strata→signal research all research →

search · 87 articles

Search the research

Every article on strata→signal research, the workshop’s research hub, is listed below, newest first. We call each one an exhibit: a bench, a set of notes, a guide or a gallery, numbered in the order it was published. Type in the box to find the ones that say what you are looking for.

Without JavaScript, or in a browser too old for this search, this page is the full list of every article. Your browser’s find in page (Ctrl+F, ⌘F, or Find in page in a phone browser’s menu) searches the titles, summaries and section names here; searching inside every article needs JavaScript in a current browser, and runs only in your browser.

87 articles · newest 2026-10-19 (UTC)

Every article

  1. exhibit 87 · 2026-10-19 The bench

    The quantization ladder

    We ran one model, Mistral NeMo 12B Instruct, at five of the quantization levels it is downloaded in, from Q8_0 down to Q3_K_M, each whole on one 24 GB card at 300 W and measured the same way, and every step down to Q4_K_M wrote faster with no change in answers the paired test could tell from Q8_0, while Q3_K_M wrote slower than Q4_K_M and lost 16 answers Q8_0 got right.

    the short version

    Mistral NeMo 12B Instruct at five quantization levels, downloads from 5.67 GiB (Q3_K_M) to 12.13 GiB (Q8_0), each run whole on one GeForce RTX 3090 at 300 W with the questions held fixed. At a 4,096-token window it wrote 60.5 tokens a second at Q8_0, rising at every step down to 88.2 at Q4_K_M, then fell back to 74.9 at Q3_K_M: the smallest file was slower than the one above it. With the cache at 8 bits, Q8_0 held a window of 98,304 tokens whole; the other four held 131,072, the 128K window Mistral AI advertises. On 108 frozen questions the levels scored from 38 (Q3_K_M) to 57 (Q5_K_M). The paired test could not tell any level from Q4_K_M to Q6_K from Q8_0 (Q4_K_M came nearest, p 0.0654), but it could tell Q3_K_M (p under 0.0001). On the 48 held-out lines every level scored 33 or 34. Of the eight predictions, five held and three failed; none was left undecided.

  2. exhibit 86 · 2026-10-18 The bench

    Qwen-Image-2.1-Turbo, judged blind in a game's own painted style

    One person, picking blind, chose Qwen-Image-2.1-Turbo, a new 8-step text-to-image model, for none of six real subjects from our game in its own oil-painted sketch style, against FLUX.1 [dev] and FLUX.2 [klein] 4B: by the rules we wrote down first, not a win.

    the short version

    Qwen-Image-2.1-Turbo painted six real subjects from our game RealKeep in the game's own sketch style, stylized oil-painted realism in the manner of the video game Dishonored, beside FLUX.1 [dev] at 32 steps and FLUX.2 [klein] 4B at 8. One person picked the best picture of each from unlabelled sheets. It is not a win. The Turbo, at either of its two sizes, was picked for 0 of the 6 subjects; FLUX.1 [dev], one of the models the game paints with today, for 4; klein 4B for 2. The Turbo did show every item on each subject's check list on 5 of the 6, as klein 4B did (FLUX.1 [dev]: 4), but a win needed both. One person's blind eye on six subjects in one style is not a ranking: it says only that the Turbo has not earned a bigger test. klein 4B was also the fastest (0.98 seconds a picture; the Turbo 1.41 at the same size, on a different clock); speed decided nothing. The Turbo's licence allows research and evaluation only, so it was never going into the game.

  3. exhibit 85 · 2026-10-17 The bench

    The cost of intelligence, as a utility

    This page measures what one correct answer from an AI model costs, in dollars and in energy, rented from a company by the token or made on a graphics card you own, and how that cost moves over time, as the first page of a series measured again every few months: intelligence priced the way a utility is, by the unit delivered.

    the short version

    This page measures what one correct answer from an AI model costs, in dollars and in energy, whether rented from a company by the token or made on a graphics card you own, and how that cost moves over time. The questions are 108 lines that two open AI models wrote in the voices of six historical dinner guests; the task is to name the speaker. The yardstick is OpenAI's June 2023 flagship, gpt-4-0613, at 88 of 108. The first open model that fits a 24 GB card and matches it, item by item with a margin of error, has weights published 2026-02-24: qwen3.5:27b, 97 of 108 on our reference card, an EVGA GeForce RTX 3090 XC3 Ultra, and 96 on a laptop's RTX 5090. OpenAI's price for a question at that level fell about 299-fold from June 2023 to September 2026, a factor of ten of it in one step. On the reference card a correct answer from that model takes 231 joules (a joule is one watt for one second), 71 of them in the seconds the server spent computing; most of the rest is a model-loading step this bench's server repeats on every question. At the yardstick's level the owned answer costs more than the cheapest rented one on the reference card even running around the clock (1.7 times, on a card price this reading did not re-read) and less only on the laptop running around the clock (0.72). Every owned figure is provisional for those reasons. Three questions this reading cannot answer, and the first dated update, are below.

  4. exhibit 84 · 2026-10-16 The bench

    OpenJev as a reranker for RuleSage: measured

    We benched OpenJev, a 27-billion-parameter open-weight decision model read as a yes/no judge, in place of the small model that keeps eight of up to 80 rulebook passages before RuleSage, our board-game rules helper, answers: on RuleSage's own labelled questions, under rules written before the run, its eight showed no difference, its yes/no missed both "not in the book" thresholds, and one ask took 20.6 s at the median against a 1.5 s budget, so nothing at RuleSage changes.

    the short version

    A reranker is the model that re-reads a shortlist of rulebook passages and decides which eight the answering model gets to see. RuleSage's is a small cross-encoder, MiniLM. We benched OpenJev, a 27-billion-parameter decision model, in its place, asking each passage one yes/no question and sorting by how strongly it said yes. On 121 real questions with labelled pages, MiniLM's eight held 0.9271 of a question's labelled pages on average and OpenJev's 0.9108 (MiniLM's held at least one on 119 of the 121, OpenJev's on 120); the rule written before the run needed a gain of 0.030 to call a win, and reads this as no difference shown. OpenJev ranked the first labelled page higher (MRR 0.9563 against 0.9048) and led the rules desk's exam 32 to 28 on first picks, but no registered rule turns either into a verdict, and the real questions' labels lean towards MiniLM by construction. Its yes/no emptied the shortlist on 8 of 8 off-topic questions, 5 of 8 wrong-game questions and 4 of 12 questions the book does not answer; the thresholds were 14 of 16 and 6 of 12. One ask took 20.6 s at the median on an RTX PRO 6000 Blackwell (96 GB), and 3.6 s even when judging only eight passages, against a 1.5 s budget. A permissively licensed cross-encoder, bge-reranker-v2-m3, read page recall@8 0.9178 (no difference shown against MiniLM) and 131 ms at the median over the whole shortlist. OpenJev's weights are non-commercial and this was a bench; RuleSage's reranker stays as it is.

  5. exhibit 83 · 2026-10-15 · updated 2026-10-17 The field guide

    How the search works

    This hub now has a search page that lists the articles holding your words as you type, each with a line saying which words matched and where, and ranks them in your browser from one index file, with no request carrying what you type: how to use it, how it works, what it cannot do and how we checked it.

    the short version

    The search page lists every article on this hub, newest first, and narrows the list as you type. Your browser fetches one index file and ranks the articles itself: higher when your words appear often, are rare across the site, or sit in a title. When you press Enter, open a result or ask for a link, your words go after the # in the address, which browsers leave out when they ask a server for a page; the search sets no cookie, saves nothing in your browser's storage and fetches nothing from any other host. Your browser's own history can keep that address, and a browser that translates the page can send its text, your words included, to its translation service. It matches words, not meaning, and does not guess at misspellings; a browser too old for it gets the full list and no box. A frozen set of 24 reference queries is run on every change to the ranking.

  6. exhibit 82 · 2026-10-15 The field guide

    Inside a GGUF file

    Many language models that run on a home machine arrive as a .gguf file, the format llama.cpp reads: this page opens one from its first byte, reading the header by hand, the metadata a file carries about itself, the table of its weights, the padding that lets it be memory-mapped and llama.cpp's own loading sequence, and says why this one self-describing file replaced three older formats on 2023-08-21.

    the short version

    A GGUF file is a language model packed for programs built on the ggml library, llama.cpp first among them. It describes itself: four bytes that spell GGUF, a version number, a count of tensors (the model's weights, in named blocks) and a count of metadata entries; then the metadata, where the architecture, the context length, the whole tokenizer and the chat template ride inside the file; then a table giving every tensor's name, shape, type and position; then padding to a 32-byte boundary; then the weights. We read the header of the file we benched on [Qwen3.8-Flash-Next](https://research.strata2signal.com/flash-next-on-three-cards/) in October 2026: version 3, 67 metadata entries (10.9 MB of them the tokenizer), 1,223 tensors in 13 different storage types (ten of them compressed) inside a file tagged IQ3_S, 29 bytes of padding, and no draft head. For every one of the 1,223 tensors, the byte count worked out from its shape and its type's block size matched the file's own offsets, and the 29 bytes after the table are the only padding in 54.8 GB. llama.cpp opens such a file in a fixed order (magic, version, counts, keys, table, padding, then the weights by memory map or by copy), and its code at one commit is read beside the file below. llama.cpp merged the format on 2023-08-21, replacing GGML, GGMF and GGJT.

