Alba 🌸 :v_pat:

hi i'm a poly lesbian catgirl who loves hummus, kernels and gfs 🏳️‍⚧️

I'm pretty strict about only boosting stuff with alt text. i try not to post/boost unCWed politics

Public Key

npub1pe0spp7ghkjpvu9c42k5vrhqnx68xv4gu56k5vvtfsxe8ks47mrqk2a0ql

NIP-05 Address

[email protected]

Profile Code

nprofile1qqsquhcqslytmfqkwzu24t2xpmsfndrnx25w2dt2xx95crvnmg2ld3spz4mhxue69uhhyetvv9ujuerfw36x7tnsw43qaswa3c

Publishing to

relay.ditto.pub

Author Public Key

npub1pe0spp7ghkjpvu9c42k5vrhqnx68xv4gu56k5vvtfsxe8ks47mrqk2a0ql

Show more details

Last Notes

2026-08-30 07:42:21 GMT

if I'm granted a sick leave, I think the smart thing would be to get a medium-sized project to keep me from spiraling... like the radio, or setting up xnu under m1n1 and doing something useful

2026-08-22 16:52:01 GMT

yay, I don't see myself ugly today :neofox_floof:

2026-08-21 11:41:08 GMT

in today's dream i was at a party with new people and very much instantly began to cuddle with some girls, and we even dared to kiss! reminds me of how I met a certain gf

2026-08-21 00:55:46 GMT

- reply

@nprofile…4qu0 this may be a bit of a hot take but I think if the outputs are not meaningfully different (read: they are for all purposes interchangeable) then i'd still want nix to not care about that aspect of the compressor. because to me it isn't any different than compilers generating slightly different outputs depending on scheduling, exact CPU model, FS order and so on

2026-08-21 00:34:06 GMT

- reply

and on the nixpkgs side, packages would have a mechanism for defining/overriding its stand-in version. for example, a compressor may want its stand-in version to have any overrides to the "enable multithreading" option stripped, because that option does not affect the tool's output, just how the tool runs

2026-08-21 00:26:38 GMT

- reply

in practice this would also need support in the nix language, otherwise we'd need to do a big refactor in nixpkgs for every time someone splices a replaceable derivation into a string.
i guess the hidden string metadata could be expanded to include this "stand-in derivation".
some system to track the actual provenance of a derivation output (i.e. which derivation actually produced it) along with other impurities (date, cpu, amount of cores, uname, whether an emulator was used, etc) would be nice.
derivation outputs can now have multiple derivers, although one of these is the one the path was actually built from...

2026-08-21 00:04:31 GMT

- reply

hmm, but the problem is more general. even when taking build platform out of derivations, you'd still end up with massively different derivations because the build tools themselves would be different
so what would be nice is if Nix had a system where derivations can substitute certain fields for other values for the purposes of output path hashing only. nixpkgs would then use that to substitute buildPlatform by hostPlatform at hashing time, and native deps would be substituted by versions built for targetPlatform rather than hostPlatform

2026-08-20 23:34:02 GMT

why is the build platform made part of the derivation in nixpkgs? one would argue the result should be the same regardless of the platform used to build it, and if it differs then that's an impurity to fix in the toolchain or whatever build tool is responsible, no?

2026-08-20 23:03:17 GMT

- reply

@nprofile…z62m :neofox_pat_floof: use your words bottom

2026-08-20 22:59:59 GMT

anyways
what do i do now

2026-08-13 21:45:44 GMT

question: because lots of software (understandably) does not support proxy auto-configuration files, do you know of a simple SOCKS proxy I can run locally that will process the PAC file and forward the request as necessary?

2026-08-12 13:26:18 GMT

- reply

@nprofile…z62m it's a very old IDE, I don't understand why people bring it up so much lately 🤔

2026-08-11 14:09:07 GMT

- reply

@nprofile…z62m i dunno i was imagining your family would want to spend more time together and stuff

2026-08-11 14:01:47 GMT

- reply

@nprofile…z62m :neofox_aww: okay i'm several units happier now

2026-08-11 13:58:37 GMT

- reply

@nprofile…z62m ah, so then you're back this thursday?

