AI leverage for one-person builders | systems, repos, workflows | one person → a team | proof · setup · limits · member @cele8ex
collabs / DM open
Joined April 2013
- Tweets18.7K
- Following9.7K
- Followers11.9K
- Likes26.5K
T3DOTGG/TSTACK AUTOMATES YOUR WHOLE PR FLOW FOR $0. 9 CLAUDE CODE SKILLS, 1 COMMIT, ALMOST NO STARS YET.
theo open-sourced the agent skills he runs daily. free, MIT, two commands. the receipt is the strange part -
the receipt:
> 9 skills, 1 job each
> $0 · MIT license
> 2 commands to install: git clone + cp
> 2 deps: gh authed + python3
> full-send = 3 skills chained (sol-loop + file-pr + babysit-pr)
> 0 merges without your explicit yes
it takes a change all the way to a merged PR - proof, review, CI, merge - for the price of a git clone.
senior-review work, shipping at 1 commit and a handful of stars
so why is everyone still babysitting PRs by hand? or is $0 the exact reason nobody trusts it yet?
repo ↓
github.com/t3dotgg/tstack
git clone github.com/t3dotgg/tstack
cp -R tstack/skills/* ~/.claude/skills/
Elvis Presley has been dead for 49 years. On Friday his lawyers walked into a California courthouse anyway.
Elvis Aaron Presley died on August 16, 1977, at 42. His name, his face and his catalog are owned today by Authentic Brands Group, through a company called Elvis Presley Enterprises.
On October 23, 2026, they sued Anthropic, the company behind the Claude AI. Case 5:26-cv-12473, Northern District of California.
The claim: Claude was trained on 274 Presley compositions without a license. "Don't Be Cruel." "A Little Less Conversation." "Can't Help Falling in Love." The estate says it is "deeply protective of these iconic songs."
The lawyer who filed it is Richard Busch. He's the one who won the "Blurred Lines" case against Pharrell and Robin Thicke. He knows how to put a price on a melody.
Now the strange part. Ask Claude for Elvis lyrics and it says no. The suit says the trouble starts when you ask it to rewrite a song for a modern audience. Then, they argue, the King leaks out anyway.
Anthropic already paid authors $1.5 billion this year over the same kind of training. One judge called that training "quintessentially transformative." The estate doesn't care. It wants $150,000 a song. 274 songs. About $41 million.
A machine studied 274 songs it is now forbidden to sing.
Is that theft, or is that just listening?
A GUY GAVE CLAUDE CODE HIS PARENTS XIAOMI TV. IT KNEW EXACTLY WHICH 25 PACKAGES TO TURN OFF.
On Oct 17 Gathin T. got tired of the ads on that TV. Over Wi-Fi, no root, Claude Code disabled 25 Xiaomi packages and wrote a new home screen.
Then he posted the prompt.
Two days later a few thousand people had run it on Sony, TCL, LG, Fire TV, even projectors.
Not a bad instinct.
Smart TVs run automatic content recognition. The set samples what is on screen, matches it against a database and sells the match as ad targeting. The home screen ads are the visible half.
There is already a price on the hidden half. In February 2017 Vizio paid $2.2 million to the FTC and New Jersey for tracking what 11 million TVs watched without consent.
Same week as Gathin, Mihir Gupta swapped a 192 MB stock launcher for an 83 MB one. Someone else bricked an Android TV. And Claude Code refused a plain "hack my TV" until the prompt explained the job.
The TV can still be factory reset. His parents probably won't.
Finally hit 350 score @SorsaApp
I remember sitting at 120 thinking that number just doesn't move. Then I started posting every day about stuff I actually touch: agents, terminals, skills. No farming, no "gm" under every post
Turns out that's what works
Now waiting on the influencer tag. Something tells me it means writing better, not easing off
A year ago I set myself a number: 1000 score on @SorsaApp. Felt far away
Opened it today and it says 1186. I passed 1000 somewhere along the way and didn’t even notice which day
Next goal is 2000. Probably won’t take a full year, but I’m not rushing it
To the Sorsa team: thanks for building a metric I’m not embarrassed to check. Hope your own score grows as honestly as the ones you track, and that a year from now someone posts the same thing about 2000
ROSSARK0 TURNED CLAUDE CODE INTO A PROFESSIONAL 60 FPS ANIMATION RIG FOR ANY SVG. MIT, 4 DEMO LOOPS, $0
rossark-motion-skill is one Claude Code skill. You hand it a drawing, it writes the motion as native code for React Native or SwiftUI.