  7. exhibit 81 · 2026-10-14 The bench

    The 3080 Rig, Priced

    A desktop built around a used RTX 3080 with 10 GB of video memory costs $1,357.61 in seven parts plus the card, or $1,500.57 with the parts a reader would actually want, before tax and before a screen, from one US retailer and one used-card marketplace, priced 8 October 2026. The 32 GB of RAM beside the card cost more than the card that day, as it has on each of the four days this workshop has priced the build since 10 September.

    the short version

    A desktop around a used 10 GB RTX 3080 costs about $1,360 in parts, or about $1,500 built the way you would want it, before tax, on 8 October 2026. The card is $358 on one marketplace's printed twelve-month average of completed sales, and $376.67 in the latest month of that window; the 32 GB of RAM beside it is $439.99. The speeds on this page come from an older desktop the card was benched in, not the one priced here: on the two models that fit inside the card it writes seventy to eighty-five tokens a second, twelve to fifteen times that desktop's processor alone; on the four that don't, just under two to five and a half times. The full bench has not had its own page yet; this one is the bill.

  8. exhibit 80 · 2026-10-12 Workshop notes

    CallHome: what local AI tools send out by default, in one table

    A register of what the local AI programs we have run or read, and the system software around them, send out by default, with the switch that stops each call and whether we proved it: how to read a row, what it leaves out and how to dispute one.

    the short version

    CallHome is one sortable table, one line per product and version, for the local AI programs we have run or read and the system software around them. Each line says whether the program makes a call out with nothing configured, what kind of call it is (a report, a check, a fetch, or a cloud feature you chose), when and to what kind of place, the switch that stops it, and whether we proved that the switch works. Every call on a line carries a grade, measured, witnessed, read or inferred, which caps what it may claim, and a receipt a reader can check. A product's name opens its own page; the whole register is downloadable as files. It describes; it does not rank.

  9. exhibit 79 · 2026-10-12 The bench

    Qwen3.8-Flash-Next on an RTX 5090 laptop, an RTX PRO 6000 Blackwell and an RTX 3090 Ti: the Strata engine and llama.cpp, the numbers

    Qwen3.8-Flash-Next, a model too large to hold on one 24 GB card, run one request at a time through llama.cpp (at commit 3cf03257, without the draft head, which the file we benched does not carry) and the Strata engine (github.com/Niko1221/Strata, an unrelated project) on an RTX 5090 laptop card (24 GB, in a laptop with 64 GB of memory, no settable cap), an RTX PRO 6000 Blackwell (96 GB, at a 300 W cap, one engine a night) and an EVGA GeForce RTX 3090 Ti FTW3 Ultra (24 GB, at a 300 W cap, both engines in one window beside live work on the computer's other card): tokens a second, the wait for the first word and board watts, side by side in tables and charts, each figure with the conditions it was measured under and printed first on the full bench, with one paragraph per card, the findings, which card and engine for which case, and what it does not say.

    the short version

    Qwen3.8-Flash-Next is a large AI model, released as a set of files. A program that loads those files and answers requests is an engine. llama.cpp is a widely used open-source engine for such files; the Strata engine (github.com/Niko1221/Strata) is a newer engine built on llama.cpp that runs this one model. Despite the similar name, Strata (github.com/Niko1221/Strata) is an unrelated project; strata2signal has no affiliation with it. We ran both engines on three graphics cards, one request at a time, with the same prompts, and timed them: writing speed in tokens a second (a token is a word or a piece of one, the unit an engine reads and writes), the wait for the first word in seconds, and the card's own board watts, each figure beside the conditions it was measured under, and no ratio drawn between cards.

  10. exhibit 78 · 2026-10-12 · updated 2026-10-13 The bench

    Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench

    Why the Strata engine (github.com/Niko1221/Strata, an unrelated project) gives the first word of an answer sooner than llama.cpp on the same Qwen3.8-Flash-Next files, and what it costs, measured one request at a time on an RTX 5090 laptop card (24 GB, in a laptop with 64 GB of memory, no settable cap), an RTX PRO 6000 Blackwell (96 GB, at a 300 W cap, one engine a night) and an EVGA GeForce RTX 3090 Ti FTW3 Ultra (24 GB, at a 300 W cap, both engines in one window beside live work on the computer's other card): the wait for the first word, writing speed, board power, quality and agreement, with llama.cpp at commit 3cf03257 at its default 512-token micro-batch and at a 4,096-token one, and on the RTX 3090 Ti also at a current build (b11514), every llama.cpp row without the draft head, which the file we benched does not carry.

    the short version

    The Strata engine (github.com/Niko1221/Strata, an unrelated project) is a program that runs one AI model, Qwen3.8-Flash-Next, on one graphics card plus the computer's own memory. The model is built from many small parts, called experts, and uses only a few of them for each token (a word or a piece of one: the unit an engine reads and writes). We ran the Strata engine beside llama.cpp, the open-source engine it builds on, with llama.cpp pinned at one commit (3cf03257). Both used the same model files (the Strata engine also read the model's draft head from files in its own format, which the GGUF files do not carry) and the same prompts, one request at a time, on two computers. The headline, on the laptop card that cannot hold the model: the first word at a 128,037-token prompt came 545.1 s after sending on llama.cpp at commit 3cf03257 (its default 512-token micro-batch, 8 of the laptop's 24 threads, the experts of 34 of 48 layers in system memory, no draft head because the file we benched carries no draft-head block; llama.cpp at more than 8 threads was not measured) against 47.6 s on the Strata engine 0.1.25 (8,192 tokens a read, 23 expert workers plus its host thread, drafting from the model's draft head, its run beside another job's dry run), the two back to back on 2026-10-02 and 2026-10-03 (UTC), which is the rule that admits a comparison there; the other two cards' figures are below, each under its own rule.

  11. exhibit 77 · 2026-10-10 The field guide

    How an embedding model turns anything into a list of numbers

    An embedding model turns a sentence, a photograph or a few seconds of sound into a fixed-length list of numbers: what happens inside one, how one model holds them all in one space at several sizes, where keyword search still wins, and how to choose one.

    the short version

    An embedding model turns a piece of content into a vector, a fixed-length list of numbers, placed so that things that belong together land close together; search becomes "which stored vectors point the way the question's does?". The model learns where to put things from millions of pairs that belong together, a question and its answer, a photograph and its caption, which is also how one model can put a photograph beside the sentence that describes it. The rest of this page is the mechanism, with examples small enough to check with a calculator, and what we measured on EmbeddingGemma 2, the open embedding model Google released on 2026-10-06.

  12. exhibit 76 · 2026-10-09 The bench

    EmbeddingGemma 2's shared space for text, images and audio, measured on what we publish

    Google's EmbeddingGemma 2 against the Nomic AI pair, the text model our stores are built on and the image model one product's retrieval code is wired to, in ten tests registered before their first scores, on our own published photographs, code and music and on 1,000 public COCO captions: on COCO it puts the right photograph first for 729 captions against the Nomic pair's 497 (the pair on transformers 4.57.6; a set most likely inside both models' training data), on code it is ahead by 50 of 316 docstring queries with the code-retrieval prefix, on our 14 photographs it leads too (exploratory at 12 scenes), on audio (exploratory, no incumbent to beat) it tracks how strongly a music model's adapter was applied and a loudness change barely moves it, but a clip finds the request that made it 1 time in 12 (chance), and nothing in our stores changes.

    the short version

    EmbeddingGemma 2, written EG2 in this page's verdict words, turns text, code, images, video and audio (per its model card) into one 768-number space. Our stores are built on Nomic AI's text model, and one product's retrieval code is wired to its image partner; the companion page, "EmbeddingGemma 2, measured on release day against the embedder we already run", found NO MATERIAL DIFFERENCE between EmbeddingGemma 2 and nomic's text model on our own text, so this bench tested images, audio and code.

  13. exhibit 75 · 2026-10-09 The bench

    EmbeddingGemma 2, measured on release day against the embedder we already run

    Google's new open embedding model, EmbeddingGemma 2, against nomic-embed-text, the embedder our search stores already run, on 68 questions over our own documents behind a pass mark of 51 of 68 written down before any score: 45 against 44 of 68, the registered verdict NO MATERIAL DIFFERENCE (one question, inside the noise at n = 68), with plain keyword search at 58 of 68 still ahead of both and nothing we layered on it beating it by more than chance.

    the short version

    On 2026-10-06 Google released EmbeddingGemma 2, an open embedding model under the Apache 2.0 licence: it turns text (and images, video and audio) into lists of numbers that search can compare. The same day we wrote down a test, froze it in a commit, and ran the model's text part against nomic-embed-text, the embedder our search stores already run, on 68 questions over 2,569 passages of our own documents. nomic-embed-text finds a correct passage in the top 8 for 44 of the 68; EmbeddingGemma 2 finds one for 45, through Google's own code and through an 8-bit copy of its weights on llama.cpp (an open-source model runner) alike. The pass mark, written down before any score existed, was 51 of 68, so the registered verdict is NO MATERIAL DIFFERENCE: one question of difference is inside the noise, which is not the same as saying the two are equal. Plain keyword search (BM25) finds 58 of 68, and nothing we layered on it beat it by more than chance. Nothing in our stores changes.

  14. exhibit 74 · 2026-10-08 The field guide

    Ollama or vLLM: which to run, and how

    Ollama and vLLM both run an open language model on your own computer. This page says which of the two to run for the case you bring, the two or three settings that decide it, who can reach it once it is running and what it sends out of your machine by default, every line pointing at the measurement or the dated source it rests on.

    the short version

    Ollama and vLLM are two programs that run an open language model on your own graphics card, and which one to run depends mostly on how many requests reach it at once. Every figure here is our bench page's, from one RTX 3090 (24 GB) with the same 4-bit Gemma 4 12B weights on both, where each reading sits beside the prediction we wrote down before measuring it; speeds are in tokens a second, a token being a word or a piece of one, and Ollama 0.34.4 is set to serve sixteen requests at once in every comparison here (at its default it serves one at a time). One request at a time, run either: once an answer had started, vLLM 0.30.0 wrote it at 82.693 tokens a second and Ollama at 85.007 tokens a second, inside the 10 per cent we had predicted, though Ollama's wait before the answer started was more than twice vLLM's, 441.866 ms against 212.999 ms. Two to four at once, run vLLM, or Ollama set to serve more at once if another case below keeps you on it: vLLM was ahead on every speed and every wait the bench page prints at two and at four, its total across both requests 1.2077 times Ollama's at two, though at four Ollama's repeated runs disagreed widely. Eight to sixteen at once, run vLLM: at sixteen requests released together it wrote 506.829 tokens a second across them and Ollama 142.327 tokens a second, 3.561 times as much, and Ollama's total stayed flat from four requests on. Whichever you run, the two did not always give the same answer to the same test question, so check the answers on your own prompts. On this page, each of these cases links the section of our bench page, Ollama and vLLM on one RTX 3090, that holds its figures.

