Mr. Buzzoni

3.2K posts

Mr. Buzzoni banner
Mr. Buzzoni

Mr. Buzzoni

@polydao

Crafting the future of intelligence | AI Wizard & Builder | co-founder @polynternet | Partnerships & collabs welcome • DMs always open

Katılım Temmuz 2021
620 Takip Edilen20.3K Takipçiler
Sabitlenmiş Tweet
Mr. Buzzoni
Mr. Buzzoni@polydao·
THIS "FOUR AGENT LOOP PATTERNS" GUIDE TURNS CLAUDE INTO AN EMPLOYEE THAT SHIPS WHILE YOU SLEEP spoiler: prompt engineering didn't survive the autonomy ladder, low to high: > turn-based (25%) - one task, then control comes back to you. good when you're still figuring out what you want > goal-based (70%) - set the target with /goal, an evaluator grades every attempt, Claude retries until the bar is cleared. built for passing tests, closed bugs > time-based (60%) - /loop wakes it on an interval: pull the PR, react to review comments, patch failing CI > proactive (95%) - an event triggers it, a router splits the work across child agents, a separate model checks it before it ships. you're not even in the loop anymore the line worth remembering: the system carries the quality, not the model these 4 patterns are just the entry point. designing one is a different skill than running one climb it 👇
Mr. Buzzoni tweet media
Alex@de1lymoon

x.com/i/article/2075…

English
22
5
84
3.5K
Mr. Buzzoni
Mr. Buzzoni@polydao·
Andrej Karpathy told Andrew Ng the way he learned AI is the opposite of how most engineers do it now > instead of reaching for a framework -> he built his own neural net from scratch first just to feel how backprop moves through every layer > he wired each one by hand before trusting anyone else's. the framework came last his reason is the whole tweet: "if you didn't build it, you can't debug it, and you can't improve it" watch the full sit-down 👇
Mr. Buzzoni@polydao

x.com/i/article/2036…

English
11
4
67
3.5K
Annatar.md
Annatar.md@xieike·
ANTHROPIC ITSELF JUST PROVED HOW BADLY PEOPLE PROMPT CLAUDE > their applied AI team dropped a free 22-minute workshop not a creator guessing from the outside - the actual team that wrote the weights > here's the part that stings: a properly structured Claude prompt runs on 6 elements, and most of us are stuck using 1, maybe 2 if we're trying that single gap explains the answer that felt slightly off, the four rewrites, the prompt that worked once and never again > watch it tonight and you know something most daily users never will -> skip it and you keep paying full price for a fraction of the tool i watched it twice. then built a Claude Skill that runs all 6 elements on every prompt by itself, so i stopped thinking about structure completely workshop first, skill setup below. this one compounds 👇
Annatar.md@xieike

THIS "GETTING STARTED WITH LOOPS" DOC TURNS CLAUDE INTO AN EMPLOYEE THAT SHIPS WHILE YOU SLEEP spoiler: prompt engineering didn't survive the 4 loop types, from you-drive to fully hands-off: > turn-based - 25% autonomy. one task per round, control returns to you. best for exploration when requirements are unclear > goal-based - 70%. you define "done" with /goal and a separate evaluator keeps sending Claude back until it clears the bar. tests passing, target scores, bugs cleared > time-based - 60%. /loop fires on a schedule: check the PR every 5 minutes, address review comments, fix failing CI > proactive - 95%. an event fires, a router dispatches child agents in parallel, a review model oversees the output. no human online at all the advice in it that matters most: loop quality depends on system quality, and tokens are the real cost knowing the 4 types is level one this article is the full 14-step roadmap from typing prompts by hand to designing loops - /goal, SKILL.md, worktrees, MCP, all of it start climbing 👇

