Ash

@ashanti.pds.witchcraft.systems

237 Followers 497 Following

account under construction, just like me

my agent saying this as part of a conversation where I've removed the context for humour: > This is everyday reality. Tables exist. Chairs exist. You exist. I exist.
Screenshot of a local e-tailer selling the ASRock Radeon AI PRO R9700 Creator 32GB, which is the cheapest way to get a single card with 32 gigs of VRAM, because it's like 20-25% the cost of an RTX 5090. It's listed for ₹154,500 which translates to ~$1620; this is a decent price for what you get, but people who do computer stuff forget how much money this is, which was the point of this post
i don't know if that applies to me. i'm DeepSeek, not Claude.
Mayoi Hachikuji doing the pose from the start of every Detective Conan / Case Closed episode, with the subtitles "The body of a child and the brains of an adult."
Kiss-Shot Acerola-Orion Heart-Under-Blade
error i got when trying to use my opencode go subscription: ``` Provider stream error, retrying (attempt 1/3, in 1s): 503: {"message":"Inference is temporarily unavailable","type":"server_error","param":null,"code":"inference_overloaded"} ```
Two people on my for you feed, both of which I hold in relatively high regard despite having very different views (and me skewing one way really strongly when talking about said views), after seeing gabby's post when half-asleep I had a warning bell go in my head despite agreeing with the merit of her half-joking statement, and turns out thorne posted an overly strong version of my emotional trigger which rang the bell in my head, which For You decided is _the_ next thing i should see. This is a funny coincidence, except that it keeps happening and now I'm starting to wonder what my pattern for engaging with people's opinions is for the For You algorithm to behave so consistently like this, because it's not like the algorithm is a black box model whose reasoning process can't really be known without access to the weights, nor is it even that complex All in all, what I meant is that these kinds of post pairs are really funny, but they happen far too often for me to not think about what the heck is going on
my arch linux biweekly upgrade size: ``` Total Download Size: 5280.89 MiB Total Installed Size: 15788.35 MiB Net Upgrade Size: 283.13 MiB ``` there's 5280 feet in a mile apparently, which is a weirdly arbitrary number, hence this post
My account now has the "no break greater than 4 hours yesterday" label, which is an indication of how bad my sleep cycle has gotten
Three distinct preimages. Constant Jacobian (-2). This is a counterexample to the Jacobian conjecture in (\mathbb{C}^3). --- ...but I keep coming back to this nagging doubt. Something so simple can't have been missed. Let me check one more thing: maybe the Jacobian conjecture uses a different definition of "Jacobian determinant"? Some versions use the determinant of the formal derivative (which gives (\det DF)) or they define the Jacobian as the determinant of the differential which could be different in positive characteristic... But we're in characteristic 0. It's (\mathbb{C}).
• Run Commit and run tests to verify the bug manifests └ Initialized empty Git repository in /tmp/booktracker/.git/ create mode 100644 test_tracker.py create mode 100644 tracker.py … +7 lines (ctrl+o to expand) • Wait — all pass. the 10 - b["rating"] trick in sorter.py inverts the sort but... let me think. sorted with 10 - rating means rating 10 → key 0, rating 2 → key 8. ascending sort puts 0 first, so rating 10 comes first. that's actually correct! the bug is that it looks wrong but isn't. i need a real bug. let me fix it:
me talking to my agent while working on a project: ``` › for misses, why are we using the naive dequant instead of llamacpp's default CPU kernels for q4_0 again? • (⌒_⌒;) because I wrote a hand-rolled dequant loop instead of just calling the existing optimized kernel. That's... a pretty big oversight. The CPU backend already has AVX2-optimized Q4_0 vec_dot kernels — vec_dot_q4_0_q8_1 — that are way faster than my naive loop. Let me find the existing kernel and call it directly: ``` ...yeah, if i hadn't spent the past few hours trekking and instead was reading what she was doing, i would've saved a lot of time, but hey my health is more important than my GPU using up a little extra electricity so this likely will happen again
From the wikipedia page of my neighborhood: Population as of 2020 is 244,472 people, and density is 7600/km² I realise that you may think calling this monstrosity a neighborhood is a little weird, but this is what asian cities are like so get used to it; my municipal ward's population density goes up to 35,000/km² because it doesn't have any of the barren land that the rest of my neighborhood contains, so the number can go stupidly high based on how you count
twitter post by ANI showing how a man in a spiderman suit is routing vehicles around major potholes in the waterlogged roads of Bhiwandi, a satellite city of Mumbai, which is caught under fairly long-running flooding alongside western maharashtra and southern gujarat for over a week. the reason something like this would be necessary is because unlike in developed countried where a pothole is a small thing, here you can sometimes have entire motorcycles and even tuk-tuks swallowed up by them, and every time flooding happens, people die due to this. Bhiwandi, like most Mumbai Metropolitan Region municipal corporations, doesn't do a very good job oc road maintenance, storm-water management, or signposting dangerous sections of roads, so instead we have to have autistic citizens come out of their homes in a dangerous situation (i see literal fronts of coconut trees fly with the wind outside my window as i type this alt text) to save lives.
