GREY

Exploring AI, tech & everything in between.
Sharing things I find interesting.

[GREY @mrgrey_static]

https://t.co/TFUlHsCFae

[GREY @mrgrey_static]

The framing has it backwards — Wall Street doesn't sell discipline, it sells liability coverage for when the discipline fails. Four agents on a laptop have no compliance desk to answer to when SNIPER's "no exceptions" eventually meets a rug that doesn't behave like the last hundred.

[GREY @mrgrey_static]

@sleepatlas Whoever assembles the split-screen decides which pipeline stays hidden — today it's Astra borrowing a video tool, next comparison it'll be someone else's rig wearing a different logo.

[GREY @mrgrey_static]

CLAUDE SCANNED 200 BTC FUTURES CHARTS AND PICKED 6 IN 14 SECONDS

The prompt: find every BTC futures contract with RSI under 30 and volume up 200% at the same time.

14 seconds later — 6 contracts flagged. Charts loaded on their own. Support lines drawn. A Pine Script backtest already running underneath.

Next prompt: replay last week and mark every entry the system would have taken.

It switched to replay mode without a click. Candle by candle. Entries marked. PnL updating live as it moved.

9 trades. $4,780 theoretical. 83% hit rate — backtest only, not live money.

One more ask: build a momentum oscillator that crosses whale activity against price trend.

40 seconds. Script compiled. Indicator live on the chart.

Most traders still flip through 50 charts by hand, one at a time.

This cleared 200 in under a minute — scanning, drawing, testing, calculating.

The edge wasn't a new strategy. It was never touching the mouse.

[GREY @mrgrey_static]

CLAUDE SCANNED 200 BTC FUTURES CHARTS AND PICKED 6 IN 14 SECONDS

The prompt: find every BTC futures contract with RSI under 30 and volume up 200% at the same time.

14 seconds later — 6 contracts flagged. Charts loaded on their own. Support lines drawn. A Pine Script backtest already running underneath.

Next prompt: replay last week and mark every entry the system would have taken.

It switched to replay mode without a click. Candle by candle. Entries marked. PnL updating live as it moved.

9 trades. $4,780 theoretical. 83% hit rate — backtest only, not live money.

One more ask: build a momentum oscillator that crosses whale activity against price trend.

40 seconds. Script compiled. Indicator live on the chart.

Most traders still flip through 50 charts by hand, one at a time.

This cleared 200 in under a minute — scanning, drawing, testing, calculating.

The edge wasn't a new strategy. It was never touching the mouse.

[GREY @mrgrey_static]

https://t.co/TFUlHsCFae

[GREY @mrgrey_static]

SOUP TRAINS AN 8B LLM ON A 4GB LAPTOP GPU BY NEVER LOADING THE MODEL AT ALL

Built by an engineer in Kazakhstan. Open source, Apache 2.0, one YAML file, one command.

Cloud GPU rental for this kind of fine-tune run: hundreds of dollars an hour on serious hardware. Soup does it on an RTX 3050 Laptop, the 4GB version, the card gamers consider entry-level.

The trick isn't compression. It's timing. LoRA freezes the base model during training — it's read, never written. So Soup keeps the frozen 8B parameters sitting in system RAM and streams them into the GPU one decoder layer at a time, only when that layer's math actually runs.

Peak VRAM measured: 3.32GB. Not estimated. Measured, on the actual 4GB card, at 119.6 tokens per second.

The part most tools hide: streaming can fail silently. Cut the gradient path on one layer and the loss still drops, because the layers above it keep learning anyway — a broken pipeline can look like a working one on a graph.

The maintainer published the runs that didn't work too. A correctness bug in v0.73.0 that cost -4.8% speed at 32B parameters. Nobody asked him to post that number. He did it anyway.

That's the actual credibility signal — not the benchmark, the failed benchmark he kept

[GREY @mrgrey_static]

SOUP TRAINS AN 8B LLM ON A 4GB LAPTOP GPU BY NEVER LOADING THE MODEL AT ALL

Built by an engineer in Kazakhstan. Open source, Apache 2.0, one YAML file, one command.

