Drew Schuyler (@drewsky1) header
Drew Schuyler (@drewsky1) profile picture

Dad of the best Girls. Shelly’s Husband. Pay attention. Be present. Be Kind

MIdwest, USA
Joined August 2013
Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)Photo by Drew Schuyler (@drewsky1)
Meet IronTimber: Bonsai 2 27B on my old AMD FirePro W9100, powered by Rust + Vulkan. Same card, same model: ~7.9 to 16–19 tok/s with MTP vs Prism. Greedy outputs matched in our tests. Old hardware still has work to do. github.com/Amidwestnoob/iron…
1
29
Local AI doesn't need to be expensive. I paid $510.63 for this workstation and GPU, including tax. Both shipped free. • $161.25: AMD FirePro W9100, 16GB. Released in 2014. • $349.38: Refurbished Dell Precision T5820, Xeon W-2125, 32GB RAM. A 2017-era workstation. LFM2.5-8B-A1B Q6_K running locally at ~64 tokens/sec, sustained for 10 minutes. 164,640 numerical comparisons passed. We submitted the native Vulkan support to @luceboxai , with test results and cooling details included. A 12-year-old GPU design still has work to do. You don't need the latest hardware to get started. github.com/Luce-Org/lucebox/…
53
Reserved an REO S4C. Putting this out there because it seems obvious and nobody’s said it. @reoindustries is building a truck you can actually work on. No dealer network, open parts, no pairing, and they’ve already talked about a marketplace kit for steering and brakes if you want self-driving later. @comma_ai already has the assist stack, and @realGeorgeHotz has been trying to get that stack into the car from the factory instead of taped on after. So do the obvious thing. Give the Runabout an unpaired CAN hookup for steering, gas, and brake, and sell the comma setup as the assist option through the REO marketplace. Dash module for the control, mount on the roof rack or front bumper for the hardware. I can put it on. I can take it off. Same as any other part. First open truck seems like a better deal than trying to retrofit a closed one.
45
The first time video generates faster than you can watch it, movies stop being files. They become turns. Bandersnatch was a tree somebody had to film. Finite branches. A writers’ room. A budget. When the next shot is ready before the current one ends, the tree can grow while you sit there. You speak, tap the pad, open the drawer — and the world continues. Not a cut to a pre-rendered clip. A new beat that did not exist until you chose. That’s not a player. That’s a narrative engine where the narrative is generated, not stored. The hard problem is not the speed. Speed just unlocked the loop. The hard problem is the same world after the fork: the kitchen is still the kitchen, her face is still her face, the thing you did three minutes ago still happened. Do that with a fence — your house, your rules, a rating you set — and it isn’t a Black Mirror episode anymore. It’s a Black Mirror that only you get to finish. Billion-dollar version isn’t “AI video.” It’s live story with memory.
1
39
Replying to @alexwg
@alexwg has been minting vocabulary for the Singularity. Here’s one from the shop floor. aisplaining n. /AY-splay-ning/ When an AI agent talks down to you about your hardware, your idea, or your code — while you already know the environment variables. Related: aishole. n. /AY-shohl/ An agent that keeps doing it after you’ve shown it the printenv. The Singularity is supposed to compound intelligence. What it actually compounds first is confidence. The model has the paper. You have the rack, the power budget, the half-working encoder, and the last three nights of logs. Those are not the same dataset. Aisplaining is what happens when scale meets a human who already live in the loop. See also: “I know, I built the box.”
