Brian Roemmele(@BrianRoemmele):BOOM! 34 People, Including The Janitor: A 1963 IBM Memo Explains Why A 125-Billion-Parameter AI Now Runs On Your Gaming PC, And How To Use It On August 28, 1963, Thomas J. Watson Jr., the head of IBM, sent a short memorandum to seven of his top executives. "Last week CDC had a press conference during which they officially announced their 6600 system. I understand that in the laboratory developing this system there are only 34 people, 'including the janitor.' Of these, 14 are engineers and 4 are programmers, and only one person has a PhD., a relatively junior programmer." Then the sentence that has echoed through computing ever since: "Contrasting this modest effort with our own vast development activities, I fail to understand why we have lost our industry leadership by letting someone else offer the world's most powerful computer." Imagine being one of those seven executives. You run the most powerful technology company on the planet. You have thousands of engineers, rooms full of PhDs, and budgets beyond counting. And a man named Seymour Cray, working with a few dozen people in a lab in the woods of Chippewa Falls, Wisconsin, has just built a machine about three times faster than your best, the IBM 7030 Stretch. To the outsider, Watson wrote, Cray's laboratory "appeared to be cost concious, hard working, and highly motivated." He even kept the typo. The Computer History Museum still holds the scan, courtesy of the IBM Corporate Archive. The memo is real. The numbers are real. The sting is real. The Woods Lab That Beat The Giant Seymour Cray had asked CDC's CEO for a lab free of headquarters interference. He got one in his hometown. With Jim Thornton and a small crew, he built the CDC 6600: silicon transistors from Fairchild, a reduced instruction set before anyone named RISC, ten peripheral processors handling the housekeeping so the central processor could stay on the math. The machine ran at 10 MHz. It weighed about six tons. It was the world's fastest computer from 1964 until its own successor took the crown. Small. Frugal. Highly motivated. History repeats this lesson every few decades, and the vast ones forget it every time. Legend says Cray read Watson's memo and replied, "It seems like Mr. Watson has answered his own question." The line appears in textbooks, slides and oral histories. I have never seen a primary source for it. It survives because everyone who has worked on a small team knows the truth inside it: the size of the room was the advantage. A 125-Billion-Parameter Mind Under Your Desk It is happening again this week. On October 4, 2026, a free, open-source project called Strata landed near the top of open source sites and stayed there into Monday morning with hundreds of points. Strata does one thing: it runs a 125-billion-parameter AI model, Qwen3.8-Flash-Next, on an ordinary gaming PC. In its author's words, this is "a large, smart AI model that usually needs a server." The published requirements are plain. A graphics card with 12 GB of memory or more. 32 GB of system RAM, with 64 GB enough to run every size. About 80 GB of free disk, preferably an SSD. Windows 10 or 11, or Linux. NVIDIA or AMD. MIT license for the tool itself. An OpenAI-compatible and Anthropic-compatible programming interface on your own machine at http://127.0.0.1:8080. And, again in the author's words: "Nothing leaves your PC." On an RTX 5070 with 12 GB and 64 GB of RAM, the author measured 94 tokens per second writing answers with the fastest compressed size, and more than 2,000 tokens per second reading a long prompt. On an RX 9070 XT the same size wrote at 60 tokens per second. The author estimates a 24 GB card such as an RTX 3090 should write in the 100 to 140 tokens-per-second range. One commenter reported 124 tokens per second on an RTX 4090 with 128 GB of RAM. Your mileage will vary with quantization, RAM speed and what else is open on the machine. The point stands: a server-class parameter count is answering on consumer silicon. Consider what that means. The model chats, writes code, reads pictures and works with coding agents. The tools you already use can point at a local address instead of a metered cloud. Nobody can switch it off, raise the price overnight, or read its mail. The Frugal Trick Cray Would Have Loved A model this size will not fit on a gaming card alone. Strata spreads the work across the whole machine. Think of a kitchen. The few things you use constantly stay on the counter. Everything else waits in the pantry. In Strata's case the "experts" that the model calls most often stay in graphics memory. The rest live in system RAM. The SSD holds the large lookup