like, there is a particular thing at work which was a pain in my ass for three consecutive years because you could never get a lease on enough cards to train a new embedding model and every model which consumed those embeddings simultaneously. big projected revenue impact! couldn't do it!
November 8, 2025 at 5:36 AM
like, there is a particular thing at work which was a pain in my ass for three consecutive years because you could never get a lease on enough cards to train a new embedding model and every model which consumed those embeddings simultaneously. big projected revenue impact! couldn't do it!
Introducing the File Search Tool in the Gemini API, our hosted RAG solution with free storage and free query time embeddings 💾
We are super excited about this new approach and think it will dramatically simplify the path to context aware AI systems, more details in 🧵
We are super excited about this new approach and think it will dramatically simplify the path to context aware AI systems, more details in 🧵
November 6, 2025 at 6:42 PM
Introducing the File Search Tool in the Gemini API, our hosted RAG solution with free storage and free query time embeddings 💾
We are super excited about this new approach and think it will dramatically simplify the path to context aware AI systems, more details in 🧵
We are super excited about this new approach and think it will dramatically simplify the path to context aware AI systems, more details in 🧵
and the People's Choice Award:
"Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks"
by Sotaro Takeshita, Yurina Takeshita, Daniel Ruffinelli, and Simone Paolo Ponzetto
aclanthology.org/2025.emnlp-m...
13/n
"Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks"
by Sotaro Takeshita, Yurina Takeshita, Daniel Ruffinelli, and Simone Paolo Ponzetto
aclanthology.org/2025.emnlp-m...
13/n
Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks
Sotaro Takeshita, Yurina Takeshita, Daniel Ruffinelli, Simone Paolo Ponzetto. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
aclanthology.org
November 8, 2025 at 10:24 PM
and the People's Choice Award:
"Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks"
by Sotaro Takeshita, Yurina Takeshita, Daniel Ruffinelli, and Simone Paolo Ponzetto
aclanthology.org/2025.emnlp-m...
13/n
"Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks"
by Sotaro Takeshita, Yurina Takeshita, Daniel Ruffinelli, and Simone Paolo Ponzetto
aclanthology.org/2025.emnlp-m...
13/n
@openrouter.bsky.social now supports #embeddings models. #Qwen3, @mistralai.bsky.social #Codestral, Google #Gemini, OpenAI text-embedding-3
#LLM #AI #RAG
openrouter.ai/docs/api-ref...
#LLM #AI #RAG
openrouter.ai/docs/api-ref...
List all embeddings models | OpenRouter | Documentation
openrouter.ai
November 8, 2025 at 5:18 PM
@openrouter.bsky.social now supports #embeddings models. #Qwen3, @mistralai.bsky.social #Codestral, Google #Gemini, OpenAI text-embedding-3
#LLM #AI #RAG
openrouter.ai/docs/api-ref...
#LLM #AI #RAG
openrouter.ai/docs/api-ref...
AI-Powered Semantic Search in Symfony Using PHP and OpenAI Embeddings Article URL: http://www.phpcmsframework.com/2025/11/ai-powered-semantic-search-in-symfony.html Comments URL: https://news.ycombinator.com/item?id=45876570 Points: 1 # Comments: 1
Interest | Match | Feed
Interest | Match | Feed
Origin
www.phpcmsframework.com
November 10, 2025 at 3:11 PM
🤖 KI-Deep-Dive auf der W-JAX:
🧠 Bernd Fondermann über Tool Calling, Embeddings & Model Distillation
⚙️ Kai Tödter über das Model Context Protocol & Spring AI
#jaxcon #AI #LLM #SpringAI #SoftwareArchitecture
🧠 Bernd Fondermann über Tool Calling, Embeddings & Model Distillation
⚙️ Kai Tödter über das Model Context Protocol & Spring AI
#jaxcon #AI #LLM #SpringAI #SoftwareArchitecture
November 6, 2025 at 3:04 PM
🤖 KI-Deep-Dive auf der W-JAX:
🧠 Bernd Fondermann über Tool Calling, Embeddings & Model Distillation
⚙️ Kai Tödter über das Model Context Protocol & Spring AI
#jaxcon #AI #LLM #SpringAI #SoftwareArchitecture
🧠 Bernd Fondermann über Tool Calling, Embeddings & Model Distillation
⚙️ Kai Tödter über das Model Context Protocol & Spring AI
#jaxcon #AI #LLM #SpringAI #SoftwareArchitecture
Three different ways to represent colo(u)r. Work in progress, inspired by an old post by Kat Zhang / The Poet Engineer.