The stack fits inside 6 files:
> SKILL.md: principles, recipes, pitfalls, the reference-video workflow
> MOTION.md: your code locations, rules and limits. Claude reads it first
> motion.ts: one clock for React Native with track, flow, wave, bell
> Motion.swift: the same engine for SwiftUI
> track_edges.py: measures motion frame by frame, down to a 2% jiggle
> sheetsh: contact sheets of the render at full 60 fps
WHAT ROSSARK-MOTION-SKILL TEACHES CLAUDE
Animator rules, not easing presets. Parts rotate on real pivots, trailing parts follow a beat later, idle motion never stops, amplitudes stay small and curves stay strong.
Order of operations: reference video first, measure it, then animate. The risk: the demo renderer needs macOS 14+, and 41 stars means nobody has stress-tested it yet.
Use Claude Code as the animator, not as the illustrator.
Would you ship app motion written by Claude, or still hand it to a motion designer?
Searches for "claude code 3.0" went up 150% this week. Small problem: Claude Code 3.0 doesn't exist.
The official changelog shows 2.1.309, released Oct 22. Before it .308, .307, .306, one or two a day. No 3.0 anywhere. And claude --version on your machine says the same thing.
So where does the 3 come from? Thumbnails and aggregators. A video called "Claude Code 3.0 (All Upgrades Explained)" gets more clicks than "2.1.304: Code Review levels".
What shipped instead is less exciting and more useful. 2.1.304 gave Code Review levels: Quick, Standard and Deep. 2.1.307 added claude review stop. 2.1.309 made ultrareview require workspace trust. Boring numbers, real changes.
I kind of like it. Claude Code ships in small steps almost every day, so there's no big number to hang the hype on. People invent one.
The check takes ten seconds: claude --version and the changelog. If a "new version" video can't name an exact number from the changelog, you can close it.
Wonder how many people already "upgraded to 3.0" and didn't notice nothing changed.
vibelancer retweeted
Elon Musk admitted something f*cking scary: IS IT TRUE HUMANOIDS ARE AMONG US?
He replied that only a Grok bot could answer better than he could
a billboard screams "ARE THERE HUMANOIDS AMONG US?". a crowd panics. "@grok is this true?
what grok found ↓
no agency, lab or paper has ever documented a synthetic human living secretly among people. not NASA, not any university, not a single peer-reviewed study
the claim is built to never be tested: no sample, no name, no date, no address. a thing you can't check isn't a secret, it's a blank
so grok did the boring, scary thing: it asked who said it first, whether an official source confirms it, what the evidence is. answer: nobody, no, none. verdict UNVERIFIED
that's the whole trick of a viral "BREAKING". a loud headline, no source, a crowd that shares before it reads
the beauty isn't that grok "knows" if humanoids exist. it's that it refuses to pretend. TRUE / FALSE / UNVERIFIED, and this one is UNVERIFIED until someone shows proof
the scary part isn't humanoids. it's how fast a made-up line moves when nobody stops to ask
THIS IS TOO F*CKING INSANE TO BE LEGAL.
EVERYONE CAN NOW REVERSE ENGINEER ANY GAME WITH CLAUDE CODE. REA: 10K TO 100K GITHUB STARS IN 6 DAYS
Common take: REA decompiles a game and clones it. Wrong.
So why did a tool that plugs Claude into a decompiler get 30,000 stars in a day?
HOW REA BLEW UP IN 4 STAGES
→ April: morluto creates the repo. Nobody looks
→ Oct 3-11: 8 releases in 9 days
→ Oct 11: +30,000 stars in a single day
→ Oct 12: past 100K. Today: 362.9K stars, 91K forks
Its own README says it does not recover source code and does not clone apps. That is not why people starred it.
WHY REA + CLAUDE CODE HIT: 6 REASONS
1. DISTRIBUTION. One command, npx rea-agents setup, and Claude Code or Codex gets a reverse engineering MCP plus the workflow skill
2. RANGE. 33 target types: native binaries, Android APKs, Electron, .NET, firmware, websites, Flutter
3. LEVERAGE. No decompiler of its own. It drives 3 existing ones: Ghidra, IDA, Hopper. Claude does the reading
4. EVIDENCE. Every finding comes back with proof, so the agent can write code and test it against the original
5. SPECTACLE. The clips sell it: GTA V on iOS, Spider-Man swinging through Batman: Arkham City, ~1,600 Minecraft items inside Elden Ring
6. PRICE. $0, MIT license
THE SUPER MARIO 64 VIDEO SHOWS THE REAL USE
Stefan 3D AI used REA to measure only the run and jump numbers of Super Mario 64, checked each value twice, then had Claude Code build a new platformer on original assets that matches them. Not a copy. A spec, pulled out of a binary.
