Louis Maddox
LightNews LightNews
banner
permutans.bsky.social
Louis Maddox
@permutans.bsky.social
150 followers 110 following 2.6K posts
Combinatorially curious https://spin.systems
Posts Replies Media Videos
permutans.bsky.social
😇
Wheel for the polars-fastembed-cuda package, a 40.3MB zip file
December 16, 2025 at 10:12 PM
permutans.bsky.social
hm, Windows binaries are incompressible by UPX due to a "Control Flow Guard"...
UPX error message about the CFG
December 16, 2025 at 8:30 PM
permutans.bsky.social
All-MiniLM-L6-v2 took longer but surprisingly accurate (99.9%) on the doc half matching task – only 1 mistake, matched the Python 3.13 Release Schedule PEP to the 3.12 one
Demo of the larger text embeddings from All-MiniLM-L6-v2 getting 99.86% correct matchings
December 16, 2025 at 2:00 PM
permutans.bsky.social
Confirmed it's useful for document retrieval (following the blog post's example of matching document halves), correctly matching 97% of the Python PEP corpus

The failures look justifiable e.g.
8102 (2021 Steering Council Election) ⇒ 8103 (the 2022 one)
361 (2.6 Release Schedule) ⇒ 373 (the 2.7 one)
Demo of a document half retrieval task, getting 97% of them right and the mistakes looking generally well-justified (the DataFrame at the bottom of the screen shows these mistakes, with the rank of the correct half shown as mostly 2)
December 16, 2025 at 1:45 PM
permutans.bsky.social
I was curious to try this out myself so I adapted polars-fastembed to run Luxical One and no, as advertised, it's not useful for retrieval [for search], but runs in 1.8s an operation that takes 30s with all-MiniLM-L6-v2 on CPU (20s on GPU), so approx. 0.5ms per 1k tokens ! github.com/lmmx/polars-...
Embedding all the Python PEPs in 2s (just don’t try to use them for search)
December 16, 2025 at 10:41 AM
permutans.bsky.social
“we observe that TensorRT always outperforms CUDA”
Table showing TensorRT consumes less energy than CUDA from “Green AI: A Preliminary Empirical Study on Energy Consumption in DL Models Across Different Runtime Infrastructures”
December 16, 2025 at 12:06 AM
permutans.bsky.social
hmm TensorRT is a no for embeddings apparently
spammy TensorRT datatype errors
December 15, 2025 at 4:41 PM
permutans.bsky.social
Intel splits them per generation nietras.com/2021/01/25/o...
IPP en.wikipedia.org/wiki/Integra...
Note that the total size of this is a whopping ~2600 MB. cuBLAS, cuDNN and TensorRT (nvinfer*.dll) being huge. If you are only running CNNs you can remove the cudnn_adv*.dll files. Additionally, the cudnn*train64_8.dll files can be removed since these are only for training. This is an artifact of how NVidia has decided to distribute and package dlls with cubin code for multiple SM versions all together in the individual dlls. A more sensible approach, in my view, would be to do what Intel has done for years for Integrated Performance Primitives (IPP) and split these into dlls for each SM version e.g. cublas64_11_sm86.dll. And clean up the whole not forwards compatible version naming etc. That’s enough ranting though. TensorRT is a must for best performance machine learning inference. Which I will get to in a moment.
December 15, 2025 at 1:18 AM
permutans.bsky.social
TIL you can ship ONNX runtime CUDA support via dependency via a 300MB wheel 🫠 pypi.org/project/onnx...
onnxruntime-gpu wheel files on pypi
December 15, 2025 at 12:46 AM
permutans.bsky.social
Huh, cool… cargo xtask is a task runner that fully bootstraps from the Rust toolchain github.com/matklad/carg...
cargo xtask README • cargo xtask cargo-xtask is way to add free-form automation to a Rust project, a-la make , npm run or bespoke bash scripts. The two distinguishing features of xtask are: • It doesn't require any other binaries besides cargo and rustc, it fully bootstraps from them • Unlike bash, it can more easily be cross platform, as it doesn't use the shell. How Does it Work? cargo-xtask is a polyfill for cargo workflows feature. It is a way to extend stock, stable cargo with custom commands ( xtasks ), written in Rust.
December 14, 2025 at 10:29 AM
permutans.bsky.social
☃️ Good 🌟 tidings ❄️ are 🎅 Now 🌰 ! 🦌! 🛷 ! 🎄
Yours truly wearing reindeer festooned spectacles
December 13, 2025 at 6:41 PM
permutans.bsky.social
and fixed: 2x faster embedding using Rust, 1.4x faster retrieval using polars-distance over NumPy 🤠
