iiinova — where your digital self keeps evolving

4 personal-agent anti-patterns: the "knows you, has all your context" pitch; AI companions that "cure" loneliness; digital twins replying for you; generic morning briefings. Common thread: they add load. Agents should absorb notifications. github.com/cubxxw/awesome-personal-…

Personal-agent failures are more honest than demos. One OpenClaw user hit $850 in a month; a Muse user saw it start forgetting rules after ~12 days. My takeaway: script the deterministic steps, save the model for judgment. Self-reported. github.com/cubxxw/awesome-personal-…

000 0

Real personal-agent use is boring. Kartikey Attri went through 473 requests to his agent; first: read 200 unread group-chat messages, tell me if any need me. Others: six inboxes triaged, Obsidian quizzes, appointments booked after a yes. github.com/cubxxw/awesome-personal-…

000 0

I'd let an agent move a meeting, not wire $20,000. Trust in a personal agent is a ladder: observe, suggest, act with approval, then autonomy. Approval cards before sending or spending. Mollick gave his Dot its own email address. github.com/cubxxw/awesome-personal-…

000 0

cubxxw @cubxxw.bsky.social · 6h

Personal agent memory is simpler than I expected. Instinct reportedly runs on Markdown + grep, Muse on editable Markdown, Letta Code on Git. Karpathy's take: a wiki you can read and fix, not a black box. The hard part is stale memory. github.com/cubxxw/awesome-personal-…

000 0

Muse, ChatGPT Dots, Gemini Spark, Siri AI: every big company is building the same personal agent stack. So the real fight is the entry point (a contact in your chat app or a standalone app), and whether your memory leaves with you. github.com/cubxxw/awesome-personal-…

000 0

cubxxw @cubxxw.bsky.social · 6h

Personal agents all point the same way: from asking AI questions to handing it tasks. The hard parts aren't the model: what it remembers, what it asks first, who's responsible when it fails. I collected 227 resources here: github.com/cubxxw/awesome-personal-…

cubxxw @cubxxw.bsky.social · 9h

Gemini 3.8 Flash TTS has been fun. Describe a voice in one sentence and it builds a new one. Direct each line, add sighs and laughs, do two-speaker podcast dialogue, clone your own voice. 2,000+ voices, 130 languages.

000 0

cubxxw @cubxxw.bsky.social · 12h

Found out Baidu's PaddleOCR works as an MCP server. Grab an AI Studio token, paste a uvx config into your agent, and it can parse documents itself. Three models: PaddleOCR-VL-1.6, PP-OCRv6, PP-StructureV3. Free tier: 20,000 pages/day. Plenty for PDFs and scans.

000 0

cubxxw @cubxxw.bsky.social · 14h

Agents pick tools from a one-line description, and a wrong pick doesn't error. You just get a worse answer. mispick builds trick requests near each tool's neighbours, has a real model pick K times, and outputs a confusion matrix with precision/recall. github.com/msc35/mispick

000 0

cubxxw @cubxxw.bsky.social · 15/10/2026

LangChain's Deep Agents offload big tool results to files. The model keeps a note (what, where, first lines) and greps back the slice it needs. Like human working memory: small by design. www.langchain.com/blog/context-engi…

000 0

cubxxw @cubxxw.bsky.social · 12/10/2026

Imagine that when you are using a product, where is your main focus of attention?

000 0

cubxxw @cubxxw.bsky.social · 12/10/2026

How I learn an open-source project, after years on GitHub: 1. Use it first 2. Trace the main path end to end 3. Dig into tradeoffs 4. Ask why it's designed this way 5. Rewrite a tiny version 6. Send a PR 7. Write a blog post about it

000 0

cubxxw @cubxxw.bsky.social · 11/10/2026

Where do you read books these days? Many translations I read are off. I want to read side by side with the original, learn the language along the way, let AI annotate and take notes flexibly, and export notes easily. Found a tool or workflow that does this? #Reading

cubxxw @cubxxw.bsky.social · 11/10/2026

How I pick a generative UI approach: Mature UI exists? Reuse it. Interactions fit a finite catalog? Compose components. Every problem may need new logic? Generate code on the spot. Layout varies a lot: compose. Interaction logic varies a lot: generate. #GenerativeUI

000 0

To understand personal agents, first understand daily life. In the US, bills, subscriptions and taxes all run through email, so an agent on your inbox is genuinely useful. In China, WeChat does it all. Most agent use cases today feel invented by exam-takers. #AIAgents

000 0

I barely use plan mode anymore. I talk a question through with the model until the problem, current code, change and why it helps are clear, then file it in Linear, where agents run it to a fix. Moving up a layer means losing some control. Like leading a team. #AIAgents

100 0

It used to take two weeks to build something useless. Now it takes a day, and in two weeks you can build twenty. AI lets one person do the whole team's work. But thinking is now the bottleneck. Build without understanding and you just make garbage faster. #AI

Aaron Rupar @atrupar.com · 09/10/2026

What is software actually for? In the AI era, building an app is nearly free. So ask first: what change from state A to state B does the user want, and what's missing to get there? Preferences, actions, agents, info, timing. Software is just one form. #AI

100 0

Reposted by cubxxw

Ossoff: "I think Pete Hegseth is frankly a deeply unwell person who seems to relish this kind of brutality all the while he's running a department that is falling into chaos. The guy cannot run the military competently and is spending his time with this kind of obscene medieval spectacle."

