Vel (nuclear-powered)

@hiway@mastodon.social yeah, that's something i've noticed in general as well. anything negative that's LLM-related is treated as entirely plausible and real on fedi confirmation bias... @ShadowJonathan@tech.lgbt

@kkarhan the overlap between sysadmins and plumbers in terms of piping shit is concerning @d_rift @rysiek

@badkeys GLM-OCR is a very impressive model for this (note: do not run it on lower quantization, it ruins the quality for small details - perhaps running in original precision might improve quality even further actually, my issues were likely due to the MLX quantized version i chose also quantizing the vision encoder. most GGUF versions do not do this, which is good). its unquantized 16-bit weights are only 2.65 GB, and the various quants are even smaller. it's also very fast once it's downloaded and loaded - the whole procedure with loading took maybe 10s, while the actual prompt processing and decoding took ~3s $ mlx_vlm.generate --model EZCon/GLM-OCR-8bit-mlx --image SHehPTYg.jpg --prompt 'Text Recognition:' [SNIP] -----BEGIN EC PRIVATE KEY----- MHcCAQEEIETqS1+8E6OQP7nUCxLFqa1MCWIlshvsGpluxCynDBMEoAoGCCqGSM49 AwEHoUQDQgAEfbsO6BAdHVm3w0KWEHReqrmXMoMUikalZTG0a3nhN9niYfRCVgqR tRfREeitDrYPcMdkqwvaQcsXczCKTF52NA== -----END EC PRIVATE KEY----- [SNIP] Peak memory: 2.206 GB the only mistake is MCWII, which it recognized as MCWIl (with a lowercase L). however, i myself have no clue which is even correct here, so i think the model gets a pass i ran a properly quantized version (mradermacher/GLM-OCR-GGUF:Q4_K_M on llama.cpp - llama-cli --mlock -hf mradermacher/GLM-OCR-GGUF:Q4_K_M --no-warmup -c 2000 --prompt 'Text Recognition:' --image SHehPTYg.jpg), which did not make this mistake (was entirely flawless). the whole thing took ~6.5 seconds and used about 1.5 GB of memory

@memdmp @theresnotime

@coolbean@brain.worm.pink Half-Life (>=18), GTA IV (16-18), Minecraft (13-16), Doom 2016 (<=13) 4 age brackets. perfectly AB 1043 compliant. @rose @rzbrk@social.tchncs.de

@hsza@social.tudbut.de even with a terrible setup consisting of like, opencode with some ultra-cheap/free LLM, i feel like one has to be incredibly stupid to have the model do this with claude code + sonnet/opus.. like? how? @ShadowJonathan@tech.lgbt

@TeflonTrout@beige.party Take a deep breath, and look at the context of what you are replying to. Does either of the messages in this thread "promote" LLMs? What you are replying to is commentary on the Fediverse's LLM-related discussion, not promotion of LLMs. (and by the way, none of the above was written with an LLM. i just usually don't write "formally".) @hiway@mastodon.social @ShadowJonathan@tech.lgbt

@ShadowJonathan@tech.lgbt i'm inclined to describe this as either: A) bullshit, or B) accidentally committed files modern LLMs just.. don't write millions of unnecessary LoC. claude code is also not a terrible agent/harness

@mort@floss.social First of all.. Zed is under a FOSS license. It doesn’t matter what they wrote in the ToS. One can use Zed the editor under its license(s) - apparently some mix of AGPLv3/GPLv3/Apache-2.0. Besides… (emphasis mine, quotes from the current (March 7th, 2026) ToS with links to the exact positions) “and related subscription service (the “Service”). These Terms of Service are a legally binding contract between you and Zed regarding your use of the Service” (https://zed.dev/terms#:~:text=and%20related%20subscription) "Subject to these Terms, Zed will permit Customer to access and use Zed’s AI-enabled software-as-a-service offering (the “Service”), which enables certain additional features and functionality (including artificial intelligence and collaboration features) in Zed’s open source code editing software (“Software”)." (https://zed.dev/terms#:~:text=Subject%20to%20these%20Terms) “Customer must be at least 18 years old to use the Service” (https://zed.dev/terms#:~:text=Customer%C2%A0must%20be-,at%20least%2018%20years%20old,-to%20use%20the) to reiterate: “Service” is referring to their subscription service. the terms only apply at all when using it, and the 18 years old clause reiterates on that. if just using the editor, or downloading the editor, nothing here applies. disclaimer that i’m not a native English speaker and definitely not a lawyer, but this seems quite explicit. feel free to dispute my points if the reader has a different understanding of the terms. @marekfort@mastodon.online @amberage@eldritch.cafe @kelset@mastodon.online

