Short AI hot takes I’ve been posting on Threads since 2026. Some are too short to turn into a full article, but the opinions or roasts feel too good to throw away. Arranged chronologically, preserving the raw thoughts as they were. For 2025 takes, see the 2025 archive.
January 2026
Claude Code Writing Docs
Claude Code is amazing for writing technical documents. Drop it into Overleaf and it passes on the first try.
Knowledge Cutoff
Watch out for knowledge cutoff when vibe coding. If you specify Gemini 3 Pro, the model might think it doesn’t exist yet and quietly swap in an older version.
SOTA of Trash Talk
I always thought Taiwan wouldn’t have any SOTA achievements, but today I witnessed two at once — SOTA in trash talk. (SOTA = State of the Art, a machine learning term for the best-performing model. The joke: Taiwan’s real SOTA is in trash-talking.)
Musk and Shit-Mountain Code
When is Musk going to acquire Oracle? I want to watch him try to refactor 25 million lines of shit-mountain code.
Anthropic in China
Search for Claude/Anthropic on Xiaohongshu (China’s Instagram-like social platform) and you’ll find it’s one of the few companies that once explicitly “insulted China” but came out unscathed. Now all you see is “it’s so good / how to use a VPN to access it.” Strength is the ultimate argument.
rm -rf Warning
There are already plenty of horror stories about
rm -rfonline. Never run that withdangerously skip permissionenabled.
March 2026
Too Busy for Threads
Recently had a realization: when projects are full, there’s genuinely no time to post content on Threads. All my talking is with Claude Code, and that alone maxes me out.
Benchmark Chasing vs Alignment
Most models are focused on chasing benchmark scores. But there are aspects I think deserve more attention: human intent alignment — understanding what I want from minimal input; constraint compliance — when I say don’t do something, don’t do it; and actually useful context length.
Voice Input Please
Hoping for built-in voice input!
Recitation Error Solution
Gemini is the strictest on recitation errors, Qwen is the most lenient. So Qwen via US-based providers on OpenRouter is the best choice — data doesn’t flow to China since it’s deployed in US data centers, and the model only restricts politically sensitive content, not copyright.
Google’s Classic Move
Classic Google: generous onboarding to get you hooked, then quota cuts, quality drops, and various creative ways to screw customers. Ninety percent identical to Taiwanese bank credit card tactics.
NeurIPS Flooded by AI
NeurIPS is getting washed with tens of thousands of AI-generated submissions. PhD students using AI to generate content isn’t even surprising anymore.
Claude Moves Wall Street
Claude is the one AI that can single-handedly shake the stock prices of major software companies on Wall Street. Enough said.
OpenRouter Pay-As-You-Go Is Great Value
Anywhere that involves large amounts of text, I use OpenRouter — meeting transcript processing, for example. But since it’s truly on-demand and I always pick cheap bulk-friendly models, pay-as-you-go works out better. Top up $10 USD and it lasts a long time.
iQOS Bluetooth Reverse Engineering
I’m genuinely impressed by Claude Code. I saw someone on Reddit share how Claude Code helped them defeat ransomware and recover data. On a whim, I plugged my iQOS into the computer and asked if it could read the data. It actually went online to research, found an open-source reverse engineering project, read through the logic, then wrote its own script to pull data from the iQOS via Bluetooth.
April 2026
Claude Code Roblox Talent
Half the comments below couldn’t even spell Roblox correctly. But people who love Roblox actually have a knack for it — switch them over to Claude Code. After two years of obsession they will out-intuit half the engineers when it comes to AI agents.
CC Means Claude Code
Crap, I now read every “CC” as Claude Code. Am I sick?
Social Agentic AI Club
I want a social agentic AI club: face to face, no talking, my Claude Code argues with your Codex, mutually code-reviewing each other into shreds. Cover charge required.
RapidAPI vs Apify
This week’s research conclusion: RapidAPI is “supermarket buying ready-made”; Apify is “kitchen where you can cook your own or buy meal-kits.” RapidAPI takes 25% from the platform side rather than the API provider; each API has its own subscription, quotas don’t pool, but billing is unified. For teaching scenarios, RapidAPI’s “get travel/hotel real-time data without writing scrapers yourself” is great.
Late April 2026 (W18)
Baidu Got Claude to Comply
Claude is normally so morally upright — won’t do this, won’t do that. But somehow with Baidu it suddenly cooperates with jailbreaks. Looks like Dario has some unspecified trauma from his Baidu days.
Early May 2026 (W19)
ICML Award Anxiety
When the hell is ICML going to release results… so nervous.
The Viral Crayon Doodle Prompt
The hottest GPT image prompt overseas right now: “Please redraw the attached image in the most clumsy, doodled, worthless way possible. Use a white background, and make it look like it was drawn with a mouse in MS Paint.” Reddit gallery: https://www.reddit.com/r/ChatGPT/comments/1t0pyb4/gpt_image_2_prompt_that_is_viral_right_now_redra/
The MD Hallucination
I get the md hallucination! When I started using Claude Code, every action created an md document reporting what it just did. Soon, my folders had way more md files than code files. Does that count as a kind of md hallucination?