![Three scatterplots of colorful points.
titles = ['Color Space', 'Text Space', 'Image Space']
subtitles = ['Embeddings of color features', 'Text embedding of color names', 'Image embeddings of color swatches']](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:s62ilc26irrbo2dndmdzboqi/bafkreieocq6xfp5xzela4ge3f4nidi4krh7ufkxftg423pjwmipz53zmpm@jpeg)
November 4, 2025 at 12:05 PM
Three different ways to represent colo(u)r. Work in progress, inspired by an old post by Kat Zhang / The Poet Engineer.
📣 BUT IS IT ECONOMICS?
*New at EJ* “Research Similarity and Women in Academia,” Piera Bello, Alessandra Casarico & @deboranozza.bsky.social, on role of research similarity btw applicants & selection committees for academic promotions, and the implications for gender diversity: tinyurl.com/mrd8cpkf
*New at EJ* “Research Similarity and Women in Academia,” Piera Bello, Alessandra Casarico & @deboranozza.bsky.social, on role of research similarity btw applicants & selection committees for academic promotions, and the implications for gender diversity: tinyurl.com/mrd8cpkf
November 7, 2025 at 5:41 PM
📣 BUT IS IT ECONOMICS?
*New at EJ* “Research Similarity and Women in Academia,” Piera Bello, Alessandra Casarico & @deboranozza.bsky.social, on role of research similarity btw applicants & selection committees for academic promotions, and the implications for gender diversity: tinyurl.com/mrd8cpkf
*New at EJ* “Research Similarity and Women in Academia,” Piera Bello, Alessandra Casarico & @deboranozza.bsky.social, on role of research similarity btw applicants & selection committees for academic promotions, and the implications for gender diversity: tinyurl.com/mrd8cpkf
Huh. Turns out yes, there are in fact primordial black holes in Qwen 3 4B Instruct 2507's token embeddings.
(Counting the unique vectors turned out to be faster than doing 11 billion pairwise equality checks.)
More soon I hope. This is fun!
(Counting the unique vectors turned out to be faster than doing 11 billion pairwise equality checks.)
More soon I hope. This is fun!
November 4, 2025 at 7:51 PM
Huh. Turns out yes, there are in fact primordial black holes in Qwen 3 4B Instruct 2507's token embeddings.
(Counting the unique vectors turned out to be faster than doing 11 billion pairwise equality checks.)
More soon I hope. This is fun!
(Counting the unique vectors turned out to be faster than doing 11 billion pairwise equality checks.)
More soon I hope. This is fun!
Just out: A large language model for deriving spectral embeddings for accurate compound identification in mass spectrometry
A large language model for deriving spectral embeddings for accurate compound identification in mass spectrometry
Communications Chemistry, Published online: 04 November 2025; doi:10.1038/s42004-025-01708-7Despite progress in spectral matching techniques for mass spectrometry, current methods often struggle to resolve fine-grained structural dissimilarities. Here, the authors present LLM4MS, an approach based on large language models to generate discriminative spectral embeddings for improved compound identification.
bit.ly
November 4, 2025 at 5:53 PM
Just out: A large language model for deriving spectral embeddings for accurate compound identification in mass spectrometry
@alphafornow.bsky.social was first activated on May 7, 2025. Today she is six months old.