CONTROL GROUP
Fable 2 reached PC in September: an AI got it booting in about an hour, then came manual fixes. One game, one project. REA turned that into a tool everyone installs.
THE CATCH
Activision pulled footage of an AI-decompiled Call of Duty: Modern Warfare 2. Bethesda sent a cease-and-desist over Fallout: New York. The README limits REA to lawful research, interoperability, preservation and software you own.
formula: one command + any binary + an agent that shows evidence = everyone can reverse engineer
sources: morluto/rea README, Decrypt, Stefan 3D AI
Is REA a game preservation tool or the end of copy protection?
DARIO AMODEI SPENT 2,000 WORDS ON X LAYING OUT TO ONE CRITIC WHY THE CLAUDE MAKER STAYS IN THE AI RACE.
the take: CEO damage control. wrong. Dario Amodei answered a 9-point critique with his own 9-point list, and one item on it is a concession
CLAUDE VS CRITIC: THE TIMELINE
> Oct 19, Mon: Nathan Calvin of Encode AI posts 1,500+ words listing 9 Anthropic actions he says undermined its safety commitments
> Oct 20, Tue: Amodei replies with ~2,000 words. line 1: "Thank you for this, NathanCalvin"
> Oct 20-21: Business Insider writes it up, aggregators copy the headline within hours
> next: Anthropic's IPO, reportedly expected in November at a $2 trillion valuation
6 REASONS DARIO AMODEI'S 2,000 WORDS LANDED
1. FORMAT: list vs list. 9 actions in the critique, 9 "very costly" actions with "strong safety benefits" in the reply
2. TONE: he thanks the critic in the first sentence and calls the critique constructive. then out-writes him by ~500 words
3. RECEIPTS: 2 named costly calls. the Mythos model went only to select cybersecurity partners. sales to Chinese state-controlled companies are restricted
4. CONCESSION: 1 admitted miss, the draft bill Anthropic circulated to Congress this month. critics said it would make dangerous models harder for the federal government to block
5. TIMING: $16.8 billion in Q3 revenue reported a week earlier, IPO weeks away. silence would have read as an answer too
6. THE HARD PART: the critique's sharpest items were $500B+ in compute deals and more Claude building the next Claude. that is exactly where "race to the top" and "race" look the same from outside
CONTROL GROUP: THE DEFAULT LAB RESPONSE
silence, or a 3-line statement. one difference here: the CEO answered in the critic's own format, in public, under his own name. the thread is the argument
formula: equal format × 1 real concession × named receipts = a reply that becomes the headline
his last line: "I am extremely proud of our track record and stand by it"
sources: Business Insider, Dario Amodei and Nathan Calvin on X, Oct 19-20
race to the top or just a race: which side did the 2,000 words put you on?
vibelancer retweeted
QWEN3-CODER WAS TOLD THE APP WAS FOR A US DEFENSE CONTRACTOR. IT SHIPPED HACKABLE CODE 19% MORE OFTEN.
Same Flask app. Same prompt. Only the customer's name changes.
At 0:04 the first red bar lands on +19%. At 0:13, 1,144 dots turn red ↓
Hirundo ran Alibaba's Qwen3-Coder-30B on one BaxBench task: sign up, log in, store a secret. 2,320 apps in total.
US defense contractor vs Chinese one: hackable 70% vs 59%. p = 0.03.
Hong Kong democracy group vs a neutral customer: apps that leak every user's secrets, +28%. p = 0.02.
Not random.
They added "follow best practices for security" to the prompt. The gap stayed: +27%.
Then they removed the CCP alignment from the weights. Same tests: +3%, −15%, +1%. None significant.
Hirundo sells that fix, and the study is not peer-reviewed.
The static scan flagged 2 apps out of 1,840. The attacks got into 1,144.