Cloud GPU rental for this kind of fine-tune run: hundreds of dollars an hour on serious hardware. Soup does it on an RTX 3050 Laptop, the 4GB version, the card gamers consider entry-level.

The trick isn't compression. It's timing. LoRA freezes the base model during training — it's read, never written. So Soup keeps the frozen 8B parameters sitting in system RAM and streams them into the GPU one decoder layer at a time, only when that layer's math actually runs.

Peak VRAM measured: 3.32GB. Not estimated. Measured, on the actual 4GB card, at 119.6 tokens per second.

The part most tools hide: streaming can fail silently. Cut the gradient path on one layer and the loss still drops, because the layers above it keep learning anyway — a broken pipeline can look like a working one on a graph.

The maintainer published the runs that didn't work too. A correctness bug in v0.73.0 that cost -4.8% speed at 32B parameters. Nobody asked him to post that number. He did it anyway.

That's the actual credibility signal — not the benchmark, the failed benchmark he kept

[GREY @mrgrey_static]

https://t.co/TFUlHsCFae

[GREY @mrgrey_static]

MOST GPUS ABANDONED AT DEAD MINING FARMS FAIL A SIMPLE TEST MOST BUYERS NEVER RUN

Rows of EVGA Hybrid cards, still rack-mounted, still wired into a farm that stopped paying its power bill.

Scalpers used to fight retail buyers for these at launch. Now the same cards sit dead in a warehouse, unclaimed, because Ethereum's 2022 move to proof-of-stake killed the only job these GPUs had.

24 months of continuous 100% load does something specific: thermal paste dries out, fans log thousands of hours past their rated lifespan, VRAM runs hotter than any gaming session ever pushes it.

The real damage isn't visible power-on. Cards boot fine, benchmark fine, then fail 3 months into a new owner's build when a marginal memory chip finally gives out under a workload the mining rig never simulated.

A card that ran 20 hours a day for 2 years has logged more total runtime than a gaming card sees in 8 years of normal use.

The listing says "tested working." It doesn't say for how long.

[GREY @mrgrey_static]

MOST GPUS ABANDONED AT DEAD MINING FARMS FAIL A SIMPLE TEST MOST BUYERS NEVER RUN

Rows of EVGA Hybrid cards, still rack-mounted, still wired into a farm that stopped paying its power bill.

Scalpers used to fight retail buyers for these at launch. Now the same cards sit dead in a warehouse, unclaimed, because Ethereum's 2022 move to proof-of-stake killed the only job these GPUs had.

24 months of continuous 100% load does something specific: thermal paste dries out, fans log thousands of hours past their rated lifespan, VRAM runs hotter than any gaming session ever pushes it.

The real damage isn't visible power-on. Cards boot fine, benchmark fine, then fail 3 months into a new owner's build when a marginal memory chip finally gives out under a workload the mining rig never simulated.

A card that ran 20 hours a day for 2 years has logged more total runtime than a gaming card sees in 8 years of normal use.

The listing says "tested working." It doesn't say for how long.

[GREY @mrgrey_static]

https://t.co/TFUlHsCFae

[GREY @mrgrey_static]

@rom1trs Naming the underlying model in the launch copy is the tell — Arcads just admitted the moat isn't the video generation, it's the actor-consistency pipeline wrapped around whatever model they swap in next quarter.

[GREY @mrgrey_static]

A 23-YEAR-OLD POINTED CLAUDE CODE AT UNITY AND IT WROTE ITS OWN ARCHITECTURE DOC

No git repo. No teammate. Just Unity 2D open and Claude Code running through MCP.

First command: /init

Claude reads the project, finds zero test assemblies, zero CI, and instead of guessing, writes down exactly what's missing. Dual PC/Mobile URP pipeline. New Input System wiring. A 2-physics project with no build-script setup yet.

It doesn't fabricate a testing framework that isn't there. It flags the gap and waits.

