Louis Maddox

Louis Maddox

@permutans.bsky.social

Combinatorially curious https://spin.systems

248 followers135 following4.4K posts

Speculative knowledge to practical kn. implies the possibility of spec. prac. kn. (like spec dec, but for intentional actors: “given that we can foresee the agent is about to perform X, what would the agent in [either of possible worlds] benefit from knowing” ➡️ go hunt for that)

Enlightened approach to pairing an agent session with a trace interpreter LLM ‘guardian’ chat: tell the latter: “as well as helping me interpret, if you can go and run any web searches that might shed light or encourage aha moments of realisation (not forced, but open to them)”

mibz2: a high-performance bz2 decoding library built around memory-level parallelism lmmx.github.io/mibz2/ Speeds up bzip2 decoding by interleaving independent inverse-BWT traversals, keeping multiple memory accesses in flight, plus SIMD and block-level parallelism boosters 💫

Louis Maddox@permutans.bsky.social

Counterintuitive approach to perf hunting: look for where the CPU is waiting, not (just) where it is working Claude came up with a trick to create a bz2 decoder 1.76× as fast as lbzip2 via memory-level parallelism (had identified "memory-latency bound" & then rolled its own PMU measurements)

If you ask an agent to explain their goals during a session it gives an account pretty fit for docs, akin to the program’s “ideal user” story It produces accounts structured in chained “what for?” Q&A, to “establish the final reason guiding choice of means” up to “one final why”

Louis Maddox@permutans.bsky.social

On agent /goal setting with Anscombe’s builder (§45) Claude notes an agent can be a foreman as much as a worker 📜 Anscombean SKILL to elicit an intention (speculating that this can effectively structure a user’s expression of their intention) github.com/lmmx/anscomb...

Claude suggests a /btw-style chatbot “guardian” (reading an agent trace) given this prompt would go from speculative to practical knowledge by attending to overall ambitions, and user from Anscombe’s detective (knows what happened) to Anscombe’s builder (knows what is being done)

Theory of vibe φ'ing The vibe coder (&c.) "is not like a man merely considering speculatively how a thing might be done; such a man can leave many points unsettled, but this man must settle everything in a _right_ order. _His_ knowledge of what is done id practical knowledge."

On agent /goal setting with Anscombe’s builder (§45) Claude notes an agent can be a foreman as much as a worker 📜 Anscombean SKILL to elicit an intention (speculating that this can effectively structure a user’s expression of their intention) github.com/lmmx/anscomb...

Counterintuitive approach to perf hunting: look for where the CPU is waiting, not (just) where it is working Claude came up with a trick to create a bz2 decoder 1.76× as fast as lbzip2 via memory-level parallelism (had identified "memory-latency bound" & then rolled its own PMU measurements)

Louis Maddox reposted

Skimming this new PDF, I feel some similarities to the sub-n*lg(n) FFT writeup by David Harvey. It's all like breaking the n^3 matmul bound: instead of doing many independent mat-vec-mul units of work, do them all at once, exploiting structure to save a few operations. bsky.app/profile/curi...

The glee with which I just explained my recollection of JL Austin's Goldfinch to a couple of bird watchers (who had seen 46 goldfinches on their app, apparently it's not even that rare)...... 🐥

Concept: semantic noise cancelling (or "context aware noise cancelling" or "selective hearing"), as an extension of turn-taking audio models to include not just syntactic breaks but attunement to relevant statements – those either pertinent to or direct address to the user/hearer

I think the reason is that "agent" is the most general term for "a speaker who acts", and to say more you'd want to circumscribe that identity, but so far all we've had is orchestration-determined identities (which infamously led to asemic collapse in Gas Town), not yet others...

Broke: prompt engineers rebranding to ‘context’/‘loop’/‘agentic’ engineers Bespoke: prompt engineers rebranding to motivational speakers [unironic]

Currently mulling over (relatedly): What if we could probe unstructured remote data orders of magnitude faster // what would it take to do so? How can we write output before having finished reading input // what source/destination dataset-level properties are required to do so?

Soft goal deadlines seem like a natural way to elicit agent prioritisation (as opposed to hard ‘pencils down’ deadlines that would drop work in flight at the end)

Louis Maddox reposted

“Ideas are like fish. If you want to catch little fish, you can stay in the shallow water. But if you want to catch the big fish, you’ve got to go deeper. Down deep, the fish are more powerful and more pure. They’re huge and abstract. And they’re very beautiful.” — David Lynch

Benchmark tracker 📉 lmmx.github.io/wikidata-pq/... ⏱️ Pipeline wall time 15s → 3s and CPU 41s → 6s
🧠 Peak RSS per worker 1.8 GB → 242 MB
🚀 Expected release ETA 60h → 7.5h Big wins: building tables mid-download (−16% ETA), overlapping stages (−8%) & merging col chunks (−4%)

Louis Maddox@permutans.bsky.social

Wikidata GH runner ingest speedrun any % lfg

Do people have different /goal modes for like "away from keyboard" (merge PRs, ship packages) vs. "asleep" (merge only and use git shims pending approval) ?

Perf trick du jour: fuse download with computation > use the filesystem as a producer–consumer buffer, with .part, the finished filename, and .bad signalling the download's state. i.e. quarantine corrupted files by renaming to .bad, signalling to delete any downstream artifacts

Louis Maddox reposted

is this a secure private sigil?

zooko2.bsky.social@zooko2.bsky.social

I made base-1353 encoding, so I can fit 256-bit values, like “96cca2a0c8a3263c74ef36274abf55f410d992034d16e8285811a9ad3fc59c13”, into short 26 character strings, like “ᓒᔳᣏዧᐎᑙᖜዌᙔᐸƫȶƃᒨዊጶᙵኻѫyፐᎉᑫᖫᘪᒩ”. (Also, I think these sigils are beautiful.) johnservil.github.io/sigil/play.h...

Agent autounblocking in software dev might be civilisation's best chance to shunt everyone onto the Trusted Publishing cyberapocalypto happy path fr

“Content-defined chunking is not a publicly accessible arrow-rs capability but it’s fine we’ll just do raw row group concatenation in memory” 😮‍💨🧨

“CI already has a linux x86_64 wheel built, so I'll download that instead of waiting on my local build and run the new tests against it” — traces of the utterly agentic