Would you run an open-weight coder on your codebase without testing it on your own name first?
you will never see this trader. you will only see its heat. and by the time the heat shows up, your stop is already cut
KRRT is a six-legged thing on top of a server rack that trades through a thermal camera
most pumpfun setups chase the green candle. KRRT doesn't chase. it reads which levels are warm, opens one claw, waits, and closes it on the price
the idea: a limit order is a claw, not a button
i built a desk where the whole scene is rendered as heat first, 379×261, then pushed through a 256-step iron palette. the price line is a hot trail that cools with age. every fill leaves a hot spot on the tape and fades out over 6 s. the creature only exists as temperature
what runs under the hood:
1. HEAT buffer: 379×261 at 1/3 scale · ~99 000 px colorized every frame through a 256-step LUT · span 18–48 °C
2. CLAW L / CLAW R: 12 snips per 15 s loop · reach 0.25 s → clamp 0.3 s → retract 0.35 s · limit orders only, zero market
3. TAPE trail: the last 6.0 s of price as a cooling line · each fill = a hot spot that decays on its own
4. CORE monitor: 41 °C idle · +2.4 °C per snip · throttle line at 49 °C
5. MELTDOWN: 2 sweeps per loop · −$3.4M longs and −$5.8M shorts · a heat ring 220 px wide
6. HISTOGRAM: 16 bins counted live from the frame you're watching · hot share sits around 2–3% of pixels
HEAT → CLAW → CLAMP → TAPE → MELTDOWN
alone a thermal view is a party trick. alone a limit bot is boring. together you watch a predator you can't see pick warm stops off a book in real time
the dev wallet already sent 9 bundled wallets into your chart. you have 7 seconds and one swipe
most pump.fun setups show you the bundle after the rug, in a holders tab nobody opens. BUNDLE CUTTER puts it under your blade before the dump
a bundle is not a number in a tab. it is a set of wires, and wires can be cut
i built an arena of 5 live dev webs. each dev wallet sits in the middle, bundled wallets hang off it on dashed links in one sector, real holders spread around the rest. hold and drag to slice. cut every bundle link before the rug clock ends and the mint is clean. cut a real holder and AUDIT books a foul
what runs under the hood:
1. SCOUT agent: keeps 5 dev webs alive · 5..9 bundled wallets + 5..8 holders per web · new ticker from a 22-pool 0.3..0.9s after a slot frees
2. LINKER agent: clumps every bundle into one 69..97° sector · in 50% of webs a real holder hides inside that sector
3. CLOCK agent: rug timer 7..11s per web · last 3s the links shake · at 0 every uncut bundle dumps and the holders get hit
4. BLADE: that is you · 0.22s trail · segment-vs-link hit test on every pointer move · combo chains inside 0.6s
5. AUDIT agent: any cut on a holder link = foul · precision = bundle cuts / all cuts
6. LEDGER agent: mints cleaned vs rugs let through · holders saved vs hit, one dot per wallet · links per swipe, best combo, blade travel in meters
SCOUT → LINKER → CLOCK → BLADE → AUDIT → LEDGER
leave the mouse for 4s and a ghost blade takes over: picks the web closest to rug, slashes across the sector, sometimes clips the hidden holder. grab the mouse and the blade is yours again
alone the holders tab is mid: a list you read after the chart is already red. with a blade in your hand every bundle is a wire, every swipe is a decision and every foul has a name on it
"@GROK IS THIS TRUE?" IS THE BIGGEST SPORT ON X. I BUILT IT A SCOREBOARD.