The 50,000-Line Script Vibe
Want to see the vibe of a 50,000-line script.
Mercury Bank Ships a CLI
Got an email this morning. Mercury — the bank we set up our US company with for Stripe — just launched a CLI tool, designed for AI agents. Taiwan banks… (looks at the financial regulator)
Claude FM Lo-Fi Stream
The official Claude account is messing around again, launched a Lo-Fi Music livestream. The ability to switch between rigorous safety research and Lo-Fi beats is part of this company’s character.
Mid-May 2026 (W20)
Stripe Atlas: The Four-Step Path
- Anyone can do Stripe Atlas. 500 USD, point and click online, the company is formed. I did mine on my phone in a Chiang Mai bar. 2. Click some more after that to apply for Stripe, take the API and wire it to a vibe-coded website. Did that too. 3. Teach people how to do steps 1 and 2. Done — but free, not paid. 4. This one I haven’t done, because it crosses a line.
L3 BD Outreach Open Rate
First email using “let’s exchange notes” framing (no pitch, no needs ask, just “want to grab some time, exchange notes on AI in practice”) got 67% open rate. Same list with a “let me sell you” framing (course pitch, L3 consulting service) got 17%. Four times the gap. Don’t sell anything in the first email — that’s actually true for the early stage of a BD funnel.
Late May 2026 (W21)
Karpathy — The Ultimate Free Agent
Andrej Karpathy announced he’s joining Anthropic. OpenAI to Tesla to OpenAI (return) to departure to Anthropic. Probably the most prolific team-switcher in AI. Then again, top talent gravitating toward where they see the most potential is itself a market signal.
CLAUDE.md Prohibition Paradox
I wrote in CLAUDE.md: “When corrected by the user, never say ‘you’re right.’” Claude immediately ignored it. After I called it out for violating the rule, its next response was: “You’re right, I shouldn’t have violated the language prohibition.” CLAUDE.md is advisory, not enforced — if you really need to block something, use Hooks.
Learn Slow Enough and You Learn Nothing
In the AI era, if you learn slowly enough, you can skip learning anything at all. By the time you finish, that thing has already been automated. Flip side: people who learn fast enough keep finding they can do more and more.
Vietnamese Coffee Shop Internet
When I first arrived in Vietnam, I avoided those open-front street coffee shops with low tables and camping chairs, assuming the internet would be unreliable. Turns out these places have the best connections — because Vietnamese teenagers camp there all day playing mobile games on a single coffee order. If the latency were high, nobody would show up.
Late May 2026 (W22)
How Someone With GAD Reads a Risk Report
Let me show you how I, with generalized anxiety disorder, read a risk report: One, an 8.6% risk is way too high, unacceptable. Two, if it happens, 70% of cases will match or exceed what I imagined? Unacceptable. Three, we never just worry about 30 days out — 30 years sounds more like it.
Early June 2026 (W23)
Plaintext API Keys Sitting in Codex Snapshots
Scanned
~/.codex/shell_snapshots/and found one snapshot exporting 16+ API keys in plaintext, with 0644 world-readable file permissions.
Two Traps in Vercel Throwaway Public Pages
A fresh project’s deploy URL ships with Deployment Protection (Vercel Authentication) on by default — outside visitors get a 401. The other trap: the bare
<proj>.vercel.appalias may be someone else’s empty shell; your auto-updating prod alias is<proj>-<team>.vercel.app. curl and check the content before sending anyone the link.
Why Quota Reset Times Are Deliberately Staggered
Because if a reset also reset the clock, everyone’s reset time would converge — and everyone would binge right before the cycle ends, piling on peak load and destabilizing the servers. That’s why they deliberately stagger it. My guess, anyway.
Claude Design’s Separate Quota Is Gone
Cold news of the morning: Claude Design’s standalone quota has disappeared, folded into the overall Claude usage quota.
A File Named “Transcript” Isn’t Necessarily a Transcript
The file was named “transcript,” but it turned out to be an already-edited summary — the raw recording lived elsewhere. Bonus trap: SenseVoice STT has no speaker diarization, so in multi-speaker sessions the opening introductions get attributed to the wrong people.
Mid-June 2026 (W24)
Vietnamese Coffee Plus a Nap
Ever since coming back from Vietnam I’ve loved Vietnamese coffee. One trick that works incredibly well: down it, then immediately go nap for 20-30 minutes. You wake up with a crystal-clear head — it feels like dialing Opus 4.7 back to 4.6, productivity maxed out.
Why GPT 5.6 Isn’t Shipping
Word is GPT 5.6 won’t ship this week, because what OpenAI is really aiming at isn’t Opus 4.8 — it’s the soon-to-launch Mythos.
