Carat

@andy_neon_ most people optimize for speed, smarter play is optimizing for cost efficiency

A CAMERA WITH NO DRONE FLIES THROUGH A CANYON AT SUPERSONIC SPEED No aircraft in frame. No cockpit. No visible rig - just pure first-person motion, inches above the rocks. 1 prompt. 9 minutes of render time. $4 in credits. A shot that would've needed a helicopter crew, 3 weeks of permits, and a 6-figure production budget, built for less than a coffee run The path: through narrow rock arches, skimming waterfalls, into a dark canyon tunnel, then bursting back into daylight as shockwaves ripple across the cliffs. Made in Dreamina AI No helicopter. No permit. No stunt pilot Do you still unimpressed?

He makes $38,000 a month building Obsidian vaults for ex-Roam users He's 30. Spent two years turning Obsidian into what he calls a real second brain - visual reading modes that turn linear notes into navigable galaxies, automatic connection detection, timeline views showing how his thinking on a subject evolved. He open-sourced the plugins on GitHub Last October a former Roam user asked him to install the setup on her machine. He charged $2,200. Two more people asked the same week Four months later: 22 clients paying $2,200 for setup and $1,500 monthly for ongoing tuning and reviews. Around $38,000 recurring. Every client came from word of mouth in knowledge-worker circles frustrated with subscription second-brain tools Roam Research raised at a $200 million valuation in 2020 on the premise that bidirectional linking required proprietary infrastructure. Users paid $15 a month for what Roam framed as the only tool that could support real second-brain workflows He's not competing with Roam. He's proving Roam's premise was correct - then building it on free open-source anyone can own He didn't build a second brain. He built the first productized version of a tool people had been trying to configure themselves for four years Every second-brain business priced on "our proprietary graph is the moat" now competes with an operator who assembles the same graph on top of software the user already owns

@Pakero8x Exactly. The bottleneck is moving from coding speed to managing complexity. The winners will be those who build better systems around AI agents. 🚀

@Caarat1 That's wild. The idea that enough personal data can predict things like job changes or relationship shifts is kinda wild and a little unsettling. But then scaling it to public events... wow. What kind of data was most predictive for the public events?

He made $47,000 in three months betting on public events using a system he originally built to predict his own behavior He's 29. Last August he started an Obsidian vault to log everything he did. What he googled. What he bought. When he woke up. Which podcasts made him change his mind. Which meetings he cancelled and why. Six months of entries, roughly 40 per day By February the vault could predict what he would do tomorrow with 78 percent accuracy. It knew he was about to change jobs three weeks before he applied. It flagged that he was going to break up with his girlfriend before he had the thought consciously. The pattern-detection was not magic It was just having enough data on one person to see the shape of their next decision Last March he pointed the same architecture at public data. Fed communications, corporate earnings calls, satellite imagery of shipping ports, cross-language political sentiment. Same retrieval logic. Different subject Six trades in three months. Two on rate decisions. One on a currency intervention. One on an earnings beat. Two on election outcomes. $47,000 net across event contract markets that priced binary outcomes on public events WHOOP runs a $3.6 billion self-quantification business selling wristbands and a subscription that tells 3 million users when to sleep, when to train, and when their body is under recovery stress. The premise is that structured data about one person reveals patterns that person cannot see themselves He's not competing with WHOOP. He's proving WHOOP was right about the mechanism, then using the same mechanism to predict something more expensive than sleep He didn't build a self-tracker. He built the first vault that noticed the same pattern-recognition that read him could read the world Every self-quantification business priced on "we know your body" now competes with an operator who built the same system on his own notes and turned it outward

@LearnStochastic AI coding’s next bottleneck is coordination, not intelligence. As agents improve, the advantage goes to people who can build systems around them-not just write better prompts. 🚀

Those who wanna create a retarded list, go through the engagement of this post, you will find out. Anyone who has worked with Claude, Codex lately, and has tried to push it to limits, share a singular bottleneck: Themselves. Yes, claude code or codex works great on coding problems, but organisation level engineering is a different rocketship altogether. There's ever changing design requirements, UX evolutions, User feedback, Scaling issues, and the list goes on. No matter how efficient your prompting is, your throughput gets brain-bound. You can't prompt and stay on terminal 24x7 to manage all those agents. But what if, those agents don't need management at all? What if they had access to everything they need, and they never fall into the rabbit-hole of context-rot? A fleet that just integrates in your engineering or SDLC as a team. If you have never been bottlenecked by what I mentioned above or similar, either you have figured out a solution, or you are living under the rocks. A sub-minute glimpse on @myprasanna's life would provide enough signal to the fact that he lives and breaths in code, since decades. And people are still fixed on "#1 coder in India"!!!! Amused how ignorant and dense people get in no time.

@Virexontic The big shift is that enterprise AI is less about “better memory” and more about access control + company context.

