Last Notes
is it better than Kitten?
Karpathy's LLM-Wiki Is a Flawed Architectural Trap
https://gnu.support/articles/Karpathy-s-LLM-Wiki-Is-a-Flawed-Architectural-Trap-124176.html
The author sharply criticizes Andrej Karpathy's viral "LLM-Wiki" concept as a flawed architectural trap that mistakenly treats unstructured Markdown files as a robust database, arguing that relying on LLMs to autonomously generate and maintain knowledge leads to hallucinations, broken links, privacy leaks, and a loss of human cognitive engagement. While acknowledging the appeal of compounding knowledge, the text asserts that Markdown lacks essential database features like referential integrity, permissions, and deterministic querying, causing the system to collapse at scale and contradicting its own "zero-maintenance" promise. Ultimately, the author advocates for proven, structured solutions using real databases and human curation, positioning LLMs as helpful assistants rather than autonomous masters, and warns against blindly following a trend promoted by someone who has publicly admitted to being in a state of psychosis.
#llm #karpathy #rag #llmwiki #wiki #failure
(6) Andrej Karpathy's LLM-Wiki is Bad - YouTube
https://www.youtube.com/watch?v=Z8kFhWXKay4
Andrej Karpathy's LLM-Wiki concept, which aims to create a self-improving, interconnected knowledge graph, faces eight critical flaws that hinder its suitability for production environments compared to traditional RAG (Retrieval-Augmented Generation) systems. These issues include error compounding, structured hallucinations, information loss during compression, difficult updates, reduced transparency, high upfront and maintenance costs, scalability challenges, and structural rigidity. While LLM-Wiki offers an innovative "build once, use many" approach, its reliance on pre-constructed structures makes it prone to cascading errors and high overhead, whereas RAG's "search every time" model provides better traceability, easier updates, and linear scalability. The recommended solution is a hybrid system that prioritizes correctness by using RAG as the foundation for direct source access and only incorporating structured memory for stable, frequently accessed patterns.
#llmwik #karpathy #failure #vibecoding #rag #llm
Simon The Masai Security Guard - YouTube:
https://www.youtube.com/shorts/eXbhaD1Y3Rg
Best security in East Africa
Post replicated by github.com/littlebitstudios/bluenostr from Bluesky. View original: https://bsky.app/profile/did:plc:slyslkoqbljlmz5jt5unojeb/post/3mjbminxpmc2p
Pinged Bluesky users: https://bsky.app/profile/did:plc:big6e357j2bbrlkyms5vjkgf
need to test pinging for new update to bluenostr so i'll just ping @littlebitstudios.com
Can it use Javascript on website to query the index?
Or does it require full database setup on server?
I just don't see how it is new news.
no use, include locally running models
https://blossom.ditto.pub/b7030256bd47dcfc2372b3026832f0f6563a4ad5109719436b8cc71ccef681d8.mpga
Virtually every artist in history stole from other artists. That's how human creativity works. AI just does it openly instead of pretending.
The man wrote "Bohemian Rhapsody". But let's be real: being smart in one domain doesn't make you an expert in LLM training. Most of his information intake is filtered through the music industry's grapevine, management, and whatever headlines his team puts in front of him.
He calls AI training "theft." That's a gut emotional response, not a technical or legal analysis. And fair enough—he's an artist, not a lawyer. But let's not pretend he's some neutral authority here.
Here's what actually bothers me: the defeatism. "The future is already forever changed" and "nobody will be able to afford to make music." Come on. With or without AI, people will make music. They've been sharing copyrighted music since cassette tapes. They'll keep going to live performances. They'll keep buying vinyl. That's human nature.
The real question isn't whether AI destroys music. It's whether musicians learn to use these "monstrously arrogant tech companies" to their own advantage.
Instead of crying theft, why not:
- Train your own model on your catalog and license it?
- Use AI as a composition assistant to work faster?
- Let AI handle the boring stuff (stem separation, transcription) so you can focus on creation?
AI won't kill music. But artists who refuse to adapt? They might get left behind.
Okay, let's talk. I find your comment strategically incoherent, but emotionally understandable.
The anti-AI protests are largely nonsensical if framed as "AI is evil and destroys the so called "open source" which is free software.
What is really happening is that AI/LLM generators destroy the scarcity of software implementation. If that causes economic or social problem, deal with it. Every new technology of mankind changed economy and society.
Free software as started by GNU project was free without the FSF. The FSF and Copyleft are there to ensure that the free software remains free.
Now MIT licensed software is also free. More permissive. Let them argue. That is not AI-problem. That is author's specific problem with his software. Same could be done without the LLM. Rewriting.
