Hacker News: electroglyph

New comment by electroglyph in "Good results fine tuning a local LLM like Qwen 3:0.6B to categorize questions"

electroglyph — Mon, 22 Jun 2026 04:04:09 +0000

existing embedding models like alibaba's modernbert tune or one of the jina v5s would probably map query to category automatically. (i.e. store embeddings of each category and calculate cosine sim for each incoming query vs. categories and pick the closest)

also, you could stick a classifier head on a BERT model as another option.

New comment by electroglyph in "The time the x86 emulator team found code so bad they fixed it during emulation"

electroglyph — Tue, 16 Jun 2026 05:58:50 +0000

heh, when Raymond Chen dunks on the MSVC team =)

New comment by electroglyph in "AI is code – and can't be prompted into being smarter"

electroglyph — Mon, 15 Jun 2026 01:37:03 +0000

it's probably a lie

New comment by electroglyph in "DeepSeek V4 Pro beats GPT-5.5 Pro on precision"

electroglyph — Mon, 08 Jun 2026 03:18:39 +0000

deepseek 4 pro is insanely good for the price

New comment by electroglyph in "Nvidia is proposing a beast of a CPU system for Windows PCs"

electroglyph — Sat, 06 Jun 2026 22:49:17 +0000

that link actually recommends not doing it from UEFI and doing it via software

New comment by electroglyph in "KVarN: Native vLLM backend for KV-cache quantization by Huawei"

electroglyph — Thu, 04 Jun 2026 21:33:54 +0000

any divergence (even if the benchmark is better) from full precision is error

New comment by electroglyph in "Gustav Klimt and Egon Schiele in Conversation (2018)"

electroglyph — Sun, 31 May 2026 05:40:23 +0000

this is better than TFA

New comment by electroglyph in "FBI Arrests CIA Official with $40M in Gold Bars in His Home"

electroglyph — Thu, 28 May 2026 01:45:33 +0000

sometimes i wonder if the left hand knows what the right is doing. it looks like we arrested our own spy in this case: https://www.politico.com/news/2026/05/25/american-journalist...

New comment by electroglyph in "Norway's 2 petabytes of Huawei flash storage and LLM training"

electroglyph — Mon, 25 May 2026 23:16:10 +0000

absolutely. somebody online was wanting an LLM with Georgian language support, and that's exactly what i suggested: start digitizing Georgian text.

New comment by electroglyph in "Wake up! 16b"

electroglyph — Sun, 24 May 2026 04:39:16 +0000

i'll upvote this each time it's submitted

New comment by electroglyph in "Qwen3.7-Max: The Agent Frontier"

electroglyph — Wed, 20 May 2026 23:09:37 +0000

you should be using dflash with that model, look it up

New comment by electroglyph in "DeepSeek-V4-Flash means LLM steering is interesting again"

electroglyph — Sat, 16 May 2026 22:30:26 +0000

heretic maintainer: https://github.com/p-e-w/heretic

the fun bits are in another branch or PRs

New comment by electroglyph in "DeepSeek-V4-Flash means LLM steering is interesting again"

electroglyph — Sat, 16 May 2026 20:56:10 +0000

p-e-w was just talking about this the other day in his Discord. seems doing the one neuron method is quite bad for KLD and that's why the newer techniques have stuck.

New comment by electroglyph in "Seeing Birdsong"

electroglyph — Mon, 11 May 2026 07:21:04 +0000

site has so little information there doesn't seem to be much to discuss

New comment by electroglyph in "A polynomial autoencoder beats PCA on transformer embeddings"

electroglyph — Fri, 08 May 2026 10:22:27 +0000

this looks awesome. i've been struggling with vector compression, and have been trying PCA + all sorts of rotations. looking forward to trying this out

New comment by electroglyph in "Making LLM Training Faster with Unsloth and NVIDIA"

electroglyph — Thu, 07 May 2026 09:19:14 +0000

nice writeup! looking forward to doing some more training as soon as i get some more data sorted. it'll be a custom arch, but i'll probably shoehorn it into unsloth for a speed boost.

New comment by electroglyph in "Train Your Own LLM from Scratch"

electroglyph — Tue, 05 May 2026 06:41:57 +0000

you can train it, but not fully

New comment by electroglyph in "Opus 4.7 knows the real Kelsey"

electroglyph — Fri, 01 May 2026 08:24:18 +0000

that's in the ideal scenario where it's only seen a single copy of it tho

New comment by electroglyph in "New copy of earliest poem in English, written 1,3k years ago, discovered in Rome"

electroglyph — Fri, 01 May 2026 08:21:49 +0000

it was 1.3e-6 billion years ago!

New comment by electroglyph in "Microsoft and OpenAI end their exclusive and revenue-sharing deal"

electroglyph — Tue, 28 Apr 2026 05:31:48 +0000

i'm doing inference on a free mi300x instance from AMD right now. not sure if the software stack is just old or what, but here's what i've observed: stuck on an old version of vllm pre-Transformers 5 support. it lacks MoE support for qwen3 models. oss-120b is faaaar slower than it should be.

int8 quantization seems like it's almost supported, but not quite. speeds drop to a fraction of full precision speed and the server seems like it intermittently hangs. int4 quantization not supported. fp8 quantization not supported.

again, maybe AMD is just being lazy with what they've provided, but it's not a great look.

right now the fastest smart model i can run is full precision qwen3-32b. with 120 parallel requests (short context) i'm getting PP @ 4500 tokens/sec and TG @ 1300 tokens/sec