They won't tell you this, but the AI most companies use internally is basically a glorified notes app pretending to understand the org chart. A builder put together a fully custom enterprise AI brain in Claude Code and put it next to a personal AI brain in Obsidian. The difference isn't the interface — it's permissions. In the custom build, every employee, their manager, their projects, their credentials, and the strategy behind those projects are mapped into the system itself, so the chat only ever knows what that person is cleared to know. Sales can't touch HR data. A new hire's assistant already knows their team and their active projects on day one. Obsidian has no concept of any of that — it just remembers what you personally wrote. Building this kind of access-aware AI brain for a company runs about $15,000–$25,000 in setup work. Licensing it out to a handful of mid-size clients puts a single builder at $60,000–$100,000 a year, without touching a single line of client data themselves. Personal AI brains are for people working alone. The moment there's a team, you need infrastructure — not a vault.

@AlxTurovski The interesting part is that AI tools are becoming workflow choices, not just model choices. A subscription is only worth it if it consistently removes friction-limits, reliability, and how well it fits your building process matter more than the brand name. 🚀

Yesterday I cancelled my first ever AI coding sub. Even though I was grandfathered from Codeium to Windsurf to Devin with $10/m, it's still not worth it. Limits are unacceptable and the product it turned to is not for me. So I got Codex instead. 😁 And now I'm a proud owner of two major subs: > Claude > Codex Let's see if I can build proper workflows with them.

@0xRafy AI development is shifting from “which model is smartest?” to “how do you build the right system around it?” The biggest gains often come from agent architecture: context management, tool use, feedback loops, and knowing when a model actually improves the workflow. 🤖🚀

Anthropic released a full session on how to build with Opus 4.6: "I couldn't sleep. this is the moment that's going to change everything" • 00:00 - sub-agents, context compaction, adaptive thinking - Opus 4.6 • 5:39 - what's improved since Opus 4.5 • 11:55 - how Opus 4.6 catches data - demo • 35:27 - Opus 4.6 builds a full financial model from scratch • 43:26 - how to know if a new model actually makes your product better this 1-hour watch will save you $500 on paid courses about building AI products with Opus 4.6 Watch it today, then read the article below on how to build efficient agentic loops from scratch.

@pixclipper Simple ideas with strong emotional hooks can still win. The chalkboard style works because it turns a widget from just information into a tiny daily moment. Personal, visible, and repeatable experiences are often what make consumer apps stick. 💛📱

chalkee is a new couples widget app using that chalkboard handwritten message format i hadnt seen this style popping for a few months now and it works perfectly for these little daily love notes or inside jokes that show up right on the home screen the format feels fresh again because its so warm and personal they are on track to 10k downloads per month

A strong investing lesson: sometimes the best move is not making a move. Having the discipline to wait for high-conviction opportunities can be a bigger advantage than constantly trying to stay invested. Patience is a real edge-especially when everyone else feels pressure to act. 📈

Blackstone founder Steve Schwarzman: "We only shoot when we have an open shot. We don't have a shot clock forcing us to shoot." at london business school in 2019, the investor who cut blackstone's buying 75% right before the 2008 crash explained why sitting still is the edge. no shot clock, so no pressure to buy 20 targets narrowed to maybe 4 worth closing cutting the investment rate 75% in 2006 because it felt about to go wrong $13 trillion of bonds at a negative yield as the warning light "the biggest risk is political, not business" - that's the setup. 00:23 - the cracks he sees in the global economy 02:41 - how he saw 2008 coming and pulled back 03:32 - why not being fully invested is the edge 04:34 - the graph he shows his students 05:06 - 8,000 private-owned companies vs 3,400 public 08:32 - the one risk that actually keeps him up Bookmark this.

@i_am_chayyy Day 3 and already shipping real progress 🔥 Building the database + validating the core user experience early is the right move. A global cafe/coworking map with check-ins and community data could become really useful for remote workers. ☕️💻

Day 3 update: build in public of Remo App 🔥 2960 cafe/coworking spots added across the globe ☕️ 💻 Checkin feature and test verify wifi ✅ Checkin experience with points animation. 💯 Signup for the waitlist: forms.gle/uPywAnUhmUn7Bx… So I can notify you when the app goes live! 🚀

@kiitanEth The strongest airdrops usually come from products that create real value first. A token can reward early users, but sustainable growth comes from demand, usage, and a community that believes in the product-not just farming incentives. 🚀

Most airdrops leave users feeling used and burned. Oz from Moonberg reveals the smarter way. Announce the airdrop early. Build real product demand first, let users trade and benefit as the token grows. Hyperliquid proved it works, six and seven figure wins for early believers. Real utility beats hype every time. This is how you turn market noise into signal.

@imfuryfist LLM inference is becoming one of the most important skills in AI engineering. Knowing how to serve models, optimize latency, manage costs, and scale workloads is what turns an AI demo into a production system. 🤖⚙️ Great resource for anyone getting deeper into AI infrastructure.

