This article isn't available in English yet — showing the English home instead.
quidproquo
Tech, climbing, surfing, coffee, and everything else.
Quid pro quo/ˌkwɪd proʊ ˈkwoʊ/ — Latin for "something for something," a fair exchange.
This is where I document AI, tech, and product thinking, along with the process of building products — plus climbing, surfing, and coffee. Not just hoarding knowledge, but turning it into something useful for others, and then giving a little more.
BossAI in Hualien, a self-taught AI attacker in Kaohsiung and the ARTEX case in Seoul show that automated attack systems no longer require a security background, and one operator can sweep hundreds of targets in weeks. Run an AI security audit on your own source code, then self-test with an ARTEX-style framework; both cost far less than an outsourced pentest.
Seven days, 85 sessions: 104 compactions, 20 SIGKILLs, 24 type-check OOMs, 72 user corrections. The fix: move no-exception rules from CLAUDE.md into PreToolUse hooks, and drop 16 MEMORY.md entries that git already knew or that had gone stale.
Google's Professional Agentic Architect is assessed as a hybrid two-part exam: a proctored multiple-choice exam with Pearson, then hands-on labs on Google Skills. Both parts are required. The GA spec is a 2-hour exam, 50-60 multiple-choice questions, $200, English, one-year validity; the beta (Sept 8-30, 2026) was 3 hours, about 80 questions, $120.
When several questions share one context, Flide computes the shared state once and forks it per question with answers identical to asking alone; up to 9.5x faster than vLLM and SGLang at best, roughly 1.3x to 3.3x on average for shared-state workloads, and only about 1x to 3x with one question per input.
Run several Claude Code, Codex or Cursor sessions at once and their dev servers, test workers and Playwright MCP browsers keep running as orphans after the sessions end, until swap hits 30 GB and tests hang. macOS has no per-process memory cap, so it takes three layers: a SessionEnd hook, a periodic orphan reaper and a memory gate.
To try letting an AI use a cloud computer, start with Simular box: the computer is free and one command connects it to Claude Code or Codex. Its specs are unpublished, each account gets one machine, and it is still in beta. For several machines or always-on use pick Orgo (from $29/month) or Runloop; for commands only, a sandbox such as E2B; for web pages only, a browser service such as Browserbase.
'Natural Mounjaro' is a Threads nickname for a wrap of canned tuna, boiled egg, cucumber, plain yogurt and nori. Its ingredients have no pharmacological counterpart to the drug. It helps with weight loss by cutting one meal from about 800 kcal to about 250 kcal, so swap in one meal a day at most.
Japan requires the giteki mark. For a short trip, rent a digital simple radio (RadiGo from ¥4,000 for 2 days; the renter does not register), or buy two 10 mW specified low-power radios locally (about ¥5,000–10,000 each). Korea requires a KC-certified digital personal radio (424/448 MHz, 0.5 W), about 4,400 won a day to rent; analog units stop being legal after 2026. For SOS: bring your inReach to both; a PLB may not be activated on land in Japan and is not permitted in Korea; iPhone satellite SOS works in Japan but not in Korea.
Install OpenCode 2, select a built-in free model, and complete your first everyday task. Connect OpenRouter later if you want more models or a free API for your own application.
The only licence-free radios in Taiwan are NCC-certified FRS units: 14 channels, 467.5125–467.675 MHz, 1 W or less. You can bring US FRS radios (such as the Rocky Talkie Mountain Radio) or EU PMR446 radios into Taiwan, but transmitting with them there uses frequencies not assigned to the public; that's what the statute says, and the NCC hasn't published a ruling that addresses it directly. For the Dolomites take PMR446 (e.g. the Rocky Talkie EU version, €102 each), for the US take FRS, and for Japan a 'tokutei shodenryoku' radio with the Giteki mark. A radio only backs up voice calls: keep the same calls, say one thing at a time, and agree beforehand what you'll do if it dies.
In Taiwan the most straightforward choice is the Garmin inReach Mini 3 (NT$14,990): get a registration code from Chunghwa Telecom with two forms of ID, then subscribe with Garmin; every plan includes SOS. iPhone and Pixel satellite SOS don't work in Taiwan, but do in Italy. A PLB has no monthly fee, but in Taiwan it must carry the 416 country code and be registered with the Maritime and Port Bureau; Japan bans land use and Korea doesn't allow PLBs at all. In Backpacking Light's deep-valley tests, Iridium-based inReach units delivered in 1–2 minutes on average, the Globalstar SPOT X in about 15 minutes and an iPhone in about 11. In three recent Taiwanese inReach rescues it took 14–20 hours from alert to reaching the casualty, so your battery has to last the night.
