Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

2,934 MB/s write lands on the RAID 0 volume before the second drive even finishes mounting. Two Samsung 990 PROs sit in the top bays of the silver chassis. The cover screws down. One cable carries everything. Blackmagic fills every ProRes and H.265 checkbox green at 2,881 MB/s read.... show more

HodlReaper

4,687 subscribers

14,811 views • 9 days ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

A Mac Mini M4 hits 2,867 MB/s the moment two 990 PROs lock into the dual-bay chassis. The Mini slides into the base. Both drives drop into the top bays. Cover seals. Blackmagic peaks at 2,867 MB/s write and 2,793 MB/s read. Every ProRes and H.265 box turns green. 31 GB lands on the RAID 0 volume in 10 seconds. Three 10Gbps USB-A ports stay live. SD and TF mount as separate volumes. The silver box sits under the monitors while the full library scrubs without a single dropped frame.

A Mac Mini M4 hits 2,867 MB/s the moment two 990 PROs lock into the dual-bay chassis. The Mini slides into the base. Both drives drop into the top bays. Cover seals. Blackmagic peaks at 2,867 MB/s write and 2,793 MB/s read. Every ProRes and H.265 box turns green. 31 GB lands on the RAID 0 volume in 10 seconds. Three 10Gbps USB-A ports stay live. SD and TF mount as separate volumes. The silver box sits under the monitors while the full library scrubs without a single dropped frame.

HodlReaper

37,278 views • 8 days ago

Acasis 8-in-1 40Gbps dock turns a Mac Mini into a 2,910 MB/s workstation with one 4TB SSD slot. Hands open the silver case. A Samsung 990 Pro 4TB drops into the single M.2 bay. Lid screws down. The Mac Mini slides in from the side. Two USB4 ports. Three 10Gbps USB-A. HDMI. DisplayPort. SD 4.0. TF 4.0. Power Delivery. Blackmagic Disk Speed Test hits 2,910 MB/s write and 2,945 MB/s read. Every ProRes and H.265 checkbox turns green. Card slots appear as separate drives. 3.82 GB copies in about 5 seconds. Infrared thermometer reads 25.8°C under load. The dock sits on the desk with three monitors running and the original Mac Mini case still closed.

Acasis 8-in-1 40Gbps dock turns a Mac Mini into a 2,910 MB/s workstation with one 4TB SSD slot. Hands open the silver case. A Samsung 990 Pro 4TB drops into the single M.2 bay. Lid screws down. The Mac Mini slides in from the side. Two USB4 ports. Three 10Gbps USB-A. HDMI. DisplayPort. SD 4.0. TF 4.0. Power Delivery. Blackmagic Disk Speed Test hits 2,910 MB/s write and 2,945 MB/s read. Every ProRes and H.265 checkbox turns green. Card slots appear as separate drives. 3.82 GB copies in about 5 seconds. Infrared thermometer reads 25.8°C under load. The dock sits on the desk with three monitors running and the original Mac Mini case still closed.

HodlReaper

16,198 views • 20 days ago

Acasis 8-in-1 40Gbps dock turns a Mac Mini into a 2,765 MB/s workstation with one 2TB SSD slot. Hands open the silver case. A Samsung 990 Pro 2TB drops into the single M.2 bay. Lid screws down. The Mac Mini slides in from the side. Two USB4 ports. Three 10Gbps USB-A. HDMI. DisplayPort. SD 4.0. TF 4.0. Power Delivery. Blackmagic Disk Speed Test hits 2,765 MB/s write and 2,812 MB/s read. Every ProRes and H.265 checkbox turns green. Card slots appear as separate drives. 2.45 GB copies in about 4 seconds. Infrared thermometer reads 27.3°C under load. The dock sits on the desk with three monitors running and the original Mac Mini case still closed.

Acasis 8-in-1 40Gbps dock turns a Mac Mini into a 2,765 MB/s workstation with one 2TB SSD slot. Hands open the silver case. A Samsung 990 Pro 2TB drops into the single M.2 bay. Lid screws down. The Mac Mini slides in from the side. Two USB4 ports. Three 10Gbps USB-A. HDMI. DisplayPort. SD 4.0. TF 4.0. Power Delivery. Blackmagic Disk Speed Test hits 2,765 MB/s write and 2,812 MB/s read. Every ProRes and H.265 checkbox turns green. Card slots appear as separate drives. 2.45 GB copies in about 4 seconds. Infrared thermometer reads 27.3°C under load. The dock sits on the desk with three monitors running and the original Mac Mini case still closed.

HodlReaper

13,316 views • 19 days ago

A $199 aluminum cube turns a Mac Mini into an 8TB 40Gbps editing station. The case slides under the Mini. Four feet. Two handles. One circular vent. Inside: a single M.2 slot. Samsung 990 PRO drops in. Cover screws shut. No tools beyond a screwdriver. Blackmagic reads 2844 MB/s write. 2881 MB/s read. Three USB-A ports at 10Gbps. SD/TF 4.0 card reader on top. Two DisplayPorts pushing 4K60 each. Camera card inserts. Drive mounts as SSD 4T. Footage lands. Timeline plays without stutter. Two monitors. One timeline. The cube sits quiet between keyboard and speaker. Surface temperature stays low enough that a thermal gun barely notices. The Mini stays clean. The storage stays external. The ports stay plentiful.