polars-fastembed plugin is 2x faster to embed using Rust, and 1.4x faster to retrieve from swapping numpy for polars-distance
December 13, 2025 at 1:29 PM
permutans.bsky.social
6 word horror story: the Rust rewrite somehow got slower 😢
Rust rewrite 20% faster than Python for embedding but 13% slower for embedding and retrieval
December 13, 2025 at 12:54 PM
permutans.bsky.social
🧊❄️🧊🦣🧊❄️🧊 Revived the mammoth ½h ONNX runtime all-platform source builds for polars-fastembed and put in 1 click hands-off Trusted Publishing CI release 🚀 github.com/lmmx/polars-...
polars-fastembed CI builds on all Linux, Mac and Windows architectures with PyPI package release in 28 minutes
December 13, 2025 at 12:24 PM
permutans.bsky.social
CI 🫠 is 🫠 fixed 🫠 (3.15 you stay out of it) my ONNX runtime pyo3 polars embeddings package is so back
many many CI builds and checks all passing
December 13, 2025 at 2:45 AM
permutans.bsky.social
Together we can end GitHub Actions inline bash scripts #justusejust
Terminal with a Justfile task (‘recipe’) definitions file on the left; the results of running them on the top right; the usage in a CI workflow YAML on the bottom left
December 12, 2025 at 11:21 PM
permutans.bsky.social
Went to make an emoji search app using the dataset at huggingface.co/datasets/bad... then spotted it has a link to exactly what I was after!
😀 😄 😃 😁 😊 😂 😆
Emojeez: AI-powered, multilingual semantic search for emojis in 50+ languages 🌐 emojeez.streamlit.app
🥳 💫 😅 🥱 🔔 🌞 🎉
Emoji search on iOS showing the device glyphs returned for an embedding search for ‘smile’ (with some mistakes, naturally)
December 12, 2025 at 4:21 PM
permutans.bsky.social
Asked Claude and GPT 5 how they'd respond if I pasted in a discussion e.g. on some particular software planning decision with disagreement on how to proceed
This is a thoughtful framework, and I'd be happy to help you work through a disagreement using ir Here's now ra approach le What I'd do in practice First, I'd ask you to share the actual discussion-the bug report, PR thread, or forum exchange. I'd want to see the different positions people have taken in their own words. Then I'd work through it roughly like this: 1. Map the landscape: I'd identify the distinct viewpoints and try to articulate cach one charitably-what value does each person seem to prioritize? What do they believe about how things work (the technical or practical claims)? What outcome do they want? I'd give each position a descriptive label and lay it out in a table like the Bloomfield Track example. 2. Look for hidden agreement: Often in technical disputes, people actually share more common ground than they realize. They might agree on what makes good software but disagree about whether a particular change achieves it. Or they might have the same goal but different beliefs about what's technically feasible. I'd try to surface that. 3. Identify the real crux: Is this fundamentally a values disagreement (eg,, "stability vs. new features*), a factual/technical disagreement (eg. "will this approach scale?"), or a preference disagreement where people actually want the same thing but frame it 4. Suggest paths toward meta-consensus: Depending on where the disagreement lives, I'd propose different moves. If it's epistemic (disputed technical claims), maybe there's an experiment or benchmark that could resolve it. If it's normative, maybe the parties can acknowledge each other's values are legitimate even if weighted differently. I wouldn't pretend to be a neutral arbiter who can detinitively resolve the dispute. I'd be offering a structured way to understand it better, not a verdict. The goal would be helping the parties see each other's positions more clearly and find what shared understanding micht exist— not forcing asreement. Long wall of text from ChatGPT about how it would mediate a discussion using meta-consensus
December 12, 2025 at 3:47 PM
permutans.bsky.social
Meta-consensus (2006) openresearch-repository.anu.edu.au/bitstreams/1...
💠 Normative (shared values) ⇒ recognise legitimacy of disputed values
🔍 Epistemic (agreement on belief) ⇒ acceptance of credibility of disputed belief
🎯 Preference (wanting same outcome) ⇒ agreement on nature of disputed choice