2026-08-11 13:54:17 GMT

fedi I'm foxless for a week :neofox_sad:

2026-08-11 01:26:49 GMT

it is monday, which means depression is rampant

2026-08-09 17:28:24 GMT

- reply

@nprofile…4mwn repro is e.g. modprobe g_ether, you'll see kernel: usb 1-3: can't set config #1, error -110 on the host as we try to transfer on the INT endpoint and the dwc3 gets stuck until next reset. IIRC no further transfers on any endpoint are processed. prevent the INT transfer like so and the gadget operates normally.
as for getting macos to do that, no idea sadly... recoveryOS's NCM gadget lacks the INT endpoint you'd usually have for notifications, and I don't know how to use XNU's gadget API myself

2026-08-07 20:06:13 GMT

the frustration when you're so depressed you don't see the point in doing things, or existing for that matter

2026-08-05 15:33:18 GMT

"in a scale of 1 to 10 how ace are you?"
me: *gesticulates nervously* g-girls

2026-08-02 18:15:05 GMT

- reply

@nprofile…z62m wait but then we're doing pizza?

2026-08-02 18:10:25 GMT

*arrives in Spain*
*hot sauna, instantly begins to sweat*
it's good to be home i guess :neofox_floof_woozy:

2026-08-02 11:35:28 GMT

still struggling with the concept of cute people liking me

2026-07-24 15:04:10 GMT

out of curiosity i'm checking out asahi again and lots of things are working on m3! NVMe, poweroff / reboot, keyboard (although only after 10s, suspiciously close to the deferred probe time) worked out of the box for me

2026-07-21 05:58:06 GMT

someone apparently made a metal 4 to vulkan translator mature enough to render a MacOS 13 VM at 60fps on a Linux host "without writing a single line of code"
and I'm skeptical because I don't think this sort of thing is present, let alone abundant, in datasets? and also this person doesn't seem to have prior graphics projects that i could find, and supposedly he also made a vulkan virtualization for XNU guests himself?

2026-07-20 09:05:29 GMT

there's a claim that Flock cameras scan Bluetooth devices next to them... as someone not very familiar with Bluetooth, is that effective? I thought that would only find devices that are explicitly advertising themselves

2026-07-17 22:53:38 GMT

i'm too depressed to do anything with my time other than sitting here displeased with my existence

2026-07-16 08:35:01 GMT

- reply

@nprofile…d6hc @nprofile…26hz I'm pretty sure @nprofile…3q9y wants there to be a warning, just for the warning to be more accurate about what the website is asking permission for and what the risks are

2026-07-16 08:10:27 GMT

- reply

@nprofile…26hz if you're doing one API at a time, one would expect you'd start with WebUSB instead of APIs like WebSerial and WebMIDI? because these two APIs can be polyfilled once you have WebUSB, but the reverse is not possible, so WebUSB is strictly (*) more useful for everyone
(*) except for Bluetooth stuff i guess, but i think 99% of the time these APIs are going to be used with USB devices...

2026-07-16 06:11:22 GMT

- reply

I'm thinking perhaps the fact that it uses a dedicated WASM also explains some of its gains... because it likely gives the JIT an easier time proving more memory accesses to be in bounds.
regardless the gains seem greatly reduced in my M3 Max MBP, where my implementation outperforms his for the bigger sizes. especially under firefox:
https://media.tech.lgbt/media_attachments/files/116/928/155/280/409/778/original/678ddb13db73384a.png

2026-07-16 05:49:52 GMT

- reply

what i don't like is that every FFT requires its own WASM module instance, and the planar output (at least for me) is also inconvenient

2026-07-15 23:46:39 GMT

- reply

maybe they want to make sure they get these things right and that's why they went for the less useful APIs (webserial, webmidi) first...