A $199 aluminum cube turns a Mac Mini into an 8TB 40Gbps editing station. The case slides under the Mini. Four feet. Two handles. One circular vent. Inside: a single M.2 slot. Samsung 990 PRO drops in. Cover screws shut. No tools beyond a screwdriver. Blackmagic reads 2844 MB/s write. 2881 MB/s read. Three USB-A ports at 10Gbps. SD/TF 4.0 card reader on top. Two DisplayPorts pushing 4K60 each. Camera card inserts. Drive mounts as SSD 4T. Footage lands. Timeline plays without stutter. Two monitors. One timeline. The cube sits quiet between keyboard and speaker. Surface temperature stays low enough that a thermal gun barely notices. The Mini stays clean. The storage stays external. The ports stay plentiful.

HodlReaper

19,978 views • 27 days ago

APPLE CHARGES $200 TO GO FROM 256GB TO 512GB ON THE MAC MINI. $400 FOR 1TB. $600 FOR 2TB. AND YOU HAVE TO DECIDE BEFORE YOU BUY. Pick wrong and you're stuck forever. Apple soldered the storage shut. That's their entire business model. Sell the box cheap, charge a premium for every upgrade you can't do yourself. One dock just broke the whole system. The Acasis 8-in-1 plugs into the Mac Mini through a single Thunderbolt cable. Slide any NVMe SSD into the top. 2,819 MB/s write, 2,857 MB/s read. Faster than Apple's own base storage. A 2TB NVMe SSD costs about $100 on Amazon. Apple charges $600 for the same upgrade. That's a 6x markup for storage you could add yourself in 10 seconds. But it doesn't stop at storage. SD card reader, DisplayPort, USB ports, power delivery. Every port Apple left out, this dock puts back in. Surface temp under load: 26.7°C. Same aluminum finish. Sits right next to the Mac Mini like Apple designed it. That's when it clicks. The $599 Mac Mini was never the cheap option. It was the entry fee. The real cost was every upgrade Apple locked behind the checkout page. One $150 dock and a $100 SSD just replaced $600 worth of Apple tax. Still think you need the upgraded model?

APPLE CHARGES $200 TO GO FROM 256GB TO 512GB ON THE MAC MINI. $400 FOR 1TB. $600 FOR 2TB. AND YOU HAVE TO DECIDE BEFORE YOU BUY. Pick wrong and you're stuck forever. Apple soldered the storage shut. That's their entire business model. Sell the box cheap, charge a premium for every upgrade you can't do yourself. One dock just broke the whole system. The Acasis 8-in-1 plugs into the Mac Mini through a single Thunderbolt cable. Slide any NVMe SSD into the top. 2,819 MB/s write, 2,857 MB/s read. Faster than Apple's own base storage. A 2TB NVMe SSD costs about $100 on Amazon. Apple charges $600 for the same upgrade. That's a 6x markup for storage you could add yourself in 10 seconds. But it doesn't stop at storage. SD card reader, DisplayPort, USB ports, power delivery. Every port Apple left out, this dock puts back in. Surface temp under load: 26.7°C. Same aluminum finish. Sits right next to the Mac Mini like Apple designed it. That's when it clicks. The $599 Mac Mini was never the cheap option. It was the entry fee. The real cost was every upgrade Apple locked behind the checkout page. One $150 dock and a $100 SSD just replaced $600 worth of Apple tax. Still think you need the upgraded model?

Framez

49,948 views • 19 days ago

This Chinese developer linked two $2,999 NVIDIA DGX Sparks into one box and runs the full Qwen3-235B at home, after dropping his $1,999-a-month cloud bill to zero. He wired 2 small boxes into a single computer, split a giant 235-billion-parameter model in half between them, and serves it across his own network at about 10 tokens a second, with no internet, no cloud, right there on the desk. No data center, no thousand-dollar graphics cards, no monthly cloud bill. Just him, 2 gold boxes the size of a sandwich, one cable between them, and 1 power strip. And here is the whole payoff. He used to pay the cloud $1,999 a month for the same model, and the meter ticked on every request. Now he paid $5,998 once for 2 boxes, they covered their cost in 3 months, and after that he sends as many requests as he wants for free, only electricity. The two Sparks talk over one fast cable, each holds 128GB of memory, and together they carry the whole model, about 73GB loaded per box, with the chip inside pinned near the limit at 96%. Both boxes work as one and keep trading data over the cable, with no cloud in the loop and no single word leaking out. The ready model sits on one local address, and any app on his network calls it as easily as ChatGPT. And here is how he described, in plain words, what this pair of boxes does: "this is a pair of boxes that holds the huge Qwen3-235B model and serves it to one network. the model is split in half, and each box owns its half. parts: // Box 1 (holds the first half of the model and starts the answer fast, the first word appears in under a second) // Box 2 (holds the second half and writes out the rest, about 10 tokens a second) // Cable (connects the 2 boxes and moves data between them on every step, with no lag) // Address (one local address where any app sends its request, like to a cloud model) // Test (a script that runs big prompts through and measures speed and delays) // Monitor (checks temperature, power draw, and load on both boxes every 2 seconds). the model never goes to the cloud. he only steps in when a box runs hotter than 80 degrees or the cable between them starts dropping data." So the system knows exactly what it is, what it is for, and where its limits are. It knows it has to hold the whole huge model across 2 boxes on its own. It knows it has to answer every request locally, with no meter, no limits, and no internet. It knows the human is only needed when a box overheats or the link between them stalls. → The setup runs around the clock on 2 boxes, each pulling under 60 watts → However many requests he sends, the monthly bill is $0, only electricity → The first box starts the answer in under a second → The second writes text at about 10 tokens a second → One request at a time: 838 tokens in 85 seconds, first word in 0.8s → Two requests at once: 697 tokens in 108 seconds, first word in 0.7s → Both boxes sit at 96% load and warm up to 76-78 degrees And only when a chip in a box runs hotter than 80 degrees or the cable between the 2 Sparks drops data does the system call the owner. And when he himself is out on a run or in a coffee shop, he still reaches his own model at home from his phone: sends a big prompt to the local Qwen3-235B, gets the full answer back in under a minute and a half, with no token meter ticking and no limit to hit. Here is what the test shows on his screen during one of the night runs: "one request at a time: 838 tokens in 84.9 seconds, first word in 0.8s, then 0.1s per token." "two requests at once: 697 tokens in 107.6 seconds, first word in 0.7s, then 0.15s per token." "Box 1: chip at 96% load, 76 degrees, 56 watts, 73GB used in memory." "Box 2: chip at 96% load, 78 degrees, 56 watts, the Qwen3-235B model fully loaded." And while everyone around is paying for AI by the month and bumping into limits, his top-tier model just sits on the desk and works as much as he wants: his own little power plant instead of a forever meter. He has no server rack of his own and no cloud account behind it. Just 2 DGX Spark boxes on a desk, one model split in half between them, one local address, and a folder of prompts next to it. Out of everything I have seen this year, this is the cleanest way to stop paying for AI: $5,998 of hardware on the desk once, $0 a month to the cloud, unlimited forever, and between them 2 gold boxes, 1 cable, and the full Qwen3-235B answering at home with no internet.