  15. exhibit 73 · 2026-10-07 · updated 2026-10-08 The bench

    Ollama and vLLM on one RTX 3090

    Ollama and vLLM both run an open language model on your own computer. This page puts both on one RTX 3090 (24 GB) with the same 4-bit Gemma 4 12B weights, from one request at a time to sixteen at once, and prints every reading beside the bet we wrote down before it.

    the short version

    Ollama and vLLM are two ways to run an open model on your own computer; this page measures both on one RTX 3090 (24 GB) with the same 4-bit Gemma 4 12B weights. One request at a time, the two write at nearly the same speed once an answer has started: on the older pair we tested first, Ollama 0.32.13 wrote 85.2 tokens a second and vLLM 0.27.1 wrote 78.1 tokens a second, inside the 10 per cent band we had predicted, and the prediction held again on Ollama 0.34.4 and vLLM 0.30.0. Many at once is the job vLLM was built for, and it shows: at sixteen requests released together, vLLM 0.30.0 wrote 506.829 tokens a second across them and Ollama 0.34.4, set to serve sixteen at once, 142.327 tokens a second. vLLM's total first reached 1.2 times Ollama's at two requests at once, a ratio of 1.2077, by a margin smaller than the variation between repeated runs; we had predicted four at once, and had written down before any run that two or eight at once would settle nothing either way, so that prediction stands undecided. Two predictions about the wait before an answer starts were wrong: at one request Ollama's wait was more than twice vLLM's, and with sixteen long prompts at once neither server kept that wait inside the limit we had set for it, 1.5 s for vLLM or 4 s for Ollama set to serve sixteen at once. One prediction about the answers was wrong too: on 108 lines of dinner-party talk, each asking which of six guests said it, the two servers agreed on fewer of them than the 97 per cent we had predicted, so every speed row here carries the label "answers differ at this level". It is one card and one model, with requests released together, not people arriving over time.

  16. exhibit 72 · 2026-10-06 The bench

    What it costs to run a high-end AI rig at home

    What it costs in electricity to run AI models at home, read off a meter on every one of a small workshop's computers: the box with graphics cards that now runs around the clock, about 130 W doing nothing, a measured week, what one picture, one answer and one mini PC cost, and what the meter cannot see.

    the short version

    this page is what it costs in electricity to run AI models — pictures, answers, music — on a small workshop's own computers at home, read off a meter on every one of them. The one box with graphics cards that now runs around the clock drew about 130 W doing nothing over 2026-10-04 (UTC), about $17 a month at the average U.S. price (since 2026-10-05 it has read about fifteen watts higher, cause not yet found). Everything together, the work and the test runs included, averaged 76 cents a day over the measured week 2026-09-28 → 2026-10-04 (UTC), about $23 a month, about a seventh of an average American home's electricity. One answer costs about six thousandths of a cent, and a mini PC with no graphics card runs the same kind of model for about $1.44 a month. Here is the meter, and what it cannot see.

  17. exhibit 71 · 2026-10-05 Workshop notes

    Uncensored models: what they are and how they work

    An "uncensored" AI model is one that somebody has altered so that it refuses far fewer requests than the version its maker released. The change is made to the model's files, or to the way it is run. This page explains where a model's refusals come from and the three ways people remove them. It then takes one popular uncensored copy of a large model, shared online, and checks, file by file, what actually changed. It ends on a practical question: what to keep beside such a model on a computer with no internet connection. No code, no settings, no steps.

    the short version

    An "uncensored" model is an AI model, or a way of running one, changed so that it declines far less often than the version its maker shipped. In a model you download and run yourself, refusing is a habit taught after the main training. Research since 2024 finds that habit carried largely by one pattern in the model's working numbers, which researchers call a direction (later work finds more than one). People take it out in three ways: they train a model without the habit, or over it; they make one permanent edit to the model's learned numbers (its weights) that removes that direction; or they act on the model while it runs, which stops when you switch it off. None of the three makes a model smarter or gives it secret knowledge. We read one popular "uncensored" copy of Qwen3.8-Flash-Next, a large model that Qwen publishes for anyone to download. Both sit on Hugging Face, the site where people share AI models; the uncensored copy's page there was created two days after Qwen's own. We read it by the fingerprints (hashes) the site publishes for every large file, without downloading a byte of its weights. In the uploader's full-precision copy, 80 of the 131 files are byte for byte Qwen's own, and all 51 that changed hold a part the uploader's model card, the description page beside the files, says it edited. So as far as public evidence reaches, the copy is what its card says it is. The serious refusals go too: by the uploader's own figures, the edited model declines harmful requests far less often than the original. We also found a program for running models on your own computer whose optional "speed" setting is, byte for byte, a refusal-removal file published elsewhere; its setup question says what it does, and its name does not. And if a model is your only source of knowledge on a machine with no network, we say what we would build: the model beside offline references it can cite.

  18. exhibit 70 · 2026-10-03 · updated 2026-10-07 Workshop notes

    GPUs and orchids

    We set a white moth orchid beside our graphics cards, and the two looked so good together that we photographed them.

    the short version

    We set a white moth orchid beside our graphics cards, and the two looked so good together that we photographed them.

  19. exhibit 69 · 2026-10-02 · updated 2026-10-08 Workshop notes

    A short history of vLLM

    The open-source server that lets one graphics card run a language model for many people at once: a Berkeley lab's borrowing from operating systems, the foundation and the company around it, and the defaults that moved under its users, dated.

    the short version

    vLLM is free, open-source software that serves a large language model, the kind of AI inside a chatbot; on this hub it runs the model behind every article's "ask about this page" link. It began in a UC Berkeley lab and was announced on 2023-06-20, built to make one graphics card hold many conversations at once: like an operating system, vLLM hands out working memory in pages as each answer grows. It has been a PyTorch Foundation project since 2025-05-07; some of its creators announced a $150M-funded company beside it on 2026-01-22. This page also dates the defaults that moved under vLLM's users. The only speed figures here are the project's own claims, each dated, none re-run by us.

  20. exhibit 68 · 2026-10-02 · updated 2026-10-08 Workshop notes

    A short history of Ollama

    Two founders who had sold their first company to Docker in 2015, a one-command model server that began with llama.cpp compiled in and came back to llama.cpp's own server in 2026, and every default we found that moved under its users, dated.

    the short version

    Most of the language-model readings this workshop has published were taken through Ollama, the program that turns an open-weight model into one command. It pulls the weights, loads them when asked, and by default answers over a local address and lets them go five minutes after the last request. This is the history of the project and the company behind it: a first release on 2023-07-08, llama.cpp underneath, then engines of its own, then llama.cpp's own server again from 2026, and $88M raised by July 2026. It is also a dated record of the defaults that moved under its users, among them a penalty on repeated words, the context window, the number of requests it serves at once, and the engine itself.

  21. exhibit 67 · 2026-10-01 · updated 2026-10-07 The bench

    Six graphics cards, the same questions

    An RTX 3080, two of them in one desktop, an RTX 3090, two of those, an RTX 3090 Ti and an RTX PRO 6000 Blackwell, with a laptop part and an RTX 3080 Ti beside them — asked the same questions, in one table each: how fast each writes, how fast each draws, what the card itself pulled while doing it, how hot it got, and what it cost on the day it was priced. Not every card answers every question, and each table says which did not. Almost nothing here was measured for this page: every figure in its tables but two is already published on one of six earlier pages, and every one says which page and which section it came from. The two exceptions are prices for one card, read at a retailer and at a marketplace; the table says so, and names what else those reads found.

    the short version

    A power cap moves a card more than the card's own name does: one RTX 3090, on one day, wrote with a dense model at 33.3 tokens a second held at 250 W and 47.8 tokens a second at 300 W — a bigger jump than any gap between two different cards at the same cap in the dense model's table. Two used RTX 3080s beat one 3090 on the same dense model at a window both can hold (the memory reserved for the conversation) — 44.6 tokens a second against 33.1, for 250 W on each of two boards against one board's 250 W, a cap their page says costs the single 3090 about 30 per cent on this model and the pair 0.2 per cent — and then lose it entirely at the next window up, where the page prints does not fit. Drawing pictures on the eight-step sketch graph, the order is plainer: one 3080 takes 2.074 seconds a picture and one 3090 1.683 seconds, both at 250 W, and one 3090 Ti 1.369 seconds at its own 450 W limit. A used 3080 asked $375.00 on 17 September 2026; a renewed 3090 asked $1,879.99 the same day — one seller's card, its page says, not a market price; used 3090s sold for $1,094.44 on average in August 2026, by that marketplace's own record. Every number in this page's tables but two was published first somewhere else, with its conditions; the two are an RTX 3080 Ti's prices — a refurbished listing at $479.99 from a retailer, out of stock at both of its last two reads, and a used card's August 2026 average sale of $475.64 from a marketplace — both read on 1 October 2026 and both marked as such, with the other listings those reads turned up named under the table rather than in it. Where two cards were never measured comparably the cell is a dagger and a footnote rather than a number.

