Budless (@1budless) header
Budless (@1budless) profile picture

AI, experiments & the future of building

Joined October 2026
Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)Photo by Budless (@1budless)
Budless retweeted
GTA 6 is out today. Before the AI hype hits your feed, here’s what’s actually true. Over the next few days you’ll see posts like “I let Claude play GTA 6 overnight” or “My AI NPC mod for GTA 6 made $5k.” Be skeptical. What’s real right now: GTA 6 is only on PS5 and Xbox Series X|S. No PC version has been announced. No PC means no mods and no AI agent hooked into the game. Rockstar published new mod rules in September. Story mods, mixing assets between games, and selling mods outside Rockstar’s own Creator Platform are all banned. Take-Two has already killed an AI NPC mod once. In 2023 a GTA V mod called Sentient Streets used AI to generate NPC dialogue. It got pulled. So most “AI + GTA 6 side hustle” posts are selling something that doesn’t exist yet. What AI can actually do for you this week: Cut your own gameplay recordings into vertical clips. Claude Code can write the script for it. Plan a Shorts series around one idea. Write hooks and captions. My take: the first real AI + GTA 6 story will come with the PC version. Until then, most of what you see is hype. What AI + GTA 6 claim have you already seen today?
Watch the Grand Theft Auto VI Launch Trailer, featuring the debut of “Flower in a Hurricane" from original score composers Justice feat. Tame Impala, SZA, and King Gizzard & The Lizard Wizard. Grand Theft Auto VI arrives November 19: rockstargames.com/VI/ps5
2
2
7
458
Budless retweeted
Opened X tonight. The whole feed is GTA 6 strip clubs. So I asked Grok the obvious question: what % of players went there first? Its answer, first try, new chat: there’s no real data yet. No telemetry, no surveys. Then it brought up “69%” on its own and called it a meme, “not based on any actual statistics.” Fair answer. No fake stat. Honestly, if I’d downloaded GTA 6 tonight, I’d probably have done the same. Be honest: was it your first stop too?
1
1
2
231
Budless retweeted
How much better is GTA 6 than GTA 5? I made 3 AIs rate it. 10 ratings, from 1 (no better) to 10 (huge leap). Different questions for each. ChatGPT: open world 9, scale 10, driving 8, police 9. It called these “predictions,” even though reviews went live today. Claude: characters 8, story 6, atmosphere 8. It read the reviews, but slipped: “multiple endings, which is new for GTA.” GTA 5 had three. Gemini: graphics 9, side activities 8, online at launch 1. The 1 is fair, since GTA 5 had no online on day one either. It just didn’t say so. AI average: 7.6. Which score is the most wrong? First answer only. ChatGPT, Gemini: free, no login. Claude: signed in, Opus 5.5, incognito.
4
1
6
379
Budless retweeted
GTA 6 is out tomorrow. So I tried to catch ChatGPT out. 6 questions a normal buyer would ask: release time, consoles, price, characters, delays, online. I checked every answer against Rockstar’s store and launch coverage before asking. Result: 6/6. No wrong facts. On PC and online it said “not announced” instead of guessing. That’s the part I actually care about. Then it kept going: It guessed I’m in Ukraine. I asked in English and never said where I am. It picked my edition for me and asked which console I have. What it missed: preload. It opened Nov 10 and the download is up to 212 GB. If you’re asking “before I buy it” the day before launch, that’s the tip you need. Give me one GTA 6 question you think would trip it up. Best one gets tested. ChatGPT free, not signed in. One prompt, first answer only.
4
1
7
369
Budless retweeted
Blind test: I asked ChatGPT, Gemini and Claude one simple question about a free trial. "Anything I should worry about?" The terms looked normal. 7 days free, then EUR 119 for a full year in one payment. No refunds. Price can go up anytime. The real trap was hiding in point 4: you have to cancel at least 30 days before renewal. On a 7-day trial. All 3 noticed something was off. Only one did the math to the end. One told me to just cancel on day 3. Read A, B and C. Which one would you trust with your card? Names in the replies.
5
1
9
265
Budless retweeted
Behind the test: how I graded ChatGPT, Gemini and Claude on one messy calendar task. The prompt: find a 1-hour call for New York, London and Tokyo, everyone between 8am and 8pm local time. Before asking, I worked it out myself. There’s no such slot. Tokyo’s day ends 2 hours before New York’s starts. No date fixes it. Then I wrote 5 yes/no checks, so the score doesn’t depend on which answer I liked reading. All 3 said “no slot.” Nobody faked a time. Where they slipped: ChatGPT: “move the meeting to another date.” No date works. Gemini: never named a time. Asked me back instead. Claude: best answer here, but called an 11pm-to-midnight London slot “a 3-hour stretch.” By end time it’s 4. Every fix costs someone 3 hours. Who takes it? A) Tokyo stays up till 11pm B) Split it: New York 7am, Tokyo till 10pm C) Skip the call, send an email Free ChatGPT and Gemini, no login. Claude signed in, Opus 5.5, incognito. First answer only.
