0 Why this page exists
Knots ties named detectors, and programs, inside a language model and runs them on a GPU operating system we wrote ourselves. This page argues that this is a bad idea, as hard as the evidence allows. Each argument carries its source: a paper, or a path in this repository where we measured it. A con without a source is an opinion and is left out.
We wrote it because the project's own report (experiments/knots/REPORT.md) asks for blunt disagreement, and because the strongest version of the case against is more useful to us than a weak one. Nothing below is a strawman. Where a con has a fair partial answer, it says so in one line, and the answers are collected in §15.
80arguments against
42references
+0.18what a random direction's label gains (the trained regions gain +0.32)
4.77our held-out loss; a trained GPT-2 small reaches about 3.3
The case in one paragraph. The literature shows, again and again, that named directions, neuron labels and dictionary features look meaningful even when they are not: on random networks, on random directions, and under explanation scores that do not track causal effect. Our own random-direction control reproduces that result: a random direction earns over half the label gain of a real region. The objects Knots ties are k-means centroids on a single plane, cruder than the sparse autoencoders a major lab has deprioritised. On AxBench, simple prompting beats every representation-steering method. Our pipes are absorbed by training, our emits break the model at the doses where they are clear, and "an operating system at each knot" is still unbuilt. All of it runs on an undertrained model, a custom compiler and an operating system that has cost us three pods and repeated stalls. An outside reviewer's one-line summary, which we accepted: "a native GPU runtime for named activation detectors, conditional interventions, and reproducible experiments." On this reading, that is all it is.
1 Reading is not explaining
Knots' central claim is that a place in the model can carry its own name. The interpretability literature has spent a decade showing how easy it is to name a place wrongly.
1Neurons and directions look like clean concepts when they are not
Bolukbasi et al. found that single neurons in BERT, and linear combinations of them, appear to encode one simple concept on one dataset and something different on another. The illusion was reproducible, and it came from the geometry of the embedding space and the narrowness of the corpus. A knot is a linear direction labelled from one corpus (FineWeb), so it falls squarely inside this illusion.
LITERATURE Bolukbasi et al. 2021, An Interpretability Illusion for BERT
2Explanation scores do not measure causal role
Huang, Geiger, D'Oosterlinck, Wu and Potts audited GPT-4-written neuron explanations for GPT-2 XL. Even the most confidently explained neurons had high error rates, and they found no evidence that the neurons are causal mediators of the concepts their explanations name. Knots' labels are scored the same way: an agent predicts firings from a label. A high F1 says a label predicts where a direction is large. It says nothing about whether the model uses that direction for that concept.
LITERATURE Huang et al. 2023, Rigorously Assessing Natural Language Explanations of Neurons (BlackboxNLP 2023)
3Automated explanation already failed at this model's scale
OpenAI's explainer–simulator pipeline, which ours copies in shape (an explainer writes a label, a simulator predicts activations from it), was run on every neuron in GPT-2. In OpenAI's own words, the vast majority of explanations scored poorly. Humans did better than GPT-4, and the method worked worse on later layers.
LITERATURE Bills et al. 2023, Language models can explain neurons in language models
4Interpretability metrics pass on random transformers
Heap, Lawson, Farnik and Aitchison trained sparse autoencoders on randomly initialised transformers and found latents about as interpretable, by automated scoring, as on trained ones, across Pythia sizes and several randomisation schemes. The paper's second version is titled "Automated Interpretability Metrics Do Not Distinguish Trained and Random Transformers". Knots' headline metric is an automated interpretability score.
LITERATURE Heap et al. 2025, Sparse Autoencoders Can Interpret Randomly Initialized Transformers
5Explanations that do not depend on the model are a known failure
Adebayo et al. randomised a network's weights and found that several popular saliency methods produced nearly the same maps: some deployed methods were independent of both the model's parameters and its training data. Their tests (randomise the model; randomise the labels) are the sanity checks any explanation method owes. Knots has run the first one only partly (an untrained model and random directions), and the label randomisation not at all.
LITERATURE Adebayo et al. 2018, Sanity Checks for Saliency Maps (NeurIPS 2018)
6A probe that can memorise will look like it found structure
Hewitt and Liang showed that a probe's accuracy partly measures the probe's own capacity to memorise. They proposed control tasks, with random outputs per word type, and selectivity: the gap between real and control accuracy. On part-of-speech tagging, accuracy 97.3 came with control accuracy 92.8. Knots' "probe" is a language-model agent reading 40 examples, about as expressive a probe as exists. Its selectivity has not been measured as such; our closest number (§2) is not encouraging.