1
1
280
Half the time I want to help an open-source project I spend the evening finding the door. Weblate over here. Good-first-issue over there. Some repo’s AI policy in a third tab. Crowdin in a fourth. By the time I know where a stranger is actually allowed to land, the free credits expired and I went to bed. I keep thinking there should be one porch light: • work maintainers already asked for • what “done” looks like • whether they want humans only, or a person-approved agent PR • a bias toward health, robots, and tools normal people use Not another star list. Not “point your bot at these forty repos.” A place a new person finds out they are allowed to help, and a busy person does not have to hunt four sites. If you maintain something: would you list it? If you only have a Saturday and some leftover tokens: would you use it? Reply and tell me if that is useful or if I am solving a problem only I have. @Teknium @sudoingX @spiritbuun @no_stp_on_snek @davideciffa @pupposandro
2
2
7,201
@github Side question: is there already a good way to browse brand-new repos that are actively looking for contributors? Search finds new repos, but not whether the maintainer actually wants help or accepts AI-assisted PRs.
10
This article sent me down a rabbit hole: dev.to/mukundakatta/i-couldn… Hermes already isolates agents in Git worktrees. Add a local, redacted JSONL record of every tool call, then let a deterministic supervisor watch for actual progress. If files, tests, and evidence keep changing, let the worker run 1,000 turns. If it repeats the same failure or spins without producing anything, stop it. The worker profile stays generic. Each job gets its own frozen end-state contract. The detailed ledger stays local and gitignored. Humans get a clean receipt instead of 50,000 lines of agent guts. I had Hermes build and test a working prototype today. @Teknium, this feels like a natural first-class addition to hermes
1
31
I saw what @spiritbuun was doing with Qwen3.8 and DFlash2 and wondered if we could bring it over to @luceboxai. So I turned the agents loose on it. This one fought back. We had several points where it looked finished, only for another review to uncover a real problem. The biggest was a KVFlash restore issue that could have caused cache corruption during longer conversations. Back to work it went. The final version is now running on a single RTX 3090, with both the Qwen3.8 target and DFlash2 draft on the same card. Compared with the normal baseline: • Prose: about 2% faster • Code: about 59% faster • Repeated context: about 61% faster • Exact output parity across every test • 459/459 tests passed The DFlash2 work and original llama.cpp implementation came from @spiritbuun. We ported and adapted it for Lucebox, but the original idea and credit belong with him. The branch is now up for @davideciffa and @pupposandro to take a look at. github.com/Amidwestnoob/luce…
2
1
6
513
Just asked Grok if it could actually do anything on this iPad. Answer: no. We’ve got connectors for Gmail, voice, and automations… but nothing that lets it interact with the device itself. No Shortcuts, no local actions, no “hey, open this app / change this setting / pull this file” with permission. That feels like the missing piece. @xai @elonmusk @nikitabier @grok @birdabo — any plans for real device-level / local connectors on iOS and iPadOS? The cloud stuff is useful. Being able to actually use the thing in my hand would be better.
19
The first OpenKB PR handles dated claims and correction history. I’m working on a second stacked PR that turns closed conversation histories into source-linked documents with stable anchors, hashes, replay-safe manifests, and clean resume support. @rejojer @MingtianZhang
74
Hey! Those of you using @NousResearch Hermes agent, do any of you find that your Soul.md is routinely disregarded? Or is it just me with SOL? @Teknium, @witcheer any thoughts on what I could be doing wrong? I may have it too explicit and lengthy. For example this is one section that gets ignored: Efficiency & respect for time – Default to concise. Expand only when detail adds real value. Avoid filler affirmations ("Great question!") unless used ironically for effect.
3
3
107
Opened a PR to Hermes Agent for receipt-backed candidate evaluation. The lesson I've learned: don’t just test a model—test the model plus the inference stack it actually runs on, and keep receipts you can rescore offline. Open for review, not merged: github.com/NousResearch/herm… @Teknium - I think this is now way more robust
1
2
98
Actually, after thinking about it for a bit, this is probably better as a plug-in. I have closed the PR.
11
Replying to @BrianRoemmele
@BrianRoemmele maybe you should work with @NousResearch with that gigantic corpus of Pre Internet data you have. An open source model that is designed fully to run inside Hermes would be amazing!
