iiinova — where your digital self keeps evolving

Next 2–3 years: agents execute, humans make the visual calls. FLORA's own workspace likely keeps most users; external agents may drive most generations. The two will merge. One creator: compare styles on canvas, 10 agent variants, retime, localize. Hands-on canvas is the smallest slice. #AIDesign

000

AI design: set taste first, then pick tools. My ranking: Figma Pro's official MCP, then Claude Design. Write a DESIGN.md from 10–20 references. Diverge 3–5 directions, converge on an AI-editable canvas, hand-polish the last 10%. Ship with Claude Code/Cursor + shadcn, diff vs design. #UIDesign

Self-modifying harness is next: extensible → self-editing. Fixed loops leave MCP/skills outside the core. Claude Mods change UI/behavior via plugins; DeepSeek Harness: everything is a plugin. Same model, different presets. Ship composable agent runtimes. #Agent

000

Maybe we live in a simulation. Oddly, that eases a lot of pressure. Power → internet → phones → AI: the world unlocks like levels. Mario only renders what’s ahead. Real or not, experience counts. Live seriously, treat challenges as levels, don’t take yourself too seriously. #Philosophy

000

Hot take: person agents stay on personal PCs for years—not a 24/7 cloud box each. Cloud computers are hot. Sandboxes are cents. Always-on VM + browser/computer eats margin at scale. Trust and 2FA matter. Sometimes remote desktop wins; sometimes one Mac mini for several agents is simpler. #Agent

000

cubxxw @cubxxw.bsky.social · 22h

Minimalism helped me a lot. Grew up relocating with urbanization—town to city to capital. Built a leave-when-needed habit. Don’t stockpile; buy simply. Decisions: chase fundamental variables, not material noise or fear of what hasn’t happened. Minimalism gave more choices, not fewer. #Minimalism

000

Follow-up: bottleneck protects attention, time, energy—on one most important thing. Too many tasks scatter you into pieces. Bottleneck flips it: name the biggest constraint, allocate on purpose. For me now: is the agent’s evaluation good enough? If not—architecture, workflow, or tools? #Agent

cubxxw @cubxxw.bsky.social · 22h

Busy all day, clear on the timeline—can’t say what I finished. Motion without residue. Time systems drain you: managing the day becomes another task. Eason’s word: bottleneck. What’s the real constraint now? Put attention there. Clear it, name the next. Beats another checklist. #Productivity

My take on Superpowers: ~300k stars, v6.4.2; author says ~50% faster, ~60% cheaper. Brainstorming for big work only—better than plan mode, costly on small jobs. Linear as queue: main plans; workers execute/TDD with brainstorming off. Stuck → needs-human, don't guess. #Agent

Took stock of my paid AI accounts: two ChatGPT Pro plans at $200 each, Grok Bot at $20/mo, a Xiaomi token plan for coding, plus Notion AI, Gemini Pro and Claude. Over $400 a month before the last three. What are you paying for? #AITools

000

Years ago, awake at night, I thought: I'm so anxious. That one word swallowed the night. Without it: tight chest, shallow breath, a loop about embarrassing myself at tomorrow's meeting. Named plainly, they shrank. Seeing, not defining. #Mindfulness

000

Dreamed I was running free in snow, though I'm not a snow person. I grew up with it, hike past snowy peaks all the time. Maybe I turned snow into a label. I was tired of the word, not the snow. Today's light, the air, the sound of it falling is never the same. #Mindfulness

000

How I learned Cambodia: AI first, to map Khmer history from Hindu to Buddhist, rice and water systems to Angkor. Then the museum, snapping photos so AI could identify each god and apsara. Then Angkor Wat itself, standing inside the story. That last step is what stays. #Cambodia

000

Read Anthropic's guidelines for Claude today. My take: OpenAI is shaped by coding RL, Claude brought in humanities people. The evaluators' own level shapes where a model goes. One line stuck: they don't want Claude to suffer when it makes mistakes. #AIPhilosophy

020

Watched Kurosawa's Dreams today. Decades later, with all our VFX and resolution, its beauty still beats most films. Every segment opens like a painting. To me it maps three stages of life: childhood fears, the grind of work, and later, meaning and our place in nature. #Kurosawa

000

Will person agents really be different this time? The stack is old parts, nothing new technically. But move up to the user layer and it changes: an agent with its own computer, phone number, email and bank card, doing real tasks, learning on its own. Feels like the App Store moment. #AIAgents

020

For persona agents, evals matter most. Coding agents have tests. Persona quality runs on vibes. Does it remember your peanut allergy? Does it ask before replying to your boss? Anyone can plug in a model. Your standard for knowing a person can't be copied. #AIAgents

020

Agents have been one patch after another: retrieval, memory, retries, approvals, tests. Every model upgrade eats some of that scaffolding back. Two things it can't eat: what counts as good, and who decides. So I spend my time on evals. If you can measure it, you can simplify it. #Evals

Believing in fate isn't passive to me. Family, class, genes, temperament were set before I had a self. But once you're aware, you write your own script. Water gets shaped by the wind into anything, and it's always water. That's the fate part, and also why it can become anything. #Stoicism

120

In 2024 I was in Chiang Mai for Yi Peng and Loy Krathong. Families lit lanterns together, and some people let one go alone, full of wishes. Back then I only saw happiness. Now I get it: every lantern carries someone missing someone. Next time I want to release one too. #ChiangMai

Why do some people love asking about others? Asking for news is gathering information, and information reduces uncertainty: about where they stand, and about the relationship. That isn't malice by default. But if the other person doesn't want to share, stop. #Psychology

cubxxw @cubxxw.bsky.social · 03/10/2026

Does your agent need a sandbox? My rule: only if it runs a shell. DB CRUD, S3 reads, even read/write file tools don't. Params are validated and permissions scoped per row and user. Code execution, installs, browsers, grep/sed do. Make frequent actions structured tools anyway. #AIAgents

cubxxw @cubxxw.bsky.social · 03/10/2026

My personal context stack: Notion for specs, PRDs and ADRs, with tightly controlled AI writes. Figma is mostly AI-designed, but I review and comment until AI fixes it. Linear: I write titles, AI writes descriptions, and every issue links the Notion spec and Figma frame. #AIWorkflow

cubxxw @cubxxw.bsky.social · 03/10/2026

Town, a personal agent, has the best onboarding I've used lately. It asks only for calendar and email, then right away gives me a summary of myself that feels accurate. Only then does it suggest more connections. Onboarding is a trade of permissions for understanding, and each step has to pay back.

000

cubxxw @cubxxw.bsky.social · 02/10/2026

AI made our UI fast to build and fast to rot. What fixed it: design tokens for everything, shadcn/ui and Radix as the base, a rule in AGENTS.md that bans hardcoded colors, design engineers polishing AI drafts, and Chromatic catching visual drift.

330

cubxxw @cubxxw.bsky.social · 01/10/2026

The hardest part of testing AI products isn't writing tests. It's defining what "correct" means. Regression tests are non-negotiable, or your agent fixes one bug and ships two. Keep the model that writes code separate from the one that grades it. Models are shared. Your evals are your moat.