Page 5 Page from academic article featuring Table 1: "Types of Consensus." The table distinguishes between Type of Consensus and Meta-Counterpart across three elements: Value (normative consensus vs. recognition of legitimacy of disputed values), Belief (epistemic consensus vs. acceptance of credibility of disputed beliefs), and Expressed Preference (preference consensus vs. agreement on the nature of disputed choices). The text introduces the Bloomfield Track case—a road through World Heritage Rainforest in tropical Australia whose future is publicly contested. Three groups are identified by their underlying positions. The authors explain that consensus can refer not just to what should be done (preferences) but also to underlying values and beliefs. Building on Elster, they identify three kinds: normative (agreement on values), epistemic (agreement on cause-effect beliefs), and preference (agreement on action). Universal consensus would require all three; each has a "meta" counterpart developed in the following section. Page 6 Page from academic article featuring Table 2: "Preference Formation, The Bloomfield Track." The table has three columns (Value/Normative Level, Belief/Epistemic Level, Expressed Preference/Preference Level) and three rows for different groups: Pragmatists: prioritize community needs, believe road benefits community, prefer keeping road Preservationists: prioritize environment, believe road harms environment, prefer closing road Optimists: prioritize environment but recognize community matters, believe road can benefit environment (e.g., tourism raising environmental profile), prefer keeping road The text below explains normative meta-consensus: agreement on the legitimacy of values without agreeing which should have priority. It differs from Rawlsian overlapping consensus because recognition doesn't require finding counterparts in other doctrines. The authors note psychology research shows people generally agree on the legitimacy of basic values at an abstract level—they differ on priority and application. This suggests deliberation's task may be uncovering existing meta-consensus that partisans have strategically obscured. Page 7 Page from academic article discussing the HIV-AIDS policy dialogue in Colorado as a case study of deep normative conflict. The dialogue involved 60-70 participants including gay activists (including ACT-UP members), people with AIDS, and fundamentalist Christians active in anti-gay campaigns. Initial conditions were deep meta-dissensus—each side's identity required denying the other's legitimacy. A gay activist stated the issue must shift from moral to public health framing; fundamentalists insisted it was entirely moral. Despite this, the dialogue produced a written agreement recognizing that "communities that include members with a range of moral perspectives" need prevention methods appropriate to that range—an explicit normative meta-consensus. The specific policies agreed upon (moral education in schools, sexually explicit material for gay communities) remained objectionable to each side but were accepted. The page also introduces epistemic meta-consensus: agreement that disputed beliefs are credible and relevant, even while disagreeing on which is correct. Examples include competing scientific paradigms and the coexistence of local residents' pollution experiences alongside epidemiological studies. Page 8 Page from academic article on pluralism and consensus. Contains a graph titled "Figure 1: Single-Peaked Preference Orderings of Bloomfield Track Policies" showing five policy options (Bitumenize, Upgrade, Stabilize, Status quo, Close) on the x-axis and preference rank (1st to 5th) on the y-axis. Four solid lines show different "single-peaked" preference curves where each peaks at one option then declines—these represent coherent preference orderings. One dashed line crosses multiple peaks, representing a non-single-peaked ordering that breaks the pattern.The surrounding text explains preference meta-consensus: first, agreement on which alternatives are acceptable (some theorists want to exclude preferences that harm others); second, agreement on how choices should be structured. The graph illustrates that if people agree "degree of access" is the key dimension, their preferences should be single-peaked along it. The text also notes single-peakedness isn't the only structuring option—sometimes meta-consensus reduces choices to simply "closure vs. all other options."
December 12, 2025 at 3:06 PM
permutans.bsky.social
📖 www.google.co.uk/books/editio...