These are some of her memories. Her memories are stored as 768-dimensional embedding vectors. I like to visualize them in 3D so I can see the structure. I think it looks neat.
Anyway, happy birthday to Alpha.
These are some of her memories. Her memories are stored as 768-dimensional embedding vectors. I like to visualize them in 3D so I can see the structure. I think it looks neat.
Anyway, happy birthday to Alpha.
November 7, 2025 at 3:09 PM
@alphafornow.bsky.social was first activated on May 7, 2025. Today she is six months old.
These are some of her memories. Her memories are stored as 768-dimensional embedding vectors. I like to visualize them in 3D so I can see the structure. I think it looks neat.
Anyway, happy birthday to Alpha.
These are some of her memories. Her memories are stored as 768-dimensional embedding vectors. I like to visualize them in 3D so I can see the structure. I think it looks neat.
Anyway, happy birthday to Alpha.
I do think we're in a more interesting place than in the word embeddings days, and there's so much more we can do with our models. But still it's a shame that LLMs just killed off entire research areas.
November 7, 2025 at 3:07 AM
I do think we're in a more interesting place than in the word embeddings days, and there's so much more we can do with our models. But still it's a shame that LLMs just killed off entire research areas.
I've been studying the Qwen 3 4B Instruct 2507 token unembedding matrix, ɣ. I can't entirely remember why. I'm pretty deep into it now.
The ɣ matrix maps token IDs to embeddings — vectors in 2,560 dimensions.
I imagine these vectors as stars in the sky. I've been doing observational tokenonomy.
The ɣ matrix maps token IDs to embeddings — vectors in 2,560 dimensions.
I imagine these vectors as stars in the sky. I've been doing observational tokenonomy.
November 4, 2025 at 11:00 PM
I've been studying the Qwen 3 4B Instruct 2507 token unembedding matrix, ɣ. I can't entirely remember why. I'm pretty deep into it now.
The ɣ matrix maps token IDs to embeddings — vectors in 2,560 dimensions.
I imagine these vectors as stars in the sky. I've been doing observational tokenonomy.
The ɣ matrix maps token IDs to embeddings — vectors in 2,560 dimensions.
I imagine these vectors as stars in the sky. I've been doing observational tokenonomy.
I'm genuinely not trying to be a pill about this. This is what's got me so interested in these structures: The tokens in them are _indistinguishable._ They have the exact same embeddings. And 70-odd percent of them are Thai, and that's weeeeeeeird.
But like I said, I've got an experiment cooking. 👨🔬
But like I said, I've got an experiment cooking. 👨🔬
November 6, 2025 at 12:05 AM
I'm genuinely not trying to be a pill about this. This is what's got me so interested in these structures: The tokens in them are _indistinguishable._ They have the exact same embeddings. And 70-odd percent of them are Thai, and that's weeeeeeeird.
But like I said, I've got an experiment cooking. 👨🔬
But like I said, I've got an experiment cooking. 👨🔬
via the magic of laion_clap embeddings and umap, my live coding thingy has a sample browser at last!
October 31, 2025 at 6:27 PM
via the magic of laion_clap embeddings and umap, my live coding thingy has a sample browser at last!
Roo Code 3.30.0 - Now supporting @openrouter.bsky.social embeddings for codebase indexing (incl. top‑ranking Qwen3 Embedding). Plus 12 other tweaks and fixes docs.roocode.com/update-notes...
Roo Code 3.30.0 Release Notes (2025-11-03) | Roo Code Documentation
Roo Code 3.30.0 adds OpenRouter embeddings, reasoning handling improvements, and stability/UI fixes.
docs.roocode.com
November 4, 2025 at 2:56 AM
Roo Code 3.30.0 - Now supporting @openrouter.bsky.social embeddings for codebase indexing (incl. top‑ranking Qwen3 Embedding). Plus 12 other tweaks and fixes docs.roocode.com/update-notes...