I was reading interview experiences on Twitter and noticed a lot of questions around: * How to host a local LLM * How to scale it for 10x or 100x more users * How to reduce latency and increase throughput * And similar production topics I wanted to learn these too, and found out they all fall under one topic: LLM Inference Found this free course while learning. Thought it might help you guys too:

The best debugging moments are often the painful ones 😅 A small mistake can hide a big lesson. Finding the real issue means the system gets better-and you understand your own product more deeply. Those uncomfortable moments are usually where the biggest improvements come from. 🚀

Day 24 building in public: Got a bad day yesterday, but a happy finding. Being uncomfortable indeed helps you improve. In my case, after some debugging, I found the problem was that I was redirecting the traffic accessing my analytics endpoint, so I was just tracking the signed up users😐😅 What is the happiest finding you got from being uncomfortable?

The interesting shift is that local AI isn’t replacing creators-it’s becoming a creative multiplier.For game studios, the value isn’t just saving API costs. It’s having a private system that can constantly generate ideas, test variations, and maintain world consistency while humans focus on the final creative decisions. 🎮🤖

A BERLIN INDIE GAME STUDIO BUILDS AN AI NPC DIALOGUE LAB FROM THREE USED 3080s — TOTAL SPEND €2,700, REPLACING ENDLESS API TESTING the studio is four people, one room, terrible chairs, the gpu rig sits near the window because it heats the office better than the radiator setup: — 3× used RTX 3080 cards — local roleplay model with lore documents — vector memory for quests, factions, and item names — nightly batch generation for NPC variations what it replaced: — API bills during prototyping — repeated prompt tuning in cloud tools — writers manually checking every goblin for lore errors the model generates first-pass dialogue, tavern rumors, quest alternatives, and failure-state lines. the writers still rewrite everything. but now they rewrite from material instead of blank pages. cloud AI gave them tokens. the local rig gives them a weird little dungeon master that never logs off.

Some projects stop feeling like apps and start feeling like ecosystems.With AI agents and simulations, one person can now create systems that behave more like living worlds-constantly generating data, improving, and evolving. The interesting part is not just the code, but the vision behind what they choose to build. 🚀

I've been shedding random AI subscriptions over the last year and just going all in on GPT Pro and CC Max. ...this is the fastest i've clicked buy like I've just been given the thing from the future that's always been in my head but never materialized Congrats @FarzaTV! This is one of the most impressively thought-through onboarding and UX's I've seen in this space, ever! 👏 🫡

Some projects stop feeling like apps and start feeling like ecosystems. With AI agents and simulations, one person can now create systems that behave more like living worlds-constantly generating data, improving, and evolving. The interesting part is not just the code, but the vision behind what they choose to build. 🚀

I genuinely don’t understand how some people live like this. You look at the screen time and think: “14+ hours coding?” Then you realize he isn’t just writing code. He’s building an entire world. AI agents collecting metrics. Different hero archetypes. Training pipelines. Arenas where bots fight for treasure. Endless simulations running while he’s improving the system. At some point it stops looking like software and starts looking like its own universe. Offline would honestly feel boring compared to what’s happening inside his laptop. Some people build startups. Some build things that deserve a Netflix documentary.

The biggest shift is that building software is becoming more about clarity of ideas than writing every line manually AI lowers the barrier to prototyping, but the real advantage still comes from knowing what problem to solve, what users need, and how to turn an idea into a useful product. 🚀

A workout tracker app got built from 1 prompt in 41 seconds. No template. No file. Just a sentence, typed once, hit send. The screen fills itself. Exercise cards appear. Bench Press, 3 sets. Squats, 4 sets. Reps already tracked, layout already clean. 0 lines of code. 0 design skills. 1 phone, 1 prompt, 1 person. This is not a demo built for engineers. ANYONE with a phone can do this. You describe the app, the AI writes the screens, the logic, the state. 41 seconds start to finish. That's the entire build. The only skill required is knowing what you want.

@SimranKaur_b @Palworld_EN That’s how the best games get you 😂“Just one more session” turns into hours because the progression loop is too satisfying. Good luck surviving the 1.0 grind 🔥🎮

Biggestt mistake I’ve everr made…🥲 Opening @Palworld_EN because a friend recommended it. Thought I’d play for a bit. Ended up in an all-nighter😭, taking down my first boss in an epic fight, and loving every second❤️. Now I’m dead tired but this game is dangerously calm & satisfying 🔥 Rate my fighting skills 1-10 😂 @Palworld_EN Huge props to the team..!!👏 1.0 drops tomorrow… pray for me😖🙏.. Who else got caught by Palworld?

@xioamijii @CapitalMintMkt Prop trading platforms are getting more attention, but always check the rules, payout conditions, drawdown limits, and company reputation before buying a challenge. A good offer matters-but risk management matters more. 📊

I was scrolling through my X feed and got to know about @CapitalMintMkt . Decided to research more about the platform and really impressed with the Offers and benefits it is providing. Those who are really into Trading, must try this prop firm to start your Trading Journey. > There are no hidden rules and no complexity you have to face like other platforms. > Flexible challenges for every type of trader. > profit withdrawal in 24hrs and many more awesome features. You can also try and buy your first funded account on capitalmintmarkets.com