This Chinese developer linked two $2,999 NVIDIA DGX Sparks into one box and runs the full Qwen3-235B at home, after dropping his $1,999-a-month cloud bill to zero. He wired 2 small boxes into a single computer, split a giant 235-billion-parameter model in half between them, and serves it across his own network at about 10 tokens a second, with no internet, no cloud, right there on the desk. No data center, no thousand-dollar graphics cards, no monthly cloud bill. Just him, 2 gold boxes the size of a sandwich, one cable between them, and 1 power strip. And here is the whole payoff. He used to pay the cloud $1,999 a month for the same model, and the meter ticked on every request. Now he paid $5,998 once for 2 boxes, they covered their cost in 3 months, and after that he sends as many requests as he wants for free, only electricity. The two Sparks talk over one fast cable, each holds 128GB of memory, and together they carry the whole model, about 73GB loaded per box, with the chip inside pinned near the limit at 96%. Both boxes work as one and keep trading data over the cable, with no cloud in the loop and no single word leaking out. The ready model sits on one local address, and any app on his network calls it as easily as ChatGPT. And here is how he described, in plain words, what this pair of boxes does: "this is a pair of boxes that holds the huge Qwen3-235B model and serves it to one network. the model is split in half, and each box owns its half. parts: // Box 1 (holds the first half of the model and starts the answer fast, the first word appears in under a second) // Box 2 (holds the second half and writes out the rest, about 10 tokens a second) // Cable (connects the 2 boxes and moves data between them on every step, with no lag) // Address (one local address where any app sends its request, like to a cloud model) // Test (a script that runs big prompts through and measures speed and delays) // Monitor (checks temperature, power draw, and load on both boxes every 2 seconds). the model never goes to the cloud. he only steps in when a box runs hotter than 80 degrees or the cable between them starts dropping data." So the system knows exactly what it is, what it is for, and where its limits are. It knows it has to hold the whole huge model across 2 boxes on its own. It knows it has to answer every request locally, with no meter, no limits, and no internet. It knows the human is only needed when a box overheats or the link between them stalls. → The setup runs around the clock on 2 boxes, each pulling under 60 watts → However many requests he sends, the monthly bill is $0, only electricity → The first box starts the answer in under a second → The second writes text at about 10 tokens a second → One request at a time: 838 tokens in 85 seconds, first word in 0.8s → Two requests at once: 697 tokens in 108 seconds, first word in 0.7s → Both boxes sit at 96% load and warm up to 76-78 degrees And only when a chip in a box runs hotter than 80 degrees or the cable between the 2 Sparks drops data does the system call the owner. And when he himself is out on a run or in a coffee shop, he still reaches his own model at home from his phone: sends a big prompt to the local Qwen3-235B, gets the full answer back in under a minute and a half, with no token meter ticking and no limit to hit. Here is what the test shows on his screen during one of the night runs: "one request at a time: 838 tokens in 84.9 seconds, first word in 0.8s, then 0.1s per token." "two requests at once: 697 tokens in 107.6 seconds, first word in 0.7s, then 0.15s per token." "Box 1: chip at 96% load, 76 degrees, 56 watts, 73GB used in memory." "Box 2: chip at 96% load, 78 degrees, 56 watts, the Qwen3-235B model fully loaded." And while everyone around is paying for AI by the month and bumping into limits, his top-tier model just sits on the desk and works as much as he wants: his own little power plant instead of a forever meter. He has no server rack of his own and no cloud account behind it. Just 2 DGX Spark boxes on a desk, one model split in half between them, one local address, and a folder of prompts next to it. Out of everything I have seen this year, this is the cleanest way to stop paying for AI: $5,998 of hardware on the desk once, $0 a month to the cloud, unlimited forever, and between them 2 gold boxes, 1 cable, and the full Qwen3-235B answering at home with no internet.

