jonah
1.6K posts
@matospiso another one has been hinted at in a linkedin post, hopefully will drop on monday? @zongyuxuan3 :D
English
QME
I'm getting a lot of (well-appreciated) credit for this but it was very much a team effort too! Every single person was essential to getting this over the line: @xmlee97, @joeldierkes, @RuiTheBaker @nainglk, @drexalt, @hannaLicht, @rikiyatakehi (👋🥲) and, of course, @aaxsh18
Ben Clavié@bclavie
I'm so excited to introduce this! We've worked on a million different moving parts to produce this. I'm fairly confident it's the best multimodal model that exists, period -- and it's not too shabby at pushing back the LIMITs of retrieval either...
English
Whole grains are good for health and for SoTA retrieval 🍞🍞
Mixedbread@mixedbreadai
Introducing Mixedbread Wholembed v3, our new SOTA retrieval model across all modalities and 100+ languages. Wholembed v3 brings best-in-class search to text, audio, images, PDFs, videos... You can now get the best retrieval performance on your data, no matter its format.
English
@tomaarsen Sparse updates very exciting, was pumped to see the PRs from @yjoonjang
English
Slowly I am fucking cooked. This is why i am learning to bake salt breads and cookies
himanshu@himanshustwts
a significant % of ml researchers might be hooked by what happened in ONE day. ai seems to be doing a research loop fascinatingly well (understand the problem + propose a change + train/test it + measure results + keep the better version + repeat) and genuinely reducing research friction. we are early to automated experimentation, frontier scale could be an interesting watch.
English
@antoine_chaffin if we only ever eval colbertv2, splade-v3 can keep looking competitive with multi vector :)
English
Different day, same results
Stop living in the past, have a taste of the future
Small models outperforming billion parameters models, the same code as dense models, easy to use indexes…
PyLate has it all, at this point, you are just missing out
lightonai.github.io/pylate/
Omar Khattab@lateinteraction
Wow. It’s absolutely preposterous that ColBERTv2, a 100M parameter retriever, still fricking outperforms Qwen3-Embed-8B, an 80x bigger dense retriever. ColBERTv2 was trained by one dude in 2021 on 4 A100s for 4 days, on top of puny BERT-base. Single-vector models hold IR back.
English
@candyflipline outting myself but 30 mins for the last few weeks on every all-in pod too
English
wait, no ones gonna talk about how artificial the entire openclaw push was?
- jan 26: explodes in stars
- early feb: big sf event
- early feb: lex podcast
- early feb: yc and openai vid
- mid feb: acquisition by openai (for 30m, right?)
it glows in the dark.
stonechat@thestonechat
wait, no ones gonna talk about how shit openclaw's code is?
English
i overdid the steak and she had to drown out the horrible taste with kimchi. worst valentine’s day ever it’s fucking over for me
Air Katakana@airkatakana
none of these racist comments are going to stop me from enjoying a steak and red wine valentine’s day dinner with my extremely attractive (even by korean standards) korean girlfriend who i love very much and maybe if you all took my advice you could find this happiness too
English
this is the only egregious cope i will comment on out of oai
ricky b@rbranson
openai cafe might have the best slice in sf
English
Releasing NextPlaid today at @LightOnIO. It’s a production-ready multi-vector database 🎉
NextPlaid lets you deploy an API in seconds using our pre-built containers. It embeds a multi-vector database and an inference engine for late-interaction models based on ONNX.
English
Look at Moonlight, its a great AI paper reading tool. I like some of the Moonlight features but why the fuck cant I use my API key?
Lazarz@Laz4rz
Huge chance for native “select text to explain and put the rest of the paper in context” tiny tool that would go extremely hard if done well
English
jonah retweetledi
I'm personally very bearish on MUVERA due to its many, many, failure cases, but I have a lot of respect for the @weaviate_io folks, so I gave this another deep read to see if there were things that could change my mind.
However this has me a bit puzzled, if I'm reading the graph below right, it means that MUVERA itself produces a ~50+% incompressible performance degradation at commonly used indexing parameters, and still a ~20% degradation at near-bruteforce search tier parameters (ef=1024), meaning that the degradation would be purely due to MUVERA itself.
For most retrieval uses, this would make the method completely unusable, as this degradation for many workflows is almost similar to the one we'd experience from replacing semantic search with pure bm25/keyword search.
I feel like I'm missing something here so I'm very happy to be corrected if I'm misinterpreting the results!
Femke Plantinga@femke_plantinga
Multi-vector embeddings (ColBERT, ColPali) are budget killers. But MUVERA can cut your memory footprint by 70%. Multi-vector models offer incredible retrieval but suffer from massive memory overhead and slow indexing. MUVERA (Multi-Vector Retrieval via Fixed Dimensional Encodings) compresses these into single, fixed-dimensional vectors. How it works: MUVERA condenses a sequence of vectors (e.g., 100x96d) into one vector via: 1️⃣ Space Partitioning: Groups vectors into buckets using SimHash or k-means clustering. 2️⃣ Dimensionality Reduction: Applies random linear projection to compress each sub-vector while preserving dot products. 3️⃣ Repetitions: Repeats the process multiple times and concatenates results to improve accuracy. 4️⃣ Final Projection: Optional final compression (not used in Weaviate's implementation). The impact (LoTTE benchmark): - Memory: 12GB → <1GB. - Indexing: 20+ mins → 3-6 mins. - HNSW Graph: 99% smaller. There’s a trade-off: You trade a slight dip in raw recall for massive efficiency gains. However, by tuning the HNSW `ef` parameter (e.g., `ef=512`), you can recover 80-90%+ recall while keeping costs low. When should you use MUVERA? → Large-scale production RAG → Systems where memory/infrastructure costs are the direct bottleneck → Use cases requiring fast indexing MUVERA in @weaviate_io 1.31+ takes just a couple of lines of code. You can tune three parameters (k_sim, d_proj, r_reps) to balance memory usage and retrieval accuracy for your specific use case. Read the full technical deep-dive here: weaviate.io/blog/muvera?ut…
English
@staghado @Dorialexander thanks for the eval, was wondering about this one. nice to see lighton keeping the crazy lead 💪
English
@Dorialexander a small improvement from v1 but overall still behind most recent models on OlmOCR-bench including PaddleOCR-VL, OlmOCR, Chandra, Mistral OCR 3, very close to dots.ocr, and way below LightOnOCR-2-1B 👑
English
So DeepSeek is just opening gradually older model artifacts (1.5 years rolling basis?) and still hitting SOTA for size range/inference speed. Good flex.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
DeepSeek OCR 2. First reaction: it uses Qwen2-0.5B. Qwen 2 came out in May-Jun 2024. If they started this work ≤1 year ago, they'd have used 2.5 at least (July 2024). To me this confirms that OCR series have been done in DeepSeek V2 era. It's uncanny.
English