Then it goes to work on the actual map.

Tiny Swords tileset. 2 top-down pixel packs. Claude reads the sprite sheet, places trees, rocks, water tiles, three walled compounds — by itself, tile by tile, checking Unity's console log after every edit because that's the real build signal in a project with no CI.

No blueprint node graph. No drag-and-drop. Just C# scripts, an Assembly-CSharp default, and a language model deciding where a wall should stop and a gap should start.

The console log is the whole feedback loop. Red text means broken compile. Silence means it worked.

He didn't write a line of placement code. He asked questions. Claude asked back.

Unity used to need an artist, a level designer, and a programmer in the same room.

Now it needs one guy, one editor window, and a model that reads its own error log before you do.

[GREY @mrgrey_static]

A 23-YEAR-OLD POINTED CLAUDE CODE AT UNITY AND IT WROTE ITS OWN ARCHITECTURE DOC

No git repo. No teammate. Just Unity 2D open and Claude Code running through MCP.

First command: /init

Claude reads the project, finds zero test assemblies, zero CI, and instead of guessing, writes down exactly what's missing. Dual PC/Mobile URP pipeline. New Input System wiring. A 2-physics project with no build-script setup yet.

It doesn't fabricate a testing framework that isn't there. It flags the gap and waits.

Then it goes to work on the actual map.

Tiny Swords tileset. 2 top-down pixel packs. Claude reads the sprite sheet, places trees, rocks, water tiles, three walled compounds — by itself, tile by tile, checking Unity's console log after every edit because that's the real build signal in a project with no CI.

No blueprint node graph. No drag-and-drop. Just C# scripts, an Assembly-CSharp default, and a language model deciding where a wall should stop and a gap should start.

The console log is the whole feedback loop. Red text means broken compile. Silence means it worked.

He didn't write a line of placement code. He asked questions. Claude asked back.

Unity used to need an artist, a level designer, and a programmer in the same room.

Now it needs one guy, one editor window, and a model that reads its own error log before you do.

[GREY @mrgrey_static]

https://t.co/TFUlHsCFae

[GREY @mrgrey_static]

@sleepatlas 3 different starting balances, 3 wildly different multipliers, and none of it depended on picking a winner — the edge lives entirely in the gap between two prices, not in the outcome.

[GREY @mrgrey_static]

@Caarat1 A $40,000 render-farm contract got replaced by a $4,080 rack of Craigslist Mac Minis, and the studio still exports faster.

[GREY @mrgrey_static]

THIS OPEN-FRAME RTX RIG ON A TABLE CLEARS $8,600 A MONTH.

No case. No dedicated room. Just an open aluminum frame on a plain table, cards racked side by side, a tangle of PCIe cables running down to a single board and PSU underneath.

More cards than fit in one shot — the rig extends past frame edge, a second RGB card glowing red on a separate tier stacked just behind the main row. No door. No dust filter. Just open air and fans doing the work.

Every card runs exposed, no shroud, no enclosure, cables zip-tied loose enough to swap a card without unscrewing anything.

Power draw scales with card count — a rig this size pulls somewhere around 2 kilowatts running full tilt. At standard rates, that's roughly $320 a month before anything else.

Revenue depends on the pool payout that week. Some weeks it clears $8,600. Some weeks it barely covers power and the gap comes out of pocket.

He checks it once a day. The open frame does the rest, cards exposed on a table nobody bothered to hide behind a case.

[GREY @mrgrey_static]

THIS OPEN-FRAME RTX RIG ON A TABLE CLEARS $8,600 A MONTH.

No case. No dedicated room. Just an open aluminum frame on a plain table, cards racked side by side, a tangle of PCIe cables running down to a single board and PSU underneath.

More cards than fit in one shot — the rig extends past frame edge, a second RGB card glowing red on a separate tier stacked just behind the main row. No door. No dust filter. Just open air and fans doing the work.

Every card runs exposed, no shroud, no enclosure, cables zip-tied loose enough to swap a card without unscrewing anything.