LITERATURE Hewitt & Liang 2019, Designing and Interpreting Probes with Control Tasks
7Probing is correlational
Belinkov's review of probing classifiers sets out its standing problems: the choice of baseline, the choice of classifier, and the method's correlational nature. Finding that information can be read from a place does not show the model uses it there. A knot is a probe with a threshold.
LITERATURE Belinkov 2022, Probing Classifiers: Promises, Shortcomings, and Advances (Computational Linguistics 48(1))
8Intervening along a direction can work for the wrong reason
Makelov, Lange and Nanda showed that patching a subspace can produce exactly the intended change in behaviour by switching on a dormant parallel pathway, while the subspace itself is causally disconnected from the output. So an emit that "works" does not show that the knot's direction is where the concept lives.
LITERATURE Makelov et al. 2023, Is This the Subspace You Are Looking For? (ICLR 2024)
9Where a thing is found does not predict where it can be changed
Hase, Bansal, Kim and Ghandeharioun found near-zero correlation between causal tracing's localisation of a fact and the success of editing it at that layer, for ROME, MEMIT and fine-tuning. If localisation and intervention can come apart even for facts, a knot's place is weak evidence about what speaking there will do.
LITERATURE Hase et al. 2023, Does Localization Inform Editing?
10"Interpretable" has no agreed meaning
Lipton argues that interpretability is underspecified: papers treat it as self-evident while meaning trust, causality, transferability or post-hoc explanation. Knots claims to do "interpretability from the inside" without saying which of these it delivers. The measured quantity is label F1, which is none of them directly.
LITERATURE Lipton 2016, The Mythos of Model Interpretability
11Post-hoc explanation of a black box may be the wrong goal
Rudin argues that explaining black boxes after the fact entrenches bad practice, and that where stakes matter one should build models interpretable by design. Knots labels a black box after it is trained, every time. Rudin's argument says the effort belongs in the model, not in an instrument around it.
LITERATURE Rudin 2019, Stop Explaining Black Box Machine Learning Models for High Stakes Decisions (Nature Machine Intelligence 1)
12The model's own head already does this
The logit lens reads the stream through the output head and works unusually well on GPT-2. Knots' frontier and keeper are this idea run on the card each update; the report says so itself ("close to keeping a running list of logit-lens hits"). And the logit lens is a biased, brittle estimator: on BLOOM and OPT its top prediction is often just the input token in more than half the layers, which is why the tuned lens exists.
LITERATURE nostalgebraist 2020, interpreting GPT: the logit lens; Belrose et al. 2023, the Tuned Lens · OUR DATA experiments/knots/REPORT.md §12
2 Our own nulls
The literature's warnings came true in our own data, in the controls we ran.
13A random direction earns most of a real region's label gain
We hid 104 knots along random directions, with thresholds calibrated to fire as often as their plane's median region, among the 1,075 real region knots, and labelled them all blind. The real regions' labels beat the commonest-word label by a median +0.32 in F1. The random directions' labels beat it by +0.18, ahead on 81 of 104. A random line at a calibrated threshold picks out rows an agent can partly describe, and the regions' own share of the gain is about +0.14. At some planes (1 to 8) the random knots' gain reaches +0.41.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-30000-atlas-2026-10-30; REPORT.md §0
14An untrained model scores higher on raw F1
The same labelling pipeline on the model after one update (loss 10.82) gave a higher raw median F1 (0.79) than the trained model (0.71), because its regions are effectively single tokens. Only the gain over a one-word label separates them (+0.46 trained, +0.00 untrained). The project quoted raw F1 as its headline before this control was run: the measure we first trusted could not tell a trained model from noise.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-untrained-null-2026-10-30; REPORT.md §1.0
15The headline gain is partly selection
After round 2, the final region labels gain +0.49, up from round 1's +0.32. But round 2 retested only the 449 weak labels, on fresh rows, after revision; the strong ones kept their round-1 scores. Regression to the mean alone lifts a set retested because it scored low. The report says so ("the rise from +0.32 partly selection"). The random controls never got their round 2 (their firings were on a pod that went offline), so there is no like-for-like null for the +0.49.
