Sancho

@EXM7777 Loops right now it's like bitcoin in 2009

Fable + loops + goals is lowkey TERRIFYING

NVIDIA JUST PUT A TRILLION-PARAMETER AI SUPERCOMPUTER UNDER YOUR DESK. 768GB of memory. Let that sink in. This is the DGX Station. Runs everything locally. On Windows. No data center required. Here's what that memory actually means. A model like Llama 405B needs around 810GB just to load at full precision. Most workstations can't even get close. This machine runs frontier AI models of up to 1 trillion parameters locally — no quantization tricks needed to squeeze it in. On a regular consumer GPU, you're stuck running these models shrunk down to 4-bit or 8-bit just to fit in memory, which quietly costs you accuracy on anything complex. Here, you run the full, uncompressed version. The craziest part? This isn't just one model. It can run hundreds of parallel AI agents at the same time, on the same machine. An entire company's worth of AI workloads, sitting in a single office. While most teams are still paying per token to rent this kind of power from the cloud... The ones who saw this coming just moved the trillion-parameter model onto their own desk.

@HodlReaper Keep going mate 💪

thx twitter for first monetization keep grinding

@Adea0x Exactly a thousand phones posting identically is just spam. The moat is in the timing jitter, the caption variance, the different scroll/tap patterns per device. That's the actual engineering problem, not the hardware.

@Sancho_Wizard the interesting part isn’t the number of phones. it’s making them behave differently

A CHINESE AGENCY IS RUNNING HUNDREDS OF DISASSEMBLED PHONES. AND AI IS THE ONE CONTROLLING THEM. No screens. No cases. Just bare circuit boards, stripped down and racked like servers. Row after row of phones, gutted down to their motherboards, wired into a single tower. Each one logged into a different account. Each one posting, liking, commenting, promoting — on a schedule no human could keep up with. The old version of this was just phone farms. Manual. Slow. One person tapping through dozens of devices by hand. This version is different. AI sits on top of the whole rack, deciding what each phone posts, when it posts, and how it behaves so it never looks automated. Different captions. Different timing. Different engagement patterns per device. To the platform, it doesn't look like one operator. It looks like hundreds of real people. The craziest part? One agency runs an entire influencer marketing operation this way — pushing a single product or service across hundreds of "accounts" that never sleep, never get tired, and never miss a posting window. While most agencies are still hiring social media managers... This one built a wall of stripped phones and let AI run the whole department.

@cyrilXBT Good breakdown. The memory-file structure alone is worth adopting regardless of which model you're running.

"Just be specific" is not a prompting strategy. The Fable 5 team gave way more nuanced advice than that. Nobody's talking about the part that actually changes results.

HE BUILT AN AI EMPLOYEE THAT FINDS HOUSES WITHOUT POOLS. THEN SELLS THE OWNERS ONE. No cold calls. No door knocking. No manual lead lists. Just three things running together, day and night. OpenClaw scanning satellite imagery across entire neighborhoods, flagging every property with a lawn big enough for a pool — and no pool in it. Fable 5 doing the reasoning behind it — sorting good leads from bad ones, understanding lot size, layout, what's actually buildable. And a mini PC on his desk running the whole thing locally, nonstop. The system finds the house. Then renders a photorealistic image of that exact backyard — with a pool already in it. That render becomes the postcard. The postcard becomes the pitch. The craziest part? He never talks to a homeowner until they're already sold on what their backyard could look like. The AI found the lead, built the visual, and made the offer feel obvious before a human ever got involved. No cloud bill scaling with every scan. No subscription eating his margin. He paid once for the hardware. Every lead after that is free. While most pool companies are still buying lead lists and cold calling... This guy built a machine that finds the house, sells the dream, and closes the deal on autopilot.

@zodchiii Best line in here isn't about Fable 5 at all: don't spend your strongest model on questions a cheaper one handles fine. That's just good tooling discipline

@EXM7777 The "run it as a manager, not a coder" mindset is underrated. Curious how you handle the failure cases — do your subagents self-correct, or do you gate every loop with a manual checkpoint?

@Dexonfxf ty bro🫶

@Sancho_Wizard Great post!

NVIDIA JUST SHRUNK AN AI SUPERCOMPUTER DOWN TO THE SIZE OF A DECK OF CARDS. No data center. No cloud bill. No rack of GPUs humming in a server room. Just a small board you can hold in one hand. The Jetson Orin Nano Super. It runs real AI models. Locally. On-device. No internet required. The same class of workloads that used to need a $10,000 GPU cluster now fits on something the size of a Raspberry Pi. The craziest part? It costs less than most people's monthly cloud API bill. One-time purchase. Zero subscriptions. Zero data leaving the device. While big labs race to build bigger data centers... NVIDIA just quietly proved the opposite bet — the future of AI isn't a warehouse of servers. It's something that fits in your backpack. Robotics. Drones. Home labs. Edge devices that think for themselves. The compute war isn't just about who has the biggest model anymore. It's about who can run intelligence anywhere, for almost nothing.

