My $0.02 is that a lot of the future research/work there will be figuring out how to identify effective sub-graphs to provide additional context, to avoid having to pass in the entire graph. As well as trying to identify ontology-less structures in real-time, which includes NER and RE, as well as named entity/relationship resolution/de-duplificatio... See more

Text embeddings are a critical piece of many pipelines, from search, to RAG, to vector databases and more. Most embedding models are BERT/Transformer-based and typically have short context lengths (e.g., 512). That’s only about two pages of text, but documents can be very long – books, legal cases, TV screenplays, code repositories, etc can be tens... See more

Developers can now generate human-quality speech from text via the text-to-speech API. Our new TTS model offers six preset voices to choose from and two model variants, tts-1 and tts-1-hd . tts is optimized for real-time use cases and tts-1-hd is optimized for quality. Pricing starts at $0.015 per input 1,000 characters. Check out our TTS guide to ... See more

Unitxt is a python library for getting data fired up and set for utilization. In one line of code, it preps a dataset or mixtures-of-datasets into an input-output format for training and evaluation. We aspire to be simple, adaptable and transparent.

Unitxt builds on separation. Separation allows adding a dataset, without knowing anything about the m... See more

GPT-4 Turbo can accept images as inputs in the Chat Completions API, enabling use cases such as generating captions, analyzing real world images in detail, and reading documents with figures. For example, BeMyEyes uses this technology to help people who are blind or have low vision with daily tasks like identifying a product or navigating a store. ... See more

We ship often

Shipping early is not enough. We also need to keep shipping tiny iterations daily instead of waiting months to ship something big.

Having a strong sense of urgency , and making hard decisions fast is how we enable that process. Our sprints only last 1 week and are quite short when compared to the rest of the industry - this is by desig... See more

Agrawal et al. argue that the framing of AI automation versus augmentation is wrong. Rather than being distinct they are often one and the same. They say that AI, initially intended for automating tasks, inadvertently acts as a force for augmentation of the broader workforce. For example, automating diagnostic skills in healthcare could diminish de... See more

We can detect factually inconsistent summaries via the natural language inference (NLI) task. The NLI task works like this: Given a premise sentence and a hypothesis sentence, the goal is to predict if the hypothesis is entailed by, neutral, or contradicts the premise.