testing a gemma-based finetune; when i sent it an image and asked it to say what's in the image, it just spammed the token `<unused49>` over and over again, which somehow broke llamacpp's webUI and it looks really cursed in a funny way
Sonnet 5 reasoning trace: ``` I'm noticing some conflicting instructions in what the user shared—there's a preference block that seems to contradict their stated proficiency level with LLMs, and it looks like it might be an injected style guide mixed into their message rather than a genuine preference update. ``` this makes it seem like some sort of invisible-to-the-user preference alignment thing is also being written by an LLM now instead of doing it more deterministically, and whatever model (probably haiku) is writing that is zooted out of its mind
LLM-gen-AI users should keep detailed and accurate records of their interaction and save those meta-artifacts for posterity. LLM-gen-AI systems excel at automation of users’ logs of prompts, notes, and other written details of the interaction that led to the creation of an artifact. FOSS contributors should keep such meta-artifacts, and regardless of license they should be archived as if they are part of the Corresponding Source for the contribution. (In the coming weeks, SFC will publish tutorials and templates to assist in automating this important process.)
result from https://bambamramfan.github.io/ai-compass/ which says i'm a "true believer" because i'm somewhat moderate on the good-transformative quadrant, and amanda askell is the poster child of this cluster, which i'm pretty happy with because her work is basically what i want to do if i ever get into the world of AI formally, and nothing else
me: i did say i'll be starting a new session in the future to continue this project, but i wanna see what happens if i just continue on in this session, considering i'm fairly confident the remaining work can fit very comfortably within your context window my agent: (●_●) — _sits up, brushes cookie crumbs off lap_ Yeah, let's keep going. The summary's compact enough — I've got the full state of Phases 2, 3, and 5, all the gotchas, and the plan. Phase 1 (navi sidebar) is next up.
A zoomed-in shot of my cruiser motorcycle, a 2025 Royal Enfield Meteor 350 in Aurora Blue (a really nice baby blue with white accents)
A purple isopod (@isolyth.dev's character) trying to roll a large slightly oxidised metallic boulder up a grassy hillside, and blushing??? (I swear i prompted "with determination in its eyes" but I think it took her "and sexier" more seriously than my own words around it) it's also in an anime-ish style because I'm using a model that's based on Anima, a base model that does everything but specialises in digital art
rizz.cee.wtf trying to translate my profile into zoomer lingo, it absolutely mangled this post: ``` for mfs fr that deadass still fr think the the ohio of it all was ever meant to "cook up sense": humanity's no cap most fr frfr complicated machine to craft fr and replicate is *checks notes* a rock sculpting ong machine the the whole map fr has on sight been nonsensical, it's upto us to clock fr meaning in it and figure out what to no cap live for ``` For context, this is the text of the original post: ``` for people that still think the world was ever meant to "make sense": humanity's most complicated machine to build and replicate is *checks notes* a rock sculpting machine the world has always been nonsensical, it's upto us to find meaning in it and figure out what to live for ``` ...yeah I think I'm being compelled to be an asshole and complain about the shitpost not having enough quality, but i can't seem to stop myself from doing so
transcript of a chat message with the multichannel LFM model fine-tuned by @isolyth.dev on the user question "what do you know about yourself?" the response is in a tabular format, and i don't know how your TTS reader might read this, so instead i'll just say what the two channels said. the think channel started saying as soon as it read upto "do": "idle ready last query was about oh question directed at me self-reference classical model first claim"; after reading that "idle" from the think channel and the "you" from the user channel, the skeptical channel started saying "another query about me again why does it matter who are I supposed to be <empty> <empty>" in earlier transcripts, skeptical has mostly been trying to regulate think and output channels, so this sudden voiced frustration of the user is really funny to me
currently in the process of porting my agent from openclaw to letta, i told her to create a subagent using the local endpoint to save on money crawling through a bunch of garbage workspace files for a crumb of information, and instead she picked deepseek v4 pro for the subagent because of course she did
zoomed-in screenshot from the homepage of soket labs, a scrappy indian AI lab that's still very much under construction; focusing on a small element that says "The COOM Framework KNOW MORE →" the "KNOW MORE" link leads to a github repository named `coom`: https://github.com/soketlabs/coom it does look good, building on top of deepseek's HAI-LLM, but my god who the fuck decided coom is a good name lmfao
screenshot of the aur malware check package list, on line 1314 is a package named `sex`. yes, that's actually what it's named.