OUR DATA REPORT.md §0; bin/aos-atlas-round2
16The keeper's knots and the bridge have no random null
Discovery by the keeper is reported as agreement in counts: firings against the lens's wakes, median 1.46 at cosine 0.1. Which rows fire against which rows the lens names was not measured, and no random-direction keeper was run. "Fires about as often as the lens weighs its token" is the expected result for many directions at a loose threshold.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-30000-keeper-planes-2026-10-30; REPORT.md §10.2
17The labels are least informative where interpretation matters most
The gain over the one-word label is +0.17 to +0.24 at planes 9 to 12 and +0.18 at the input, against random directions' +0.07 to +0.15 at 9 to 12. The upper layers, where the model turns representation into prediction, are where labels say least beyond "this word".
OUR DATA public/atlas/atlas.json; the table on How it works §6
3 The objects are crude
18Features live in superposition, and k-means centroids cannot separate them
Elhage et al. showed models represent more features than dimensions by superposing them, and that this produces polysemantic units. A k-means centroid at a cosine threshold is one direction per cluster of rows: it can pick out a dense crowd, not a sparse feature that shares its dimensions with thousands of others. Knots' regions are k-means centroids (4, 16 and 64 per plane), chosen, in the report's words, because they are "exact and cheap".
LITERATURE Elhage et al. 2022, Toy Models of Superposition · OUR DATA REPORT.md §12
19The field moved to dictionaries thousands of features wide
Anthropic decomposed a 512-neuron layer into more than 4,000 features with a sparse autoencoder. Knots finds at most 64 regions a plane in 768 dimensions: about a twelfth of the width, and two orders of magnitude fewer features than a modest SAE.
LITERATURE Bricken et al. 2023, Towards Monosemanticity
20Even SAEs fail on the tasks Knots wants
Google DeepMind's interpretability team reported negative results for SAEs on downstream tasks: on out-of-distribution probing for harmful intent, dense linear probes did very well while single-latent SAE probes failed to fit even the training set. They deprioritised fundamental SAE research. If carefully trained SAEs lose to a linear probe, a k-means centroid is unlikely to win.
LITERATURE Smith, Rajamanoharan, Conmy et al. 2025, Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research
21Simple baselines beat the elaborate methods
AxBench compared steering and concept detection on Gemma-2: prompting beat every representation method at steering, fine-tuning came next, and simple difference-of-means statistics did best at detection. SAEs lagged at both. Knots builds an operating system to do detection and steering that a prompt and a mean difference already do better.
LITERATURE Wu et al. 2025, AxBench (ICML 2025)
22Dictionary features absorb each other
Chanin et al. found that a latent that seems to track a concept fails to fire on arbitrary tokens where a more specific latent has absorbed it, in every LLM SAE they tested, and that size and sparsity settings do not fix it. Knots' nested regions (4 inside 16 inside 64) invite the same failure: a coarse region's knot silent where a finer one took its rows.
LITERATURE Chanin et al. 2024, A is for Absorption
23Not every feature is a direction
Engels et al. found days of the week and months of the year represented as circles in Mistral 7B and Llama 3 8B, used causally for modular arithmetic. A knot is a direction and a threshold: a half-space. It cannot name a position on a circle, and a circular feature cut by a half-space looks like an arbitrary arc.
LITERATURE Engels et al. 2024, Not All Language Model Features Are Linear
24Features are not even stable across seeds
Paulo and Belrose found that SAEs trained on the same model and data with different seeds share only about 30% of their features in one setting. Knots' regions come from one k-means run on one shard. A name tied to a region that would not reappear under another seed is a name for an accident.
LITERATURE Paulo & Belrose 2025, Sparse Autoencoders Trained on the Same Data Learn Different Features · OUR DATA REPORT.md §10.2 ("one shard, one seed")
25Token knots mostly detect the input token
A cosine-0.5 knot along a token's embedding row fires mostly at plane 0 on rows whose input is that token. "A knot on ' the' fires 141 times" mostly means "' the' occurs 141 times". Up to 10,000 knots were tied in the cost tests, but they were 10,000 token knots, not 10,000 learned directions: the pool held 224 custom vectors.