Blaze

93,871 views • 3 months ago

🚨 S&P 500 CRASHES EVERY TIME THIS LINE IS CROSSED 🚨 Look at the bottom indicator on this chart It tracks US inflation. And it has one consistent pattern Every time it crossed 4% - the S&P 500 crashed It happened in 2001, 2008 and 2022. Drops ranged from -27% to -57% Right now inflation sits at 3.8% and is accelerating Before the Iran war started in February it was at 2.4% Three months later it's already at 3.8%. Top economists are forecasting 6% by end of quarter The Fed can't cut rates with inflation rising. Which means borrowing stays expensive. Which means earnings fall. Which means the market follows The line is not a prediction. It's a track record. Get ready! NOTIFS ON!

🚨 S&P 500 CRASHES EVERY TIME THIS LINE IS CROSSED 🚨 Look at the bottom indicator on this chart It tracks US inflation. And it has one consistent pattern Every time it crossed 4% - the S&P 500 crashed It happened in 2001, 2008 and 2022. Drops ranged from -27% to -57% Right now inflation sits at 3.8% and is accelerating Before the Iran war started in February it was at 2.4% Three months later it's already at 3.8%. Top economists are forecasting 6% by end of quarter The Fed can't cut rates with inflation rising. Which means borrowing stays expensive. Which means earnings fall. Which means the market follows The line is not a prediction. It's a track record. Get ready! NOTIFS ON!

NoName

47,377 views • 3 months ago

we made the simplest & cheapest cloud computer for agents that can still do complex work: box "--type small" at $0.000,005/s for 2vCPU, 4GB ram, 40GB nvme it can even run chrome while streaming its desktop at 30fps! and the best power for cost deal in the sandbox market: box "--type large" at just $0.07/h for 8vCPU 16GB w/ 100gb nvme it can run minecraft, have chrome open, all while streaming itself at 60fps! the cherry on top: you can stop/resume or fork any box of any type into any other type, conserving all files, in just a few seconds learn more at

we made the simplest & cheapest cloud computer for agents that can still do complex work: box "--type small" at $0.000,005/s for 2vCPU, 4GB ram, 40GB nvme it can even run chrome while streaming its desktop at 30fps! and the best power for cost deal in the sandbox market: box "--type large" at just $0.07/h for 8vCPU 16GB w/ 100gb nvme it can run minecraft, have chrome open, all while streaming itself at 60fps! the cherry on top: you can stop/resume or fork any box of any type into any other type, conserving all files, in just a few seconds learn more at

Anicet

34,383 views • 26 days ago

Qwen 3.8 27B on Atomic Agent was 2x faster than on Hermes and Prime agents! Outputs: Atomic: 5,535 lines, 19 modules, 2h35min Hermes: 10,424 lines, 109 modules, 4h15min Prime: 12,226 lines, 129 modules, 4h42min We gave three agents the same tasks, rebuild these games: ✦ Hole Io ✦ Tower Bloxx ✦ Subway Surfers Results: ✦ Atomic Agent gave us a real city going down the hole, skyscrapers tipping in, cars on the crossings, shops eaten one by one. In the tower every block swings in on its cable and lands where you drop it. Its runner is the cleanest of the three, though the world behind the boy stands still while he runs. ✦ Hermes dropped its hole onto a bare brown field with a few crates, and the hole grew while nothing ever fell in. Its tower buildings hang in the air with no cable at all, the logs reporting 12/12 tests passing on a game it had just broken. Its runner is an empty corridor with almost nothing in it. ✦ Prime built the prettiest town of the three, and five seconds in the whole thing tore off the ground into a tornado spinning around the hole, cars and houses circling forever and never falling in. Its tower lags on every drop, colours flickering on the buildings, and the lag is what brings the whole thing down by the fifth floor. In its runner the coins move at exactly the boy's speed, so the gap never closes and you can never pick one up. Run Qwen 3.8 27B in Atomic Agent!

Qwen 3.8 27B on Atomic Agent was 2x faster than on Hermes and Prime agents! Outputs: Atomic: 5,535 lines, 19 modules, 2h35min Hermes: 10,424 lines, 109 modules, 4h15min Prime: 12,226 lines, 129 modules, 4h42min We gave three agents the same tasks, rebuild these games: ✦ Hole Io ✦ Tower Bloxx ✦ Subway Surfers Results: ✦ Atomic Agent gave us a real city going down the hole, skyscrapers tipping in, cars on the crossings, shops eaten one by one. In the tower every block swings in on its cable and lands where you drop it. Its runner is the cleanest of the three, though the world behind the boy stands still while he runs. ✦ Hermes dropped its hole onto a bare brown field with a few crates, and the hole grew while nothing ever fell in. Its tower buildings hang in the air with no cable at all, the logs reporting 12/12 tests passing on a game it had just broken. Its runner is an empty corridor with almost nothing in it. ✦ Prime built the prettiest town of the three, and five seconds in the whole thing tore off the ground into a tornado spinning around the hole, cars and houses circling forever and never falling in. Its tower lags on every drop, colours flickering on the buildings, and the lag is what brings the whole thing down by the fifth floor. In its runner the coins move at exactly the boy's speed, so the gap never closes and you can never pick one up. Run Qwen 3.8 27B in Atomic Agent!

Atomic Agent

93,468 views • 15 days ago

This cooler hack is one of those “why didn’t we always do it this way” things. Most of us dump a bag of ice in, throw the drinks on top, and spend the next three hours fishing through slush for a can that’s now floating sideways with a soggy label. In Texas heat that gets old fast. The fix is almost too simple: ice on the bottom, a cheap oven rack or grill grate right on top of the ice, then the drinks sit on the rack. Melt water stays down where it belongs. Everything stays organized, nothing gets waterlogged, and you can actually see what you have without digging. Two extras that make it even better for long days: • Freeze a few water bottles the night before. They act as extra ice and turn into drinking water as they melt. • Pre-chill the drinks. The rack works a lot better when you’re not starting with warm cans. We’ve used this for cookouts, lake days, and family gatherings ever since. No more slush-dive. Just grab and go. Simple change. Big difference.

