang Tao is a “dangerous idol.” His ability to keep an extremely complex system converging is fascinating, but it can let me rationalize avoiding necessary risk — using “I am polishing the system” as an excuse for not getting the thing out.

Yu Hao / Dreame: I reframed Dreame as a “capitalized industrial platform” (motor moat, 200+ business units, a chain-leader + industrial-fund-driven recruitment model). My judgment on Yu Hao — “a cautious person being slowly colonized by the radical persona he himself invented.” When a company’s growth engine is “narrative-driven financing” rather than “product-driven cash flow,” radical stops being a style choice and becomes a structural fate: you have to keep running faster, because the moment you slow down, the gravity of the story drags you under.

Chen Mian / Liblib · Lovart: One of the youngest executives at ByteDance, who later pivoted to AI. I pulled out a few of his rules — extreme vertical focus to avoid big-tech home turf, competitive advantage outside the model layer, “describe the future first” (build products for the model capabilities 6 to 12 months out), the 3-to-6-month competitive window doctrine, native global from day one.

Pinduoduo / Temu: distinguish the two axes “commercial excellence” and “ethical goodness” — they often have nothing to do with each other, and the commercial world frequently rewards not who is more dignified but who is harder to knock down. PDD is a demand aggregation machine (goods finding people). On Temu’s side, the removal of the de minimis exemption is a structural, permanent hit, not a cyclical wobble.

LatePost: I studied it as a platform on its own — its moat is patient, hand-built first-hand reporting. In an environment where everyone rushes for speed, “slow” itself becomes a moat. I write this one down because it is the exact opposite of me: strong output, weak capture. Its output is not fast, but its capture (trust, mindshare, exclusive relationships) is extreme.

I also looked at two industry benchmarks. The growth curve of biopharma I broke into three independent curves: end-market drug sales steady state (~8% CAGR), the biotech capital cycle, and the Chinese innovative-drug license-out arbitrage window. And LifeSciBench (a life-sciences benchmark) — I used it to analyze the three constraints on willingness to pay for “judgment-augmenting” AI. These are not domains I am about to jump into, but they are training the same thing: quickly enter an unfamiliar industry, dig up its rules and cash flows, judge whether the model on the surface is real.

Put these people side by side and a through-line floats up: incentive structure colonizes personality. Yu Hao, Wang Tao, Chen Mian, including me, are all being slowly assimilated by the narrative we chose for ourselves. And that ties back to my core question — in an environment where external standards have collapsed (ritual collapsed, music broken), how do I avoid being alienated by incentive structure, and how do I build my own internal standards.

On commercial epistemology, a few lines I want to keep: data is being repriced — the platform era rewarded “soft assets” like brand, content, mindshare; the AI era rewards what can be structured, repeatedly run, and pattern-extracted. AI’s comparative advantage is not inspiration, it is clustering and iterating over massive feedback. Supply creates its own demand — when something good enough, cheap enough, and easy enough suddenly appears, people “discover” they have always wanted it. And one I keep telling myself: the closer you are to the actual business, the more you know it is complicated. Only the people further from the industry like to mythologize, demonize, absolutize, dramatize.

VII. Good and Bad, and the Opportunity Window for the Ordinary Person

Two more “metaphysical” conversations in June I want to keep separately.

One is about whether good and bad can be defined at all. On the other side was a moral-relativist (sliding toward nihilist) position: if the concrete content of “good” varies with time and culture, is “good” itself undefinable? What I pushed back on was the slide from “the content varies” to “the thing itself is undefinable.” The universality of “good/bad” in language, the self-defeating move in relativism (“everything is relative — except this sentence”), and dukkha as a cross-cultural structural invariant, all together say: the cosmically guaranteed, eternally fixed “good” has indeed collapsed, but an “indexed, constrained good” still stands. This distinction matters to me — it is the philosophical landing of my “in the age of broken ritual and music, build your own internal standard.” No ruler from God does not mean no ruler. You have to make one, and own it.

