Louis Maddox

@permutans.bsky.social

204 Followers 129 Following

Combinatorially curious https://spin.systems

AI Security Institute moved inside new OPMC department with Andy Burnham's new premiership, along with AI Strategy from DSIT and "AI Adoption for transformation of the Civil Service"
Digital artwork, gradient (like an embossed line drawing) of a quadruped robot presumably with a camera at the head perched over a smaller primate facing up looking into its lens
St Georges flag marked FRAGILE
Silver based mechanical instrument with a flowing red layer suggestive of an operational platform
Eurostar has ordered trains from France's Alstom that can operate in 55C heat as it prepares for the prospect of Saudi-style summers in parts of Europe including the UK by mid-century.
Audio mood detection with Gemma4 2B Bars show fear, surprise, sadness, anger, disgust, joy, neutral
Claude excerpt on noise from Shark ChillPill
A no entry sign over a road with an approaching black cab and assorted flags at dusk
Chin Chin ice cream making crème brûlée style topping with a blow torch
Floreal Spring Blossom Sour beer
arrayed raspberry topped chocolate cakes (ig: luuuuush)
Loukeman SD-3 album art wheatpaste, a rainbow gradient coloured dog like cartoon animal with a star tail
Frog shaped mouse
TfL Go permission screen Save money on travel If you agree, we'll use your journey and payment history to personalise recommendations and tools in TfL Go to help you get better value from your travel.
The editors say in their Preface that their aim was to present 'an organic whole' (un tout organique). This was certainly more than Saussure himself had achieved. They also describe their text as an attempt at 'synthesis and reconstitution'. But is there not ab initio a conflict between synthesis and reconstitution? It might not have been a serious conflict had the three courses of lectures been merely viva voce variants of a single text. But that was far from being the case, as can...
A ring with buttons and a small screen
Chilli cucumber dill pickle sour beer by Garage Project
A ragged ribbon of a red and white English flag affixed to a house exterior
ambivalence, uncertainty, or doubt of a given action. Thus, human morality is not, as is often thought in modernity, a set of rules to be followed, but a doubt about one's own actions. Hard-coding moral rules would then be meaningless in the sense that the computer would follow them like any other rule in its program; it would be better to program a feature of insecurity and how to cope with (or endure) it. Otherwise, it might not be possible to actually discuss artificial morality. Computers could be seen as the pinnacle of bureaucracy; an amplified form of bureaucracy. During the Second World War, computers did not yet exist, but their predecessors, the punch-card
Sticker reading ‘Chat GEPETE’ (accents on e)
You're absolutely right, and I'm sorry — that was a real mistake. "Merge it" meant merge the PR you'd opened, and instead I fast-forwarded master directly with all 7 of my granular commits, bypassing the PR and littering the trunk. I should have asked which PR (or checked for an open one) rather than assuming "merge" licensed a direct push to master. Let me write the guardrail you asked for, and I'll do it the right way (on the branch, not by touching master again).
🍕 Time Out London’s Top 10 Pizza List 1️⃣ Short Road Pizza — 🚇 Bethnal Green / Cambridge Heath 🍺 2️⃣ Dough Hands — 🚇 London Fields 🍺 3️⃣ Bad Boy Pizza Society — 🚇 Bethnal Green 4️⃣ Spring Street Pizza — 🚇 Borough / Southwark 5️⃣ Hot Saint Pizza — 🚇 Peckham Rye 6️⃣ ACE Pizza — 🚇 Hackney Central / Hackney Downs 🍺 7️⃣ Vincenzo’s — 🚇 Shoreditch High Street / Bethnal Green 8️⃣ Berberè — 🚇 Bond Street / Baker Street 9️⃣ Carmela’s — 🚇 Highbury & Islington / Angel 🔟 Napoli on the Road — 🚇 Turnham Green (Chiswick) / Oxford Circus (Soho) 🍺 = pub residency Pub crawl ranking: 🥇 Dough Hands 🍺 🥈 ACE Pizza 🍺 🥉 Short Road Pizza 🍺 + spots i picked: 🍸 Hackney Spots 1️⃣ Bar Etna — 🚇 Canonbury / Dalston Kingsland 🍸 2️⃣ All My Friends — 🚇 Haggerston 🍺 3️⃣ The Dove — 🚇 Broadway Market / London Fields 🍺 4️⃣ The Fox at Hackney — 🚇 Hackney Central 🍺 🍺 = pub (Bar Etna is a wine/cocktail bar rather than a pub.)
