Семантический поиск vs полнотекстовый: сравниваем три embedding-модели на 10 000 категорий Ozon

Сравнил полнотекстовый поиск PostgreSQL ( tsvector / tsquery + GIN-индекс) с семантическим поиском через pgvector ( cosine distance ) на датасете из 10 019 товарных категорий Ozon. Три embedding-модели: GigaChat EmbeddingsGigaR (Сбер, 2560-мерные векторы, API) Qwen3-Embedding-0.6B (Alibaba, 1024-мерные, локальный инференс через HF Text Embeddings Inference на GPU) text-embedding-3-small (OpenAI, 1536-мерные, API) Прогнал 18 запросов в пяти категориях: синонимы и сленг, intent-запросы в свободной форме, подарочная тематика, cross-lingual (EN-запросы к RU-данным), абстрактные формулировки. Замерил латентность и top-5 с cosine similarity score. Разбор каждого запроса, таблицы и код — под катом.

https://habr.com/ru/articles/1010200/

#искусственный_интеллект #gigachat #openai #qwen #embeddings

Fourth week of "#Deepseek v4 dropping this week". I wonder if they are having the same problems as the #Qwen team.

Pretty sure that the pace of high quality open weights Chinese models is not sustainable in the long run. At some point the execs who love monetization will take control. Probably what happened in Qwen's case.

#ai

Current coding setup:

- Editor: Zed
- Local Inference Engine: LM Studio
- Models: qwen3.5:35b-a3b and qwen3-coder-next

Using AI for questions and suggestions regarding code, not for vibe coding.
This constellation works pretty well for me and the new qwen models definitely are pretty usable for my case \o/
And: no worries about code leaving my device :)

#ai #coding #lmstudio #zed #qwen

Adding image rendering to agent-shell

Maybe it'll finally make #gnuplot use stick for me :)

#emacs #mistral #qwen #claude #anthropic #ai #google #goose #gemini #linux #macos #oss #foss #opencode #codex #vibe #windows

February's about to end tomorrow so here's to all the fellow lonely male people out there in the world lololol :p

#ai #llm #aiSlop #stablediffusion #chatgpt #grok #claude #mistral #qwen #deepseek #glm #kimi #flux

ALT

February's about to end tomorrow so here's to all the fellow lonely male people out there in the world lololol :p

#ai #llm #aiSlop #stablediffusion #chatgpt #grok #claude #mistral #qwen #deepseek #glm #kimi #flux

ALT

Chińska bezpieka i ChatGPT

Co w tym wszystkim jest dla mnie naprawdę interesujące, to fakt, że chińska bezpieka uznała, że ChatGPT jest lepszy niż chińskie modele.

https://wp.me/p3fv0T-i7s #Japonia #Chiny #ChRL #dezinformacja #bezpieka #OpenAi #AI #Qwen #DeepSeek #ChatGPT #POLECANE

Latxa VL eredu berriak probatu ditut. Euskarazko dokumentuekin OCR modura erabiltzeko oso interesgarriak!!

https://xezpeleta.github.io/blog/latxa-vl/

#latxa #latxavl #hitz #adimenArtifizialara #adimenArtifizialaEuskaraz #qwen #qwen3 #qwen3vl

Really impressed with Qwen3-TTS and voice cloning. A 10 second sample was all it took to have my very own AI-Casey (@Casey) voice, wishing a Hardfork listerner happy birthday.

With #claudecode this only took 20 minutes from idea to working output. Most of that time was actually me waiting for the model download.

Not the fastest on my M4, but still faster to generate a @kevin less episode than waiting a whole week for the next Hardfork

https://www.wunsch.dk/blog/voice-cloning-with-claude-code/

#claudecode #qwen #tts #ai