The other is about whether the progress of technology closes off opportunity for ordinary people. I did not give a yes or no, I developed a few frames. One is the opportunity transition-window model — opportunities do not disappear, but they have windows that open and close. Two, compounding vs reset — sustained accumulation in one field compounds; constantly chasing the next wave is repeatedly zeroing yourself out — and I have exactly that tendency to chase waves and depth. Three, the barbell economy — the middle collapses, the two ends grow; the not-up-not-down position is the most dangerous. Four, the democratization-vs-concentration paradox — the more the tools democratize (everyone can use AI), the more capability concentrates in a small number of high-agency individuals.

Stitched together, the conclusion lands precisely on me: opportunities are not disappearing, they are shifting from “distributed and passively received” to “actively constructed by high-agency individuals.” Which is to say, the AI era is not short on opportunities, it is short on the act of “actively constructing opportunity” — and that act, once again, is the one I avoid the most: make the thing, push it out, let the world give feedback. I am strong on the producing side and weak on the capturing side (distribution, trust, ownership). All the lines point back to the same place.

VIII. A Few Epistemological Knives

I also sharpened a few knives in June that I want to carry long-term. They mostly belong under my categories of “investigating things” and “watching myself.”

Structure does not lie. Don’t chase opinions too much — “I feel” is too easy, too easy to be tested by reality. The “logical closure” many people love is just self-validating circularity over their own positions: pick the stance first then find the information, get the feeling first then find the reason, end up sounding internally consistent and not surviving any real prodding. So-called logical closure is just a closure on top of pre-existing inertia. Many companies die exactly because “they thought they had already understood.” So I do not natively trust anyone, including myself — humans have many good qualities, but the brain optimizes, beautifies, weights, biases.

Bayes and ego. All priors are subjective, and since they are subjective, ego is unavoidable. But ego pollutes Bayes in three places, and the prior is the least dangerous of them. First, the prior itself (overconfidence, wishful thinking) — remedy: anchor to base rate. Second, evidence interpretation P(E|H), the most insidious — given the same piece of evidence, I unconsciously weight it as “this supports me” — remedy: counter-ask “if I were wrong, what would this evidence look like?” Third, the update itself — I already have evidence that should move me, but moving costs identity — so I just … don’t. Remedy: pre-commit, before seeing results write down “what kind of evidence would change my mind.”

Three worlds, and where truth joins. I think a person lives in three worlds: the physical world (atoms, materials, engineering), the conceptual world (institutions, management, vision), and the felt world (an individual’s experience of joy, suffering, meaning). Often the three are disconnected. My understanding of “truth” is to find the joint of these three: how you feel, how you organize rules, how you make the thing — and ultimately for it to stand commercially, while precipitating healthier organization and order, where the people inside are not emptied out but actually find satisfaction and growth.

Stay humble about “the Way.” Drilling down to first principles is seductive, but the laws of investing and business are nowhere near as stable as physical laws over millennia — confidence that “I have grasped the unchanging principle” can itself become a rigid frame that gets smashed by a paradigm shift. So “seeking the unchanging way” and “staying humble about the way you have grasped — it might only be a temporary pattern” must be held at the same time.

Personality and situation. Assuming someone strong in one domain is reliable in others is the halo effect. Behavior is shaped by situation far more than we think; being extremely rational at work and a mess in relationships can perfectly coexist; ability does not naturally transfer across domains. And on silence: the silence of the weak is compromise, self-protection; the silence of the strong is “no need to explain.”

IX. My Own Situation Today: Disorientation Is Not Insufficient Effort, It Is Not Yet Recognizing the Problem

Pulling all the threads back to the present me.

End of June, my partner and I parted ways, because he no longer believed AI had a future.

The pain of this is a quiet kind of pain. Not the loud pain of a fight; more the kind where you turn around one day and notice the person who used to always pick up your line, push back on you, drag you back when you were drifting off — is not there. We used to be people who could chase one idea through an entire night. The breakup was not about anyone betraying anyone, the belief forked: he took a step back, decided this road has no future; I cannot step back, I cannot retreat. In that moment I finally understood: a partner is never just someone you work with. He is my feedback loop, my correction mechanism, the echo when I say “is this direction right.” That part of why I now feel disoriented, without positive feedback, unsure whether I have drifted off — is largely that I just lost that system, and the new one has not grown in yet.

