Datalab’s Lift is a targeted doc extraction instrument with a particular promise: give it a PDF or picture plus a JSON Schema, and it returns schema-shaped JSON immediately. Instead of changing a doc to Markdown first and then asking one other mannequin to extract fields, Lift reads rendered web page photos and makes an attempt…
Netflix’s engineering crew printed a method for handling wide partitions in Apache Cassandra. The analysis work targets Netflix’s TimeSequence Abstraction, a platform for temporal occasion knowledge. TL;DR Dynamic partitioning splits vast Cassandra partitions per TimeSequence ID, asynchronously and transparently, with no utility adjustments. Detection runs on the learn path by way of byte counting and…
Today, OpenAI launched GPT-Live. It is a brand new era of voice fashions. GPT-Live now powers the ChatGPT Voice expertise. The said purpose is pure, real-time dialog with AI. Two variations ship first: GPT-Live-1 and GPT-Live-1 mini. Both roll out to ChatGPT customers globally in the present day. TL;DR GPT-Live is a full-duplex voice mannequin…
NVIDIA has launched Audex (Nemotron-Labs-Audex-30B-A3B), a unified audio-text giant language mannequin. It understands and generates each audio and speech. It additionally retains the textual content intelligence of its spine. The checkpoints, together with a smaller Audex-2B, are launched underneath a noncommercial license. Most multimodal fashions pay a textual content tax. When labs add audio or…
Liquid AI has launched Antidoom, an open-source methodology that targets a typical failure mode in reasoning fashions. That failure mode is the doom loop. In a doom loop, a mannequin emits a span. It then repeats that span many times. The output continues till the context window is exhausted. Small reasoning fashions are extra susceptible…
Tencent’s Hy workforce launched Hy3. Hy3 is a 295B-parameter Mixture-of-Experts (MoE) mannequin. It prompts solely 21B parameters per token. The weights ship below the Apache License 2.0. Hy3 is geared toward reasoning, agentic workflows, and long-context duties. What is Hy3? Hy3’s structure incorporates a sparse MoE with 192 consultants and top-8 routing. Only 8 consultants…
OpenAI has launched two new Realtime fashions in its API. They are named gpt-realtime-2.1 and gpt-realtime-2.1-mini. Both goal low-latency voice and multimodal experiences. The mini mannequin is the notable a part of this launch. It is a mini reasoning mannequin for realtime voice. It ships at the identical value as the earlier gpt-realtime-mini. OpenAI additionally…
Sakana AI has added a new characteristic referred to as Sakana Translate to its chat service, Sakana Chat. It handles bidirectional translation throughout Japanese, English, and Chinese. The translation engine is Namazu, the corporate’s mannequin collection tailored for Japanese. Sakana Translate ships as a free net app. A single account unlocks all three of its…
LlamaIndex has revealed legal-kb, a public reference utility on GitHub. It is described as a data base for authorized paperwork, powered by LlamaIndex Index v2 (the LlamaParse Platform). The venture demonstrates a sample the crew calls a Retrieval Harness for agentic retrieval. The method differs from single-shot retrieval. Instead of 1 embedding search per question,…
In this tutorial, we construct an end-to-end accounts-payable extraction pipeline with lift-pdf, utilizing artificial bill PDFs as managed take a look at paperwork and a structured JSON schema because the goal output format. Instead of treating bill parsing as a easy OCR job, we body it as schema-guided doc understanding: we generate lifelike invoices, outline…