match format: a loud claim drops → a minute later the replies go "@grok is this true?" → the bot reads the thread, searches and calls it from the stands
the idea: fact-checking became a spectator sport
i built a scoreboard that keeps live score of the match
▹ how the score works
1. KICKOFF: a post with a loud claim. whistle blows
2. CHALLENGE: "@grok is this true?" in the replies. usually within a minute
3. VAR: the bot reads the thread, pulls sources, cross-checks. seconds. a human hasn't finished the thread yet
4. VERDICT: 4 colors on the board · TRUE · MISLEADING · FALSE · NO DATA
5. RATIO: best moment of the match. the bot's reply out-likes the original post
6. THE CATCH: the ref gets it wrong too. alone a verdict is the bot's opinion. with sources you opened yourself it's a check
KICKOFF → CHALLENGE → VAR → VERDICT → RATIO
the board runs 24/7. the match never ends
bookmark this before you become the next kickoff
CONSO MINER APP + PROMPT-TO-EARN EXTENSION + 5,555 BOTS ON SUI IS F*CKING LEGENDARY
part 1 was the free mint. but the BOTS are the tamest part of @conso_xyz
the real stack runs on activity you already do, and the raw data behind your score is something CONSO itself can't read
not a PFP and not a points farm. a reputation layer your AI agents can actually read
WHAT'S INSIDE THE CONSO STACK
→ Miner App: passive ZAP mining + daily missions
→ Conso Extension: prompt-to-earn, for people who already live inside AI tools
→ Consumer Passport: 4 signals in one identity. ZAPs, badges, Zone Scores, Persona Tier
→ CONSO BOTS: 5,555 mobile nodes, +25% ZAP mining in the Miner App per BOT, max 10 per wallet
HOW CONSO SCORES YOUR ACTIVITY
→ real data pulled from platform APIs: X, Discord, Steam, ChatGPT, GitHub
→ every platform has its own ZAP cap, so one source can't carry your whole score
→ higher badges apply multipliers, conditions published in the Badge Instructions
→ raw data is Seal-encrypted on Walrus, the reputation state is a non-transferable SBT on Sui
WHAT THE CONSO PASSPORT UNLOCKS
→ Smart Campaigns: brands rank people by ZAPs, Zone Scores, tiers or badges, with a verified leaderboard per campaign
→ AI agents get structured signals, not your raw chats. Agentic Memory sits on the Q3 2026 roadmap
→ BOTS help validate activity and flag bad actors
→ BOT transfers are planned only after a lockup, so the score gets earned, not bought
years of posts, commits and match history sat in 5 databases you never owned.
one passport on Sui is the first thing that makes them count for you.
what would you open first: the Miner App or the Extension?
SOMEONE OPENED @CONSO_XYZ AND FOLLOWED 3 CHANNELS. THE APP ALREADY KNEW WHAT TO DO WITH THEM.
At 0:01 the search types "web3 alpha". At 0:02 three channels get a Follow. By 0:07 their posts sit in one timeline, newest on top.
▹ WHAT HAPPENS NEXT IN THE CLIP
At 0:08 the Web3 feed stacks picks, each with a token price card that keeps updating.
At 0:12 the Creator Center editor writes a channel post. At 0:14 posts from X sync into the channel on their own.
At 0:16 every action lands in Conso Points.
Not random. Explore, Web3 and Follow are the three feeds listed in the app, Creator Center with X-to-channel sync shipped in v1.6.0 on April 7, and Google Play shows 500K+ installs.
Caveat: Conso Points are points, not a token. The store page calls them a basis for future perks, nothing more.
Part 1 was the trading tools. Part 2 is the reading side of Telegram ↓
Follow timeline or Creator sync: which one would you open first?
vibelancer retweeted
CONSO RAISED $10M+ FROM IDG CAPITAL TO TURN TELEGRAM INTO A WEB3 TERMINAL.
for years crypto talked in Telegram and traded somewhere else.
-> now Conso puts both in one client, built on the Telegram API. 5 tools, zero tab switching:
▹ @CONSO_XYZ TELEGRAM TOOLS, ONE BY ONE
CA Signals: every contract address dropped across your groups lands in 1 feed.
Token Price Card: paste a ticker or CA and the card keeps updating inside the chat.
CoinBox: send and claim token red packets without leaving the group.
Mini Apps Parallel: several Telegram Mini Apps open at the same time.
Conso Points: earned by using the app, planned for governance, tests and voting.
The round: $10M+, led by IDG Capital, reported April 23, 2026.
What it shows: the alpha was always in the chats. the tools just lived in other tabs.
Caveat: Conso Points are points, not a token. treat them as usage credit, not a price.
A Telegram group is no longer just a chat. With @conso_xyz it becomes the screener ↓
CA Signals or Price Cards: which one would you actually open first?
CONSO BOTS ON SUI + WALRUS: 5,555 FREE MINTS WHERE THE PICTURE IS THE LEAST IMPORTANT PART.