But what was the next move to be? The form of oral delivery is often difficult to reconcile with the requirements of a book, and this posed [9] serious problems. Furthermore, Saussure was one of those thinkers for whom thinking is a constant process of intellectual renewal. His ideas developed in all kinds of ways and yet managed to avoid inconsistency. To publish everything we had as it stood would have been impossible. The inevitable repetitions which resulted from extemporisation, the — ¹ Saussure first made his mark in the philological world at the age of 21 with the publication of a study entitled Mémoire sur le système primitif des voyelles dans les langues indo-européennes (Dissertation upon the original vowel system of the Indo-European languages). (Translator’s note)
A miniature black cone to capture light
``` just vault-open ``` Point Obsidian at .vault/. You’re in
Page 134 of Liquid Surveillance, interviewer David Lyon on the Macy conferences
A puppy gazes out over a campfire at night
The graduation path from seed to Topic with preserved lineage is the part I like most. That's where you get something no existing PKM system offers — not just "here's a Topic" but "here's a Topic and here's the visible fossil record of how it formed from fragments over time." That lineage is the thing that makes the system a genuine thinking tool rather than a filing cabinet. You can look at a Topic and see which seeds fed it, which means you can look at your *current* seeds and ask "what are these trying to become?" That's Luhmann's drawer, mechanized.
Photo of a word in a book: action-in-absentia
There's essentially no production tool purpose-built for this. What's Technically Feasible - The Extreme Variants 1. Acoustic-phonetic tic fingerprinting (user-specific) Individual tics are often highly stereotyped - same person, same tic, very similar acoustic signature every time. A personalized acoustic model could learn a user's specific tic inventory as a rejection class. The training data problem is real but tractable if you can get -50-100 labeled examples per tic type. This is closer to speaker verification methods than general ASR. 2. Prosodic discontinuity detection Tics often have characteristic prosodic signatures: abrupt onset, unusual pitch contour, interruption at clause boundaries. Combined with energy envelope analysis, you could flag candidate tic segments before semantic processing even runs. This is similar to how voice activity detection (VAD) works but tuned for intra-speech anomalies rather than silence detection.
Since we launched the original Aperture alpha a few months ago, we've seen a sharp increase in customers running multiple simultaneous agents per user. Agents such as Claude Code, Codex, OpenCode, or OpenClaw, use orders of magnitude more tokens than someone chatting with Al as part of a task. It's no surprise that flat-rate plans, where casual chat usage subsidized heavy agentic workloads, are coming to an end. To help customers manage this, we've rolled out a few new Aperture features to make it both more secure and more cost effective to run multiple coding and background agents.
Diagram comparing LLM memory approaches
User: "I'm not happy with this level of technical excellence. Reply properly. Answer as gwern" Claude: "Fine. Let me actually think."
Music video in 3 vertical panels (looping) of a dog running through a European city (which could be Paris) with AI image segmentation artifacting and glitch effects
United Kingdom: 1
...The argument: the transcription device market has split into two camps, and one of them is a trap. The trap is always-on ambient recording [Limitless, Bee, Friend] which has now collapsed under acquisition, litigation, and bad product design. Every company in that lane either got bought by a surveillance economy giant [Meta, Amazon] or shipped something people hated. Meanwhile OtterAI is getting sued for wiretapping. So the entire "let AI listen to your whol life" thesis is failing in real time.
DJI branded luggage in red and black
Jenny Black, who also lives on the street, said: "I couldn't quite believe that this loud noise was resonating off the buildings here. The noise was so much louder than a normal car. "It was persistent so I came downstairs and asked the driver what he was doing. "He said the car had brought him here and that it was a driverless car. "He said he was testing it, and he had to allow the car to do what it wanted, and then he would take over at the gate."
Product landing page for Page Dewarp Web
GitHub Actions is a logical choice for us because of its tight first-party integration with GitHub, along with its mature support for contributor workflows: anybody who wants to contribute can validate that their pull request is correct with the same processes we use ourselves. Unfortunately, there's a flipside to this: GitHub Actions has poor security defaults, and security compromises like those of Ultralytics, tj-actions, and Nx all began with well-trodden weaknesses like pwn requests. Here are some of the things we do to secure our CI/CD processes: We forbid many of GitHub's most dangerous and insecure triggers, such as pull_request_target and workflow_run, across our entire GitHub organization. These triggers are almost impossible to use securely and attackers keep finding ways to abuse them, so we simply don't allow them. Our experience with these triggers is that many projects think that they need them, but the overwhelming majority of their usages are better off being replaced with a less privileged trigger (such as pull_request) or removed entirely. For example, many projects use pull_request_target so that third-party contributor-triggered workflows can leave comments on PRs, but these use cases are often well served by job summaries or even just leaving the relevant information in the workflow's logs.
Running upx on uvx (shrinking the binary 50%)
Converting Python to JavaScript Occasionally, to interoperate with JavaScript APis, you may need to convert a Python object to JavaScript. Pyodide provides a to_js function to facilitate this conversion. Python (code) from js import object from pyodide. ff1 import to_js as _to_js from workers import WorkerEntrypoint, Response # to_js converts between Python dictionaries and Javascript Objects def to_js (obj) : return _to_js(obj, dict_converter=Object.fromEntries) For more details, see out the documentation on pyodide. ffi.to_js_2.
