One cool thing about ColBERT-based search compared to the... cosine-based vector retrieval is that you get interpretability for free as a byproduct of the MaxSim computation. It's kind of like the Lucene highlighter, letting you grab the most relevant snippets from a long document to show users where their query matches. With Jina-ColBERT-v1, which supports up to 8K token length, released by us earlier this Feb., the visualization of the late interaction between a query and a document is almost... artistic. The video shows the late interaction between the query "Elephants eat 150 kg of food per day." and the Wikipedia article about "Indian Elephant". Darker colors indicate stronger semantic matches. The darkest area corresponds to "The species is classified as a megaherbivore and consume up to 150 kg (330 lb) of plant matter per day." from the original article.show more

Vector Database by Hand ✍️ Vector databases are revolutionizing... how we search and analyze complex data. They have become the backbone of Retrieval Augmented Generation (#RAG). How do vector databases work? [1] Given ↳ A dataset of three sentences, each has 3 words (or tokens) ↳ In practice, a dataset may contain millions or billions of sentences. The max number of tokens may be tens of thousands (e.g., 32,768 mistral-7b). Process "how are you" [2] 🟨 Word Embeddings ↳ For each word, look up corresponding word embedding vector from a table of 22 vectors, where 22 is the vocabulary size. ↳ In practice, the vocabulary size can be tens of thousands. The word embedding dimensions are in the thousands (e.g., 1024, 4096) [3] 🟩 Encoding ↳ Feed the sequence of word embeddings to an encoder to obtain a sequence of feature vectors, one per word. ↳ Here, the encoder is a simple one layer perceptron (linear layer + ReLU) ↳ In practice, the encoder is a transformer or one of its many variants. [4] 🟩 Mean Pooling ↳ Merge the sequence of feature vectors into a single vector using "mean pooling" which is to average across the columns. ↳ The result is a single vector. We often call it "text embeddings" or "sentence embeddings." ↳ Other pooling techniques are possible, such as CLS. But mean pooling is the most common. [5] 🟦 Indexing ↳ Reduce the dimensions of the text embedding vector by a projection matrix. The reduction rate is 50% (4->2). ↳ In practice, the values in this projection matrix is much more random. ↳ The purpose is similar to that of hashing, which is to obtain a short representation to allow faster comparison and retrieval. ↳ The resulting dimension-reduced index vector is saved in the vector storage. [6] Process "who are you" ↳ Repeat [2]-[5] [7] Process "who am I" ↳ Repeat [2]-[5] Now we have indexed our dataset in the vector database. [8] 🟥 Query: "am I you" ↳ Repeat [2]-[5] ↳ The result is a 2-d query vector. [9] 🟥 Dot Products ↳ Take dot product between the query vector and database vectors. They are all 2-d. ↳ The purpose is to use dot product to estimate similarity. ↳ By transposing the query vector, this step becomes a matrix multiplication. [10] 🟥 Nearest Neighbor ↳ Find the largest dot product by linear scan. ↳ The sentence with the highest dot product is "who am I" ↳ In practice, because scanning billions of vectors is slow, we use an Approximate Nearest Neighbor (ANN) algorithm like the Hierarchical Navigable Small Worlds (HNSW).show more

HOLY SHIT...you can "inference" LLMs in Sora 🤯 the... prompt was: "Open chatgpt and send a message!" How insane is it that the generated audio is not only a relevant response to the query that Sora made up out of nowhere, but the haiku is syllable-accurate?! 🥲show more

I tested Nutrient on a CMS-1500 health claim, which... is exactly the kind of document that can trip up extraction systems. The red grid alone can get mangled by scanners, and the form is packed with dense fields where one wrong value can matter. What stood out to me was the grounding. Hover over any extracted field and Nutrient shows the exact place on the original claim where that value came from. So instead of asking “did the AI probably get this right?”, you can check the evidence yourself. For healthcare documents, that feels like a much better standard: not just extraction, but extraction you can actually verify.show more