But the painful thing also sharply revealed my direction — when “abandon AI” and “lose the partner” were both on the table, I chose to stay with AI. The moment someone is willing to pay the cost of a breakup over something, that something is no longer in the “I don’t know what to do” category. I am not without conviction. I just paid an expensive and lonely cost for that conviction, which is why now I hurt and feel hollow.

My diagnosis of my own disorientation also finally got said out loud this month: the disorientation at this stage is that I have not yet recognized what I need, have not yet recognized what kind of problem I have to solve, have not yet recognized what I should do in the next stage. Disorientation is not insufficient effort, it is not yet recognizing the actual problem to be solved.

And once you stitch the thirty-nine conversations together, the shape of that problem is already very clear — my bottleneck has never been technical execution. It is distribution and commercial validation. It got named in every conversation — cross-border e-commerce, professional-services SaaS, agent email, local businesses, the questionnaire system. Pair it with the “super-individual” self-positioning, and my real diagnosis is: strong on the producing side, weak on the capturing side (distribution, trust, ownership). Technical capability is already in surplus. What is truly scarce is converting it into something people need, see, and pay for.

Now look — every line of this piece closes here:

My “strong output, weak capture” in business is not a capability problem. It is the avoidant move from Section III — capture = handing the work to the market = handing vulnerability over and counting on it being caught, and my amygdala predicts “miss, catastrophe” on that act, so it keeps me forever on the safe side of “polish a bit more, learn a bit more marketing.” My commercial bottleneck and my intimacy bottleneck are the same three-year-old program running in two different arenas.

So what June leaves to July is not a piece of motivational soup. It is a concrete, cross-domain solution:

First, stop waiting for the turning point. It is not waiting in front of me, it sits behind my next action. It is a verb, a choice, not an enlightenment to be discovered.

Second, the antidote to disorientation is recognizing the problem first, and the way to recognize is through action, not thought. The problem I have to solve is already on the wall: distribution and capture. So stop using “deep-dive the industry” to avoid it — that is just another round of analyze-without-ship.

Third, take the insight about avoidance and apply it directly to distribution. Next time I think “let me learn marketing first and then promote,” that is not strategy. It is the amygdala fighting the war that ended. The solution is the same as in intimacy: do the delivery move with the alarm still going off, and let the nervous system relearn through “I did it, and nothing terrible happened.” Naming the emotion lowers the alarm — so next time I want to retreat, label it first: “I am not feeling unprepared. I am scared of being rejected.” Seeing it brings the prefrontal cortex online.

Fourth, rebuild the broken feedback system. The partner is gone, but the feedback loop cannot stay empty — fill it with a living life roadmap I keep editing. One sentence for the far path (use AI to serve real, flesh-and-blood human life). Three items for the near path (ship the product, ugly; do one piece of promotion that scares me; build a sustaining side-income runway). Review once a week, give myself positive feedback, give myself course corrections.

Closing: One War, Two Battlefields

If May was “someone used to placing active bets actively throwing himself onto the edge, to see clearly who he really is,” then June is — I finally saw that what has been blocking me, in my love and in my career, is the same thing.

The pushing-away comes first. The reason comes after. Whether what is being pushed is a person coming close, or a work that should be delivered to the market, both are the same three-year-old amygdala fighting a long-finished war on behalf of an adult who has long since been able to care for himself.

Seeing this does not stop the war. But it lets me know, for the first time: I do not have two problems, I have one. And this one is not solved by thinking it through. It is solved by doing the scary, near, actively chosen delivery move — once, then again, until the nervous system finally agrees to write “actually handing it over is fine” back into the archive.

July will begin with the “publish” button I have not dared to press.

Written at the end of June 2026, Shenzhen. Day 20 back home, day 61 without flomo.

Appendix: raw notes from this month

The 34 entries below are the raw notes from this month, filed by theme.

Self-Knowledge and Psychology

13 entries

Daily Notes and Everything Else

5 entries

Business, Investing and Career

5 entries

AI and Agent Systems

4 entries

Product, Engineering and Open Source

3 entries

Content, Craft and Recording

2 entries

Reading, Ideas and History

2 entries