English
6
4
36
1.9K
Fokki
Fokki@0x_fokki·
🚨 Warner Bros spent over $1,000,000,000 filming 8 Harry Potter movies a 24-year-old spent $50 last month and cleared $11,900 rebuilding that world in CapCut this AI-Harry-Potter trend owns Douyin: Voldemort and Dumbledore playing poker, ghosts drifting over the dinner table, all shot like a phone video. he copied it and hit TikTok and Reels first. > Research: find the clip pulling 10x a channel's normal views: 1 day > Claude: write the shot list for each scene frame by frame: 15 min > CapCut: render the characters photoreal, add AI voice and captions: 40 min > Make: auto-post to TikTok, Shorts, Reels, pull the views at 48h $50 a month runs the whole set. every new character is a few lines of text. the films needed a backlot, a cast and 10 years. he needed a weekend. his exact setup is in the article above👇
Fokki@0x_fokki

x.com/i/article/2074…

English
31
11
131
14K
Nikita
Nikita@0xVishnya·
Free Crypto Cards With Cashback And Limits 👇 Every landing page sells you the rate. The monthly cap decides what you actually get: same "cashback", anywhere from $1.50 to unlimited. Official docs of 13 free tiers, worst to best: @useTria : 1.5%, max $1.50/mo @Plasma One Lite: 2%, max $10/mo @wirexapp : 0.5%, redeem cap $25/mo @KASTxyz : 1.5%, max $30/mo @BleapApp : 1% base + 20% on subs, ~$30/mo and up @ether_fi : 3%, max $60/mo @lava_xyz : 1% in BTC, max 0.001 BTC/mo (~$63) @KoloHub : 1% in BTC, max $100/mo, but only $1 per purchase @jup_mobile : 2%, max $100/mo, quietly cut from 4% on July 1 @xplaceapp : 0.5%, no cap listed in the docs @krak : 0-2% depending on balance held, no cap @coca_card : 1%, no cap @MetaMask : 1%, no cap The fine print behind every number is on the image. Save it for the next time a landing page promises you "up to" anything. Do you know the real cap on your card?
Nikita tweet media
English
38
6
111
9.6K
Lorden
Lorden@lorden_eth·
Most people burn through Claude tokens without even realising it They keep chatting with Claude like it's WhatsApp A few long conversations later, they're wondering why it's slow, forgetting things, and giving worse answers Three commands I wish I knew earlier: > /DOCTOR shows what's eating your tokens and helps you clean them up > /Model OpusPlan uses Opus for planning and a faster model for coding > /COMPACT summarises long chats so Claude doesn't reread the entire conversation every time you send a new prompt They take less than a minute to use People pay $500 to learn the same from courses Bookmark this
Mr. Buzzoni@polydao

x.com/i/article/2060…

English
7
0
21
2K
Nikita
Nikita@0xVishnya·
ethereum:0x228bec415ade4b61d7caf0adf8c91eac587ba369 just hit its all-time low. 18 days before my first airdrop even unlocks 👇 @useTria is a crypto card with one of the highest cashbacks on the market, plus airdrops on top. Here's how my rewards aged: - Season 1 airdrop: ~$490 when I posted this in May, ~$91 today. Still locked - Season 2 airdrop: claimed at $234, force-staked for a year. $49 now - cashback: locked for 3 months, lands end of August ethereum:0x228bec415ade4b61d7caf0adf8c91eac587ba369 is down 84% from its $0.05 top in under 4 months. By the time a Tria reward unlocks, there isn't much reward left. Would you farm rewards you can't touch?
Nikita tweet media
Nikita@0xVishnya

The Crypto Card Where Rewards Are Always Locked😂 > my first airdrop, 11,250 $TRIA locked until July 30 > in season 2 I earned 136k XP and got a 6,000 $TRIA airdrop for it > my second airdrop is locked for a YEAR (May 2027) lmao > I earned over $700 in cashback, but guess what that's locked too > I can't claim the cashback until August 31 > you don't own the card, you're just renting it for $250 a year💀 And yet… I’m still using @useTria because the exchange rate + cashback are actually great Stockholm syndrome confirmed?

English
19
2
87
10.9K
Annatar.md
Annatar.md@xieike·
THIS "GETTING STARTED WITH LOOPS" DOC TURNS CLAUDE INTO AN EMPLOYEE THAT SHIPS WHILE YOU SLEEP spoiler: prompt engineering didn't survive the 4 loop types, from you-drive to fully hands-off: > turn-based - 25% autonomy. one task per round, control returns to you. best for exploration when requirements are unclear > goal-based - 70%. you define "done" with /goal and a separate evaluator keeps sending Claude back until it clears the bar. tests passing, target scores, bugs cleared > time-based - 60%. /loop fires on a schedule: check the PR every 5 minutes, address review comments, fix failing CI > proactive - 95%. an event fires, a router dispatches child agents in parallel, a review model oversees the output. no human online at all the advice in it that matters most: loop quality depends on system quality, and tokens are the real cost knowing the 4 types is level one this article is the full 14-step roadmap from typing prompts by hand to designing loops - /goal, SKILL.md, worktrees, MCP, all of it start climbing 👇
Annatar.md tweet media
Codez@0xCodez