tables. A small helper guesses the next few words; the big model checks them in one pass. The result is the same answer, delivered sooner, on hardware you already own. Vast data centers are one way to run a mind. The modest machine under your desk is now another. Why This Matters To You Earlier this year I founded a Zero-Human Company on a 12-year-old MacBook in my garage: AI agents as the staff, local models as the regulators, salvaged archives as the raw material. The limit has always been the size of the mind I could run locally without asking permission. That limit just moved. Personal AI is the new invention of fire. The people who own the minds on their own desks will write their own rules. A reader with a gaming PC can now staff a small company of AI agents with a model in the same weight class as systems the giants rent out by the token. This is the worst it will ever be. How To Hire It Here is the playbook, drawn from Strata's public README and install docs as of this writing. Follow those docs for the latest flags and model menu. Licenses for the model and each compressed build are separate from Strata's MIT license; read them before you rely on the system for anything important. Step 1. Check your hardware. You need a supported NVIDIA GeForce RTX 20/30/40/50-series card, or a supported AMD Radeon card, with at least 12 GB of VRAM. Confirm 32 GB of system RAM or more. Free about 80 GB on an SSD. Update your graphics driver from NVIDIA or AMD. Step 2. Get Strata. Clone or download the repository at https://t.co/FBXmn4NVpf. Unzip it. On Windows, double-click START-HERE.bat. On Linux, run ./setup.sh inside the Strata folder. If you already use a coding assistant on that machine, you can paste the project's AI setup prompt and let it drive the install from docs/AI_SETUP.md. Step 3. Answer the installer's questions. It will ask which model size to download, how much context to keep, and whether you want image input. Press Enter for the recommended answers if you are unsure. The download is roughly 70 GB; if it stops, run the installer again and it continues. Step 4. Pick a size that fits your RAM. With 32 GB, start with the Coder build, which drops half the experts and is tuned for code. With 48 GB, IQ2_XS or Q2_0. With 64 GB, IQ2_XS is the author's recommended default; IQ3_S is smarter and slower. With 96 GB or more you can try the larger Unsloth builds. Smaller sizes are faster. Larger sizes keep more quality. Step 5. Wait through the first start. The first launch can make the PC slow or unresponsive for one to three minutes while Strata loads tens of gigabytes into RAM. That is normal. Do not close the window. Your browser should open the Strata app at http://127.0.0.1:8080. Step 6. Point your tools at home. In any app that accepts an OpenAI-compatible provider, set the base URL to http://127.0.0.1:8080/v1. Any API key and any model name will work for the local endpoint. For Anthropic-style tools, use http://127.0.0.1:8080/v1/messages. For Codex CLI and the OpenAI Responses API, use /v1/responses. Thinking effort can be set to off, low, medium or high; off is fastest, high is for hard questions. Step 7. Keep it on your desk. By default Strata binds to localhost. Leave it there unless you know why you need otherwise. If you ever expose it to another machine, the docs say to set an API key. For a Zero-Human Company roster, treat this local endpoint as the shared mind: one CEO agent for strategy, one engineer agent for code, one researcher for archives, one archivist for files, and one night janitor that cleans logs and memory while you sleep. Wire them through the same OpenAI-compatible address. The org chart fits on one machine the way Cray's lab fit in the Wisconsin woods. Step 8. Know the limits. Heavy compression trades some quality for speed. The largest sizes may page from the SSD and slow down if RAM is tight. You are running a third-party binary and a third-party model on your own hardware; that is the point, and it is also a responsibility. Close other memory hogs if the disk light never stops blinking. Read the troubleshooting doc before you open an issue. Watson had thousands of people and could not understand how 34 beat him. You may need one gaming PC and one evening. The Pattern Every few decades a frugal team builds something the giant said required a cathedral. In 1963 it was a supercomputer in Chippewa Falls. In 2026 it is a 125-billion-parameter mind under a glass side panel, glowing cyan and magenta on a desk that also holds a coffee mug and a slide rule of a different kind. The memo is sixty-three years old. The question is the same. Who will you hire, and on whose machine will they work?