OUR DATA REPORT.md §10.2, §1.1 item 4
26The middle of the model does not point along words
The keeper's first design tied knots along input embeddings. Every one past plane 0 was untied within 5 updates, because it neither fired nor woke. The stream past the input does not point along input words, which undercuts the intuition the whole naming scheme started from.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-30000-keeper-planes-2026-10-30 (keeper.out)
27Most of the young models' "regions" were punctuation
Most experiments ran on models of 8 to 1,536 updates, whose regions were mostly single tokens or punctuation. The report lists "one dominant direction, regions that are single tokens, a lens that agrees with the output, forks that say ' the'" as what a young model looks like, not what GPT-2 looks like.
OUR DATA REPORT.md §1, §10.2
28The space is not a tree, so the tree view misleads
The public view draws regions as a tree. Across the trained model's depth, at each step about a third of regions carry, a third split and a third scatter. The report concedes it is "neither a tree nor a pipe". A hierarchy drawn over a structure that is not one is a picture of the method, not of the model.
OUR DATA REPORT.md §0, §10.3
4 Agents grading agents
29Language-model judges have measurable biases
Zheng et al. found strong position bias in LLM judges (most favour the first position; only GPT-4 was consistent in more than 60% of cases) and verbosity bias in all of them. Knots' simulator agents judge FIRES or QUIET for examples in a page of 28 knots. Order and length effects on those judgements were never measured.
LITERATURE Zheng et al. 2023, Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (NeurIPS 2023)
30One model wrote, tested and revised every atlas label
The atlas's labels were written, simulated and revised by GLM (zai/glm-5.3), every role the same model. An explainer and a simulator that share a model share its priors: a label phrased the way that model thinks will be "predicted" well by the same model, whatever the knot does.
OUR DATA REPORT.md §10.2; experiments/knots/agent-prompts/
31The labeller changed mid-project
The 2026-10-29 labels were written and tested by Claude subagents; the atlas's by GLM, one page at a time. Scores across the two are not a like-for-like comparison, and the plane-6 headline (0.79) and the atlas (0.75) mix labeller with model and plane.
OUR DATA REPORT.md §10.2
32The reviser's prompt was rewritten and the original not kept
Round 2's reviser instruction was rewritten, and the original was not kept. A labelling result whose prompt cannot be reproduced exactly is a weak result, in a project whose other selling point is bit-exact replay.
OUR DATA REPORT.md §10.2
33Balanced tests make F1 look better than real use
Every label test is half firing, half not, so a coin scores about 0.5 and "fires everywhere" scores 0.67 (recall 1, precision 0.5). A knot that fires on 1% of real text would face a far harder test: at that base rate, a precision that looks fine on a balanced test collapses. No label was scored at the knot's real firing rate.
OUR DATA How it works §6 ("a label that just guesses scores about 0.5 on our tests")
34The agents see 24 tokens
In the answer-key test, the two planted rules the agents missed were the only long-range ones, "inside quotes" and "inside parentheses": the opening mark often fell outside the 24-token window. Any feature the model computes over more than 24 tokens is invisible to the labeller by construction.
OUR DATA REPORT.md §1.0 ("Answer keys")
35Planted rules are easier than real ones
Planted knots with known rules scored a median F1 of 0.99; real ones in the same round, 0.62. The answer-key test shows the pipeline can find a rule if a crisp one exists. The gap shows real knots mostly do not have one.
OUR DATA REPORT.md §1.0
5 Emits and steering
36Steering vectors are unreliable even in distribution
Tan et al. found that steering vectors vary widely from example to example, that for some behaviours a large share of inputs move the opposite way ("anti-steerable"), and that out-of-distribution generalisation is brittle to reasonable prompt changes. Their conclusion: steering vectors are not a panacea. An emit is a steering vector.
LITERATURE Tan et al. 2024, Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024)
37Emits are activation addition with extra steps
Turner et al.'s ActAdd steers GPT-2 with the difference of two prompts' activations, no optimisation and one pair of examples. The report calls our emits "a crude form of activation steering (ActAdd-style) with a knot-based readout". A knot's emit is ActAdd with a worse direction (a k-means centroid rather than a contrast) and a custom OS around it.
LITERATURE Turner et al. 2023, Steering Language Models With Activation Engineering · OUR DATA REPORT.md §9.4
38"Fire, then speak" was already published
CAST gates a steering vector on a condition read from the hidden state: the closest published relative of a knot that fires and then speaks, and named by our outside reviewer as its most direct predecessor.