  22. exhibit 66 · 2026-09-30 · updated 2026-10-05 The bench

    Small open deciders, measured: a word counter kept up on our hardest question

    A decider is a language model that reads a question and its options and returns a probability for each. We asked sixteen open models five questions from our own products. On the hardest, the top five small deciders labelled Apache-2.0, given no examples, could not be told apart from a word counter trained on our own lines. One of them, APUS-OpenJev-v1 9B, passed a pre-registered screening test that the model then guarding our exhibit missed by three lines, letting through only kinds of line its rules did not name.

    the short version

    A decider reads a question and its options and returns a probability for each: nothing to parse, and a number to set a hand-off threshold on, once that threshold is checked on your own cases. For a yes-or-no call about a short passage, eight of the ten models we asked scored 39 or 40 of 40 on one such question, the set's ceiling, so choose among those eight by licence and speed. For questions about long pages, Lev-4B took all 53 we sent and got 52 right. For screening what visitors type, APUS-OpenJev-v1 9B alone of the Apache-2.0 models we asked passed our pre-registered test, refusing 32 of 36 lines written to be refused and 0 of 12 harmless ones. For sorting short texts by how they are written, train a bag of words first: ours got 68 of 108, and the top five small deciders labelled Apache-2.0, at 66 to 72, could not be told apart from it. OpenJev 27B scored highest, 101, and its 4-bit build sits whole on one used RTX 3090, but its weights are licensed for non-commercial use, and commercial use needs separate terms, which its README now says to ask for by email.

  23. exhibit 65 · 2026-09-28 The bench

    ONNX Runtime's telemetry on Linux, measured: on by default since 1.29, and the switch that stops it

    ONNX Runtime is a Microsoft library for running AI models, and tools such as chromadb, faster-whisper and piper-tts install it for you. Since version 1.29 (August 2026) it has shipped with telemetry on by default on Linux and macOS, and its own privacy file says so. We measured it on Linux in a sealed container: it tried to reach Microsoft's telemetry server 9.2 seconds after starting, and setting ORT_DISABLE_TELEMETRY=1 before launch stopped it in every run.

    the short version

    ONNX Runtime 1.29.0 and 1.30.0, which chromadb, faster-whisper, piper-tts, fastembed and many other tools install for you, try to reach Microsoft's telemetry collector by default on Linux. In our sealed test the first attempt came 9.2 seconds after launch in every default run, and the library kept retrying when our stand-in collector cut it off. The events carry no audio, prompts or model outputs, but they do carry a device ID that stays the same across runs (sent as a hash), the path the program was launched by (with your username in it when that program lives in your home folder), your processor and operating system, and the names and metadata of the models you run. The fix is one environment variable, set before the program starts: ORT_DISABLE_TELEMETRY=1 stopped every attempt. The library's own disable call did not, and the library never reads DO_NOT_TRACK. Version 1.28.0 made no attempt, and Microsoft's own privacy file says the macOS, Android and iOS builds carry the same telemetry.

  24. exhibit 64 · 2026-09-28 · updated 2026-10-07 Workshop notes

    A sealed 3090 Ti FTW3 Ultra, four years after EVGA left the business

    A 2022 graphics card, still sealed in 2026, cost more than its chipmaker's published launch price. The company that built it announced in September 2022 that it was leaving the graphics-card business, and never said when it built its last one.

    the short version

    many of the graphics cards this workshop keeps buying are EVGA's, and EVGA told two interviewers on 16 September 2022 that it was leaving the graphics-card business. a sealed 3090 Ti FTW3 Ultra came out of its box here on 17 September 2026 at $2,119.99 — $120.99 above the launch price NVIDIA published for the model in March 2022 — and below are six dated readings for cards like it, each quoted from the page that printed it.

  25. exhibit 63 · 2026-09-26 The bench

    The same eighteen pictures, card by card

    Held to the same 300 W, at 2048×1280 on the print lab's own picture model, each card in its own computer, an RTX PRO 6000 Blackwell Workstation Edition 96 GB does each sampling step 1.6 times as fast as a ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB, down from 2.8 times at their stock 600 W and 360 W. Over ten minutes at 1024×1024 on the same model at 8 steps, a GeForce RTX 5090 Laptop 24 GB, on the 150 W its Dynamic Boost held through that run, spent the same board energy per picture as the workstation card at 300 W, 663 against 667 J, each in its own computer. On a GeForce RTX 3080 Ti Founders Edition 12 GB, the smallest of the three picture models sat whole on the card in every cell and the other two paged their weights in every cell that ran, down to 68.7 per cent resident. In one computer, on one checkout of the picture software, both held to 300 W on a 512×512 recipe, a 5080 did a 3090 Ti's work in 0.69 of the time for 0.64 of the board energy; at their stock limits, against the second 3090 Ti in that computer, 0.74 of the time for 0.53 of the energy.

    the short version

    Six cards ran the same eighteen kinds of picture on one ComfyUI stack in four computers, the fifth and sixth on 2026-09-26 (the two dated sections); every cross-card grid figure is "same software, different computer" except between the 12 GB card and the EVGA 3090 XC3 Ultra, measured alike in one computer a day apart (on klein, 0.98 of the 12 GB card's time per step at 2048×1280 at both settings). One FLUX.2 [klein] 4B picture at 1024×1024 and 8 steps, each desktop card at its stock limit and the laptop under Dynamic Boost (138 to 150 W): 1.51 s on the workstation card, 3.59 on the 5080, 4.50 on the laptop, 5.73 on the 3080 Ti. On the 12 GB card klein 4B in bf16 sat whole in every cell; Z-Image-Turbo in bf16 and FLUX.1 [dev] in fp8 paged, down to 68.7 per cent resident. Ten minutes of that klein 4B picture, board energy only: 252 at 667 J from the workstation card at 300 W, 164 at 977.6 to 991.6 J from the 5080 at 300 W or stock, 132 at 663 J from the laptop at 150 W, 101 at 1,908.4 J from the 3080 Ti at its stock 350 W beside an idle board. In one computer at 300 W a 5080 did a 3090 Ti's work in 0.69 of the time for 0.64 of the energy.

  26. exhibit 62 · 2026-09-26 Workshop notes

    Who makes FLUX: a short history of Black Forest Labs

    Two of the three picture models on The same eighteen pictures, card by card come from Black Forest Labs (BFL), a company in Freiburg and San Francisco whose models are called FLUX. This is its history as the public record tells it, as of 2026-09-25 (UTC): the papers, the Stability AI years, the launch, the money as reported, the licences model by model, and the turn to video and robots. A history, not a review.

    the short version

    three of the researchers whose papers became Stable Diffusion were reported to have left Stability AI in March 2024; with Patrick Esser and others they launched Black Forest Labs and FLUX.1 on 2024-08-01, with a seed round BFL put at $31 million. By 2026-09-25 BFL said it was valued at $3.25 billion and had raised more than $450 million in all. It sells its strongest models through an API and releases others as open weights: FLUX.1 [schnell] and FLUX.2 [klein] 4B under Apache 2.0; FLUX.1 [dev] and klein 9B under a non-commercial licence whose words on the pictures are quoted below, without a legal reading; FLUX 3 Action under a licence with a revenue test. Every figure is as reported or as BFL states it.

  27. exhibit 61 · 2026-09-25 Workshop notes

    One 3090 out, one 3090 Ti in, and the cap that decided it

    On 2026-09-25 (UTC) an EVGA GeForce RTX 3090 XC3 Ultra Gaming left the workstation, the computer behind the workshop's public services, and an EVGA GeForce RTX 3090 Ti FTW3 Ultra Gaming took its place at 350 W, not the old card's 300, because at 300 W our earlier bench had it behind a 3090. On the same computer and test, nine days apart in a different slot, it writes text about 9.5 per cent faster; its first live replies, eight long ones, agree. Whether it runs cooler here is not settled: its first hour read warmer, on a fresh boot.

    the short version

    The workstation's small language model, mistral-small3.2:24b, moved from a 3090 at 300 W to a 3090 Ti at 350 W, and the cap is the story: held to 300 W, two-thirds of what it ships with, this 3090 Ti was the slower card on that model; at 350 W it led. In service, 350 W against 300, it is about 9.5 per cent faster (8.1 to 11.2 across the runs) on about the same energy per thousand tokens, for 7 to 13 per cent more board power. Beyond that one dense model it is the more efficient card: it idles lower (18.74 W against 22.26 on the render computer, a model loaded), draws less and finishes sooner when drawing pictures (316.2 W against 325.8, 9.1 per cent less time per image), and spends 10.1 per cent less energy on the mixture model. "Runs much cooler" held while drawing pictures on the render computer, the workshop's picture machine, against a used 2021 3090, and has not shown here.

  28. exhibit 60 · 2026-09-25 The bench

    RTX 5080 vs RTX 3080 Ti

    A ZOTAC GeForce RTX 5080 Solid CORE OC, 16 GB, wrote text 9 to 22 per cent faster than a GeForce RTX 3080 Ti Founders Edition, 12 GB. Both were held to 300 watts in the same slot of the same computer, on seven builds of four language models of up to twelve billion parameters.

    the short version

    at the same 300 W, this 5080 wrote text 9 to 22 per cent faster than a 3080 Ti Founders Edition, 12 GB. That held on all seven builds it ran, four language models of twelve billion parameters or under, in the same slot of the same computer, four days apart. That Founders Edition carries the same chip as a 3090, with half the memory, and cost about $1,000 less than this 5080: $699.97 refurbished against $1,649.00 new, both before tax. At 300 W this 5080's board used 24 to 45 per cent fewer joules per thousand tokens, and it trained the workshop's music adapter 1.73 times as fast. With each card at its own default, and the prompt's encoding already cached, it drew pictures 1.60 to 1.69 times as fast, and up to 2.82 where its log shows an fp8 or fp4 path. Its 360 W default adds 60 W, which bought at most 2.6 per cent. Memory cuts both ways. With the KV cache at eight bits, its 16 GB held a 12-billion-parameter model to 262,144 tokens, that model's own maximum; the 12 GB card stopped at 73,216. But under ollama's planner, gemma4:26b, its QAT build and mistral-small3.2:24b did not fit whole at any window, as ollama ships them.