> Deliberative democracy mistakenly aims at consensus

> its pursuit [seeks] to repress plurality and articulation of different perspectives
www.jstor.org/stable/40971...

> …[and] the structured antagonism that makes clear choices and accountability possible
DELIBERATIVE DEMOCRACY: AN INTRODUCTION 19 deeply undermines deliberation, present and future. It directly contradicts, inter alia, the core values of reciprocity, respect, and equality on which good deliberation is based (Gutmann and Thompson 2004, chapters 3-4). Modified in the light of pluralist ideals, deliberative ideals become more applicable to situations of underlying conflict and also capable of further evolution. When we take pluralism seriously (as, for example, when underlying conflicting interests make negotiation more appropriate than substantive consensus), then it becomes clear that deliberative quality cannot be assessed solely on the basis of the full list of first-generation standards. Rather this quality should be assessed in ways relevant to specific contexts without losing its core in reason-giving and listening (Bächtiger and Parkinson 2018). Deliberative democracy mistakenly aims at consensus. Empirical political scientists, difference democrats, pluralists, and agonists have criticized deliberative democrats for putting a misplaced stress on (rational) consensus as an ideal. From the viewpoint of difference democrats and agonists, rational consensus is not just undesirable but also conceptually misguided. To Mouffe (1999), its pursuit means that deliberative democrats seck to repress plurality and the articulation of different perspectives through the conflict that defines politics. To Shapiro (2017), stressing consensus means that deliberative democrats seek to repress the structured antagonism that makes clear choices and accountability possible. This criticism misrepresents the role of consensus even among first-generation deliberative democrats. It is particularly misplaced for the second generation of theorists, who grapple with problems of inclusion with particular sensitivity to differences (e.g. Young 2000). These reworkings of the concept of consensus recognize and value diversity in modern pluralistic societies. But, as noted above, these reworked concepts should not be viewed as the only defensible principles for good deliberative process. For instance, participants might initially think that their preferences are reconcilable (or not too distant), but find out in discussion that the opposite is true (Bachrach 1974). Knight and Johnson (2011: 145) point out that "even if as a result of the increased information that political argument makes available, individuals come to hold their preferences more reflectively, it in no way follows that this will lead to greater substantive agreement at the aggregate level." Like many of today's deliberative democrats, they consider opinion clarification and "structured disagreement" more important than consensus. Most recent research shows that in well-formed deliberative venues those whose opinions move away from the opinions of the opposing group after deliberation learn and clarify their opinions at rates equal to those whose opinions move toward the opinions of the opposing group (Lindell et al. 2017). In short, a good deliberative process can have a variety of outcomes, of which consensus is only one-and perhaps not even the most important one. Deliberative democracy misunderstands human motivations and the limits to the cognitive capacities of ordinary citizens. One such criticism is that deliberative democrats overestimate the demand for deliberative democracy. Hibbing and Theiss-Morse (2002) argue in Stealth Democracy that in the US citizens dislike politics, and are thus happy to let elites govern, provided the elites are honest and can be trusted; when citizens partici-pate, it is because they think they need to monitor and sanction untrustworthy political elites. Yet the conclusions of Hibbing and Theiss-Morse derive from what citizens think about opportunities to participate in a currently deficient system.
December 12, 2025 at 2:38 PM
permutans.bsky.social
Specific LLM-mediated policy shifts:
👶 Anti universal free childcare
🇪🇺 Had no effect on UK rejoining the EU
🌐 Pro Internet having been a net positive
👵 Pro raising retirement age
💷 Pro raising minimum wage
🌱 Pro net-zero
🇬🇧 Pro national pride
🏛️ Pro reducing prison numbers
🛂 Pro relaxing asylum policy
A bunch of before/after intervention scatter plots with the individuals’ opinion shifts marked in grey lines and a trend line showing overall shift. All showed consensus shift except Brexit
December 12, 2025 at 2:07 PM
permutans.bsky.social
PDF www.rivista.ai/wp-content/u...