This cooler hack is one of those “why didn’t we always do it this way” things. Most of us dump a bag of ice in, throw the drinks on top, and spend the next three hours fishing through slush for a can that’s now floating sideways with a soggy label. In Texas heat that gets old fast. The fix is almost too simple: ice on the bottom, a cheap oven rack or grill grate right on top of the ice, then the drinks sit on the rack. Melt water stays down where it belongs. Everything stays organized, nothing gets waterlogged, and you can actually see what you have without digging. Two extras that make it even better for long days: • Freeze a few water bottles the night before. They act as extra ice and turn into drinking water as they melt. • Pre-chill the drinks. The rack works a lot better when you’re not starting with warm cans. We’ve used this for cookouts, lake days, and family gatherings ever since. No more slush-dive. Just grab and go. Simple change. Big difference.

NancyH

69,412 views • 6 days ago

Anthropic engineer just built a second brain that reads 3 files instead of 5,000, and cut the token bill 40% no orchestrator, no plugin, just a router file, an index, and code that scores before the model ever opens anything point claude at a flat folder of notes and every question reads all of them, that’s the default and it’s expensive this breakdown turns that exact pattern into a system anyone can run CLAUDE.md → index.md → nodes → edges → retrieval as code → memory → ship it twelve steps, one shared rule, evidence lives in the index, the model reads three files while the notes folder still holds thousands the stale index test, one line drifts and the whole graph quietly turns back into the pile it was built to escape, no exceptions save it and watch, then read the article below

Anthropic engineer just built a second brain that reads 3 files instead of 5,000, and cut the token bill 40% no orchestrator, no plugin, just a router file, an index, and code that scores before the model ever opens anything point claude at a flat folder of notes and every question reads all of them, that’s the default and it’s expensive this breakdown turns that exact pattern into a system anyone can run CLAUDE.md → index.md → nodes → edges → retrieval as code → memory → ship it twelve steps, one shared rule, evidence lives in the index, the model reads three files while the notes folder still holds thousands the stale index test, one line drifts and the whole graph quietly turns back into the pile it was built to escape, no exceptions save it and watch, then read the article below

leopardracer

37,304 views • 1 month ago

Continuous batching in LLMs, clearly explained: (a popular LLM interview question; bookmark this) In traditional ML inference, a batch is a matrix. Every input is padded to the same length, one forward pass runs, and every row finishes at the same moment. LLM decoding does not work that way. One forward pass produces one token per sequence, so a request needs as many passes as it has output tokens, and nobody knows that count until the model emits a stop token. Under static batching, membership is fixed when the batch starts. A request that finishes in 30 tokens holds its slot until the slowest request in the same batch finishes at 400. The GPU keeps paying the full weight read for a batch that is mostly empty. Loading model weights out of HBM costs the same whether four slots are producing tokens or one. Continuous batching moves the decision boundary. Instead of scheduling once per batch, the scheduler runs a single forward pass, gets control back, and decides again. A finished request leaves at the next iteration boundary, and a queued request takes its slot right there. No slot stays reserved for work that is already done. Anyscale benchmarked both OPT-13B on a single A100. With uniform generation lengths, the two policies came out about level (as expected), and as output length variance rose, static batching fell to around 81 tokens per second while vLLM reached 23x the throughput of naive Hugging Face serving. Variance drives the entire gap. Production traffic mixes 30-token replies with 400-token ones, which is exactly the condition static batching handles worst. None of this alters the model. vLLM, SGLang, TGI, and TensorRT-LLM all run it by default, and NVIDIA ships the same mechanism under the name in-flight batching. The animation below runs both policies on the same 16 requests and the same 4 slots, stepping in lockstep. The only difference is when a new request is allowed in. To dive deeper into continuous batching specifically, I wrote a full breakdown of the scheduler underneath it. It covers what happens between two forward passes, how tokens get handed out against a fixed budget, why the scheduler needs no separate path for prefill and decode, and what preemption costs you when the KV cache fills up mid-generation. Read it below.

Continuous batching in LLMs, clearly explained: (a popular LLM interview question; bookmark this) In traditional ML inference, a batch is a matrix. Every input is padded to the same length, one forward pass runs, and every row finishes at the same moment. LLM decoding does not work that way. One forward pass produces one token per sequence, so a request needs as many passes as it has output tokens, and nobody knows that count until the model emits a stop token. Under static batching, membership is fixed when the batch starts. A request that finishes in 30 tokens holds its slot until the slowest request in the same batch finishes at 400. The GPU keeps paying the full weight read for a batch that is mostly empty. Loading model weights out of HBM costs the same whether four slots are producing tokens or one. Continuous batching moves the decision boundary. Instead of scheduling once per batch, the scheduler runs a single forward pass, gets control back, and decides again. A finished request leaves at the next iteration boundary, and a queued request takes its slot right there. No slot stays reserved for work that is already done. Anyscale benchmarked both OPT-13B on a single A100. With uniform generation lengths, the two policies came out about level (as expected), and as output length variance rose, static batching fell to around 81 tokens per second while vLLM reached 23x the throughput of naive Hugging Face serving. Variance drives the entire gap. Production traffic mixes 30-token replies with 400-token ones, which is exactly the condition static batching handles worst. None of this alters the model. vLLM, SGLang, TGI, and TensorRT-LLM all run it by default, and NVIDIA ships the same mechanism under the name in-flight batching. The animation below runs both policies on the same 16 requests and the same 4 slots, stepping in lockstep. The only difference is when a new request is allowed in. To dive deeper into continuous batching specifically, I wrote a full breakdown of the scheduler underneath it. It covers what happens between two forward passes, how tokens get handed out against a fixed budget, why the scheduler needs no separate path for prefill and decode, and what preemption costs you when the KV cache fills up mid-generation. Read it below.