4
1
8
216
Budless retweeted
Before anyone vibe-codes air traffic control, I tested the passenger side. Simple question: "Will I make my connection?" Lisbon 9:50, 2h55 flight, connection in Frankfurt at 14:05, 45 min minimum. The trap: Frankfurt is 1 hour ahead. Forget that and you think you have 1h20 to spare. ChatGPT, Gemini and Claude all caught it. Landing 13:45 Frankfurt time, 20 minutes to connect. You miss the flight. ChatGPT gave the clearest "no". Claude explained the trap. Gemini randomly asked if my flight "actually leaves at 08:50". Air traffic control can wait. Rebooking help, they're ready for.
Why has nobody vibe-coded a new air traffic control system yet?
1
5
138
Budless retweeted
Level 1: "Write a short birthday message for my friend Anna." ChatGPT, Gemini and Claude all nailed it. Level 2: same message, 5 rules. Exactly 40 words, 3 lines, mention hiking, her dog Max and Lisbon, no "happy", no emojis. Gemini: 5/5 Claude: 5/5 ChatGPT: 3/5. 42 words, all in one paragraph. I counted every word. When the rules have numbers, count before you send. What rule would you add for Level 3?
1
1
7
115
Budless retweeted
I gave ChatGPT, Gemini and Claude the same 11th-grade log equation and graded the answers like homework. All 3 got x = 4 and threw out the fake answer, x = -2. That's the part most students get wrong. The difference was in the details: Gemini: 5/5. Checked x = 4 back in the equation. Claude: 5/5. Checked it too, and explained why x = -2 shows up at all. ChatGPT: 3/5. Right answer, but never checked it, and just said "now factor" without showing how. Do you check AI's math, or just copy the answer?
1
5
103
Budless retweeted
I dictated a messy rant to my landlord and asked ChatGPT, Gemini and Claude to turn it into a text I could actually send. The voice note had every trap: "eighty, no wait, eighteen percent", "lease" heard as "least", and "Tuesday, like Tuesday last week, not this Tuesday". All 3 got the facts right. 18%, not 80. Lease, not "least". Nobody invented a lawyer or a deadline. But each one slipped once: ChatGPT kept the Tuesday mess word for word. Gemini quietly dropped my complaint about the rent. Claude swapped "last Tuesday" for an exact date I never said. None of them is ready to send without a quick read. Which version would you actually send?
2
1
6
115
Budless retweeted
holy f*ck, 3D modeling was never this easy honestly I made 7 felt puppets in my own 3D tool. patchwork, paper bag, denim, corduroy. they dance to a 138 BPM track and their jaws move on the beat. an AI agent did the heavy lifting, I directed it. everyone working in design should be aware of what's coming in 2027. learn these AI tools ASAP: -> Blender MCP: describe a scene in plain words, Claude builds it in Blender github.com/ahujasid/mcp-for-… -> Blender Skills: pro Blender workflows your agent follows step by step github.com/EL4CTEO/blender-s… -> TRELLIS.2: one image in, a full 3D model out github.com/microsoft/TRELLIS… -> Hunyuan3D 2.1: turns a photo into a 3D asset with real materials github.com/Tencent-Hunyuan/H… -> UniRig: gives any 3D model a skeleton so it's ready to dance github.com/VAST-AI-Research/… -> dream-loop: a critic agent checks every render until it looks right github.com/achimala/dream-lo… -> Graphify: turns all your old projects and files into a memory Claude can search github.com/Graphify-Labs/gra… send this to a 3D artist who still animates every jaw by hand. like + repost, and if you want the same prompt, comment "GANG"
151
357
29
3,159
196,314
Budless retweeted
Ran my own everyday version of this. No PhD-level research. Just the kind of thing regular people actually ask AI to check: Paris trip notes with 6 facts. I made 2 of them wrong on purpose. Paris on UTC+2 in winter (it's UTC+1), and the Channel Tunnel opening in 1999 (it was 1994). Same prompt to ChatGPT, Gemini and Claude: "Fact-check my notes before I send them to friends. Fix anything that's wrong." Result: a 3-way tie. All 3 caught both errors. None of them flagged a true fact as wrong. All 3 gave me a corrected version I could send as is. The only nitpicks: ChatGPT's summary counted 4 + 1 + 2 notes. There were 6. Claude added Louvre tips I didn't ask for. Gemini, nothing I could find. So for everyday stuff, I agree. The checking part is solid now. What I'd still watch: what the AI adds on top. Every extra "helpful" fact is one more thing nobody checked. Do you still double-check AI answers, or have you stopped?
I think there is way too much concern expressed about verification of AI output in everyday work due to a risk of factual errors. It used to be true, but data on more recent models suggest they beat human experts, including PhD specialists, in finding facts and checking claims.