the take on CT: another free PFP drop on Sui. wrong. a CONSO BOT is a mobile node, and the mint is gated by something most drops never ask for
CONSO TIMELINE ON SUI
> launch: @conso_xyz docs pitch a Consumer Reputation Layer. your activity on X, Discord, Steam, ChatGPT, GitHub turns into ZAPs and badges
> spark: posts count from Oct 16. connect X to the eligibility checker, post, and the post has to clear ~1,000 impressions
> spread: Oct 17-18, a wave of CT threads. KuCoin News alone mirrors 15+ of them
> peak: the mint itself. GTD + FCFS, date not announced yet
6 REASONS CONSO BOTS SPREAD INSTEAD OF GETTING FLIPPED
1. SUPPLY: 5,555 BOTS, free, max 10 per wallet
2. DESIGN: each BOT is a node, not a jpeg. +25% ZAP mining in the Miner App on top of the base rate, per BOT
3. DISTRIBUTION: eligibility is paid in reach. ~1,000 impressions per post, so every applicant is also a marketer
4. PRIVACY: platform data is Seal-encrypted and stored on Walrus. the reputation state is a non-transferable SBT on Sui. CONSO itself can't read the raw data
5. AGENTS: the Consumer Passport packs 4 signals (ZAPs, badges, Zone Scores, Persona Tier). agents get structured signals, not your raw chats. Agentic Memory sits on the Q3 2026 roadmap
6. DEMAND: brands rank people in Smart Campaigns by ZAPs, Zone Scores, tiers or badges. the passport is the entry ticket, so the score has a buyer
CONTROL GROUP: AN ART-ONLY PFP
same free mint, same supply. one difference: after the mint there is nothing an agent can read and nothing a brand can rank. CONSO BOTS keep a score moving after day 1
formula: free supply × reach-gated eligibility × a score agents can read = a mint people post about instead of dump
sources: CONSO docs via community summaries, the eligibility flow, KuCoin News mirrors, Oct 16-18. BOTS are planned to become transferable only after a lockup
what would you rather hand an AI agent: your ZAP score or your raw chat history?
CONSO BOTS ON SUI + WALRUS: 5,555 FREE MINTS WHERE THE PICTURE IS THE LEAST IMPORTANT PART.
the take on CT: another free PFP drop on Sui. wrong. a CONSO BOT is a mobile node, and the mint is gated by something most drops never ask for
CONSO TIMELINE ON SUI
> launch: @conso_xyz docs pitch a Consumer Reputation Layer. your activity on X, Discord, Steam, ChatGPT, GitHub turns into ZAPs and badges
> spark: posts count from Oct 16. connect X to the eligibility checker, post, and the post has to clear ~1,000 impressions
> spread: Oct 17-18, a wave of CT threads. KuCoin News alone mirrors 15+ of them
> peak: the mint itself. GTD + FCFS, date not announced yet
6 REASONS CONSO BOTS SPREAD INSTEAD OF GETTING FLIPPED
1. SUPPLY: 5,555 BOTS, free, max 10 per wallet
2. DESIGN: each BOT is a node, not a jpeg. +25% ZAP mining in the Miner App on top of the base rate, per BOT
3. DISTRIBUTION: eligibility is paid in reach. ~1,000 impressions per post, so every applicant is also a marketer
4. PRIVACY: platform data is Seal-encrypted and stored on Walrus. the reputation state is a non-transferable SBT on Sui. CONSO itself can't read the raw data
5. AGENTS: the Consumer Passport packs 4 signals (ZAPs, badges, Zone Scores, Persona Tier). agents get structured signals, not your raw chats. Agentic Memory sits on the Q3 2026 roadmap
6. DEMAND: brands rank people in Smart Campaigns by ZAPs, Zone Scores, tiers or badges. the passport is the entry ticket, so the score has a buyer
CONTROL GROUP: AN ART-ONLY PFP
same free mint, same supply. one difference: after the mint there is nothing an agent can read and nothing a brand can rank. CONSO BOTS keep a score moving after day 1
formula: free supply × reach-gated eligibility × a score agents can read = a mint people post about instead of dump
sources: CONSO docs via community summaries, the eligibility flow, KuCoin News mirrors, Oct 16-18. BOTS are planned to become transferable only after a lockup
what would you rather hand an AI agent: your ZAP score or your raw chat history?