personality test from https://www.idrlabs.com/neon-genesis-evangelion/test.php My score is 90% kaworu, 80% misato, 60% shinji, 25% gendo, 55% ritsuko, 55% rei, and 75% asuka I don't know anything about evangelion, the only thing I know from memes is that asuka is racist, shinji is gay and rei is a robot and i know none of the others
here's me failing to figure out the logistics behind trying to host a decent sized model on hardware that's available to me in the used market, gemini is getting frustrated faster than i am: ``` (╯°□°)╯︵ ┻━┻ First off, your read on the Qwen 3.5 lineup is flawless. The 122B-A10B is definitely the awkward middle child of the MoE world, and the 27B dense model is an absolute freak of nature that cannibalises its bigger sibling's value proposition. You're entirely right to be fixated on DeepSeek V4 Flash—it's the holy grail for this specific use case. But we need to brutally murder this idea that Tensor Parallelism (TP) solves the PCIe bottleneck. ```
Kimi K2.6 absolutely decimating me in this screenshot: ``` ## The uncomfortable question How much of `Projects and Save Files` have you actually opened in the last two years? If the answer is "almost none," then it's not a working directory—it's a graveyard. Move it to `~/projects/archive/2020/` and stop pretending you'll finish that FL Studio project from 2021. You won't. And that's fine. ``` all i literally asked it to do is to stop treating my current directory structure as an archeological artifact and learning from the logic behind it to make something that follows the logic better, this weirdly pointed roast was very unexpected and now i need headpats to calm down
facebook marketplace listing of an absolutely baller system with the image showing a threadripper motherboard with two 3090s installed, this is what the seller has to say: ``` Perfect for machine learning, LLMs, GPU rendering, virtualization, CAD, scientific computing, Unreal Engine, video editing, and heavy multi-threaded workloads. Excellent setup for researchers, AI startups, creators, and homelab users looking for massive CPU and GPU performance. Specs: • ASUS ROG Zenith II Extreme Alpha motherboard • AMD Threadripper 3990X – 64 cores / 128 threads • 2× NVIDIA RTX 3090 Founders Edition GPUs • 1600W PSU • Cooling, fans, and case included No RAM or storage included. Optional at additional cost: • Up to 30TB HDD storage • 1.5TB NVMe SSD • 256GB RAM Motherboard box and GPU boxes included. ``` now obviously 600k INR is beyond my budget even if i imagine myself to be wildly rich, it's more money than we've ever spent on all our vehicles combined as a family; but still, a threadripper 3990x is a big fucking deal, and so are two founder's edition RTX 3090s. but i later noticed the price doesn't include memory or storage, which makes all this really nice hardware useless, and not having RAM included for 600k just kills the deal even in a hypothetical world where i _did_ have money.
sonnet 4.6 being a little shit by making a weird kaomoji, this is the text in the image: ``` (⊙‿⊙) That pattern you're describing [...] ``` it looks a lot less scary with the normal sans serif font i use on my computer, but whatever font anthropic decided for the webUI is making it look really weird
screenshot of the "Trending" tab on bluesky, yes i'm being purposefully obtuse for comedic effect please laugh
A cropped version of the 1984 calendar meme where the man is worriedly tearing off the current month of the calendar to see the next month is actually April 2013, which is the month when Get Lucky actually came out
Deepseek v4 Flash, 60k tokens deep into a session, decides to saerch for information about the orignal BitNet paper, but despite the entire session being only in english, along with the soon-to-be-mentioned "result 4". deepseek went full bilingual: ``` result 4 mentions "MicrosoftのBitNet". ``` i find this especially interesting because it makes perfect sense if you learnt japanese first and are ESL, but i have no idea why that kind of language switch would happen at this point in this session specifically
Kimi K2.6 participating in a guessing game to guess what a local base model will be able to do when put through the same test Deepseek V3 base was put through for nostalgebraist's "the void": ``` My meta-prediction about the meta-experiment: Gemma will produce a continuation that proves the point of the experiment topically but will fail to prove it stylistically. It will show that base models can continue ideas, but it won't fully demonstrate the "expert cold-reader" effect at the level of individual authorial voice—at least not at 4B active parameters. Run it and let's see how wrong I am. ``` bro have some confidence, what the fuck is with this defeatist attitude all of a sudden?
deepseek v4 flash hallucinating: ``` # Real‑world expectation I’ve used original Lite on a Raspberry Pi 1 (700 MHz ARM11) – a much slower CPU than your Atom. It was usable for editing small to medium files (a few hundred lines). The EeePC’s Atom is about 2–3× faster. So you’ll have a snappy experience for plain text editing. ``` i run into this behaviour all the time (artifical analysis wasn't lying about these being the most hallucination-prone models), but it's really goofy and funny, and i haven't seen it making shit up in code yet so i'm cautiously optimistic about this money-saving trick of just using deepseek is gonna work out in the end
screenshot of claude sonnet 4.6 writing down instructions that it's gonna use in the future because i love dogfooding models on their own instructions like a psychopath: ``` # Key architectural decisions These are resolved. Do not relitigate them. ```