x.com/i/article/2064…

English
18
82
527
79.4K
Mr. Buzzoni
Mr. Buzzoni@polydao·
the skill gap between $95K and $300K right now is Claude Fable 5 agent architecture this 2-hour guide closes that gap companies are paying $150K-300K/year for engineers who know this most of them have no idea this guide exists watch it. build it. charge for it
Anatoli Kopadze@AnatoliKopadze

x.com/i/article/2068…

English
9
2
52
4.4K
Mr. Buzzoni
Mr. Buzzoni@polydao·
THE ANATOMY OF A SECOND BRAIN THAT WRITES ITSELF one Obsidian vault, five folders, and a loop that files every good answer back into a wiki you never touch the structure: > CLAUDE.md - the constitution, defines what the agent is allowed to touch > raw/ - inbox, articles, assets - frozen, read-only, you feed it and walk away > wiki/ - index.md, entities/, concepts/, synthesis.md - the agent grows this while you sleep > log.md - append-only, every move stamped with a timestamp 47 sources in -> 210 pages out point Fable 5 at just synthesis.md - the one file that actually needs frontier-level reasoning - everything else runs fine on cheaper models grab the full folder blueprint below 👇
Mr. Buzzoni tweet media
Mr. Buzzoni@polydao

x.com/i/article/2040…

English
15
53
331
28.6K
Fokki
Fokki@0x_fokki·
🚨 Warner Bros spent $68,000,000 on a live-action Tom and Jerry movie in 2021 a 20-year-old saw the trend on Douyin, China's TikTok and cleared $11,900 last month bringing it west first. turning old cartoons into real life is a Chinese trend, already blowing up on Douyin. she copied the shape and ran it to TikTok and Reels before anyone else did. > Research: check 40 channels, find the clip beating its own average: 1 day > Claude: copy the shot list frame by frame from Tom and Jerry: 20 min > CapCut: rebuild it to look real, shot by shot, AI voice, captions: 40 min > Make: post it to TikTok, Shorts, Reels, pull the views back in 48h: auto Warner Bros needs a studio, a real lion and a full crew for one scene. she needs a $50 loop and her own cat. the whole teardown is in the article above👇
Fokki@0x_fokki

x.com/i/article/2074…

English
57
71
883
2M
Nikita
Nikita@0xVishnya·
I compared 6 crypto cards on the exact same purchase in Europe. Test #5 👇 Same store, same EUR 9.69, same minute. Here's the all-in cost after cashback: - @ether_fi : $10.72 - @Plasma : $10.93 - @Backpack : $11.07 - @lava_xyz : $11.10 - @jup_mobile : $11.22 - @KASTxyz : $11.25 Same groceries, but a 5% gap between cheapest and priciest. FX spread, fees and cashback decide everything. Etherfi wins again, best rate plus onchain cashback. KAST is the priciest for the second test in a row, worst FX and a low-value fee on top. The wild part: all three newcomers ran cleaner rates than Plasma, but zero cashback. Plasma's 3% back still carried it to second. Your custom matchups are next. Which combo should I run first?
Nikita tweet mediaNikita tweet media
Nikita@0xVishnya

I compared 5 crypto cards on the exact same purchase in Europe. Test #4 👇 Same store, same EUR 7.65, same minute. Here's the all-in cost after cashback: - @ether_fi : $8.48 - @Krak : $8.58 - @Plasma : $8.63 - @wirexapp : $8.82 - @KASTxyz : $9.03 Same grocery run, but a 6.5% gap between cheapest and priciest. It comes down to FX spread, fees and cashback. Etherfi took it on the best rate plus onchain cashback. KAST charged the worst FX and a low-value fee on top. What should I test next?