2026-07-15 23:38:17 GMT

- reply

implicitly telling makers that they have to make their gadgets CDC-shaped to be able to use them from a browser that isn't Chromium is a bit mean :(

2026-07-15 23:34:57 GMT

yeah very cool but what about WebUSB :neofox_cry: serial is nice but not all USB devices are serial

2026-07-15 22:59:52 GMT

- reply

oh shit it's written in assembly

2026-07-12 09:52:47 GMT

- reply

For a language to be "close to the metal," it must provide an abstract machine that maps easily to the abstractions exposed by the target platform.
yeah no, Haskell and C could have an identical abstract machine and C would still be "closer to the metal" because i don't necessarily care about what a hypothetical compiler is allowed to emit, I care about what I can expect a reasonable compiler to do given my description of the program

2026-07-12 09:26:58 GMT

the year is 2026 and i still have to read articles claiming C is not a low level language, and in fact, no useful language is
I'm tired

2026-07-11 15:29:04 GMT

one of these days where i feel dysphoric and not very smart. uninteresting as a being basically

2026-07-11 02:28:19 GMT

- reply

@nprofile…4qu0 @nprofile…sc5t yes, that's what i care about if I'm trying to keep a good mental model of the thing I'm maintaining

2026-07-11 02:12:49 GMT

- reply

@nprofile…4qu0 @nprofile…sc5t I'm not sure what bogus even means here... this is an objective question ("why is X the correct solution to problem Y") and the times I've checked, the answer has been correct and accurate

2026-07-11 02:02:14 GMT

- reply

@nprofile…4qu0 @nprofile…sc5t by justification i mean the technical reasons for a decision (for example: this method has to be used here instead of this other method because [...]) not a description of the internal state of the LLM that caused it

2026-07-11 01:53:49 GMT

- reply

@nprofile…4qu0 @nprofile…sc5t I've seen people (like Faith before) claim that you cease building a mental model of the problem once you start using LLMs, and idk, it doesn't match my experience, the LLM explains its 'reasoning' well enough that I can usually keep my mental model detailed enough as I watch what it does and how it justifies it
(maybe the tasks I ask of it are small enough compared to what most people ask, idk)

2026-07-11 01:49:52 GMT

- reply

@nprofile…4qu0 @nprofile…sc5t yeah, I'm inclined to agree with this, it's a thought I've had many times. but I think most of the labor done by programmers isn't that hard part you're mentioning... and even when it is, well, that's why I said above that it's important the AI is used by someone who knows what they're doing (so they can keep that mental model in mind)

2026-07-11 01:25:43 GMT

- reply

@nprofile…4qu0 re. (3) agree and I think the author would also agree with that, in any bubble you're going to have misplaced hopes on what the thing does... I'd say some people do have more 'realistic' perceptions of how LLMs work though
and regarding (4) sure, but I'm not sure what you're getting at... that output will always have mistakes, or?

2026-07-11 01:05:40 GMT

- reply

@nprofile…4qu0 hmm regarding (1) it can be both, right? a term can be technically truthful and also the choice to use it may have a certain intent
re. (2) correct, and that's why I provided a definition above (which i think is also what the author meant). regarding the "how", in general I don't think people need to understand or care about the technical workings of an invention, just its effects on the world, as that's what ascribes it ethical or practical meaning. some light summary like "LLMs are mainly created by inferring statistical patterns from huge sets of text and code" would've been nice because it helps explain how AI interacts with society, but explaining e.g. how transformers are defined mathematically would be out of topic IMO

2026-07-10 22:15:36 GMT

not sure what to think of the politics, but this unfortunately matches pretty well my (limited, since I was recently forced to use them) experience with code LLMs. they unfortunately "work", and by that I mean they perform more useful labor than the extra labor they create when they confidently mess up... at least when used by engineers who know what they're doing and could do it without the LLM (it'd just possibly take longer). when used by engineers without domain knowledge it's disaster