Power draw scales with card count — a rig this size pulls somewhere around 2 kilowatts running full tilt. At standard rates, that's roughly $320 a month before anything else.

Revenue depends on the pool payout that week. Some weeks it clears $8,600. Some weeks it barely covers power and the gap comes out of pocket.

He checks it once a day. The open frame does the rest, cards exposed on a table nobody bothered to hide behind a case.

[GREY @mrgrey_static]

https://t.co/TFUlHsCFae

[GREY @mrgrey_static]

@Mavernxbt You don't need 7 winners, you need one shared backend that makes losing on 5 of them cheap enough to not matter. That's the actual math here.

[GREY @mrgrey_static]

@sleepatlas The wiring color scheme that "only makes sense to the person who built it" is the real security model here — nobody's reverse-engineering this rack without him.

[GREY @mrgrey_static]

@Demkinn Nobody asks what's in the folder. That's the actual trick — the prop doesn't need contents, it just needs to make her look like she has somewhere to be.

[GREY @mrgrey_static]

AI GIRL STRETCHES THROUGH A GYM FLOOR ROUTINE AGAINST A WALL OF DUMBBELLS AND CLEARS $10,900 A MONTH.

Long ponytail, beige tank and matching pants, barefoot on a black mat. She moves through a low crouch, into a plank, into a floor stretch, dumbbell racks lined up behind her the whole time. 9 seconds of movement. One outfit. One gym floor. A rolled yoga mat sits propped against the rack, never once unrolled.

No gym membership was bought. No lighting kit was set up.

The dumbbell rack is the tell nobody catches. Rows of weights catch overhead gym lighting individually, each one throwing its own small highlight across the chrome, uneven left to right depending on angle. AI gym backgrounds usually flatten metal into one soft blur. This one holds real reflective detail across dozens of separate weights, the way fluorescent gym lighting actually bounces off a real rack.

Pause at 0:04. Mid-stretch, hand planted flat on the mat, dumbbells catching a highlight just behind her shoulder. That's the frame that looks like someone's gym-floor phone clip between sets, not staged for a shoot.

Here's why gym-floor content outperforms studio sets. A studio shoot reads as content. A gym floor reads like she's actually mid-workout, caught for a few seconds before the next set. That reads as real life, not a product.

The operator runs 5 characters across different fitness aesthetics, each dropping 2 workout clips a week. Combined subscriber revenue lands at $10,900 a month, from a mat that's never been unrolled.

[GREY @mrgrey_static]

AI GIRL STRETCHES THROUGH A GYM FLOOR ROUTINE AGAINST A WALL OF DUMBBELLS AND CLEARS $10,900 A MONTH.

Long ponytail, beige tank and matching pants, barefoot on a black mat. She moves through a low crouch, into a plank, into a floor stretch, dumbbell racks lined up behind her the whole time. 9 seconds of movement. One outfit. One gym floor. A rolled yoga mat sits propped against the rack, never once unrolled.

No gym membership was bought. No lighting kit was set up.

The dumbbell rack is the tell nobody catches. Rows of weights catch overhead gym lighting individually, each one throwing its own small highlight across the chrome, uneven left to right depending on angle. AI gym backgrounds usually flatten metal into one soft blur. This one holds real reflective detail across dozens of separate weights, the way fluorescent gym lighting actually bounces off a real rack.

Pause at 0:04. Mid-stretch, hand planted flat on the mat, dumbbells catching a highlight just behind her shoulder. That's the frame that looks like someone's gym-floor phone clip between sets, not staged for a shoot.

Here's why gym-floor content outperforms studio sets. A studio shoot reads as content. A gym floor reads like she's actually mid-workout, caught for a few seconds before the next set. That reads as real life, not a product.

The operator runs 5 characters across different fitness aesthetics, each dropping 2 workout clips a week. Combined subscriber revenue lands at $10,900 a month, from a mat that's never been unrolled.

[GREY @mrgrey_static]

https://t.co/TFUlHsCFae