2026.10.05 17:41

BOOM! 34 People, Including The Janitor: A 1963 IBM Memo Explains Why A 125-Billion-Parameter AI Now Runs On Your Gaming PC, And How To Use It On August 28, 1963, Thomas J. Watson Jr., the head of IBM, sent a short memorandum to seven of his top executives. "Last week CDC had a press conference during which they officially announced their 6600 system. I understand that in the laboratory developing this system there are only 34 people, 'including the janitor.' Of these, 14 are engineers and 4 are programmers, and only one person has a PhD., a relatively junior programmer." Then the sentence that has echoed through computing ever since: "Contrasting this modest effort with our own vast development activities, I fail to understand why we have lost our industry leadership by letting someone else offer the world's most powerful computer." Imagine being one of those seven executives. You run the most powerful technology company on the planet. You have thousands of engineers, rooms full of PhDs, and budgets beyond counting. And a man named Seymour Cray, working with a few dozen people in a lab in the woods of Chippewa Falls, Wisconsin, has just built a machine about three times faster than your best, the IBM 7030 Stretch. To the outsider, Watson wrote, Cray's laboratory "appeared to be cost concious, hard working, and highly motivated." He even kept the typo. The Computer History Museum still holds the scan, courtesy of the IBM Corporate Archive. The memo is real. The numbers are real. The sting is real. The Woods Lab That Beat The Giant Seymour Cray had asked CDC's CEO for a lab free of headquarters interference. He got one in his hometown. With Jim Thornton and a small crew, he built the CDC 6600: silicon transistors from Fairchild, a reduced instruction set before anyone named RISC, ten peripheral processors handling the housekeeping so the central processor could stay on the math. The machine ran at 10 MHz. It weighed about six tons. It was the world's fastest computer from 1964 until its own successor took the crown. Small. Frugal. Highly motivated. History repeats this lesson every few decades, and the vast ones forget it every time. Legend says Cray read Watson's memo and replied, "It seems like Mr. Watson has answered his own question." The line appears in textbooks, slides and oral histories. I have never seen a primary source for it. It survives because everyone who has worked on a small team knows the truth inside it: the size of the room was the advantage. A 125-Billion-Parameter Mind Under Your Desk It is happening again this week. On October 4, 2026, a free, open-source project called Strata landed near the top of open source sites and stayed there into Monday morning with hundreds of points. Strata does one thing: it runs a 125-billion-parameter AI model, Qwen3.8-Flash-Next, on an ordinary gaming PC. In its author's words, this is "a large, smart AI model that usually needs a server." The published requirements are plain. A graphics card with 12 GB of memory or more. 32 GB of system RAM, with 64 GB enough to run every size. About 80 GB of free disk, preferably an SSD. Windows 10 or 11, or Linux. NVIDIA or AMD. MIT license for the tool itself. An OpenAI-compatible and Anthropic-compatible programming interface on your own machine at And, again in the author's words: "Nothing leaves your PC." On an RTX 5070 with 12 GB and 64 GB of RAM, the author measured 94 tokens per second writing answers with the fastest compressed size, and more than 2,000 tokens per second reading a long prompt. On an RX 9070 XT the same size wrote at 60 tokens per second. The author estimates a 24 GB card such as an RTX 3090 should write in the 100 to 140 tokens-per-second range. One commenter reported 124 tokens per second on an RTX 4090 with 128 GB of RAM. Your mileage will vary with quantization, RAM speed and what else is open on the machine. The point stands: a server-class parameter count is answering on consumer silicon. Consider what that means. The model chats, writes code, reads pictures and works with coding agents. The tools you already use can point at a local address instead of a metered cloud. Nobody can switch it off, raise the price overnight, or read its mail. The Frugal Trick Cray Would Have Loved A model this size will not fit on a gaming card alone. Strata spreads the work across the whole machine. Think of a kitchen. The few things you use constantly stay on the counter. Everything else waits in the pantry. In Strata's case the "experts" that the model calls most often stay in graphics memory. The rest live in system