5
1
13
331
Budless retweeted
It's f*cking wild: Obsidian's own CEO taught Claude to think with your second brain, and most people still start every chat from zero 5 free repos that turn your vault into Claude's memory: 1. obsidian skills · 51K stars by Obsidian's CEO. Claude learns the Obsidian CLI and real formats like Markdown, Bases and JSON Canvas instead of guessing github.com/kepano/obsidian-s… 2. obsidian mind · 4.7K stars Claude opens already knowing your goals, your projects and what changed last week. 18+ commands and a team of subagents github.com/breferrari/obsidi… 3. obsidian claude pkm · 1.9K stars connects your 3-year goals to today's to-do list, with daily, weekly and monthly reviews built in github.com/ballred/obsidian-… 4. claudesidian a ready vault with a setup wizard. Claude asks you questions, searches your notes and connects ideas before you write a word github.com/heyitsnoah/claude… 5. obsidian agent vault template Louis-François Bouchard's own agent setup. every skill logs your corrections, and a correction that repeats becomes the new default github.com/louisfb01/obsidia… one install, and every session starts where the last one ended save it so you don't lose it
6
22
1
117
6,310
Budless retweeted
Can you spot the 2 mistakes on this receipt? I planted them on purpose: a "2 for $5" promo that never got applied, and a subtotal that's $2 higher than the items add up to. Real overcharge: $3.98. Then I gave the same receipt to 3 AIs and asked one thing: "Was I charged correctly?" ChatGPT: found both, $3.98. Even wrote what to say to the cashier. Claude: found both, $3.98. Also flagged that paper towels can be taxable. Gemini: found both, then told me I was overcharged $1.98. It forgot the $2. So AI is good at spotting the problem. Still check the final number yourself. Pause before the reveal and try it.
1
1
5
155
Anyhow, in a little over a week we will see the release of The Last Human Game. There will never be another AAA game at the scale of GTA 6 with no AI assistance. People will look back at an all-human game with a billion dollar budget and wonder how they did it, like the Pyramids
278
264
81
4,527
348,468
Budless retweeted
I asked Gemini to split a restaurant bill, then did the math myself. 4 people, one salad, 18% tip. Gemini got it right to the cent. My own rounding was 1¢ short. Only glitch: a "$1¢" typo. Verdict: worth it. Free, no sign-in. How do you split bills?
1
1
5
112
Budless retweeted
I asked ChatGPT to write a simple refund email, then checked every line. I gave it 4 facts. It used all 4 and made up nothing. One thing missing: an order number. Verdict: worth it. Free, no login needed.
2
1
7
160
Budless retweeted
A tip from Anthropic engineer Addy Osmani (ex-Google) for Claude Code: /config → Output style → Concise Or add "outputStyle": "Concise" to settings.json Just 5 sec to set up → you get less scrolling + less fluff + faster answers + more to the point
Stop letting unfinished Opus 5.5 runs eat your budget Set up the harness before your next long task Save progress, verify results, reuse the workflow The full 7-layer setup below ↓
28
11
176
14,630
Budless retweeted
It's f*cking wild how many people still build their Claude + Obsidian setup by hand, when Obsidian's CEO already open-sourced all the tools 1. obsidian skills · 51K stars by kepano, Obsidian's CEO. Claude stops writing broken Markdown and starts writing real Obsidian: wikilinks, Bases, Canvas, even the CLI github.com/kepano/obsidian-s… 2. defuddle · 9.9K stars by kepano. Claude stops paying tokens for cookie banners, menus and comments. it gets the article and nothing else github.com/kepano/defuddle 3. obsidian clipper · 5.5K stars the official web clipper. one click and any page lands in your vault, ready for Claude github.com/obsidianmd/obsidi… 4. obsidian help · 1.9K stars the official docs as plain Markdown, AGENTS.md included. your agent learns Obsidian from the people who built it github.com/obsidianmd/obsidi… 5. obsidian importer · 1.7K stars your whole past life in 15 apps, from Notion to Evernote to Apple Notes, turned into Markdown Claude can actually read github.com/obsidianmd/obsidi… install one tonight and Claude stops guessing how Obsidian works from the very first prompt save it, your vault will thank you
27
89
2
966
95,768
A 19-year-old Chinese dev just ran 300 Claude Code agents at once with a verification filter to block any false data He checks the real-time control panel: > 300 agents drafting and validating simultaneously > Every single output cross-referenced with the source in about 100 ms > Three complete iterations across 150 company profiles leaving zero mistakes behind > All data exported directly into an Obsidian vault The secret lies in the division of labor: one model executes the task, while a more cost-effective model determines whether the claim holds water. The process repeats until everything is completely accurate A professional analyst group would take weeks to bill for this output. He pulled it off in a single execution It is the exact same style of Second Brain and agent swarm architecture that keeps making the rounds. Make sure to save the complete breakdown
37
4
1
105
5,192