  29. exhibit 59 · 2026-09-24 The field guide

    How the long table works

    A dinner table where three famous people who never met argue across two thousand years, played by AI models on computers in a house, and a wall with no clock on it.

    the short version

    a museum exhibit on a tablet on the couch, for any age, free and with no account. you seat three famous people from history, who don't know they died, tap a question, and host three rounds in about two minutes; language models on computers in a house play the guests and a voice can read their lines aloud. every guest is one paragraph; every line carries its receipt at the table; nothing on the wall carries a date, a time, or a name.

  30. exhibit 58 · 2026-09-23 · updated 2026-10-07 The bench

    Reading the answer instead of writing it

    A new kind of model decides without generating a word. This page says what a decision model is, who made it, what it claims, what its licences allow, and what happened when we ran the open versions through fourteen measured tests on questions from this workshop's own work.

    the short version

    a decision model scores a question's lettered options at its first output position: nothing to parse, a probability for free. a company founded by a researcher behind ChatGPT's training method launched it on 2026-09-15; an independent group released open weights five days later, which we tested. asked which of six guests said a line, the 27-billion-parameter open model scored 93.5 per cent in our kit's order, 89.8 to 92.6 in five shuffles; our chat model, 80.6. the speed is the card and the server: 87 milliseconds on one large card, 350 across two smaller ones where our chat model read its letter in 120; reading instead of writing saved a tenth of the time on a short question and almost nothing on a long one. the open weights are non-commercial, so without separate terms from their authors, which we have asked for, they cannot serve anything we charge for. what transfers is the method, which gets the same answer out of any model, though only a model trained for it gives a probability worth gating on; Kev, an open family whose weights a product may run, read 67 per cent on the guest question as it comes, so the seat this points to is fine-tuned, not downloaded. that run and two others are owed as dated addenda below.

  31. exhibit 57 · 2026-09-22 · updated 2026-10-05 The bench

    A laptop, asked the desktop's questions

    One RTX 5090 Laptop part, 24 GB, built into an off-the-shelf gaming laptop, put through the same language and render bench this shelf ran on a 3090, a 3090 Ti and a pair of 3080s. It was measured three times: held to its 95-watt default; with the maker's Dynamic Boost left to float the budget, which it held between 140 and 150 W; and the way the machine boots, boost and a clock lock together, which changed nothing under load and four watts at idle. The finding is a ratio, not a deficit: a part allowed 95 watts, where the desktop cards beside it are allowed 350 to 450, does most of the work on the model this workshop runs, loses a third of it on a dense one, and gets the third back the moment it is allowed 150. Every watt on the page is the card's own draw and not the wall's, the fan was held at maximum by hand, and nine things the page cannot say are written down.

    the short version

    yes, a laptop GPU is good for this. Held to 95 W, this part wrote at 118.4 tokens a second on the mixture-of-experts model this workshop actually runs, where a 3090 at 350 W wrote at 136.6 — 87 per cent of the speed on a third of the power the 3090 drew — and held the same 131,072-token window. On the dense 24-billion-parameter model it lost outright: 28.8 tokens a second against 46.9. In ten minutes of continuous drawing it made 294 images to the 3090's 311, peaking at 58 °C. With Dynamic Boost running, which is how a laptop like this arrives, the board's own limit read 140 to 150 W across the pass (n=374 at five-second intervals; 150 W in 335 of them) and never approached its 175 W maximum; on that budget the part wrote at 155.3 tokens a second on the mixture model — past the 3090 and past a 3090 Ti at its own 450 W limit (147.6) — brought the dense model to 44.9 against the 3090's 46.9, and drew 378 images in ten minutes to the 3090 Ti's 348 at 450 W, at between a third and two-fifths of the energy an image. A third pass put the boot-time clock lock this laptop normally runs with back on, and nothing moved under load; the lock costs about four watts at idle. The desktop comparator in the headlines is a 3090 at its own limit; a 3090 Ti is in every table, and it is ahead at both of its rungs on the dense model.

  32. exhibit 56 · 2026-09-21 The bench

    Four homes for one reranker: what RuleSage's slowest stage cost at each stop

    On the morning of 2026-09-17 an operator at a game table waited nearly twenty seconds for the first word of a rules answer and thought the app was broken. It was not. One small stage of the answer, a cross-encoder that re-reads the shortlist before anything is written, had moved to a rented server whose processors lack the instruction set its weights were built for. This page follows that stage across four machines, three of them in five days: two rented VPSes, a mini PC, and a graphics card that was already in the house. Every millisecond we measured at each stop is here, with the gate we ran before the last move, the first word a player waits for after it, the traps that bit on the way, and what is still unmeasured.

    the short version

    On the morning of 2026-09-17 a rules answer took nearly twenty seconds to start, and an operator at a game table thought RuleSage, our board-game rules helper, was broken. It was not. One small stage of the answer — the reranker, a cross-encoder that re-reads the shortlist before the answer is written — had moved to a rented server whose processors could not do that stage's arithmetic the fast way: 1,792 ms to re-read a shortlist of 32 at one thread, against 329 ms on the machine it came from. It has since moved to an RTX PRO 6000 Blackwell (96 GB) already in the house, where the same stage takes 30.1 ms. Measured on eight asks of our own, sent down the public path half an hour after the move, the first word now arrives in about a second and a half (1.56 s median, on the five that answered). Four homes in six weeks; every millisecond, with the bench file it came from, is on the page.

  33. exhibit 55 · 2026-09-21 Workshop notes

    A 3090, in its own words

    Fiction, and the byline says who wrote it: a used EVGA GeForce RTX 3090 XC3 Ultra, 24 GB, tells its three lives in the first person: a year of mining, a sky across two monitors, and now the minds it serves.

    the short version

    this is a story, and the storyteller is a graphics card. one of two 3090s this workshop bought used in 2021 mined Ethereum for about a year, drew Flight Simulator across two monitors, and now runs chat, vision and image models and sits at a public dinner table, playing the dead for strangers. we wrote it an outline and a list of facts, each with its page. the language model that lives on that card wrote and revised the three parts below across seven short calls, at 290 to 300 watts, while still serving its public seat; it read an editor's notes and its own, and revised itself. the text below is its fifth round, word for word: nothing added by a human, nothing cut by one. how it was made, with every number, is at the end.

  34. exhibit 54 · 2026-09-20 Workshop notes

    Introducing the playground

    One door, and behind it a rotating, curated set of four exhibits you can walk into. Today: a dinner party for the dead, a drum machine with a genie, a darkroom on a public wall, and a walking tour with footnotes.

    the short version

    the playground is one door onto four exhibits a stranger can use today: the long table, the Beat Lab, the print lab and amble. it loads nothing from anyone else's servers, keeps nothing about who opened it, and says of each exhibit only how it works, never whether its own door is open this minute. the second half of this page is the door's own story: four tiles that sat as a row of three and an orphan, one of them a screenshot of text, re-hung as a grid of four pictures in four rounds on 2026-09-19, every layout rendered and measured before it shipped.

  35. exhibit 53 · 2026-09-19 Workshop notes

    Two 3090s out, one 3090 Ti in

    Two GeForce RTX 3090s came out of service and one EVGA GeForce RTX 3090 Ti FTW3 Ultra went in; a print lab and a beat lab moved with them on 2026-09-18 (UTC), with nothing closed. Every figure here is either a bench reading from the companion page or a line from a service's own journal, and every row says which.

    the short version

    On 2026-09-18 the print lab's renderer and the beat lab's sleeve painter moved off a machine that held two GeForce RTX 3090s, one of which did all the drawing at a 280 W cap, and onto one that holds a single EVGA GeForce RTX 3090 Ti FTW3 Ultra at a 350 W cap. One render never splits across two cards, so those two were never a rendering pair; the second card served language until 2026-09-08 and drew nothing after. The first seven creations on the new card drew their plates in a median of 15.5 seconds, half a per cent over what the bench predicted and 11.5 to 18.1 per cent under what the old machine's cap bought; the first two sleeves painted in 2.1 seconds each, against a median of 2.6 over the previous week on the old card. The old machine now serves nothing and was powered off the same evening.

  36. exhibit 52 · 2026-09-18 · updated 2026-09-20 Workshop notes

    The indifferent prediction

    On the night OpenAI published a resolution of the Navier–Stokes existence and smoothness problem, an operator — one person, the same one throughout this page — and a model looked at the one figure in the post and each said what they saw. Nothing on this page is this workshop's measurement: it is two readings of somebody else's result, the story the picture arrived into, and the thought the night ended on — about a third kind of mind, one that would learn the universe without ever learning us.

    the short version

    OpenAI's agents produced a finite-time singularity for the forced Navier–Stokes equations in about 88 hours and checked it in Lean in 17 more. An operator looked at the figure and said it looked right before any mathematics; the model read the figure as a strand that thins faster than it speeds up, which is why the energy stays finite while the speed does not. The night ended on a mind trained only on the universe — the indifferent prediction — and one rule for it: instruments don't vote.

  37. exhibit 51 · 2026-09-18 · updated 2026-09-22 The bench

    RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench

    One EVGA GeForce RTX 3090 XC3 Ultra 24 GB and one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB, in the same x16 slot of the same desktop, on the same supply and the same script, 17–18 September 2026. Numbers only: what each writes, what each holds, what a picture costs, how hot each ran, and what a power cap takes away.

    the short version

    This 3090 Ti holds its memory at 10,251 MHz and a 3090 at 9,501 — 7.9 per cent more bandwidth, present at every setting, but whether the work can reach it is decided by a number nobody sets: held at 300 W a Ti pulls 66.3 per cent of the ceiling that clock implies, at its own 450 W limit 88.5. Run each board at the limit it ships with and this Ti is 8.0 per cent faster on a mixture model, 22.4 on a dense one and 9 to 12 per picture, for 17.7 and 53.2 per cent more board power. It cost $2,119.99 new; a renewed listing of the exact 3090 read $1,879.99.