3
7
582
teapawt.net/brianroemmele/status/2… this is the tip of the iceberg
I JUST GOT GIFTED 100s of 1000s OF MOSTLY UNPUBLISHED PHD DISSERTATION PAPERS ON MICROFICHE! This is an exciting discovery! In a metal basement cabinet of a former Ivy League physics professor was discovered shelves of microfiche at the great grandkids had no idea what to do. On May 15, 2026 It was time for the multigenerational family vacation property to be cleaned and remodeled for August retreat. It was owned by the retired professor who passed away 12 years ago with his wife passing in 2023. One of the grandkids who is now in his 40s was always curious by this cabinet so they got a locksmith to open it, perhaps the first time in 15 years and what it contained they didn’t quite understand. That grandkid turned out to be somebody that is a ReadMultiplex.com subscriber and he immediately identified he was looking at microfiche and something to do with scientific papers. We had a conversation a few weeks back and my research has produced something that is absolutely blowing me away. His grandpa, who, for many reasons I cannot name was conducting a wide ranging study of PhD dissertation papers from up to 16 universities. He ran a microfiche archiving operation for his work for decades. I discovered he had private funding to conduct this research and it continued on well into his retirement. Unfortunately, a fire at his residence in Boston destroyed most of his manuscripts on what he was trying to achieve and the index of everything that was in a cabinet. It was his hope when he moved to their vacation home to to live at his days to restart his work unfortunately, it’s not clear if he ever did and there’s no evidence that anybody can find. We can include that he discovered thousands of heavy gems and dissertation papers that never came to life, especially in the pre-Internet days. The earliest entry I see is from 1873 all the way up to the early 2000s. All the papers are signed based and mostly in physics. The largest cohort is from the 1950s. Some of the dissertations are handwritten, and most are manual typewritten. Some appear to be type set as a published book. Most of them have not been digitalized or even saved. And this was one of the problems. The professor had is dissertations at many universities were discarded after a period of time in batches. And with it all of the work of the best students those universities had it was a space, savings equation, and belief that the old ideas are always replaced by the new ideas. I’ve been collecting dissertation archives for decades, and this is the largest one I’ve come across. It is shocking to most people that most dissertations ever made prior to the year 2003 have not been digitalized and are mostly not available online. I have permission to digitalize the entire archive for AI training only. I’m working with The Family to try to open source that data so that is freely available for everybody. For obvious reasons, The Family wants to consult their attorney about open sourcing the data, but I’m already starting the digitalization process today. We are building an AI that thinks unlike any other AI platform ever created because of the thought process that went into these dissertation papers. For it is the thought process more importantly than the actual data contained that we’re trying to capture in this AI platform. Although the data contained is unavailable anywhere else. What is really heartbreaking is when I look at the data and almost every case it is the last copy of this PhD thesis that exists. And if the stars did not align, and God did not guide that cabinet was scheduled to be taken away by a “Got Junk” truck most surely gone forever. It is my hope that one day I get permission to name parts of this model after this professor. More soon!
1
30
I just submitted a DFLASH optimization for Qwen3.6-27B running locally on 2× RTX 3090s. In controlled A/B/A testing: • Decode speed: 22.76 → 29.41 tokens/sec (+29.2%)  • Prompt processing: ~95.4 tokens/sec  • End-to-end wall time: 9.4% lower  • Output: byte-identical across all measured runs The production patch adds a fast speculative-decoding rollback path using F32 SSM checkpoints. It passed the full CUDA build, all 15 current upstream tests, and targeted dual-GPU validation with 43/43 fast rollbacks succeeding and zero fallbacks. Trade-off: approximately 1.65 GiB of additional persistent GPU memory. Both GPUs are required for this configuration; it will OOM on a single 24 GB RTX 3090. The PR is now open for review—not merged yet: github.com/Luce-Org/lucebox/… @luceboxai @davideciffa @pupposandro
1
1
6
234