Look at whether things go back and forth between members, not at sign-ups or message volume: does a check-in get a response within 72 hours, do people who got a response come back within 7 days, is more than one core member responding to newcomers, and is the group building up practice others can reuse. Most new communities die within three months (57% of 472,231 new Facebook groups stopped posting), and groups with a single admin whose founder added every member die sooner.
DaoDao set out to help individual learners build habits and stalled on numbers: a lone learner who checks in and gets no response stops coming back. It now runs three lines in parallel: helping teachers and community hosts who bring in whole groups at once, running its own reading and study sessions to test formats, and open co-challenges with no host where people with any goal can join. Software stands in for the host where there is none and saves facilitators' time.
LLM prose has four recurring failures: contrast framing (refuting an X nobody said), significance inflation (dressing a plain fact up as a turning point), indicative summaries (announcing the topic, withholding the conclusion) and overgeneralization (summaries that drop conditions and scope). Fix the first three by cutting the filler and writing the answer; fix the fourth by checking each sentence against the source. Contrast framing has been traced to preference data, and models revising a draft move it into another contrast form, so banning a phrase in the prompt may not stop it.
Ask two questions of every contrast: do readers actually hold X, and does rejecting X tell them anything new? If either answer is no, drop X and state Y. 'Rather than' became about seven times as frequent in NLP papers from 2019 to 2026, and five commercial models writing Chinese use the equivalent construction 2.5–9.2 times as densely as a human reference corpus. Models almost never misuse it, but more than half the uses are unnecessary, and with eight in one paper, at least one tends to annoy the reader.
The four entry points are one staircase, not a menu: add a DO when you need memory and resumption, AIChatAgent for the chat pipeline, Think or PiHarness for loop management, and dynamic agents only when you need to split work.
Read a finger crack by how deep your fingers go: only to the first knuckle is tips, to the second knuckle or the base of the fingers is a good finger crack, and in but wobbling is rattly fingers. For the good size, go thumb down: fingers in to the base, palm and forearm flat on the rock, rotate the elbow down until the forearm lines up with the crack, and pull down, not out. Too thin, flip to pinkie down; rattly, rely on torque and push your thumb on the edge; any wider is off-fingers. Knuckles locked above a constriction cost the least effort, and pin scars are ready-made constrictions, though they are hard to protect and hard on your toes. Feet go on face holds first, then toe jams or front-on smears on the crack edges. If the jam won't stick, the edge is good to grip and the wall isn't steep, layback. A millimeter changes everything, so measure cracks with your own fingers.
In a hand crack a foot jam takes four steps: drop the knee outward so the foot turns to its thinnest side (little toe down), slide the toes in, then bring the knee back in line with the crack so the foot twists tight. The most common mistake is getting only the tips of the toes in. Sources disagree on depth; Pete Whittaker's sweet spot puts the crack edge half to a full toe's length past the webbing of your toes. As you weight it, let the heel sit lower than the toes: pressure moves off the toe knuckles and the jam gets more secure. Foot pain usually means over-twisting or shoes that are too tight. To get out, reverse it: knee out, ankle relaxed. Finger cracks take only the toe; fist cracks take the whole foot straight in.
Before you climb a crack, work out how wide it is: tips, fingers, off-fingers, hands, fists, offwidth, squeeze or chimney. Each size has its own jams. But you measure the size with your own hand. Pete Whittaker's comparison: cam sizes 0.75 and 1 are perfect hand jams for small hands and thin, strenuous jams for big ones; a #4 is a fist jam for big hands and a butterfly stack for small ones. So 'this one's easy' from a friend may not apply to you, and you have to climb every size for the difficulty to even out. The series starts with hand cracks, the easiest to learn, and works outward toward narrower (fingers) and wider (fists, offwidths), with 2020-or-later videos from many crags and climbers in every post.
A hand jam's strength comes from the thumb, not the fingers: relax and flatten the hand, slide it in to the wrist, then tuck the thumb into the palm and push it toward the little finger so the muscle at its base presses on the rock. The fingers only help. The most common mistakes are not going deep enough, grabbing the crack edge with your fingers, and jamming at chest height. Thumbs up to move up, thumbs down to rest or place gear; cross hands in straight cracks, shuffle in leaning cracks and corners. A crack that's perfect for big hands may be cupped hands for small ones, so find your own 'hand number' first.