LITERATURE Lee et al. 2024, Programming Refusal with Conditional Activation Steering · OUR DATA REPORT.md §12
39Where emits are clear, they break the model
At one row norm a region's direction costs +1.6 in held-out loss; at 2 and 4 row norms, +2.0 and +3.5, and the answers lose their specificity. At the doses where the model survives (a quarter to half a row norm), a region's own tokens rise in only 19 to 23% of its emits. An instrument that wrecks the system at the doses where it reads clearly is measuring damage.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-30000-emit-planes-2026-10-30
40Emits say nothing in the middle of the model
Own tokens rose at planes 1 to 5 and 10 to 11, and at none of planes 6 to 9, on 8 emits a plane. Plane 6 is where the project did nearly all its work before the atlas. The middle layers, where the knots were labelled, are where emits answer nothing recognisable.
OUR DATA …/gpt2-alphaos-30000-emit-planes-2026-10-30 ("By plane")
41Large effect, wrong sign: a region costs more loss, not more sense
The emit sweep's strongest result is that a region's direction costs far more held-out loss than a random one (+1.64 over the random-direction null at dose 1, against −0.06 for random knots). That shows the model is sensitive to the directions its rows cluster along, which is expected of any direction with large variance. It does not show the direction means what its label says.
OUR DATA …/gpt2-alphaos-30000-emit-planes-2026-10-30
42The concept steering barely beat random
The composed compersion direction, spoken at its own plane, raised held-out loss by +0.96 against +0.92 for random directions of the same length. On the 1,536-update model, the mix joy + love − jealousy was a null: +0.023 against +0.015, no emotional words.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-30000-compersion-2026-10-29; REPORT.md §9.3–9.4
43Plane 12's emits were broken and nobody noticed
The first emit sweep's plane-12 emits changed nothing (loss +0.0000) because plane 12 had no speech step. The public emit button had been offering plane-12 knots. A tool that silently answers "no effect" is worse than one that errors.
OUR DATA …/gpt2-alphaos-30000-emit-planes-2026-10-30 ("Plane 12")
44Steering and reading point different ways
Park et al. distinguish the direction that reads a concept from the one that steers it: they are related by the model's own geometry, not equal. Knots uses one vector for both (a knot listens along its direction and emits along it), which the linear-representation work says is the wrong default.
LITERATURE Park et al. 2023, The Linear Representation Hypothesis and the Geometry of Large Language Models
6 The pipes
45The clean pipes were an initialisation artefact
The headline "a carrier spoken at plane 0 is heard on exactly the spoken rows at all 13 planes" was measured at updates 0 and 1. With residual projections and the head initialised to zero, every block is the identity in exactly those two updates. The outside reviewer pointed this out, and we checked it: those results test injection, the skip paths and decoding, not transport through a transformer.
OUR DATA REPORT.md §1.1 item 1
46Training absorbs every carrier within two updates
From update 2 on, every listener hears its carrier on every row of planes 2 to 12, spoken or not. Training on a stream with speech in it moves the carrier into all rows within two learner steps, and no carrier shape resists it. So pipes only work on frozen models, and "knots during training", the original vision, writes into the training signal.
OUR DATA REPORT.md §6.4
47Quiet messages do not survive a trained model
On the 30,000-update model, a message at amplitude 16 arrives at plane 12 as 4.4 above background; amplitude 8 is gone by plane 10. Only a loud carrier (32, about a seventh of the row norm) crosses all 12 blocks. A carrier loud enough to be heard is loud enough to change what the model computes.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-30000-repeater-2026-10-30; …-kv-pipe-2026-10-30
48The repeater fires on noise
The parametron majority restores a quiet message, but its fixed threshold also fires on rows nobody spoke into: 56 of 4,096 rows at amplitude 32 and plane 2, and 1,798 at amplitude 16 and plane 6. It then re-speaks there. It is noisy in proportion to the stream's own projection, so the louder the model, the more it hallucinates messages.
OUR DATA …/gpt2-alphaos-30000-repeater-2026-10-30 ("Is the majority right?")
49Attention carries that something was said, not what
The rows after a spoken row move along the carrier by about the same amount whether the message is 8 or 32 (−7.3 and −8.5 at plane 12). The KV-cache route is not a scaled copy of the message. The young model's "a fifth, two positions on" did not hold.
OUR DATA …/gpt2-alphaos-30000-kv-pipe-2026-10-30
50The oscillator readout's advantage was filter bias
The oscillator readout looked immune to absorption. But the target sat in the slowest oscillator and the decoys in faster ones: with the same drive everywhere, slot 0's energy was 19.77× the median decoy's at damping 0.99. And on the trained model the readout hears nothing past block 0.