@memdmp @theresnotime https://codeberg.org/robida/human.json probably. another technical solution to... what exactly?

@theresnotime can't do JSON, only tensors :neocat_pensive: array([[-0.0381165, -0.00387955, 0.0583191, ..., 0.0194702, -0.0175629, 0.0563354]], dtype=float16) or well. the full tensors are just too long (idk i just put a into a tiny embedding model. it's not even particularly funny)

@theking@mskey.nekomimi.party :neocat_nervous: oh no @spinach@girlcock.club

@garbrys FYI CoMaps is a fork of Organic Maps that was created due to concerns with the project's management. also, it apparently has additional features. @privateger @mochabeau

@garbrys there are also Organic Maps and CoMaps - forks of MAPS.ME. i don't know how well the public transport routing works however, as they all seem to lack support for Moscow's public transport @privateger @mochabeau

@lucydev Vel can see the utility of LLMs in programming for making some like. quick scripts or whatnot for solving specific problems without programming knowledge (and not wanting to become an actual programmer). this could technically be called democratization however, in this narrow use case the LLM is acting as a stopgap, essentially as an idea -> language translator. ideally, in a more "well designed" world, this would not be "needed"

@Natanox in Moscow, Moskva-2020 metro trains only have USB-A (5V@3A), while some buses have USB-A and USB-C. i'm not sure what interfaces Moskva-2024 has i have never seen either being used @whophd @alda @theresnotime @mattgrayyes

@cutesobri@cathode.church system attestation checks if the system is "secure" (well, usually that's the purpose), not if the app is authentic @GrapheneOS@grapheneos.social

@benno@eigenmagic.net wonder if the prices on everything LLM creeping/jumping up will finally teach some people that Opus and full-sized GPT are not the only models that exist..

@argv_minus_one@mastodon.sdf.org ugh. was unaware of the patent bullshit.. from my memory: Mastodon re-encodes images, but keeps them in the same format as uploaded (so, for lossy formats, making efforts to encode more efficiently worthless), unless the format is not allowed (like AVIF) - in which case it's encoded to JPEG from checking the code: 1 - this (lib/paperclip/vips_lazy_thumbnail.rb) decides when to convert (with libvips) an image, which seems to mean re-encoding it def needs_convert? strip_animations? || needs_different_geometry? || needs_different_format? || needs_metadata_stripping? end 2 - this (app/models/media_attachment.rb) specifies which image MIME types are supported and which will be converted (hence, JPEG, PNG, GIF, WebP can stay, but will be re-encoded anyway if needed to strip metadata it seems - it seems to be the same path): IMAGE_MIME_TYPES = %w(image/jpeg image/png image/gif image/heic image/heif image/webp image/avif).freeze IMAGE_CONVERTIBLE_MIME_TYPES = %w(image/heic image/heif image/avif).freeze to JPEG: IMAGE_CONVERTED_STYLES = { original: { format: 'jpeg', content_type: 'image/jpeg', }.merge(IMAGE_STYLES[:original]).freeze, small: { format: 'jpeg', }.merge(IMAGE_STYLES[:small]).freeze, }.freeze seemingly at Q=90 (both for native JPEG and previously-AVIF/HEIC/HEIF): def save_options case @format when 'jpg' { Q: 90, interlace: true } else {} end end ugh. the Markdown rendering broke and i'm not sure why it was the list @soft@normal.style