Avi Chawla

16,140 views • 16 days ago

CLAUDE + OBSIDIAN = THE ONLY AI FORMAT THAT WILL SURVIVE THE NEXT DECADE I struggled to build a functional second brain until I applied Karpathy's system -> I used to rely on walled gardens to store my thoughts, jumping from one app to the next then I moved everything to plain text in Obsidian and connected it to Claude seeing the AI read my entire folder locally completely upgraded my approach to work: > you point the AI directly at your local directory for instant context > you own your data forever because everything stays on your machine > the setup eliminates subscription fees and messy database exports > the open format is immune to startup failures because simple text lasts forever complex note apps will eventually shut down. a plain text folder works for a lifetime watch the breakdown 👇

CLAUDE + OBSIDIAN = THE ONLY AI FORMAT THAT WILL SURVIVE THE NEXT DECADE I struggled to build a functional second brain until I applied Karpathy's system -> I used to rely on walled gardens to store my thoughts, jumping from one app to the next then I moved everything to plain text in Obsidian and connected it to Claude seeing the AI read my entire folder locally completely upgraded my approach to work: > you point the AI directly at your local directory for instant context > you own your data forever because everything stays on your machine > the setup eliminates subscription fees and messy database exports > the open format is immune to startup failures because simple text lasts forever complex note apps will eventually shut down. a plain text folder works for a lifetime watch the breakdown 👇

Mr. Buzzoni

38,875 views • 2 months ago

INTEL JUST SHIPPED A WORKSTATION CARD WITH TWO GPUs ON ONE PCB AND 48GB OF VRAM. FOUR CARDS GIVE A SINGLE MOTHERBOARD 192GB OF POOLED INFERENCE MEMORY FOR THE PRICE OF TWO RTX 5090s 00:14 he holds one card up to the camera, two GPU dies side by side under the cooler, each one running its own x8 PCIe lane back to the chipset the cluster sees eight discrete accelerators in software. intel's Battle Matrix stack shards a model across all eight, so a 235B parameter network loads in slices and answers requests in parallel what 192GB of VRAM unlocks: an entire 200B class model in memory without quantization. a vision agent reading 100 invoices at once. a research box that holds three frontier models loaded simultaneously, switching between them in under a second intel is the slow side of inference. nvidia is faster per token, that is the honest tradeoff. but the only other path to this much VRAM is a $40,000 nvidia rack or three networked Mac Studios four B60 cards plus the chassis lands under $5,000. power draw averages 800 watts, $55 a month in electricity. one engineer paying $400 a month for combined ChatGPT Pro and Claude Code Max pays the hardware off in less than a year

INTEL JUST SHIPPED A WORKSTATION CARD WITH TWO GPUs ON ONE PCB AND 48GB OF VRAM. FOUR CARDS GIVE A SINGLE MOTHERBOARD 192GB OF POOLED INFERENCE MEMORY FOR THE PRICE OF TWO RTX 5090s 00:14 he holds one card up to the camera, two GPU dies side by side under the cooler, each one running its own x8 PCIe lane back to the chipset the cluster sees eight discrete accelerators in software. intel's Battle Matrix stack shards a model across all eight, so a 235B parameter network loads in slices and answers requests in parallel what 192GB of VRAM unlocks: an entire 200B class model in memory without quantization. a vision agent reading 100 invoices at once. a research box that holds three frontier models loaded simultaneously, switching between them in under a second intel is the slow side of inference. nvidia is faster per token, that is the honest tradeoff. but the only other path to this much VRAM is a $40,000 nvidia rack or three networked Mac Studios four B60 cards plus the chassis lands under $5,000. power draw averages 800 watts, $55 a month in electricity. one engineer paying $400 a month for combined ChatGPT Pro and Claude Code Max pays the hardware off in less than a year

NO1ennn

48,247 views • 2 months ago

The vitamin K shot given to every newborn has two black box warnings. Here's what the package insert actually says: The package insert for the Vitamin K shot given to newborns at birth carries two black box warnings. The first: fatal hypersensitivity reactions, including anaphylaxis, have occurred during and immediately after injection. The second: this product contains aluminum that may reach toxic levels if kidney function is impaired. That context is missing from every headline currently circulating about a rise in infant bleeding deaths linked to Vitamin K refusal. One widely shared report cited "more than 700 newborns" who died from spontaneous bleeding in 2024, then attributed a "meaningful portion" of those deaths to Vitamin K deficiency, based on the opinion of six unnamed medical specialists and one CDC official. The report attributed a "meaningful portion" of those deaths to Vitamin K deficiency based on the opinions of six unnamed medical specialists and one CDC official, none of whom are identified and none of whom point to a study. Parents who have read the package insert, seen the black box warnings, and asked questions before consenting deserve reporting with more rigor than that.