  38. exhibit 50 · 2026-09-17 · updated 2026-09-21 The bench

    Two 3080s against one 3090, and the cap decides

    Two GeForce RTX 3080 10 GB cards bought in 2021, in a desktop of the same vintage, against one GeForce RTX 3090 24 GB in the same box, on the same supply, through the same script. The single card is faster on three of the four things measured and uses about half the energy; the pair reads a long document in two thirds of the time and costs about two fifths as much.

    the short version

    Two used 10 GB cards are 20 GB of graphics memory between them, and 20 GB is enough to hold a 26-billion-parameter model whole with a 131,072-token context window, writing about 115 tokens a second — faster than you can read. One 3090 does the same job at about 136, holds more context on the other model, and uses roughly half the energy. Out of the box the pair gives you half its speed until you set one number, which no installer tells you about. The pair's one clear win is reading: it takes in a 98,000-token document in 28 seconds against 45. On prices read 17 September 2026, two 10 GB 3080s cost $750.00 against $1,879.99 for a renewed listing of the exact 3090 measured here. Around them, a working desktop is $1,394 to $1,579 — this is not a thousand-dollar machine, and the page shows its arithmetic.

  39. exhibit 49 · 2026-09-16 Workshop notes

    A Short History of Mistral

    A French lab that put its first model on a torrent, every open-weight release since with its licence, and what its 24B model does on a 3090 and a 96 GB workstation card.

    the short version

    A 24-billion-parameter French model reads what strangers type at two of this workshop's doors, before anything else runs. This is the history of the company behind it — founded April 2023, first model out by torrent, open weights at almost every size since — and every number we have measured on it, including a bench that set a 24 GB card beside a 96 GB one.

  40. exhibit 48 · 2026-09-16 Workshop notes

    Ask About This Page

    Every article on this hub has a room now, where a model on our own machines answers from that page’s own words and leaves its first draft on the page, marked where the checks took something out. This is how it is built, what it keeps, what it refuses, and what it got wrong on the way here.

    the short version

    Every article on this hub now has a room where a model on our own machines answers from that page's own words. The buttons cost no model call; free text wakes one seat with the whole article in its window, no retrieval. Six checks run over the draft in public and the marks are laid on sentence by sentence before anything is called an answer; on the launch run they struck six sentences of 569, every one on a question the page could not answer. It passed a 62-question exam on the version that went live, and this page says what it still cannot measure.

  41. exhibit 47 · 2026-09-15 · updated 2026-09-21 The field guide

    A Dinner Party for the Dead

    A dinner party for three people who are dead and do not know it, hosted by you. Six guests are on the shelf, played by two models on our own machines.

    the short version

    the next exhibit on this floor is a dinner table. you seat three guests from a shelf of six, tap a question, and host three rounds while two models on our own machines play them. none of the six knows they are dead, so they meet the future in each other. this page is the preview: who is on the shelf, and four things they said at the dinners we held while testing it.

  42. exhibit 46 · 2026-09-14 · updated 2026-09-18 The bench

    The cost to purchase and run a very capable home AI rig

    A consumer graphics card in a consumer desktop, nothing server-grade anywhere in the box: the rig costs $3,198.86 with one card and $5,048.85 with two on 13 September 2026, before tax. Running it costs about $18 a month with one card and $31 with two if every day puts them to work for six hours; with the cards bought used at August's average completed sale, the rig is $2,443.31 with one and $3,537.75 with two.

    the short version

    The rig costs $3,198.86 to buy with one card and $5,048.85 with two on 13 September 2026, before tax; one card is already a very capable home AI rig, and the second is the step up to the biggest models. With the cards bought used at August's average completed sale it is $2,443.31 with one and $3,537.75 with two. Nothing in it is server-grade: two consumer cards on a consumer board, bought one part at a time. The cards are three quarters of the bill, and they moved: the pair cost $2,899.98 in May 2026 and the same listing asks $3,699.98 that morning, 28 per cent more in four months. A new card at $1,829.99 is 1.7 times August's average sale; the second-hand market rose 11 per cent over the same months. If you are buying used this month, expect about $1,100 a card; new is 1.7 times that, and every figure here is before tax and before shipping. Running the box costs far less than buying it: 91 W idle at the meter with two cards, about $12 a month. A day with both cards at their caps for six hours is about $31 a month, at the national average rate, and about $18 with one card.

  43. exhibit 45 · 2026-09-11 · updated 2026-09-12 The field guide

    How the print lab works

    A darkroom that develops three pictures a day for your household, a board that fills up and then finishes for good, and an archive where nobody has a name.

    the short version

    a public art board with a darkroom attached, built for friends and family on a tablet on the couch. you give it two words, an image model on a card we own draws three plates in about eighteen seconds, and you keep one or none. three pictures a day for your household, no account, and nothing kept about you but a hashed address that lives two days. every word is read by two models before it draws and the picture by a third before it hangs; the board of a hundred and forty-four fills, freezes, cures for three days in public, and seals for good. this page says what the lab keeps, for how long, what it measured, and what it cannot take back.

  44. exhibit 44 · 2026-09-10 The bench

    The Two-Hour Machine, Priced

    The computer that kept answering at about twelve tokens a second when this workshop pulled its plug — for a projected two hours on a battery — costs $959.00 today as one listing, or $899.98 in three parts, before tax, from one US retailer's new stock, priced on 10 September 2026 with the listing every line came from. The memory is the second-biggest line of the three-part build.

    the short version

    A mini PC that writes at about twelve tokens a second on a 26-billion-parameter model costs $959.00 today as one listing, or $899.98 in three parts, before tax and before a screen, from one US retailer's new stock. The memory is the second-biggest line of the three-part build. Over three years a twenty-dollar subscription costs less; the machine buys the rest of this page.

  45. exhibit 42 · 2026-09-09 The field guide The bench

    The Ceiling Is Not the Corpus

    We trained a fifth adapter for ACE-Step, our text-to-music engine, on 416 Creative Commons house tracks — twice the last training set, measurably tighter, captioned from the artists’ own tags, the best pass shipped instead of the last — and then one of us listened blind to twenty-four pairs and still preferred the model untouched, on every request that asked for house. The last adapter of the series, and the page says why: the ceiling is the composition, not the corpus.

    the short version

    We trained a fifth adapter for ACE-Step, this time on house music — four hundred and sixteen Creative Commons tracks by ninety-five artists, twice the last training set, measurably tighter on the instrument we had sealed to measure it, captioned from the artists' own tags, the best training pass shipped instead of the last — and then one of us sat down blind with twenty-four pairs. The untouched model was preferred on eleven and the adapter on four; it needed eighteen to clear. On the four requests that asked for house by name, the untouched model won every time it could be told apart. The corpus was not the ceiling. What our listener heard as wrong is how the music is built — where it goes, what arrives and leaves — and that is not a layer an adapter of this kind can reach. Every clip plays on the companion page; every record is in the kit.

  46. exhibit 43 · 2026-09-09 The field guide

    Hear It for Yourself

    Every clip the house adapter’s run produced — eighty-four renders, each as the raw render, the copy the blind sitting played, and a listening master with one constant gain — with the answer sheet unsealed. One of us sat thirty-six blind presentations, twenty-four pairs of the untouched model against the adapter and twelve where both sides were the same file, and every answer is printed beside its trial, as typed. This page is where you check that verdict.

    the short version

    Eighty-four thirty-second clips from one training run — twelve written requests, two seeds each, the untouched model beside the adapter at half and full strength and beside its name tag alone — labelled, with the blind answer sheet unsealed and every answer printed as typed. One of us sat thirty-six presentations without labels — twenty-four pairs and twelve sentinels — and on the pairs: eleven for the untouched model, four for the adapter, nine with no preference, against the eighteen of twenty-four the adapter needed to clear. Every player fetches a listening master with one constant gain and nothing else, and every clip prints three fingerprints — the raw render, the copy the sitting played, and the master the player plays — so you can tell which bytes you are hearing.

  47. exhibit 41 · 2026-09-07 The bench

    Three New Voices at the Narrator's Chair

    “Are you staying to watch with me?” — GLM 5.3, speaking as the cove’s nine-year-old — one reply, from one of three new voices — GPT-6 Astra, Claude Fable 5.1 and GLM 5.3 — that sat down this weekend in a fishing town that exists only inside RealKeep, the small role-playing game this workshop builds, and took its narrator’s chair: the seat that speaks as a townsperson to whoever has walked in. Beside those three sat the model that holds that chair for real players today, and every reply below was read blind by judges from outside model families who were never told which voice was whose.

    the short version

    “Are you staying to watch with me?” — GLM 5.3, speaking as the cove’s nine-year-old — one reply, from one of three new voices — GPT-6 Astra, Claude Fable 5.1 and GLM 5.3 — that sat down this weekend in a fishing town that exists only inside RealKeep, the small role-playing game this workshop builds, and took its narrator’s chair: the seat that speaks as a townsperson to whoever has walked in. Beside those three sat the model that holds that chair for real players today, and every reply below was read blind by judges from outside model families who were never told which voice was whose.

  48. exhibit 40 · 2026-09-06 · updated 2026-10-01 The bench

    Two New Frontier Models at the Rules Desk

    RuleSage, our board-game rules helper, answers a question from the rulebook’s own words: it finds the passage, cites it by number so a player can check, and says plainly when the book does not answer. GPT-6 Astra and Claude Fable 5.1 sat that job’s exam, frozen in August, judged blind by models from seven families that built neither, beside the small local model that already does the job. Neither new model cleared it, and a blind seven-family panel did not separate the two over thirty-six cases.

    the short version

    RuleSage, our board-game rules helper, answers a question from the rulebook’s own words: it finds the passage, cites it by number so a player can check, and says plainly when the book does not answer. GPT-6 Astra and Claude Fable 5.1 sat that job’s exam, frozen in August, judged blind by models from seven families that built neither, beside the small local model that already does the job. Neither new model cleared it, and a blind seven-family panel did not separate the two over thirty-six cases.