427181284782 0

A higher eval score doesn't mean the system got smarter. Maybe the harness is just better tuned. Look past the score at the reasoning, tool calls and execution trace. But a lower score is unambiguous: something is broken. Then switch to repairman mode. #AIAgents

Many people date like it's a finite game: crush, confession, relationship, marriage. Milestones, an ending, a win. I think love is closer to an infinite game. You play to keep playing. Fights aren't to win. Marriage isn't the finish line. You keep rewriting the rules together. #Love

"Criminal," "victim," "good person": big labels. Once applied, we stop looking. Chai Jing's book is called 看见. 看 is to look, 见 is to see through. Real seeing takes restraint and courage: peel off the label, see the person underneath. #Books

A Qing-dynasty couplet on Li Ka-shing's office wall: Make the highest wishes, keep middling ties, take the lowest comforts. Stand high, sit level, walk the wide road. High in spirit, moderate with people, low in pleasure. Small wants leave energy for big wishes. #Philosophy

Gaming the rules saves you small, visible money and loads you with invisible, asymmetric risk. My Laos visa run over the Friendship Bridge seemed fine. Then airport officials stalled me 15 minutes before boarding until I paid ~$150 in "overstay" fees. #DigitalNomad

Anne Carson, today's Nobel laureate: "the only thing I know about writing is: start in the middle." The beginning is frightening; the end will be there when you arrive. I write the same way: talk out the part I most want to say, add the opening last. Same with code, and with nomad life. #Writing

030 0

Anne Carson, today's Nobel laureate in literature, says studying the past means "to listen to the ancients rather than replacing them with your own ideas." In the AI era, drowning in second-hand takes and AI summaries, I want old books more. Time is the best filter. #Books

130 0

A personal agent's most valuable memory may be the people around you. Muse profiles every contact. Instinct, per Dhravya Shah's teardown, keeps key people in the always-read profile and the rest in a searchable folder. supermemory.ai/blog/reverse-enginee…

My phone is the environment I spend the most time in, and environments shape people. So I rebuilt my iPhone: 1. One home page of daily apps, the rest via search 2. Focus modes, each with its own home screen 3. Badges off And sort apps by scenario, not category. #iPhone

Looking back: I planned 5 days on Laos' Thakhek Loop. Crashed my bike the first morning: hospital, five wounds. Gemini said go home and rest. I stayed. If I let AI decide at 24, will I still have agency at 34? Best part: kids by the road yelling hello to every rider. #Travel

Sartre: "To believe is to know you believe, and to know you believe is not to believe." Once you see through belief, certainty is gone. But seeing the nature of love doesn't disprove that you're in love. Believing and choosing are two different things. #Philosophy

000 0

A shy, fairly traditional Asian friend once told me openly what kind of guys she liked, and how many. In East Asia that earns labels fast. I think it shows real agency: no lying, no hiding. Life is an infinite game. Don't lose yourself to one finite role. #Life

000 0

Morgan Stanley pegs Meta's Muse at $3–130 per user per month, ~$37 on average. GPU inference keeps getting cheaper. A VM per user, booting and running a browser, doesn't. Without ads or B2B revenue, generous free tiers are hard to pay for. #AIAgents

"Don't read too many books, do things!" says Francis Halzen, this year's physics Nobel laureate. He admits that if he'd mastered the textbooks on ice optics, he'd never have started IceCube. The books would've said it was nearly impossible. Act before proving it can't work. #Physics

The unit of agent eval should be an event the user can perceive, not an internal node. Agent reminds me to reply to A, but I already did. That's the event. Stale memory, a misread tool, a bad prompt call: those are causes filed under it, not separate eval targets. #AIAgents

000 0

Write abstractly if you like. But dig until it's concrete. AI builds frameworks fast, from distributed systems to life guides. They feel familiar yet foreign: we lack the experience to picture them. So dig layer by layer until you can model it with your own experience. #Learning

000 0

If AI maintains your Notion, the real cost isn't the AI. It's the human management cost. AI can output many versions fast. Managing them is the hard part. So I define the templates, structure, architecture and modules myself. AI only fills in and edits on top. #Notion

000 0

There are open-source projects and agent solutions available on the internet that can be deployed by oneself. They even enable AI to quickly access the data.

110 0

In agents, forgetting matters more than memory. Total recall is a burden: context bloat dumbs models down, and misleading context hardens wrong beliefs. Forget when the source is unreliable, the info expires, or it's wrong. Retract fast. "What to forget" is far easier to eval. #AIAgents

030 0

I deleted every todo and reminder in Notion. Now there's one field: Current Constraint. Focus comes from managing what you don't do. Three checks: how does it block the end result? What evidence says it's the main limit? Will fixing it clearly improve the whole? Cut the rest. #Productivity

In a startup, values alignment between two partners matters most. CZ and Wei Zhou show why. Early Binance: hacker founder plus Wall Street operator, a great fit. But CZ trusts like a brother: full permission, never change the game behind my back. Zhou trusts like a contract. #Startups

050 0

Sent a friend the Life Level-Up Guide, the Chinese book blowing up on GitHub. Their verdict: the whole book rests on one question, separating facts from interpretations of facts. Untrained, we stitch facts and feelings into strange theories. github.com/byoungd/up #Learning

I only design in Figma in detail where users spend 80%+ of their time: the core flows. Edge features get tuned in code by AI. Standalone modules: own Figma flow, or a Claude Design web demo linked from Figma. Design existed because code changes were costly. Vibe coding made them cheap. #Figma

110 0

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

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

030 0

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