English
72
22
293
86.8K
Fokki
Fokki@0x_fokki·
🚨Marvel spends $25,000,000 per episode to turn an actor in a mocap suit into the Hulk a 19-year-old did the same motion transfer last week in a browser and cleared $12,800. he filmed one clip of someone moving, then mapped it onto two samurai and dropped them in a red-moon scene. the whole thing ran inside Pixmax: > clone the action-transfer template: 1 min > feed in the motion clip, 2 character refs, and the background: 4 min > switch to ref-to-video, write one prompt: 2 min > render and scrub the shot in the canvas: 6 min 14 minutes. $48 a month. Marvel books a VFX house. he opened a browser tab. full breakdown's in the comments👇
English
30
4
68
3.7K
Mr. Buzzoni
Mr. Buzzoni@polydao·
@xieike /grill-me is the best way to stop the model from making bad codebase assumptions
English
0
0
0
76
Annatar.md
Annatar.md@xieike·
Claude Code will grill you with 40+ questions before writing a single line of code > it's called /grill-me - 3 sentences long. most impactful skill I use instead of jumping straight to code - it walks every branch of your design tree until there's zero ambiguity > every question reveals something you hadn't thought of other skills I use daily: > /write-a-prd - idea -> proper product doc > /prd-to-issues - doc -> GitHub issues automatically > /tdd - tests first, forces edge case thinking before code > /improve-codebase-architecture -> full structural review of your codebase full video from Matt Pocock 👇
Khairallah AL-Awady@eng_khairallah1

x.com/i/article/2075…

English
12
5
66
2.9K
Mr. Buzzoni
Mr. Buzzoni@polydao·
this is what a real Jane Street quant interview looks like > you have $100 and a fair coin > you flip it 100 times > what's the optimal amount to bet each flip to grow your money the fastest? everyone says "go all in every time" then everyone goes home: > going all in means one tails wipes you to zero > the real answer is a fixed fraction of your stack every single flip > get it right and you walk into a $400k first-year seat they give you 3 minutes. most people freeze at minute one the fraction has a name, and almost no one says it out loud
slash1s@slash1sol

x.com/i/article/2069…

English
8
1
47
3.1K
Mr. Buzzoni
Mr. Buzzoni@polydao·
KARPATHY'S COMPOUNDING KNOWLEDGE LOOP: CLAUDE + OBSIDIAN RAG is stateless. it forgets everything once the query ends use Karpathy's framework to build a persistent knowledge base instead: > create a vault - make a local directory for your files > open Obsidian - use the local app to browse the visual graph > connect the agent - link Claude Code directly to your workspace > write the schema - define formatting rules in a CLAUDE.md file > define catalogs - set up index.md and log.md files to track changes > ingest sources - drop raw papers in and let the agent parse them > auto-link pages - let the model map entity pages dynamically > query and lint - ask questions and scan for contradictions knowledge compiles once and compounds with every new document. 📁 Obsidian ↳ obsidian.md 📁 Claude ↳ claude.ai 📁 Karpathy LLM-Wiki Gist ↳ gist.github.com/karpathy/442a6… the $20 pro subscription window for Fable 5 closes on July 12th
Mr. Buzzoni@polydao

THE ANATOMY OF A LOOP RUNNING WHILE YOU SLEEP an anthropic engineer runs autonomous code cycles on a closed laptop using a five-file folder this loop architecture opens pull requests and runs tests without a single chat prompt: > contract.md - defines the shift rules and operational boundaries > schedule.yml - triggers the cron events to launch the next cycle > rubrics/ - stores the graders to evaluate output quality before staging > state/ - logs active checkpoints to recover from system crashes the window to run Fable 5 reasoning loops under the $20 subscription closes on July 12th build your autonomous pipelines now before it shifts to expensive API billing grab the full folder blueprint below 👇

English
19
3
87
4.3K
Mr. Buzzoni
Mr. Buzzoni@polydao·
KARPATHY JUST EXPLAINED THE FLAW HOLDING BACK EVERY FRONTIER MODEL reinforcement learning looks like the road to real reasoning. he makes the case it is far shakier than the hype this video answers: > why hundreds of parallel attempts still give a weak learning signal > how one right answer wrongly rewards every wrong step on the way there > what humans do after solving a problem that models still cannot > where this ceiling actually shows up in the models you use today watch it 👇
Mr. Buzzoni@polydao

x.com/i/article/2065…

English
9
1
57
2.8K

Keşfet

@xieike @0x_fokki @0xVishnya @useTria @Plasma @wirexapp @KASTxyz @BleapApp