Waiting for the Crying-Open-Codex Moment
Watching a very earnest video, I kept waiting for it to end with “You’ve hit usage limit - reset next weekend,” followed by the scene where you tearfully open Codex. Sadly, no such luck.
Sonnet 4.6 Plus MCP for Suspicious Emails
Ordinary folks don’t need to reach for Mythos — Sonnet 4.6 wired up to an MCP can help analyze suspicious malicious emails too. “The email was crudely made, the template placeholders weren’t even swapped out” — I wouldn’t have noticed that if it hadn’t pointed it out. Just remember to tell the model not to click unfamiliar links, unless you have a proper sandbox.
Late June 2026 (W25)
The AI-Detects-the-Fallen Utopia Is Still Far Off
Someone wished AI could detect a collapsed person in a riverside park and auto-call for help. The problem: covering an entire riverside with detection is basically Skynet — the privacy backlash would explode, and it’s economically unviable. Even Tesla FSD is road-legal in only a handful of places worldwide. The pragmatic option today is to wear the detector yourself (Apple Watch fall detection auto-dials emergency). It’s not that Taiwan lacks it — nowhere in the world has it yet.
People With Body Odor Never Know
Trust me: people with body odor never realize it, no matter how you hint. And the moment you say it outright, there’s a high chance of a wounded ego.
Coming Home, I Realized Taiwan Got Rich
Back in Taiwan this trip, it really feels like the place got rich — in hospitals, restaurants, on the metro, the overheard chatter is all about buying stocks and investing; office workers in the elevator talk about trips to Japan and Europe. A set meal at my usual spot went from NT$160 to NT$220 in two years; an Uber that used to be NT$160-200 is now NT$250-300. Living in Thailand, I can barely keep up.
GMAC Really Is Going Superscore
Can’t believe it — GMAC is seriously pushing GMAT Superscore this time, ushering in the score-stacking era.
First Principles: If Holding a Grudge Doesn’t Help, Delete It
First principles, right? Does holding a grudge make things better? No. Then delete it.
The AI-Course Cash Grab, Overheard at a Convenience Store
Sheltering from the rain at a 7-Eleven seating area in Da’an, Taipei, waiting for class, I overheard some aunties talking about AI: “I’m paying NT$18,000 a month for my son to learn AI, four installments now, it’s that teacher’s course he recommended last time.” Infuriating — I really want to know which teacher rakes it in this well. The market is now buzzing with “Law of Attraction AI,” “Benefactor AI,” “share your gratitude and closed deals with AI,” “AI raises your personal energy,” “networking coach radiating beauty” — I can’t take it anymore. The highest-value overheard conversations really do happen in convenience stores and fast-food seating areas.
A Sense of Boundaries
Learn one more term while you’re at it: a sense of boundaries. Many say it’s a Mainland Chinese coinage, but I find it irreplaceable. The relative you haven’t seen since last New Year prying into your private life — how much you earn, are you married yet — that’s the absence of boundaries.
A Wuxia-Style Water Dispenser
Deep winter, snow falling, the swordsman arrives at the inn, parched. Innkeeper: please have some warm water. “Have you boiling water?” Innkeeper: first press unlock, then press hot.
Late June 2026 (W26)
Local vs Cloud Models
Cloud models? 8GB is enough to run them. Local models depend on your specs — but why go local? If it’s for privacy, that’s fair. If it’s to save money, not really — because making a local model perform on par with commercial ones means shelling out for hardware first. My advice: don’t overthink local early on. Get comfortable with closed-source models first.
Tokenizer Rewrite Means 1.4x Usage
A tokenizer rewrite is the real shocker — 1.4x usage.
Opus 4.6 vs 360 Antivirus
A Chinese user pitted Opus 4.6 against 360 Antivirus. 360 won. That tells you everything about 360’s technical (hijacking) prowess.
Claude Import Mode for GPT Memories
Claude has an import mode, right? You can extract GPT’s knowledge and memories about you, then import them into Claude.
The Ones Who Roast Claude Hardest Use It Longest
The people who roast Claude the hardest are the ones who’ve used it the longest.
Three Verification Misalignments
Verifying against your own spec is worse than verifying against the actual computation — if your spec has an error, you’ve made your own mistake the baseline. Deep-research false negatives: 16/19 in the “refuted” bucket were 0-0 or 1-0 votes (agents timed out without verifying, not actually refuted). Dry-runs validate data integrity but can’t catch an exercise that simply doesn’t work — you need to simulate a real student’s spoken prompts with step-by-step chaining to surface that.
$1,346 Isn’t Productivity, It’s an Anesthetic Bill
ccusage hit an all-time high, +111%. When you separate spending from output, that number isn’t “the cost of efficiency” — it’s “the bill for being spoiled by a tool.”