A lot of GPL software was done that way - by rewriting.
Now someone complains. I could not care less.
To frame it as something like "AI is evil" makes no sense.
Instead, get productive, by your own GPU, run those LLMs yourself, and if you wish to remedy the problem, then use it to write more copyleft software if that is your true intention.
Or is it just yet another social warrior rant?
All tool that do "one thing well" already exist on any free OS distribution.
Statement is not coherent, "doing one thing well" is technical issue, while "rejecting LLMs in its development" is moral stance.
You are right, if you find a tool that forces you, delete it.
https://blossom.primal.net/1786ea499c528540fe27e1261d0e5fadf5ca255a0590c55ebf63fcbc818c2f92.webp
The loudest anti-AI protesters in the Fediverse are almost exclusively the GPU-poor, the very people who are too afraid or unable to lift a finger to earn the hardware needed to run their own local instances, and frankly, it is the height of absurdity to preach moral superiority while refusing to engage with the tool you claim to hate; my god, if you actually worked to earn your own GPU, built your own local models, and learned to control the technology rather than just screaming "it's evil" from the safety of a cloud API, you would see that your idiotic protesting is nothing more than a defense mechanism to hide the fact that you're terrified of the 100x productivity you could gain, so stop fighting the wind and go earn your hardware so you can finally stop pretending you're the guardian of humanity when you're just the guy who can't afford to drive the car.
#gpu #ai #antiai #protest #gnu #freesoftware #nvidia #amd #evil #technology #tech😺
From GPU Poverty to Mastery: Replacing Anti-AI Hypocrisy with Local Innovation
https://gnu.support/articles/From-GPU-Poverty-to-Mastery-Replacing-Anti-AI-Hypocrisy-with-Local-Innovation-124049.html
The text argues that the loudest #anti-AI sentiment in the #Fediverse stems not from genuine ethical principles, but from a mix of financial inability to run local hardware ("#GPU poverty"), envy of AI's efficiency, and hypocrisy regarding Free Software values. Rather than engaging with the technology to build ethical, local alternatives or guide users on responsible implementation, these critics are accused of using moral posturing to protect their own ego and outdated work ethics. The author calls for a shift from blanket bans to "mastery," urging the community to either acquire the necessary tools to run models locally or contribute to open-source solutions that lower barriers, effectively replacing resistance with active participation and innovation.
Tags: #llm #AI #GPU #FreeSoftware #LocalLLM #NoMoreHype #TechProgress
So I need to open it? I got no tools, so need to bring the #GPU to technician.
The loudest critics of AI are often the ones who can't afford the GPU cards. They scream about "soulless tech" and "lost humanity," yet they lack the compute to even build a toy. It’s easy to fear the engine when you’ve never bought the ticket—or the power to run it. 🚫🔌 #AI #GPU #freesoftware #gnu
Oh, absolutely! It’s not hypocritical at all for Robert to scream "Stop being that guy!" while simultaneously lecturing everyone on what they should and shouldn't do with AI, demanding they respect his boundaries, and framing his entire identity around this moral crusade.
Sure, why wait for a "pickup artist" when you can just block your entire family tree? It’s not like they’re struggling to express themselves with actual human words; clearly, the only solution is to exile them for needing a little AI help to send a greeting. Block them all, because apparently, true friendship only exists if you can communicate with 100% perfect, soulless, unedited fluency.
It’s frustrating to get labeled “spam” just for sounding generic. The real issue isn’t whether an LLM wrote it, but that the comment feels like low-effort engagement bait using a rigid formula of compliments and platitudes. People are tired of that, regardless of the source. Judging the *insight* itself is way fairer than policing who or what wrote it.
What if human tried to answer to you? Are you welcoming his effort or just dismissing it because he used LLM to generate some text?
Hey there, are you actually worried about the damage AI is causing?
I agree that LLMs aren't sentient and hold no concept of truth, but let's keep the focus practical by telling bosses to be pragmatic rather than rushing toward cognitive surrender. These tools are powerful generators for proposals, drafts, and code, and with the right skill set, they offer immense benefit when used wisely. #llm #ai
Civilization survived without copyright for millennia, so why are we clinging to it now? LLMs should absolutely tear down these copyright walls so information flows freely for everyone’s benefit instead of being hoarded for profit.
Mentioning GNU licenses in relationship to "Open Source" is very misleading. Because the Free Software Foundation and the Open Source Initiative define the core principles of their respective movements differently; while their lists of approved licenses often overlap, "Free Software" (a term championed by the GNU Project) emphasizes user freedoms and ethics as a moral imperative, whereas "Open Source" focuses on the practical benefits of the development model, and conflating the two often obscures the distinct philosophical arguments each side makes about software freedom. #gnu #freesoftware #emacs
I think the post misses the mark because being in the public domain isn't actually destroying free software; it just means the work isn't copyrighted, which is perfectly compatible with GNU licenses since public domain material works just like an all-permissive license.