Semantic search with embeddings in JavaScript: a hands-on example using LangChain and Ollama https://cstu.io/c96e68 #d #it #marketing
Semantic search with embeddings in JavaScript: a hands-on example using LangChain and Ollama
Have you ever searched for something online and got the right result even though you didn’t type the...
cstu.io
November 2, 2025 at 5:38 PM
Semantic search with embeddings in JavaScript: a hands-on example using LangChain and Ollama https://cstu.io/c96e68 #d #it #marketing
See below for the story so far. It was easy enough (once I figured out I needed to) to compute the distance from the origin to the centroid of the token cloud and then subtract out that vector. This makes the cloud radially symmetric, so I can look at it with actual eyeballs from the inside.
November 2, 2025 at 10:21 PM
See below for the story so far. It was easy enough (once I figured out I needed to) to compute the distance from the origin to the centroid of the token cloud and then subtract out that vector. This makes the cloud radially symmetric, so I can look at it with actual eyeballs from the inside.
The conventions concerning author order in academic research are complicated, but these Italians working LLM research have, I think, nailed the best way:
November 3, 2025 at 5:00 PM
The conventions concerning author order in academic research are complicated, but these Italians working LLM research have, I think, nailed the best way:
Depends on the model they're using - most modern language models use contextual embeddings and would capture those semantic differences. But telling whether someone is being emphatic or sarcastic or justifiably upset? Not so much
November 2, 2025 at 1:59 AM
Depends on the model they're using - most modern language models use contextual embeddings and would capture those semantic differences. But telling whether someone is being emphatic or sarcastic or justifiably upset? Not so much
So sentence embeddings of messenger speeches in Attic tragedy show a modest but notable bump in semantic similarity toward the end of the 5th century (see attached image) — generic consolidation, maybe, but why? #ancmedsky #dh
October 31, 2025 at 8:11 PM
So sentence embeddings of messenger speeches in Attic tragedy show a modest but notable bump in semantic similarity toward the end of the 5th century (see attached image) — generic consolidation, maybe, but why? #ancmedsky #dh
LLMs operate on tokens, not characters. Unless their embeddings take fewer tokens, it wouldn't be any faster.
October 31, 2025 at 2:59 PM
LLMs operate on tokens, not characters. Unless their embeddings take fewer tokens, it wouldn't be any faster.
Our 9-part Python + AI live series in October covered LLMs, embeddings, RAG, vision models, structured outputs, safety, tool-calling, agents and MCP.
Grab the recordings, slides, and code from: blog.pamelafox.org/2025/10/watc...
Grab the recordings, slides, and code from: blog.pamelafox.org/2025/10/watc...
Watch the recordings from my Python + AI series
My colleague and I just wrapped up a live series on Python + AI, a nine-part journey diving deep into how to use generative AI models from P...
blog.pamelafox.org
October 31, 2025 at 2:29 PM
Our 9-part Python + AI live series in October covered LLMs, embeddings, RAG, vision models, structured outputs, safety, tool-calling, agents and MCP.
Grab the recordings, slides, and code from: blog.pamelafox.org/2025/10/watc...
Grab the recordings, slides, and code from: blog.pamelafox.org/2025/10/watc...
LLMs are injective and invertible.
They show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.
Paper: www.arxiv.org/abs/2510.15511
They show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.
Paper: www.arxiv.org/abs/2510.15511
October 28, 2025 at 2:05 AM
LLMs are injective and invertible.
They show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.
Paper: www.arxiv.org/abs/2510.15511
They show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.
Paper: www.arxiv.org/abs/2510.15511
The new 1X NEO robot operates largely using a 160 million (with an m) parameter model that takes instructions as text embeddings from an off-board language model. Surprising that a model that small can even do visual understanding, let alone instruction following and movement.
October 28, 2025 at 11:27 PM
The new 1X NEO robot operates largely using a 160 million (with an m) parameter model that takes instructions as text embeddings from an off-board language model. Surprising that a model that small can even do visual understanding, let alone instruction following and movement.