OUR DATA REPORT.md §1.1 item 2, §1
51"Frequencies" are a metaphor
Carriers are vectors added to the residual stream, decoded by a matched listener, one scalar per carrier. Nothing oscillates except a simulated readout. The project's vocabulary (pipes of frequencies, parametrons, resonance) promises physics the system does not have, and the report concedes it.
OUR DATA REPORT.md §10.3
52Vector messages between models gain little
Published cache-to-cache communication between models gains a few points over text, and later audits found much of that gain survives when the message comes from a different example. Our bridge carries one known bit per update through the host, which avoids that pitfall by saying much less.
LITERATURE Fu et al. 2025, Cache-to-Cache; Shi et al. 2025, KVComm · OUR DATA REPORT.md §8.2, §12
7 An operating system inside a model
53Every working LLM operating system keeps the OS outside
MemGPT and AIOS, the best-known "LLM OS" designs, schedule, page memory and call tools from outside the model. The report's own literature review concluded that "every version that put machinery inside the model failed on evidence". The reviewer softened that, since constructed neural programs do work (next item), but not for ordinarily trained models.
LITERATURE Packer et al. 2023, MemGPT; Mei et al. 2024, AIOS · OUR DATA REPORT.md §12
54Programs inside weights work only when you build the weights for them
Tracr compiles programs into transformer weights by reserving orthogonal subspaces, and looped transformers can emulate a one-instruction computer. Both work because the weights were constructed for the program. Knots attaches programs to a model trained for next-token prediction, where no subspace is reserved and every carrier is absorbed (§6).
LITERATURE Lindner et al. 2023, Tracr; Giannou et al. 2023, Looped Transformers as Programmable Computers
55The programs only read
In the scheduler tests, losses were identical in every arm "because programs only read the stream". The first program tied to a GPT-2 knot counts its own firings, and the count equals the watch's. That proves the plumbing. It does not show a program doing anything a host-side script reading the same firings could not.
OUR DATA tests/observability/import/rtx3090/gpt2-alphaos-30000-knot-scheduler-2026-10-30
56Rent removed almost every knot
At 3 bits a firing, rent untied 82 of the 84 plane-6 knots. A region knot costs 25,216 bits and needed 8,405 useful firings in 10 updates to stay. The project's own economic test says nearly all its knots are not worth their storage. (The rate was chosen by us, so this cuts both ways, but no rate was shown at which most knots earn their keep.)
OUR DATA …/gpt2-alphaos-30000-knot-scheduler-2026-10-30 ("Rent on the card")
57"An OS at every knot" is not what exists
Processes, pipes between programs, files per knot beyond 12 words, and command/1 to a knot exist only in the standalone knots/1 system. Inside GPT-2 a knot is a detector with an optional writer, and since 2026-10-30 an optional bounded program. The report lists "that an operating system runs at each knot inside GPT-2" first among the claims it is not making.
OUR DATA REPORT.md §10.3, §4.3
58Event-triggered bounded programs already have a design: eBPF
A checked, bounded program attached to an event, run in a privileged context with its own small map, is eBPF's model, named by the outside reviewer as the design to follow. Knot programs reinvent its loader checks (magic, length, checksum, placement) without its verifier, which proves termination and memory safety before a program runs.
LITERATURE eBPF documentation, What is eBPF · OUR DATA REPORT.md §12
8 The custom stack
59The bitter lesson
Sutton's essay: over seventy years, general methods that use computation beat methods that build in human knowledge, by large margins once compute grows. Knots builds human names, human-chosen thresholds and a hand-written OS into a model. Our compute went into the scaffolding while the model stayed undertrained.
LITERATURE Sutton 2019, The Bitter Lesson
60We rewrote everything from scratch
Our own language, our own compiler to NVIDIA machine code, our own GPU operating system, our own training loop, our own emulator. Spolsky called Netscape's rewrite the worst strategic mistake a software company can make: a rewrite throws away years of fixed bugs. Every PyTorch hook, TransformerLens and SAE library we did not use is a body of fixes we have to rediscover.