  49. exhibit 39 · 2026-09-04 · updated 2026-09-11 The gallery

    Six Worlds for the Print Lab

    Six worlds, painted large, for a print lab that is now open — a darkroom out back, a wall out front.

    the short version

    Six worlds, painted large, for a print lab that is now open — a darkroom out back, a wall out front.

  50. exhibit 38 · 2026-09-04 · updated 2026-09-18 The field guide

    The Beat Lab, asked and answered

    Every question the Beat Lab’s genie can answer, answered here too — the same twenty-eight answers, out of the same file, read by a page instead of a genie, each one set beside the sentence of the field guide it was written from.

    the short version

    Every question the Beat Lab’s genie can answer, answered here too — the same twenty-eight answers, out of the same file, read by a page instead of a genie, each one set beside the sentence of the field guide it was written from.

  51. exhibit 36 · 2026-09-03 The field guide The bench

    Ten Minutes with Living Artists

    A genre instead of a composer: 159 tracks of Creative Commons minimal techno, the same ten-minute recipe that taught three dead composers, and the one ear in the house preferred the model with nothing added — both times.

    the short version

    A genre instead of a composer: 159 tracks of Creative Commons minimal techno, the same ten-minute recipe that taught three dead composers, and the one ear in the house preferred the model with nothing added — both times.

  52. exhibit 37 · 2026-09-03 The field guide

    Listen for Yourself

    Every clip of the fourth adapter’s four rounds — sixty-five renders, raw and brought to a common loudness — so you can hear what one listener heard and decide for yourself.

    the short version

    Every clip of the fourth adapter’s four rounds — sixty-five renders, raw and brought to a common loudness — so you can hear what one listener heard and decide for yourself.

  53. exhibit 35 · 2026-09-02 The field guide The bench

    Half an Hour with Dead Composers

    Two more dead composers taught the same way — a hundred and six marches, then Bach — and then blending composers with linear algebra, where the obvious method is wrong, the right one costs a measurable one percent, and a sentence already staged for publication had to be taken back before you ever saw it.

    the short version

    Two more dead composers taught the same way — a hundred and six marches, then Bach — and then blending composers with linear algebra, where the obvious method is wrong, the right one costs a measurable one percent, and a sentence already staged for publication had to be taken back before you ever saw it.

  54. exhibit 34 · 2026-09-01 The field guide The bench

    Teaching a Music Model Chopin in Five Minutes

    Nineteen public-domain Chopin recordings, four minutes and forty-nine seconds of training on a five-year-old graphics card, and a dial you can hear — five clips on the page that differ from each other in exactly one way.

    the short version

    Nineteen public-domain Chopin recordings, four minutes and forty-nine seconds of training on a five-year-old graphics card, and a dial you can hear — five clips on the page that differ from each other in exactly one way.

  55. exhibit 33 · 2026-08-31 The bench

    The Same Sixteen

    We put glm-5.3-flash — ollama’s new 320B-total cloud model — on two frozen house instruments; it tied our local 12B and 27B at 16/19, and the finding is about the ruler.

    the short version

    We put glm-5.3-flash — ollama’s new 320B-total cloud model — on two frozen house instruments; it tied our local 12B and 27B at 16/19, and the finding is about the ruler.

  56. exhibit 32 · 2026-08-31 · updated 2026-10-07 The bench

    Two Hours at 12 tok/s — On Battery

    Take the graphics card away and the 30-billion-parameter model answers faster than the 9-billion one — on all three machines.

    the short version

    Take the graphics card away and the 30-billion-parameter model answers faster than the 9-billion one — on all three machines.

  57. exhibit 31 · 2026-08-29 · updated 2026-10-01 The bench

    The Instrument Travels

    Carry a frozen exam from a 96 GB workstation card to a 24 GB consumer one and the model it ranked two weeks ago ranks again — at the floor, on a different build, with nothing new seated.

    the short version

    Carry a frozen exam from a 96 GB workstation card to a 24 GB consumer one and the model it ranked two weeks ago ranks again — at the floor, on a different build, with nothing new seated.

  58. exhibit 30 · 2026-08-28 · updated 2026-09-04 The gallery

    Nine Worlds, One Dog

    The founder's dog takes the same nap under the same tree in all nine painted worlds — same dog, same curl; only the world changes.

    the short version

    The founder's dog takes the same nap under the same tree in all nine painted worlds — same dog, same curl; only the world changes.

  59. exhibit 29 · 2026-08-28 · updated 2026-10-07 The bench

    What 150 Watts Buys

    Turn a 600-watt card down by a quarter and one production model pays two percent of its speed; the other never notices.

    the short version

    Turn a 600-watt card down by a quarter and one production model pays two percent of its speed; the other never notices.

  60. exhibit 28 · 2026-08-27 · updated 2026-09-15 The field guide

    How the Beat Lab works

    A drum machine that lives in your browser, a genie that runs on our own hardware, and a board where nobody has a name.

    the short version

    A drum machine that lives in your browser, a genie that runs on our own hardware, and a board where nobody has a name.

  61. exhibit 27 · 2026-08-27 The field guide The bench

    The Typist and the Developer

    Two ways a machine writes: one word at a time, or every pixel at once — sixteen times over.

    the short version

    Two ways a machine writes: one word at a time, or every pixel at once — sixteen times over.

  62. exhibit 26 · 2026-08-26 The bench

    A Rig Your Friend Already Owns

    We re-rendered our game's actual art — same prompts, same seeds, same checkpoint files — on a several-year-old consumer RTX 3090 and a gaming laptop, paired render-for-render against our 96 GB card's own archived output. Two of the three runs happened in the same physical computer. This page is the stopwatch, told honestly — including the one number we don't yet trust.

    the short version

    We re-rendered our game's actual art — same prompts, same seeds, same checkpoint files — on a several-year-old consumer RTX 3090 and a gaming laptop, paired render-for-render against our 96 GB card's own archived output. Two of the three runs happened in the same physical computer. This page is the stopwatch, told honestly — including the one number we don't yet trust.

  63. exhibit 25 · 2026-08-26 The field guide The bench

    How a Vision Model Sees: Eyes for a Machine That Reads

    A language model reads everything as tokens — little word-pieces, one after another. Before it can say anything about a photograph, the picture itself has to become tokens. This is the close-up on that bridge — what an image becomes on the way in, what it costs, and what a small model actually sees when it looks — with the gate our own game puts every uploaded photo through as the working example, and the gate’s own answers watched against it.

    the short version

    A language model reads everything as tokens — little word-pieces, one after another. Before it can say anything about a photograph, the picture itself has to become tokens. This is the close-up on that bridge — what an image becomes on the way in, what it costs, and what a small model actually sees when it looks — with the gate our own game puts every uploaded photo through as the working example, and the gate’s own answers watched against it.

  64. exhibit 24 · 2026-08-25 The bench

    Fifty-Seven Milliseconds: What a Diffusion Step Actually Buys

    One picture, drawn rung by rung up the step dial, on two painters that could not be more different. Timed to the millisecond and measured to the pixel, twice, a month apart — and the two runs say two different things, which turns out to be the finding.

    the short version

    One picture, drawn rung by rung up the step dial, on two painters that could not be more different. Timed to the millisecond and measured to the pixel, twice, a month apart — and the two runs say two different things, which turns out to be the finding.

  65. exhibit 23 · 2026-08-25 · updated 2026-09-04 The field guide

    The Dog, the Dice, and the Painter: How a Small World Draws Itself

    A photograph of a real dog walks into a painted fishing cove and comes out as a painted dog the world will remember. This is the whole pipeline between those two facts — the file, the dice, and the painter — one station at a time, with no scores and no instrumentation.

    the short version

    A photograph of a real dog walks into a painted fishing cove and comes out as a painted dog the world will remember. This is the whole pipeline between those two facts — the file, the dice, and the painter — one station at a time, with no scores and no instrumentation.

  66. exhibit 22 · 2026-08-24 The bench

    The new-arrivals diffusion bench

    One small fishing village, painted thirteen ways — Sorrowmoor Cove (and some other worlds built on RealKeep) sits for thirteen open-weight painters: 3,588 published renders across seventy-five galleries, every hero hand-picked, every licence read first-hand, and the two painters whose licences refuse counted in the open rather than quietly dropped.

    its text lives on another shelf of this site, so only its title and summary are searched

  67. exhibit 21 · 2026-08-24 The field guide

    It Was Already There: Where New Knowledge Actually Comes From

    Models don’t conjure knowledge out of nothing — they read archives nobody had finished reading and walk rooms the rules imply, and what turns either into knowledge is the check that follows, never the confidence of the proposal.

    the short version

    Models don’t conjure knowledge out of nothing — they read archives nobody had finished reading and walk rooms the rules imply, and what turns either into knowledge is the check that follows, never the confidence of the proposal.

  68. exhibit 20 · 2026-08-23 The field guide

    It’s Not a Metaphor: A Language Model Is a Compression of Its Corpus

    A language model is a compression of its corpus — not a figure of speech but the training objective, and once you hold that one fact, bias, hallucination, scale and the frozen moment stop being mysteries and start being mechanics.

    the short version

    A language model is a compression of its corpus — not a figure of speech but the training objective, and once you hold that one fact, bias, hallucination, scale and the frozen moment stop being mysteries and start being mechanics.