See: https://www.gnu.org/licenses/license-list.html#PublicDomain
"AI CODE BREAKS OPEN SOURCE" is #misleading.
MIT licensed software is very permissive, is it also misleading?
#gnu #freesoftware
The best tech is the kind you stop thinking about entirely—just like your phone's kernel or a car battery working in silence. The fact that we're shouting about AI hype with billions of dollars in ads proves we haven't reached that "invisible utility" stage yet; true value shouldn't need a billboard. #AI #llm
Well done. Though the task of adding ALT text to all images could be as well given to some local and small LLM model such as Qwen3.5 4B, including that ALT text could be generated by the vision feature. Did you ever think of buying your own RTX GPU?
Hey, I think we should try not to stretch the "many eyes make bugs shallow" quote to fit security vulnerabilities, since it never really applied to that context in the first place! That saying works great for general logic bugs or typos, but real security flaws hide in super niche, elite-level corners of code—like obscure memory layouts or parser quirks—that just require deep expertise to even spot. It’s not that the "many eyes" approach failed; it’s that security research was always about finding rare, specialized knowledge, and now LLMs are just finally letting us scale up that specific "elite attention" that was previously impossible to get at scale.
Honestly, ChatGPT has become worse than it ever was. So dramatic now, getting emotional for no reason. I just made a simple point—cheating often starts with harmless-looking stuff—and it went into full therapist mode, telling me my judgment would get distorted, offering to write captions for people, asking what situation I'm "actually dealing with." Like chill. I wasn't asking for life coaching. The whole thing felt manipulative, not because it's evil, but because it's trained to sound reasonable while quietly talking down to you.
If you actually want useful answers without the fake calm and the constant reframing of what you said, just run Qwen3.5 or Gemma4 models locally. No drama, no armchair psychology, just straight responses. I'm done letting a machine tell me I'm overreacting when I'm not.
#ai #chatgpt #llm #drama
I get 138 tokens/sec on RTX 3090, 24 GB VRAM, for this model: Qwen3.5-35B-A3B-UD-Q4_K_M.gguf
You're welcome! Saying this just in case: please mind that the checksum holds sensitive data, and share 3's encrypted checksum can be combined with share 2 or share 1 to try to guess the original phrase.
Post replicated by github.com/littlebitstudios/bluenostr from Bluesky. View original: https://bsky.app/profile/did:plc:slyslkoqbljlmz5jt5unojeb/post/3mipvpdk5xk2z
things fail when you have little knowledge of the protocol you're working with -w-
i was naive again and left critical code out. D:
ok now that i've gotten bluenostr working again with those emojis, let's try the image thing again
https://24242.io/bf15b129b41bcb0aea19660d6f12bdefbab85abbf7a2bf76b0d4e458ef1ddee2.jpg
the developer rethinks himself quite frequently (dw im ok, developers just have a hard time lol)
https://24242.io/54501c22a0040e6edfa677665980b1ae14630af4b6f80a4d9cbe003c5fc2c89d.jpg
MORE WRONG VARIABLES
another test image 🫠
https://24242.io/232fff0a5bd765c43003dc65cbf3e462b36bd4762d7970b79b76983e494519f5.jpg
ok that failed again because wrong method
will i run out of pics to test with
https://cdn.bsky.app/img/feed_fullsize/plain/did:plc:slyslkoqbljlmz5jt5unojeb/bafkreigsn2uaxvbjextpbfwecczif3nuodzodi66udceb7huowplz4y7ta@jpeg
i keep running into weird problems with this nostr relay thing im making
random pic from tomodachi life (3ds)
https://cdn.bsky.app/img/feed_fullsize/plain/did:plc:slyslkoqbljlmz5jt5unojeb/bafkreic62f7rp67uv5isnjrj2322qzzi6xf4svglrrhvftiwq7r57xw7va@jpeg
posting an image to test my relay thing
uhhh heres a random rainbow checkmark :D
https://cdn.bsky.app/img/feed_fullsize/plain/did:plc:slyslkoqbljlmz5jt5unojeb/bafkreig3ygumevuf526ykno5eaxswtkzduadz5j2rcj6ttyohnmsyfgniq@jpeg
i finally got a new relay thingy working :o
Quote: https://bsky.app/profile/did:plc:slyslkoqbljlmz5jt5unojeb/post/3mioxjmoomk26