RAM. The SSD holds the large lookup tables. A small helper guesses the next few words; the big model checks them in one pass. The result is the same answer, delivered sooner, on hardware you already own. Vast data centers are one way to run a mind. The modest machine under your desk is now another. Why This Matters To You Earlier this year I founded a Zero-Human Company on a 12-year-old MacBook in my garage: AI agents as the staff, local models as the regulators, salvaged archives as the raw material. The limit has always been the size of the mind I could run locally without asking permission. That limit just moved. Personal AI is the new invention of fire. The people who own the minds on their own desks will write their own rules. A reader with a gaming PC can now staff a small company of AI agents with a model in the same weight class as systems the giants rent out by the token. This is the worst it will ever be. How To Hire It Here is the playbook, drawn from Strata's public README and install docs as of this writing. Follow those docs for the latest flags and model menu. Licenses for the model and each compressed build are separate from Strata's MIT license; read them before you rely on the system for anything important. Step 1. Check your hardware. You need a supported NVIDIA GeForce RTX 20/30/40/50-series card, or a supported AMD Radeon card, with at least 12 GB of VRAM. Confirm 32 GB of system RAM or more. Free about 80 GB on an SSD. Update your graphics driver from NVIDIA or AMD. Step 2. Get Strata. Clone or download the repository at Unzip it. On Windows, double-click START-HERE.bat. On Linux, run ./setup.sh inside the Strata folder. If you already use a coding assistant on that machine, you can paste the project's AI setup prompt and let it drive the install from docs/AI_SETUP.md. Step 3. Answer the installer's questions. It will ask which model size to download, how much context to keep, and whether you want image input. Press Enter for the recommended answers if you are unsure. The download is roughly 70 GB; if it stops, run the installer again and it continues. Step 4. Pick a size that fits your RAM. With 32 GB, start with the Coder build, which drops half the experts and is tuned for code. With 48 GB, IQ2_XS or Q2_0. With 64 GB, IQ2_XS is the author's recommended default; IQ3_S is smarter and slower. With 96 GB or more you can try the larger Unsloth builds. Smaller sizes are faster. Larger sizes keep more quality. Step 5. Wait through the first start. The first launch can make the PC slow or unresponsive for one to three minutes while Strata loads tens of gigabytes into RAM. That is normal. Do not close the window. Your browser should open the Strata app at Step 6. Point your tools at home. In any app that accepts an OpenAI-compatible provider, set the base URL to Any API key and any model name will work for the local endpoint. For Anthropic-style tools, use For Codex CLI and the OpenAI Responses API, use /v1/responses. Thinking effort can be set to off, low, medium or high; off is fastest, high is for hard questions. Step 7. Keep it on your desk. By default Strata binds to localhost. Leave it there unless you know why you need otherwise. If you ever expose it to another machine, the docs say to set an API key. For a Zero-Human Company roster, treat this local endpoint as the shared mind: one CEO agent for strategy, one engineer agent for code, one researcher for archives, one archivist for files, and one night janitor that cleans logs and memory while you sleep. Wire them through the same OpenAI-compatible address. The org chart fits on one machine the way Cray's lab fit in the Wisconsin woods. Step 8. Know the limits. Heavy compression trades some quality for speed. The largest sizes may page from the SSD and slow down if RAM is tight. You are running a third-party binary and a third-party model on your own hardware; that is the point, and it is also a responsibility. Close other memory hogs if the disk light never stops blinking. Read the troubleshooting doc before you open an issue. Watson had thousands of people and could not understand how 34 beat him. You may need one gaming PC and one evening. The Pattern Every few decades a frugal team builds something the giant said required a cathedral. In 1963 it was a supercomputer in Chippewa Falls. In 2026 it is a 125-billion-parameter mind under a glass side panel, glowing cyan and magenta on a desk that also holds a coffee mug and a slide rule of a different kind. The memo is sixty-three years old. The question is the same. Who will you hire, and on whose machine will they work?