The vitamin K shot given to every newborn has two black box warnings. Here's what the package insert actually says: The package insert for the Vitamin K shot given to newborns at birth carries two black box warnings. The first: fatal hypersensitivity reactions, including anaphylaxis, have occurred during and immediately after injection. The second: this product contains aluminum that may reach toxic levels if kidney function is impaired. That context is missing from every headline currently circulating about a rise in infant bleeding deaths linked to Vitamin K refusal. One widely shared report cited "more than 700 newborns" who died from spontaneous bleeding in 2024, then attributed a "meaningful portion" of those deaths to Vitamin K deficiency, based on the opinion of six unnamed medical specialists and one CDC official. The report attributed a "meaningful portion" of those deaths to Vitamin K deficiency based on the opinions of six unnamed medical specialists and one CDC official, none of whom are identified and none of whom point to a study. Parents who have read the package insert, seen the black box warnings, and asked questions before consenting deserve reporting with more rigor than that.

The HighWire

19,345 views • 3 months ago

A DeWALT 20V battery powers a Mac Mini for 5 to 8 hours through one USB-C converter. Hands snap the yellow battery onto the converter plate. A short cable plugs into the converter and into the gray side mount on the Mac Mini. Power button lights up. The Mac Mini boots. Custom shortcut turns an iPad into a wireless portable monitor. Full macOS desktop appears. Another shortcut disconnects it clean. The yellow battery sits on the desk while the Mac Mini runs. No wall outlet. Runtime depends on the workload. The test is next.

A DeWALT 20V battery powers a Mac Mini for 5 to 8 hours through one USB-C converter. Hands snap the yellow battery onto the converter plate. A short cable plugs into the converter and into the gray side mount on the Mac Mini. Power button lights up. The Mac Mini boots. Custom shortcut turns an iPad into a wireless portable monitor. Full macOS desktop appears. Another shortcut disconnects it clean. The yellow battery sits on the desk while the Mac Mini runs. No wall outlet. Runtime depends on the workload. The test is next.

HodlReaper

10,684 views • 17 days ago

The last housing crisis was subprime. This one is hiding in FHA and VA loans. Low down payment, or none at all, taken by buyers already living paycheck to paycheck. And the fraud is showing up on the ground. Investors posing as owner-occupants to grab 3 percent down, then renting the homes out. Here is the tell. When default letters go out, a chunk come back from empty houses. A homeowner clings on. An investor is already gone. In parts of Florida, track homes that traded in the low 400s two years ago now go for the low 300s. Buyers who bought the mania are underwater. It follows the K-shaped economy exactly. The bottom cracks first while the top stays flat. And the read here is that this is only the first or second inning.

The last housing crisis was subprime. This one is hiding in FHA and VA loans. Low down payment, or none at all, taken by buyers already living paycheck to paycheck. And the fraud is showing up on the ground. Investors posing as owner-occupants to grab 3 percent down, then renting the homes out. Here is the tell. When default letters go out, a chunk come back from empty houses. A homeowner clings on. An investor is already gone. In parts of Florida, track homes that traded in the low 400s two years ago now go for the low 300s. Buyers who bought the mania are underwater. It follows the K-shaped economy exactly. The bottom cracks first while the top stays flat. And the read here is that this is only the first or second inning.

Jeffrey P. Snider

18,360 views • 1 month ago

DELL ASUS ACER AND GIGABYTE ALL SHIPPED $2,999 DGX SPARK CLONES IN THE SAME QUARTER, SAMSUNG GEN 5 SSD CHOICE CAUSES 2X PERFORMANCE GAPS, THE LOCAL AI HARDWARE MAP JUST GREW BY FOUR FRONTIER TIER OPTIONS 02:13 the reviewer flips through the spec sheets, "we saw major differences in the SSD, some are using Gen 4 SSDs which will slow these platforms down, others have Gen 5 Samsung drives, Samsung is the leader by almost a factor of two" NVIDIA's GB10 Grace Blackwell reference board got picked up by Dell, Asus, Acer and Gigabyte in the same quarter, four $2,999 mini boxes with 128GB unified memory and 1 PFLOP of AI throughput from competing OEMs the Asus version runs coolest across GPU, CPU and SSD thanks to an unpainted bottom plate and a copper heatsink element, the Acer leads on price-to-perf, the Dell honeycomb intake actually restricts airflow slightly, the gigabyte sits middle of pack Samsung Gen 5 versus Gen 4 inside the same chassis causes a 2x read-write gap for the exact same $2,999 sticker, the buyer who reads the SSD spec sheet gets a 70B model loading in 4 seconds, the one who doesn't waits 8 your serie's frontier tier was the $1,700 EVO-X2, four DGX Spark OEM clones at $2,999 extend the map upward, the local AI hardware boom is now staffed by every major PC vendor not just AMD and apple, the category fully arrived bookmark this and read the article below

DELL ASUS ACER AND GIGABYTE ALL SHIPPED $2,999 DGX SPARK CLONES IN THE SAME QUARTER, SAMSUNG GEN 5 SSD CHOICE CAUSES 2X PERFORMANCE GAPS, THE LOCAL AI HARDWARE MAP JUST GREW BY FOUR FRONTIER TIER OPTIONS 02:13 the reviewer flips through the spec sheets, "we saw major differences in the SSD, some are using Gen 4 SSDs which will slow these platforms down, others have Gen 5 Samsung drives, Samsung is the leader by almost a factor of two" NVIDIA's GB10 Grace Blackwell reference board got picked up by Dell, Asus, Acer and Gigabyte in the same quarter, four $2,999 mini boxes with 128GB unified memory and 1 PFLOP of AI throughput from competing OEMs the Asus version runs coolest across GPU, CPU and SSD thanks to an unpainted bottom plate and a copper heatsink element, the Acer leads on price-to-perf, the Dell honeycomb intake actually restricts airflow slightly, the gigabyte sits middle of pack Samsung Gen 5 versus Gen 4 inside the same chassis causes a 2x read-write gap for the exact same $2,999 sticker, the buyer who reads the SSD spec sheet gets a 70B model loading in 4 seconds, the one who doesn't waits 8 your serie's frontier tier was the $1,700 EVO-X2, four DGX Spark OEM clones at $2,999 extend the map upward, the local AI hardware boom is now staffed by every major PC vendor not just AMD and apple, the category fully arrived bookmark this and read the article below