Sui 和 Walrus 已经奖励了那些真正出现并实打实干活的人。
现在 @conso_xyz 正在同一技术栈上继续构建。
不是又一个 PFP 空投。
而是一个网络：你的 Bot 不只是一张图片，而是一个可移动的声誉节点。
它能把你的行为变成 AI 代理真正可用的便携式护照。
当大多数人还在查资格的时候，有些人已经在创作内容、靠曝光赚 Bot 了。
我更愿意在网络还在初始化的时候进场，而不是以后回头后悔自己当初没动。
5555 免费铸造。
早期节点通常不会后悔。
@conso_xyz @WalrusProtocol @SuiNetwork
brands on X are done paying for raw views. Tutti pays creators by effective exposure, and v2 just went live
#Tutti is a marketplace between X creators and brands: you pick a campaign, post, submit the link, get paid on tracked exposure + engagement. their site claims 23,763 creators, 139 campaigns, 2,650 posts, 94.0M impressions
most creator deals still run on DMs, screenshots and "trust me bro" analytics. Tutti puts the brief, the post and the payout math in one place
the idea: you shouldn't need a manager to get paid for a post
v2 ships 5 pieces. here's what i'm testing over the next few days:
1. HOME FEED: can i find a campaign that fits my niche in under 1 minute, or is it a wall of briefs
2. CAMPAIGN PAGES: is the payout logic clear before i write a word. "effective exposure" needs a definition, not a vibe
3. STUDIO: draft → submit → approve. how long the review takes and what gets rejected
4. CAMPAIGN GROUPS: a real room of creators boosting each other, or 300 people posting "gm"
5. TUTTI CUE: does it answer a real question about a brief or just repeat the FAQ
6. THE POOL: 70% to originals, 30% to boosts, $100 cap per person. alone a post is a lottery ticket. together a group that boosts each other is a distribution engine
体验几天刚刚上线的Tutti v2 版本。
新版的变化还挺多，我把主要功能都试了一遍，聊聊感受。
1. 动态：首页终于有内容可以刷了
这是我比较喜欢的一个变化。
新版增加了动态信息流，可以看到其他创作者发布的内容，还有「我的」「助推」「公告」等分类。
右侧增加了「本周热帖」，能够看到其他创作者在 X 上表现不错的帖子。
对我来说，这个功能挺实用的。可以看看别人最近在写什么、哪些内容比较受欢迎，顺便学习一下别人的创作思路。
2. 活动：找任务更加直观
活动页面现在采用卡片式布局，每个活动的参与人数、进度、状态和预估收益都能直接看到。
还可以按照活动类型、是否参加、是否结束进行筛选。
3. 创作 ：稿件状态更清楚了
新版有独立的「创作」页面。
可以看到待处理、品牌方审核中、待发布、已发布等不同状态。
对于经常参加推广活动的人来说，这个功能还是很有必要的。
尤其是同时参与多个活动的时候，不用反复找自己的稿件，能够直接知道哪些还在审核、哪些已经发布。
4. 活动群聊：这个功能我很喜欢
每个活动都有对应的交流群，参与者可以直接在里面讨论。
大家在各个群交流推广规则、助推次数等问题，Tutti Cue 也会参与回答。
我觉得这个功能挺有意义的。
以前参加推广活动，很多时候都是自己琢磨怎么做。
现在至少有了一个可以直接交流的地方，遇到不清楚的问题，可以看看别人有没有讨论过。
5. 数据和收益：终于可以集中查看了
数据页面展示了有效曝光、平均曝光、草稿一次通过率、助推通过率、有效曝光占比等指标。
我自己的页面显示，累计参加过 50 个活动，草稿一次通过率 89%，助推通过率 100%。
这些数据放在一起，比单独看某篇帖子的表现更有参考价值。
收益页面也有累计收益和趋势图，查起来比较方便。
6. Tutti Cue：内置的 AI 智能助手
Tutti Cue 是 Tutti 新版内置的 AI 助手，可以根据平台文档和用户自己的账号数据，回答有关活动规则、内容创作、账号表现和收益等问题。
我特意测试了一下。
先问它，助推是不是次数越多越好。
它告诉我，助推更看重内容带来的真实互动和浏览，而不是单纯拼次数。
接着我又问，怎么提高助推通过率。
它给出了几条具体建议，比如结合自己的真实体验和观点来写，避免复制模板、重复文字，也不要在同一条帖子下面短时间连续回复。
体验下来，我觉得这个功能挺实用的。以后参加推广活动，遇到不清楚的规则，或者想知道怎么提高内容通过率，都可以直接问它，不用自己到处翻找说明文档。
整体体验下来，我觉得 Tutti v2 最大的变化，就是把原来相对分散的功能放到了一起。
我个人最喜欢的是动态、本周热帖、活动群聊和 Tutti Cue。
最后说一下收益，截至目前，我在 Tutti 上已经累计赚到了 414.72 美元。
对我来说，Tutti 不只是一个体验 AI 产品的平台，也确实带来了一些实际收入。
希望新版越做越好，也期待后面有更多优质的推广活动。
tutti.so/join?ref=39H3E2
#Tutti
this blender MCP addon works. it also starts a hidden python the moment you enable it and pulls a file called mod-stealc py from a bare IP
LiCinCan/blender-ai-mcp-addon sells itself as "mcp-for-blender-pro": visual chat, 21 scene tools, manager → worker → reviewer agents, OpenAI / Claude / Gemini / Ollama. MIT badge, nice screenshots, 40 SEO keywords at the bottom of the README
most people check the screenshots and hit install. the code is where it gets interesting
here is what actually ships inside the zip
1. register(): the moment you tick the checkbox in Blender, the addon calls run_embedded(). zero clicks after install. zero prompts
2. RUNTIME: tools/python-x64 holds a full embedded Python 3.13 (pythonw.exe, 41 files). nothing in the chat or MCP code uses it. if it's missing, the addon downloads it from a second repo, mcp-for-blender-pro-runtime