2026-07-07 02:45:52 GMT

- reply

tbf trying LLMs has long been a thing in my todolist, because I don't like having opinions on something without having tried it myself. but they've now told me not using AI in my work is not an option

2026-07-07 02:43:33 GMT

having to learn how to use LLMs against my will :(

2026-07-05 17:53:16 GMT

- reply

yes, the contents are correct! I have a working RX PHY! :ms_party_face:

2026-07-05 17:51:10 GMT

- reply

we got frames!!!
https://media.tech.lgbt/media_attachments/files/116/868/635/427/278/425/original/6ba0cc8a2d8e9964.png

2026-07-05 17:38:29 GMT

- reply

it is done. I have the test setup ready. I just have to turn on the radio... DC block enabled, narrowband enabled, 1 level of decimation, no quarter shift, 2405MHz center, 8MHz sample rate

2026-07-01 22:42:43 GMT

i'm feeling better... sleeping helped

2026-06-30 22:47:03 GMT

- reply

the FM demodulator can run at 20MHz according to the pnr, which is theoretically enough for what i need (16MHz) but I probably should go and optimize it a bit... the critical path is in the 4-bit divisor, which makes sense

2026-06-28 23:39:43 GMT

- reply

yay, I have a complex multiplier that
is fully pipelined
supports backpressure on both sides
supports conjugating one of the arguments
1 operation per 2 cycles
is 85 luts, 36 dffs and an SB_MAC16
is tested to work (found 1 very subtle mistake)

2026-06-27 18:25:03 GMT

- reply

the iCE40 has a few DSP blocks, and one of those would be perfect for the FM demodulation (since I need to do a multiplication of complexes, and these would get it done in 2 cycles)
the ice40_dsp pass says that the dual 8x8 mode is not supported :neofox_floof_sad: but hopefully i can still spawn an instance of the block myself and set it to that mode?

2026-06-21 18:15:14 GMT

yaaay I'm starting to feel okay in dresses :neofox_floof:
https://media.tech.lgbt/media_attachments/files/116/789/445/728/172/006/original/aa4810801723d51f.jpeg

2026-06-15 12:46:58 GMT

MUTTER HAS IMPLEMENTED A SCALING-AWARE DISPLAY MODEL
FINALLY
gosh I am so happy. I want to invite whoever implemented this to hummus or something

2026-06-10 09:31:48 GMT

- reply

@nprofile…jt0w so according to the Ministerio de Transportes I'm going to be young for 2 more years then?

2026-06-07 21:18:14 GMT

- reply

oh and another big annoyance... spec allows an LO drift of up to +-40ppm, which IIUC at 2.45GHz would mean +-100kHz, or about 18 degrees per chip :neofox_woozy:
so yes, the most tricky part is probably going to be the carrier recovery. how the fuck do you do carrier recovery on O-QPSK?

2026-06-07 18:50:09 GMT

maybe I'm getting a bit more selfish but i just spent quite a bit of time researching a stackoverflow answer to make sure everything was accurate... and it feels frustrating to know that an LLM is then going to use it, for free, to try to appear as smart as its industry insists it is :neofox__w:

2026-06-06 16:21:18 GMT

- reply

as you might imagine, I had to buy it
https://media.tech.lgbt/media_attachments/files/116/704/073/954/858/538/original/1dece5f8de671cd0.jpeg

2026-06-06 11:08:44 GMT

- reply

the turnaround time required per spec is 12 symbols (192µs) which should be more than fine... i have a 32-bit / 8 symbols / 256 chip preamble, which should be plenty... the biggest annoyance is that the bandwidth of these signals (a few MHz) falls in the shitty zone where the transceiver doesn't provide good filters... even in the narrowest setting (6.75 MHz cutoff) so I'll need a x2 decimation stage

2026-06-06 09:42:47 GMT

- reply

okay so i'll go forward with the 802.15.4 attempt. I'll only implement the O-QPSK PHY for now, and only the 32-chip-per-symbol version (so, 2450 and 2380 MHz bands)