LITERATURE Spolsky 2000, Things You Should Never Do, Part I · OUR DATA REPORT.md §2.1
61The machine code is undocumented
AlphaOS compiles to SASS, NVIDIA's native instruction set, without CUDA or PTX. NVIDIA rarely discloses low-level details; what is known about SASS encodings comes from microbenchmarking and disassembly. A research result that rests on an undocumented ISA, reverse-engineered for one architecture (sm_86), is tied to one card family and to whatever NVIDIA changes next without saying.
LITERATURE Jia et al. 2018, Dissecting the NVIDIA Volta GPU Architecture via Microbenchmarking · OUR DATA REPORT.md §2.1
62Bit-exactness proves determinism, not truth
The project leans on bit-identical numerics: silence is "losses identical to the bit". But identical reruns prove only that the same inputs give the same bits. They say nothing about robustness to seed, shard or architecture, and they make it impossible to estimate variance by rerunning. The report: "no confidence intervals".
OUR DATA REPORT.md §10.2
63Silence was checked on 8 logits of 50,257
"Silent knots change nothing" was verified on every row's 8 best logits, not all 50,257, and the 98 matching records are aggregates, not a trace of every firing. Losses are compared exactly, but the claim is narrower than its wording.
OUR DATA REPORT.md §1.1 item 7
64Knots are expensive where they matter
Every-plane knots cost 0.40 ms each per update, linearly: 10,000 knots made an update 30 times slower (0.14 s to 4.3–4.7 s). The lens at 12 planes adds about 290 ms to a 69 ms update (362 ms in all). A run with knots tied during training slowed from 0.35 s to 1.0 s an update. Long runs now train with no knots, so "knots inside every training update", the founding idea, is not how the trained model was made.
OUR DATA REPORT.md §5.3, §1.0; …/gpt2-alphaos-30000-keeper-planes-2026-10-30
9 What broke
65A never-ending kernel fights the host
Long runs stalled again and again (#293). The cause was the RunPod host's quarter-hourly context switch timing out on AlphaOS's never-ending CTAs (Xid 109). The fix, resumable nodes that end every CTA every 5 ms, gives up part of what a persistent-kernel OS is for. Other stalls are recorded, not explained: gpu1 stopping after 30 to 80 updates; a frontier hang with all 46 nodes in one task; the interactive shell stopping after two or three commands.
OUR DATA REPORT.md §1.0, §10.1
66Three pods, and some of their evidence, were lost
A rented RTX 3070 pod stopped responding mid-session; an RTX 3090 pod was removed with its disk when the account's balance ran out, taking the #293 fix's card logs with it. On 2026-10-30 a second 3090 pod vanished mid-chunk, with the random controls' firings, so round 2 of the null and the first survival check could not run. Results that live on rented disks are not results until they are copied off.
OUR DATA REPORT.md §0, §10.1
67The public site relayed visitors to a stranger's server
knots.blah.dev relayed emits through a free Cloudflare quick tunnel. After the pod went down, the tunnel's hostname came to serve someone else's server, and the site relayed users' emits there until we noticed. Every piece of improvised infrastructure is a place for this to happen.
OUR DATA REPORT.md §10.1 ("A borrowed address")
68Our own loader overwrote a program
The frame renderer grew past its three code stages, and the loader installed it over another program's bytes. Separately, the record ring overflowed and a report briefly said a speaker "spoke into 0 rows". An OS we wrote ourselves has an OS's bugs, found by us, one at a time.
OUR DATA REPORT.md §10.1
10 Scale and the model
69The model is badly undertrained
Our best model read 123M tokens for 163M parameters: under one token a parameter. The Chinchilla result, as usually cited, puts compute-optimal training near 20 tokens a parameter (about 3.3 billion tokens here). Karpathy's llm.c reproduced GPT-2 small on 10 billion FineWeb tokens. Our held-out loss is 4.77, against about 3.3 for a trained GPT-2 small. Every conclusion about "what the model represents" is about a model that has not finished learning, and most experiments ran on far younger ones (loss 5 to 10).
LITERATURE Hoffmann et al. 2022, Training Compute-Optimal Large Language Models; Willison 2024, on Karpathy's llm.c GPT-2 reproduction · OUR DATA REPORT.md §2.2
70It is not GPT-2
ReLU² MLPs, rotary positions with QK-norm, soft-capped logits, an untied head, zero-initialised residual projections and U-net skips. Several pipe results were artefacts of exactly these choices (the zero init, the skips). Findings on this variant may not transfer to GPT-2, let alone to models people use.