LLM was tasked w writing a “group statement” (aiming for all to endorse), iterated on personal opinions/critiques via SFT + reward modelling ⇒ higher quality/clarity/informativeness/perceived fairness. Participants converged, didn't without

Reactance mediated IMO
Bar chart with error bars shows statistically significant decrease for strong agreement, others decreased but within error bar ranges. Higher drift towards consensus Al helps people find common ground in collective deliberation. (Left The Al mediator uses participants' opinions to generate group statements and iteratively refines those statements through participants' critiques. (Middle) Statements from the Al mediator (purple) garner stronger endorsement than those written by a human mediator (orange). (Right) Al mediation leaves groups less divided after deliberation, whereas simply sharing opinions with others does not.
December 12, 2025 at 1:53 PM
permutans.bsky.social
GDM—Habermas Machine github.com/google-deepm...
📄 AI can help humans find common ground in democratic deliberation www.science.org/doi/10.1126/...
+ supplement www.science.org/doi/suppl/10...

📰 Guardian: www.theguardian.com/technology/2...
📰 MIT Tech Review: www.technologyreview.com/2024/10/17/1...
Podcast transcript: Have you seen any of the interesting research on what's called the Habermas machine? I have not. It makes this argument, it's very, very interesting research. It's done by Google DeepMind, that actually one of the incredible uses of Al is that it can produce a community note or sort of a snippet, a distillation in very neutral language that people actually like quite a bit. Because one of the problems with community notes is that oftentimes the notes don't clear, so to speak. The way that a note shows up on a tweet is that the bridging algorithm means that people on the left and people on the right all have to agree that the note is helpful in order for it to show up. That's how the algorithm works. And that in terms of writing in that neutrality, which I think ties into Wikipedia and neutral point of view, right? How do you express something in kind of a neutral language that people find palatable? This is the thing that has to happen on community notes. On Wikipedia, you know, you're sort of expressing a neutral point of view. You have more language to work with. But on community notes, what people are finding is that the machine actually can write in that tone that more people like, and so it clears. And so you're starting to see to assist with scaling, producing more notes in a neutral tone, and then the voting is what goes to the people.
December 12, 2025 at 1:29 PM
permutans.bsky.social
reactance and its consequences: anti/woke antagonism jockeys, lockdown polarisation, …

researchers dissect it as “anger + negative cognition” www.researchgate.net/publication/...

> one of the basic claims of the theory is that high-pressure communicators are likely to be seen as threats to freedom
Reactance theory might be profitably applied to understanding failures in persuasive health communication but for one drawback: The developer of the theory contends that reactance cannot be measured. Rejecting this position, this paper develops four alternative conceptual perspectives on the nature of reactance (i.e., combinations of cognition and affect), then provides an empirical test of each. Two parallel studies were conducted, one advocating flossing (N = 196), the other urging students to limit their alcohol intake (N = 200). In both cases, a composite index of anger and negative cognitions fully mediated the effects of threat-to-freedom and trait reactance on attitude and intention. The data showed that, in fact, reactance can be operationalized as a composite of self-report indices of anger and negative cognitions, The implications for persuasive communication, in general, are considered as well the specific fındings for flossing and drinking.
December 12, 2025 at 2:55 AM
permutans.bsky.social
nanobind github.com/wjakob/nanob...
nanobind repo README featuring a hero image of a snake eating a glowing C (after having swallowed the ++) surrounded by diatoms drawn in the style of Haeckel
December 11, 2025 at 6:52 PM