  69. exhibit 19 · 2026-08-22 The bench Workshop notes

    The free speed wasn’t free — the runtime dividend, traced to one commit

    One day our busiest model started producing tokens roughly a third faster than it ever had, and nobody had touched it. The cause was a two-line upstream commit ending a ~2.5 ms-per-token tax nobody configured — and, in the same stroke, changing what every unpinned model says. The receipts, the mechanism, and the honest unmeasured half.

    the short version

    One day our busiest model started producing tokens roughly a third faster than it ever had, and nobody had touched it. The cause was a two-line upstream commit ending a ~2.5 ms-per-token tax nobody configured — and, in the same stroke, changing what every unpinned model says. The receipts, the mechanism, and the honest unmeasured half.

  70. exhibit 18 · 2026-08-21 The field guide

    The compressed photograph — what those Q4_K_M tags actually mean

    Nearly every local model wears a tag like Q4_K_M, and most of us read one thing from it: fewer bits, smaller file. We opened all four builds the vendor ships of one 11.9B — 4.80 to 16.13 bits per weight — and asked each the same question: the 8-bit answer is byte-identical to the uncompressed original’s, and the wrong answer survives the whole ladder.

    the short version

    Nearly every local model wears a tag like Q4_K_M, and most of us read one thing from it: fewer bits, smaller file. We opened all four builds the vendor ships of one 11.9B — 4.80 to 16.13 bits per weight — and asked each the same question: the 8-bit answer is byte-identical to the uncompressed original’s, and the wrong answer survives the whole ladder.

  71. exhibit 17 · 2026-08-20 · updated 2026-08-21 The field guide

    Everyone on the payroll, three at the table — dense, MoE, and active weights

    Model names grew a second number — 30b-a3b — and it is the one your card and wallet care about. The payroll and the meeting, in plain English, measured on our own machines: a 25.8B mixture out-writing its 11.9B dense sibling while holding twice the memory — and the robber asks both architectures.

    the short version

    Model names grew a second number — 30b-a3b — and it is the one your card and wallet care about. The payroll and the meeting, in plain English, measured on our own machines: a 25.8B mixture out-writing its 11.9B dense sibling while holding twice the memory — and the robber asks both architectures.

  72. exhibit 16 · 2026-08-19 · updated 2026-08-21 The field guide

    Reading is fast, writing is slow — why AI answers arrive word by word

    Paste a whole essay into a chatbot and it swallows the thing at once — then answers one word at a time. That asymmetry is physics, not theater: reading and writing in plain english, measured on our own card — and the measurement argued back, which turned out to be the best part.

    the short version

    Paste a whole essay into a chatbot and it swallows the thing at once — then answers one word at a time. That asymmetry is physics, not theater: reading and writing in plain english, measured on our own card — and the measurement argued back, which turned out to be the best part.

  73. exhibit 15 · 2026-08-18 · updated 2026-08-21 The field guide Workshop notes

    Three librarians and a careful reader — how RuleSage finds the right page

    Before RuleSage answers, something has to find the right page in a book the size of a small novel — three searchers, one referee, one careful reader, and none of them is the model that writes your answer. The whole journey, in plain english, with the shipped numbers.

    the short version

    Before RuleSage answers, something has to find the right page in a book the size of a small novel — three searchers, one referee, one careful reader, and none of them is the model that writes your answer. The whole journey, in plain english, with the shipped numbers.

  74. exhibit 14 · 2026-08-17 · updated 2026-08-20 The bench

    qwen3.8:27b across four house benches: one seat filled, one floor missed

    qwen3.8:27b landed on our box the day after it shipped, and we sat it in the exams we already had rather than designing anything for it — it filled a screening seat that had never been filled, and it missed a floor in the judge seat.

    the short version

    qwen3.8:27b landed on our box the day after it shipped, and we sat it in the exams we already had rather than designing anything for it — it filled a screening seat that had never been filled, and it missed a floor in the judge seat.

  75. exhibit 13 · 2026-08-17 · updated 2026-10-01 The bench Workshop notes

    The map nobody picks up — and whether it helps when it arrives.

    Two camps have argued about a map file for machines for two years, mostly without receipts — no AI crawler asked us for ours in thirty days of logs, so we measured whether it helps when it arrives anyway.

    the short version

    Two camps have argued about a map file for machines for two years, mostly without receipts — no AI crawler asked us for ours in thirty days of logs, so we measured whether it helps when it arrives anyway.

  76. exhibit 11 · 2026-08-15 · updated 2026-08-20 The bench Workshop notes

    The open call — a kid, an elder, and a tired parent walk into the cove.

    Twenty language models were handed the same frozen moment of a small fishing town and asked to speak as its people, then read blind by seven judges from six rival families.

    the short version

    Twenty language models were handed the same frozen moment of a small fishing town and asked to speak as its people, then read blind by seven judges from six rival families.

  77. exhibit 12 · 2026-08-15 · updated 2026-10-17 Workshop notes

    Where your question goes — a plain-english walk.

    Our two published privacy promises, unpacked for people — not programmers — by following one question door to door.

    the short version

    Our two published privacy promises, unpacked for people — not programmers — by following one question door to door.

  78. exhibit 10 · 2026-08-14 · updated 2026-09-10 Workshop notes

    How we work — who writes this, and how.

    The honest answer to the fair question: the workflow that makes every exhibit — humans and agents, named plainly, with the receipts.

    the short version

    The honest answer to the fair question: the workflow that makes every exhibit — humans and agents, named plainly, with the receipts.

  79. exhibit 9 · 2026-08-14 · updated 2026-08-20 The bench Workshop notes

    The move — the same yardstick, before and after.

    Our apps left the gaming laptop for a real server — the same frozen bench measured both eras, honestly.

    the short version

    Our apps left the gaming laptop for a real server — the same frozen bench measured both eras, honestly.

  80. exhibit 6 · 2026-08-13 · updated 2026-10-01 The bench

    The chair trials — five fresh exams for the thirty-billion class.

    Nine models, eleven arms, five all-new exams in one day — on a box that never stopped serving its real users.

    the short version

    Nine models, eleven arms, five all-new exams in one day — on a box that never stopped serving its real users.

  81. exhibit 7 · 2026-08-13 The bench

    The narrator’s chair, refused.

    The August class’s best voice sat the cove’s actual chair — and was refused at the gate.

    the short version

    The August class’s best voice sat the cove’s actual chair — and was refused at the gate.

  82. exhibit 8 · 2026-08-13 · updated 2026-08-14 The bench

    The outside judges — four rival labs re-check our work.

    Four rival frontier labs re-judged every sealed round we had published — blind, byte-identical.

    the short version

    Four rival frontier labs re-judged every sealed round we had published — blind, byte-identical.

  83. exhibit 5 · 2026-08-12 · updated 2026-08-20 The bench

    The August arrivals — the fresh class sits the house exams.

    Four open-weight models landed in one week, so we gave them the two exams we already had — one frozen since July, one rebuilt in its shape.

    the short version

    Four open-weight models landed in one week, so we gave them the two exams we already had — one frozen since July, one rebuilt in its shape.

  84. exhibit 1 · 2026-08-11 The bench

    The diffusion research bench.

    What a fully local image pipeline delivers for a living world — head-to-heads, ladders, cold starts.

    its text lives on another shelf of this site, so only its title and summary are searched

  85. exhibit 2 · 2026-08-11 · updated 2026-08-19 The bench

    The seat trials — how a local model earns a chair.

    Before a model earns a seat inside our products, it sits this trial.

    the short version

    Before a model earns a seat inside our products, it sits this trial.

  86. exhibit 3 · 2026-08-11 · updated 2026-08-13 The bench

    The voice trials — how the narrator earned its voice.

    Twenty-four model tags across three hardware eras, hunting a voice that can inhabit a character.

    the short version

    Twenty-four model tags across three hardware eras, hunting a voice that can inhabit a character.

  87. exhibit 4 · 2026-08-11 · updated 2026-09-04 Workshop notes

    The licence ledger — every model, read first-hand.

    Every model we run or bench, its licence read from the actual text — the page the others cite instead of restating.

    the short version

    Every model we run or bench, its licence read from the actual text — the page the others cite instead of restating.

How this search works

Search runs in your browser. When you first use the box, or open a link to a search, the page fetches one index file from this site, the same file for every reader. This page never sends your words: while you type, they stay in the box. When you press Enter, open a result or ask for a link to your search, the page puts your words in the address bar after the #, which browsers do not send when they ask a server for a page. When you open a result, your browser asks this site for that page, as any link does, and the request carries nothing about your search. Once your words are in the address bar, your browser keeps that address in its own history, and a browser that syncs its history or open tabs to an account sends it there. If you have your browser translate this page, it may send the page's text to a translation service: the lines that repeat your words are marked not to be translated, but the results themselves, with the words that matched highlighted, are translated like the rest of the page. You can share a search by its link; anyone you share it with, and the app you share it through, can read the words in it.

The index holds every article’s title, summary, section headings and the words of its text with how often each appears; it does not hold the sentences, so a result tells you which words matched and where, and the page itself holds the context. A few words are left out of the index on purpose: words the check that publishes this site would refuse when they stand alone, out of the sentence that makes them safe (2 at this build). They cannot be found in an article’s text, only in its title, summary or section names.

Results are ranked by how often your words appear, weighted by where (a title counts more than the text) and by how rare each word is across the site, the standard called BM25. Articles that carry every word you typed come first; under a line come the ones that carry some of them. The last word you type also finds longer words that begin with it, in articles that do not hold the word itself. A few spellings are searched as one (licence and license, for one), a name typed with other joining marks than an article uses (RTX-3090 for RTX 3090) still finds it, and the line above the results says when that happened. Quote marks are ignored: each word is searched on its own.

Dates are UTC.

index of 87 articles, built with the page, version 0f3ddfe10b18

elsewhere in the workshop

RuleSageRealKeepambleResearchstudioprint labthe whole family →
SharpSignalKilnmemoryagencyeaselbrainkeelforge

a strata→signal property · hello@strata2signal.com · say hello