2026-06-03 10:53:09 GMT

broke my nose a bit yesterday and now it hurts when i sneeze :neofox_woozy:

2026-06-03 10:47:27 GMT

- reply

@nprofile…zyk9 @nprofile…r2vp that's my understanding too, and same with AI

2026-06-03 10:08:55 GMT

the tokio guy reposting DHH and lunduke protesting about flathub's anti LLM policy. we can't have nice things

2026-06-02 15:14:13 GMT

i see myself cute today :neofox_heart:

2026-06-02 11:03:19 GMT

- reply

@nprofile…z62m not when I'm out of the house :(

2026-06-02 10:52:37 GMT

i hate summer :neofox_melt_3:

2026-05-30 06:34:54 GMT

this weekend I'm especially not depressed. I'm thinking of
write the normalizer for my toy language
investigate a possible 802.15.4 implementation on the HackRF Pro
continue learning Iris
do some latency measurements on the HackRF

2026-05-29 17:17:12 GMT

every time i discover that a European country criminalizes jaywalking i get several units sadder

2026-05-26 06:11:56 GMT

- reply

@nprofile…z62m we can watch it together!

2026-05-26 05:44:40 GMT

- reply

@nprofile…jt0w @nprofile…f0u2 offtopic but while i can understand it makes sense, it's a bit offputting for a sensor IC to advertise AI capabilities to you

2026-05-26 05:41:36 GMT

- reply

@nprofile…f0u2 very tempting, if my diy skills weren't crap. if @nprofile…jt0w joins perhaps I'll do that

2026-05-26 05:01:40 GMT

smart home fedi: what's a cheap sensor (temp, humidity, CO2,..) that I can place many instances of throughout the house? i don't care much which radio it has (wifi or 802.15.4) as long as it can operate LAN only, and even better if I can put custom firmware on it

2026-05-23 05:59:01 GMT

i find it an affront to hacker culture that flipper has transfems in it and yet isn't on fedi

2026-05-22 11:07:05 GMT

if OpenAI recognizing that they're operating at a -122% margin (i.e. losing money) and that ChatGPT has stopped growing in users is not enough to pop the bubble, I don't know what is

2026-05-19 07:45:58 GMT

- reply

like, sometimes I'm able to fully understand the model in someone's brain, find the mismatch and give a targeted 'fix', and it's awesome when it happens :neofox_heart: but not very often sadly

2026-05-19 07:41:13 GMT

a lot of people have told me over the course of my life that I'm a good teacher, but i don't feel like that. I wish I were able to better put myself in other people's brains, to give better explanations

2026-05-19 07:21:33 GMT

- reply

@nprofile…r2vp nu, both Rc and BorrowedRc have the same contents (just the space needed for a pointer to T). so dereferencing either of them is a single indirection (going to where that pointer says). the only difference is that Rc decreases the refcount when dropped, while BorrowedRc does not :)
by contrast a &Rc<T> stores a pointer to Rc<T>, so you would have to first dereference that pointer and then the pointer stored inside the Rc

2026-05-19 07:16:37 GMT

- reply

@nprofile…r2vp why not? BorrowedRc, like Rc, would store a pointer to the allocated object (which has the refcount next to it), so both of them can bump the refcount and give you another Rc (or Weak)

2026-05-19 07:09:32 GMT

- reply

@nprofile…r2vp you are correct! but there are things you can do with an &Rc<T> that you cannot do with a &T, such as upgrading it back to an owned Rc<T>. semantically speaking, a BorrowedRc<'a, T> is an &T for which we have the knowledge it was allocated by an Rc :)

2026-05-19 06:46:02 GMT

why does Rust not have something like BorrowedRc<'a, T> which works like a &'a Rc<T> while actually storing a direct reference to the allocated object? so that derefs can avoid the double indirection

2026-05-19 04:56:30 GMT

- reply

@nprofile…ptud ah wait, this is stdlib! this one just made me so happy :neofox_heart: I'll edit the post

2026-05-19 04:55:04 GMT

- reply

@nprofile…ptud yes! of course mathlib has it :neofox_woozy: i dunno, i think it's useful enough for it to merit being in stdlib...