In 2000, Jennifer Lopez attended the Grammy Awards wearing... the now-infamous green Versace "Jungle Dress". The striking, plunging gown immediately became a global sensation, generating an unprecedented demand from users searching for photos of the dress. At the time, Google's search results were primarily text-based, and the company realized they couldn't directly satisfy the massive, visually-driven query. This overwhelming and specific search demand, which became the most popular query Google had ever seen, served as the direct catalyst and motivation for the company to develop and launch Google Images in 2001. . . . #jenniferlopez #grammys #google #tech #usashow more

Here is a 3D visualization of the coming solar... storm using model parameters from NASA's run, NOAA's run (shown below), and the coronagraph views from STEREO-A/COR2 and SOHO/LASCO. The storm is expected to hit Earth early December 9. The 3D visualization clearly shows the low inclination orientation of the solar slinky core inside the larger storm envelope. Of course, the earlier eruption that erupted to the northeast, might deflect this structure a little bit, but likely that will only make the structure move more towards the ecliptic plane, which will give us a larger chance of crossing through the solar slinky core. One nice thing about low inclination flux rope cores is that they almost ALWAYS have southward field in them at some point. So this is good news for aurora photographers-- as long as this solar storm stays true to its launch trajectory and inclination, we could see some organized southward pointed magnetic field. Considering this solar storm is fast-- that usually indicates the storm will not deviate much from its original launch trajectory! Here is the link to a static frame of the 3D frame visualization so you can fly around it yourself:show more

Due to a bizarre interaction between Pokeys and Turn... Blocks in Super Mario Maker 2, it is possible to create a "Jumpscare Pokey" that rises up extremely quickly and unexpectedly from the bottom of the screen.show more

DOM3 is made from HTML, CSS and Javascript, closer... to a document than a canvas. It's a strange contemporary material that shapes much of our digital lives. The rhythm of endless creation and deletion in DOM3 is a self-driving version of what happens when you browse the web. "The Document Object Model (DOM) connects web pages to scripts or programming languages by representing the structure of a document—such as the HTML representing a web page—in memory." A html element, and a rectangle on a canvas, are not the same thing. It's a different material. Even if it all ends up controlling the same pixels, the process is different. The html element kind of actually exists, in this abstract context of a document object model. And in memory, in the moment. If it was possible to measure its weight, it would be heavier than a pixel on a canvas.show more

Playphrase. me is a website where you can look... up a phrase and it'll show you the same phrase found in movies or tv shows. It shows you over 1000s of matches and it's completely free. This is great if you want to find funny clips for videos saying the same phrase.show more

In Prisma Next, you can stream query results straight... from the database, row by row and fully typed. That means exports, reports, and ETL jobs of any size run smoothly, with the same safety you get on a normal query.show more

how based are you? it’s nice to see how... more and more new apps are appearing inside the base app every day. it’s a great opportunity for users to get in early and earn rewards it’s also a big boost for the base ecosystem as a whole: active users get rewarded, the ecosystem grows, developers earn, users earn. a real win-win that’s what “base is for everyone” is all about “base me” is a fun app from the creator Fokki. you can spin the wheel and find out just how based you are. i got lucky and ended up being as based as jesse.base.eth you can join and try it yourself using the link belowshow more

I built JevSearch, search the web & validate your... results with Jev. Give a query and selection criteria, use Browserbase search to get the t25 results, then Jev scores and returns the t5 results. Jev often chooses urls outside of the initial top 5 as more relevant.show more

Head Keeper Edwin shares a tender interaction with Kitich,... one of our Nairobi orphans. As Ed gently scratches Kitich’s head and tongue, the young elephant’s eyes close in pure contentment, highlighting the bonds of love and trust that develop between the elephants and their Keepers. Learn more about our work to rescue and rewind orphan elephants at:show more

Each year, gray whales embark on a 12,000-mile round-trip... migration, moving from the Arctic to the sheltered and warm waters of the Baja lagoons to calve and breed. Averaging about 75 miles per day, the journey is one of the longest migrations of any mammal.show more

The giant oceanic manta ray is the largest type... of ray in the world. It can grow to a disc size of up to 7 m (23 ft) across with a weight of about 3,000 kg (6,600 lb). Scuba diver for scale. [📹 deepbluecoast]show more