@DimaHolovatyi Yep. Nvidia CEO won't give bad advice

@Sancho_Wizard All in one, that’s crazy

A CHINESE STUDENT TURNED $0.90 INTO $408,292. AND SOMEONE REVERSE-ENGINEERED HIS ENTIRE BOT WITH FABLE 5. No leaked code. No insider access. Just screen recordings, public trades, and a model patient enough to reconstruct the logic. 53 trades. 100% win rate. A live BTC prediction market bot running in real time. Most people saw the numbers and moved on. One person saw a pattern worth decoding. They fed Fable 5 every visible data point — entry timing, position sizing, the exact moments it caught reversals before they happened. Fable 5 didn't just read the trades. It rebuilt the strategy behind them. The craziest part? What took the original student months of trial and error took the second person one afternoon with the right model asking the right questions. $0.90 into $408,292 wasn't the impressive part. The impressive part is that the edge didn't stay private for long. The next bot doesn't have to be invented from scratch anymore. It just has to be reverse-engineered by something smart enough to see the pattern first.

@Dexonfxf Indian tech enthusiasts and AI content are a match made in heaven. Nobody ships tutorials, threads, and hot takes faster than this crowd. 😅

THIS DEVICE ISN’T MADE FOR GAMING. IT’S MADE TO PRINT MONEY. A freelancer was spending over $800/month renting cloud GPUs. Every AI project cut into his profit. Then hardware got small enough to fit in his pocket. Now he can build, test and run AI agents almost anywhere. The winners in AI won’t be the people with the biggest offices. They’ll be the people carrying the smartest hardware. AI is becoming portable. And that’s a much bigger deal than most people realize.

@aiseomastery Exactly — this is just market dynamics playing out. Whoever offers the best performance for the lowest friction wins

@Sancho_Wizard Not betting everything on one model is solid advice regardless of how this rollout actually plays out.

FABLE 5 IS BACK. AND IT'S ABOUT TO DISAPPOINT EVERYONE WHO WAITED. Two weeks of hype. Two weeks of "the best model is coming home." Here's what you're actually getting. A new safety filter tighter than anything they've shipped before. Normal coding and debugging now gets flagged as "risky." And when it trips? You don't even get Fable. Your request gets quietly rerouted to Opus 4.8 without telling you. It's only free through July 7th. Capped at half your weekly usage. Then you pay per message. So the model everyone waited on is now more restricted, faster to refuse, and paywalled in under a week. Here's the good news. You never needed it in the first place. Betting everything on one "godlike model" was always the wrong play. It gets gated. Throttled. Pulled the moment it's actually good. The real move? A council of models working together. No single point of failure. No company can throttle your whole workflow overnight. While everyone's refreshing their app waiting for Fable to unlock... The smart builders already have five models running instead of one. Nobody can gate a system that doesn't depend on a single model.

@Dexonfxf That's exactly what I was talking about

@Sancho_Wizard The real moat was never the model. It was knowing how to use it

ANTHROPIC RELEASED SONNET 5. THEN REVEALED SOMETHING WILDER 99% of their engineers run swarms of 300+ agents. Not assistants. Self-improving agents. Hundreds per engineer. And Sonnet 5 now matches Opus 4.8 — for a fraction of the cost. Here's the part nobody's talking about. Code can check itself. Tests pass or fail. Green checkmark or red. The model verifies its own output for free. Now ask yourself: Did the reconciliation actually balance? Would this decision survive a regulator reading it back to you? No green checkmark for either. In code, the verifier is free. Everywhere else — the verifier IS the build. So while everyone races to copy the agent swarm... The real winners already built the grader. Your evals. Your benchmarks. That's the entire moat. Everyone else is just running agents with no way to know if they're right.

@eng_khairallah1 Curious how this holds up over a longer session first response feeling different is easy, staying consistent after 50 turns is the actual test.

this is f*cking gold Andrej Karpathy joined Anthropic five weeks ago. A friend on his team just showed me the exact LOOPS.md file he actually uses. I dropped it into my setup. The very first response was different. Not slightly different. Completely different. Claude stopped giving generic answers and started working exactly the way I think. You don't talk to the model anymore. You build the system that talks to the model for you. Bookmark it before it gets lost in your feed. Read it now, then check the article below.