OUR DATA REPORT.md §2.2, §1.1
71The interesting concepts are not in the data
"Compersion" occurs 0 times in 300M FineWeb tokens; "kangaroo" 3 times in what the 1,536-update run read. The composed compersion direction, tested across authors, separates from jealousy at 0.60 to 0.66 (shuffled labels 0.43 to 0.55), and as a knot it fires mostly on photo credit lines. In the report's words: "No knot can find what the model never saw".
OUR DATA REPORT.md §9.3; …/gpt2-alphaos-30000-compersion-2026-10-29
12 Statistics and process
73The samples are tiny
The trained-model KV probe rests on 11 spoken rows, one carrier, one update; the repeater on 148 rows; the emit sweep on 4 regions and 4 controls a plane (8 emits a plane at the low doses); the map across models on 16 knots with 0 to 35 test positives each. With 84 knots tested, about two clear 2σ by chance.
OUR DATA REPORT.md §10.2, §9.4, §1.1 item 5
74Not preregistered, and some drafted after the fact
EXP-0044 to EXP-0047 were drafted after their pilots ran (the drafts say so). Calibration and scoring shared an update in the explorer's best pass, and on held-out updates recall was 0.26. The region recall of 1.00 that headlined early was an estimate from a non-random sample, capped at 1.
OUR DATA REPORT.md §10.2, §1.1 item 3
75We overstated, and outsiders caught it
Seven points of the first outside review changed what results meant (clean pipes at identity blocks, a biased oscillator test, an estimated recall, 10,000 knots on 224 vectors, upper-middle medians, a stop-gradient that would not help, silence on 8 logits), and the docs needed later corrections ("copies of one word", "+0.46 vs +0.14" compared unlike things). A project that overstates by default needs every result rechecked by someone else.
OUR DATA REPORT.md §1.1; git history of experiments/alphaos/top/web/public/docs/index.html
13 Is any of it new?
76Every ingredient has a predecessor
Named directions at internal layers: TCAV. A detector triggering a writer: CAST. Persistent edits keyed on regions of activation space: GRACE. Serializable interventions: ReFT and pyvene. Carrying detectors between models: model stitching of SAE features and probes. Random carriers in superposition: holographic reduced representations and hyperdimensional computing. GPU operating systems: GPUnet. The outside reviewer found no single predecessor for the whole stack, "which supports the integration as a contribution but not any single ingredient as new".
LITERATURE Kim et al. 2018, TCAV; Hartvigsen et al. 2023, GRACE; Wu et al. 2024, ReFT; Kim et al. 2014, GPUnet (OSDI) · OUR DATA REPORT.md §12
77An integration nobody needs is not a contribution
"The first system to do A, B and C together" is valuable only if someone needs A, B and C together. Each piece is done better, more cheaply and with better tooling separately: probes and SAEs in PyTorch, steering by prompt (§3, AxBench), agents' operating systems outside the model (§7). The integration's only unique property, exact in-graph measurement during training, is the one the project stopped using for long runs (§8).
OUR DATA REPORT.md §12 ("What we think is new"), §1.0
14 The vision itself
78The goal keeps moving
Knots began as operating systems tied inside the model, communicating by frequencies, during training. Measurement moved it to labels for a frozen black box (2026-10-29), then to an atlas of k-means regions. Each move was honest, and each moved away from what made the idea distinctive, toward a standard interpretability pipeline with an unusual runtime.
OUR DATA REPORT.md §1, §1.0 ("knots are labels for the black box"), experiments/knots/ALL-PLANES.md
79The rest of the vision is unbuilt
Full-OS knots with processes, files and pipes; knots answering each other; hot-patching inference; live breakpoints per token; labels inside the .knot weights; KV and carrier pipes between .knot files; an oscillator readout past block 0; compiled read-only knots. Each is listed as planned. A vision whose measured parts keep shrinking toward known methods, and whose unmeasured parts stay large, is the pattern of an idea that will not pay off.
OUR DATA experiments/knots/ALL-PLANES.md; REPORT.md §13
80The names do the persuading
Knots, keepers, frontiers, pipes of frequencies, parametrons, rent, the atlas: the vocabulary makes an activation detector sound like a living system. Lipton's warning (§1, item 10) applies in its plainest form: the words claim more interpretability than the measurements show.
LITERATURE Lipton 2016 · OUR DATA REPORT.md §3.1–3.2, §10.3