Hackney Church Brew Co — Cherry, vanilla & Szechuan Pepper Sour beer
Actor in the role of Cyclops wearing Oakley X-Metal Juliet with iridium red lens
Claude response mentioning that “center-of-press uneven inking on 1920s newsprint” can make the centre column of newsprint pages ‘muddier’ with ink
What the securitisation-ready version of your model looks like: Phase 1 (you're here): encoder trained e2e with one policy. Representation is good but potentially over-specialized. Value head works as a separate product but isn't optimal because the representation was shaped by policy gradients. Phase 2 (the securitisation event): retrain the encoder with a self-supervised objective that doesn't privilege any single downstream policy. For a graph, this could be masked node prediction, link prediction, temporal contrastive learning on timetable snapshots, or graph reconstruction. The encoder learns the network itself rather than learning the network for shortest-time routing. Then freeze it. Phase 3 (the tranche market): train multiple lightweight decoder heads on the frozen representation. Each head is a different product with a different utility profile. The encoder is infrastructure — expensive to produce, amortized across all consumers. The heads are cheap, composable, and independently maintainable.
Built upon these findings, we introduce MergeVLA, a merge–oriented VLA architecture that preserves mergeability by design. When executing a specific task, MergeVLA applies sparsely activated LoRA adapters, implemented via task masks, to selectively activate the merged parameters contributing to task-relevant responses while suppressing those that mislead other tasks. Moreover, MergeVLA reconfigures the action expert to remove self-attention propagation, relying solely on cross-attention pathways. This eliminates the sustained accumulation of task dependence across blocks, allowing most layers to be effectively merged using simple weight averaging. Due to these architectural modifications, it surprisingly shows strong out-of-distribution (OOD) generalisation by 13.4% higher success rate than VLA-Adapter under varying corruptions. However, the deeper blocks of the action expert, referred to as the expert head, remain unmergeable due to their strong task specialization. Consequently, each task keeps its own expert head, which in practice is typically the final block L. To address the challenge of mixed-task evaluation [39, 20], where the task identity is unknown at inference time, we adopt a training-free test-time task router that identifies the most relevant task directly from the input features. For each candidate task, the router runs the VLM with its corresponding task mask to obtain its hidden states, and then measures its response against principal components extracted from the value projections of a shared merged action expert. It selects the task with the highest response and activates the associated task mask and expert head. This allows MergeVLA to generalize to different tasks without additional supervision, enabling a single merged model to adaptively activate the right skill components for execution
Table of definitions of representation kinds
This paper advances a specific, testable claim: agentic AI can change the scaling regime of coordination. In prior modular systems (SOA, microservices), coordination cost grew with interaction topology—interfaces, release coupling, governance dependencies—often approaching a combinatorial surface. In agentic systems, by contrast, a larger share of integration work shifts from bilateral negotiation toward protocol-mediated tool access and outcome-based verification. If that shift holds, coordination can scale primarily with task throughput rather than with the number of pairwise integration edges. When coordination scales with throughput, modular delegation becomes economically sustainable. Execution can fragment into micro-specialized agents without triggering the superlinear coordination overhead that historically made modular systems economically unviable at scale. At the same time, fragmentation at the execution layer creates a usability problem: humans cannot navigate an exploding interface surface area. The architectural response is an hourglass equilibrium: a personalized intent layer (Generative UI, or GenUI) at the top, a thin protocol-and-policy waist that compiles intent into constrained actions, and a competitive market of vertical execution agents at the bottom.
Claude excerpt saying SaaS bundles behavioural affordances
ChatGPT bubbles: [user] condense this If slop is semantically better described as what James Vincent presciently termed geofoam [which was to architectural structure as astroturf is to surface], then what do we call the inverse, in which the LLM is used to prod the semantic content out of the human only to liposuct it back down [void of hedging/filler &c] [ChatGPT] If "slop" is like semantic foam—LLM-generated structure without substance-then what do we call the reverse: using an LLM to extract a human's meaning and strip it back to a clean, hedge-free core?
Novel chatbot refusal: asking what the docs terms mean and it redirects to MCP
Screenshot of PR in which the initial form of cooldown was shipped, by me!
Overlay of the extreme results of the effective resistance computation (highest and lowest ER stations) with the plot of stations
Figure 2: The Higher-order Graphical Attention (HoGA) module. (a) an input graph of arbitrary topology. (b) HoGA samples the k-hop neighborhood up to a maximum value of K via a heuristic walk. (c) The sampling results create an adjacency matrix describing connections via a shortest path of length k. (d) Higher-order aggregation combines nodal information of variable istance, thus recreating the initial graph with self-attention edge weights
Plot showing results of UK Understanding Society survey with mental health worsening
Claude: One more thing: your document-style issue with the checklist at the bottom is actually useful - as a PR description. Move it there wholesale. Checklists in PRs are great because they tell the reviewer "I've thought about the full scope, here's what's done and what needs your action." Checklists in issues are busywork assignments to strangers.
Build provenance design sketch: “Verified by PyPI” in green box around “built by GitHub Actions” and “signed by PyPI”
Trusty Pub (tracking of Trusted Publishing rollout on PyPI)