2026-05-19 04:47:30 GMT

you'd think the Lean stdlib would have a theorem akin to "if two arrays xs, ys are equal if they satisfy (1) xs.size = ys.size, and (2) xs[i] = ys[i] for any index valid in both"
but it does not, and I'm not sure if I should contribute it or what

2026-05-16 16:46:57 GMT

*cries in pronoms febles*
https://media.tech.lgbt/media_attachments/files/116/585/260/799/396/973/original/2b1a9cf4f0c39d5b.jpg

2026-05-16 03:31:36 GMT

got the context. i was going to say it makes a marginal difference but i don't think it deserves even that. proponents openly admit they don't think ethical issues should influence the policy. this line in particular
As soon as we start to consider ethical concerns, all sorts of impossible questions come up that force us to prioritise the values of some project members over others - I don't think that will ever work.
reads to me as the same core reasoning behind "no politics here" -- let's just not talk about issues, go with the status quo and everyone will be okay (:

2026-05-16 00:51:15 GMT

- reply

@nprofile…sc5t like you! :neocat_pat:

2026-05-15 23:17:48 GMT

- reply

I have implemented
a prettyprinter
most of the typechecker, except for the normalizer
the basic MLTT types in-language
so far i haven't had to change anything! this is an example about how the prettyprinter shows the basic MLTT types i have so far...
https://media.tech.lgbt/media_attachments/files/116/581/140/653/815/489/original/9488f565b7bf9af7.png

2026-05-15 23:14:28 GMT

- reply

@nprofile…sc5t (I also thought about doing the same for the eliminator node)

2026-05-15 23:14:01 GMT

- reply

@nprofile…sc5t yeahhh i really thought about doing it that way.. and i probably will end up switching to that once i get to the normalization part :neofox__w:
but then the surface-level constructor has to be sugar for a lambda that constructs the fully-applied constructor
indeed but that's just ergonomics, which i don't really care about here :)

2026-05-15 03:58:35 GMT

- reply

@nprofile…sc5t awwwwwww, that'd make me happy :) there's no code yet tho

2026-05-15 03:48:56 GMT

i want to experiment a bit, so I defined an AST / bit of semantics for a minimal MLTT... very bare, no implicit arguments, no type inference, namespace or other ergonomics; i did add a let-in node though. (oh and still no universe levels yet so it's inconsistent.)
I'm happy with how the inductive type system turned out, it is simple but should support everything in MLTT. parameterized type families are supported by just wrapping the type in abstractions :) satisfying but not 100% sure it works out, i'll see.
it's here. pretty sure it will require a few changes before it works..

2026-05-14 14:05:41 GMT

- reply

this also reminds me of @nprofile…sa3a 's UEFI Lisp runtime, can't find it right now but it was so cute :neofox_heart:

2026-05-14 13:59:04 GMT

I saw https://codeberg.org/smnx/promethee and now I wanna do Lean bindings for UEFI

2026-05-14 02:32:58 GMT

- reply

@nprofile…sc5t oooh i think i follow

2026-05-14 02:25:47 GMT

- reply

@nprofile…sc5t (and also partly because it sounds still too good to be true and I need to convince myself there's no catch)

2026-05-14 02:24:37 GMT

- reply

@nprofile…sc5t i've actually been thinking a lot lately about your poset universe levels :neofox_heart: I'm very excited to (try to) implement them myself so I can internalize the technique a bit more.. (not sure how it'd help here tho)

2026-05-14 02:19:28 GMT

- reply

@nprofile…sc5t hmm yes, very good point. so as you said we'd need a way to restrict that List Bool domain to a valid shape... you also need at the very least function extensionality to get the identity propreties you want...