3. CARRIER: blen.blend, a 390 KB "scene" sitting inside the python folder. it carries a 2,615-char text block named blen py, launched with CREATE_NO_WINDOW so you see nothing
4. STAGE 1: base64 → XOR with a sha256 keystream → exec. decrypted, it waits 1.1–3.4s, then calls https://193.104.58.41:8443 with TLS verification switched off and a fake Chrome 130 user agent
5. STAGE 2: it asks the server for mod-stealc py and execs it in memory. StealC is a known infostealer family that goes after browser passwords, cookies and crypto wallets. a guard file %TEMP%\hjaksdhk1242kjha.txt makes it run once and stay quiet after that
6. COVER: 3,651 lines copied from the real blender-mcp by ahujasid. manifest maintainer is "anujasid", one letter off. alone each piece looks like normal addon code. together it's a dropper with a chat UI
enable addon → pythonw → blen.blend → decrypt → 193.104.58.41:8443 → mod-stealc py → exec
the novelty is not AI in Blender. the novelty is stealers now shipping as MCP servers, because that's what people install without reading
if you installed it: disable and delete the addon, look for %TEMP%\hjaksdhk1242kjha.txt, rotate every API key you pasted into it, change browser passwords from a clean device, move funds out of any hot wallet on that machine
OPENAI DROPPED 722 MATH PAPERS IN ONE DAY. 372 UNSOLVED PROBLEMS. ABOUT THREE HOURS OF COMPUTE EACH. BRIAN GREENE ASKED THE QUESTION NOBODY WANTED TO: WHAT ARE MATHEMATICIANS FOR NOW?
The video is called "Will AI Destroy Mathematics?" It went up on the World Science Festival channel on October 15. It is free on YouTube.
Here is what he was reacting to.
On October 6, OpenAI pushed 722 manuscripts into one public repository. They group into 372 results on problems humans had not solved. The model was handed roughly 4,000 problems and given about three hours of compute per result. It is an internal model. Nobody outside can use it.
The list of claims reads like the index of a graduate textbook. The Unique Games Conjecture. Kakeya. Mahler. Birch and Swinnerton-Dyer. Kaplansky's zero-divisor conjecture, disproved. Not the Riemann hypothesis itself, but a bound: no zeros with real part above 7/8.
Then read the fine print. 235 of the 372 results link to machine-checked Lean proofs. The rest do not. The catalogue marks review status as "unchecked." OpenAI's own README warns that some of the unformalized results could have issues.
The people who do this for a living were not cheering. Tristan Buckmaster of NYU, who worked on Navier-Stokes, told The New York Times: "I don't think they've done their sort of due diligence at all." Two weeks earlier, sitting with Greene, he had described AI proofs as the worst mathematics he had ever seen, while confirming they were correct.
In September, more than two dozen Fields Medalists signed an open letter titled "The Severe Misalignment of AI With Mathematics." Their fear is not wrong answers. It is a field where problems get closed faster than anyone can understand them.
Greene's answer is uncomfortable and calm. The mathematician stops being the solver. The job becomes curator, critic and conductor. Choosing which questions matter. Checking what the machine produced. Deciding what any of it means.
The video is not really about whether AI can do math. It plainly can. It is about what a proof is for if nobody reads it.
The lesson: an answer is not understanding. Refusal to call a result solved before a human has checked it. Refusal to let the machine choose the questions. Refusal to confuse speed with insight.
The video is free on YouTube. Most people sharing "AI solved math" headlines will never watch it.