starmex

17,573 views • 2 months ago

🚨 WOMAN DECIDES TO “SEE WHAT’S REALLY INSIDE MONSTER ENERGY” — 30 SECONDS LATER HER FACE SAYS EVERYTHING A woman sits down with a can of Monster Energy and says one line: “Let’s see what’s inside.” Then she goes completely silent. She just starts rebuilding the drink ingredient by ingredient in a glass using the label on the can. First the base. Then the powders. Then the extracts. Then the acids. Then the stimulants. One scoop… One drop… One strange chemical name at a time. What starts as clear water slowly turns into a glowing neon green mixture that looks more like a lab experiment than a drink. She still hasn’t said a word the entire time. Then she finishes. Looks at the glass… Looks at the camera… And makes a face like she just realized something she wasn’t expecting. The clip is going viral because watching the entire ingredient list dumped into one glass hits very differently than just reading the label. The strangest part? Millions of people drink these every single day without ever thinking about what it actually looks like when everything goes in at once. Would you still drink it after watching this?

🚨 WOMAN DECIDES TO “SEE WHAT’S REALLY INSIDE MONSTER ENERGY” — 30 SECONDS LATER HER FACE SAYS EVERYTHING A woman sits down with a can of Monster Energy and says one line: “Let’s see what’s inside.” Then she goes completely silent. She just starts rebuilding the drink ingredient by ingredient in a glass using the label on the can. First the base. Then the powders. Then the extracts. Then the acids. Then the stimulants. One scoop… One drop… One strange chemical name at a time. What starts as clear water slowly turns into a glowing neon green mixture that looks more like a lab experiment than a drink. She still hasn’t said a word the entire time. Then she finishes. Looks at the glass… Looks at the camera… And makes a face like she just realized something she wasn’t expecting. The clip is going viral because watching the entire ingredient list dumped into one glass hits very differently than just reading the label. The strangest part? Millions of people drink these every single day without ever thinking about what it actually looks like when everything goes in at once. Would you still drink it after watching this?

HustleBitch

51,989 views • 5 months ago

APPLE SOLD THIS 39.9-POUND TOWER FOR $2,499 IN 2009 - ONE LATCH TURNED THE ENTIRE MACHINE INTO A WORKBENCH. this is the Early 2009 Mac Pro. the video shows the part spec sheets rarely capture. pull one latch and the aluminum side panel comes away. the memory, graphics card, drive bays and expansion slots are immediately visible. no hidden screws before you can inspect the machine. no loose SATA cables hanging from every drive. the base model shipped with: 2.66GHz quad-core Xeon 3GB of 1066MHz DDR3 ECC memory 640GB SATA hard drive GeForce GT 120 with 512MB of memory four PCIe 2.0 slots. the quad-core model supported up to 16GB of RAM. storage was handled by four cable-free carriers that connected directly inside the chassis. the case also left space for two optical drives. the rear panel carried hardware that now feels like an archive: dual Gigabit Ethernet optical audio USB 2.0 FireWire 800. the GT 120 offered Mini DisplayPort and dual-link DVI. install four of those cards and Apple rated the tower for up to eight 30-inch displays. the enclosure measured 20.1 inches tall and weighed almost 40 pounds before serious upgrades. this was not a compact desktop. it was a machine designed to be opened, understood and rebuilt. honest line: the Xeon, GT 120, SATA drive and USB 2.0 ports are ancient by modern standards. beautiful serviceability does not make old hardware fast. those parts aged. the enclosure did not. 17 years later, the most impressive component is still the case. bookmark & watch today ↓

APPLE SOLD THIS 39.9-POUND TOWER FOR $2,499 IN 2009 - ONE LATCH TURNED THE ENTIRE MACHINE INTO A WORKBENCH. this is the Early 2009 Mac Pro. the video shows the part spec sheets rarely capture. pull one latch and the aluminum side panel comes away. the memory, graphics card, drive bays and expansion slots are immediately visible. no hidden screws before you can inspect the machine. no loose SATA cables hanging from every drive. the base model shipped with: 2.66GHz quad-core Xeon 3GB of 1066MHz DDR3 ECC memory 640GB SATA hard drive GeForce GT 120 with 512MB of memory four PCIe 2.0 slots. the quad-core model supported up to 16GB of RAM. storage was handled by four cable-free carriers that connected directly inside the chassis. the case also left space for two optical drives. the rear panel carried hardware that now feels like an archive: dual Gigabit Ethernet optical audio USB 2.0 FireWire 800. the GT 120 offered Mini DisplayPort and dual-link DVI. install four of those cards and Apple rated the tower for up to eight 30-inch displays. the enclosure measured 20.1 inches tall and weighed almost 40 pounds before serious upgrades. this was not a compact desktop. it was a machine designed to be opened, understood and rebuilt. honest line: the Xeon, GT 120, SATA drive and USB 2.0 ports are ancient by modern standards. beautiful serviceability does not make old hardware fast. those parts aged. the enclosure did not. 17 years later, the most impressive component is still the case. bookmark & watch today ↓

Grimmer

224,481 views • 14 days ago