π€― Chances are, you've already spoken to it without even noticing... It can:
Understand emotions
Respond in any voice or language
And is already handling 1M+ phone calls simultaneously - 24/7 - for some of the world's largest businesses.
For enterprises, it's been a financial game changer across sales, operations, customer support, you name it. It's like the perfect employee, so expect major shifts in international and US jobs.
Amazon's Alexa division incurred a $10 billion loss in 2022 and laid off staff, highlighting the unsustainability of its loss leader strategy despite the high household penetration. As enthusiasm for smart assistants like Siri and Google Assistant also wanes, Amazon is banking on generative AI to reinvigorate Alexa's capabilities and user engagement. The company's focus is on enhancing conversational interactions and overcoming the "smart timer" limitation.
Sakana, a Japanese AI firm, has released a system that can autonomously perform research by suggesting hypotheses, performing experiments, writing code, and summarizing the results in well-reasoned papers. The company has provided examples of the papers the system wrote along with an open sourced version of the system.
Replika CEO Eugenia Kuyda recently discussed her vision for AI companions in human relationships, highlighting the app's role in offering friendship, therapy, or romance through avatars. With the evolution of LLMs, Replika aims to complement, not replace, human interaction, creating a new category of relationships. Despite controversies, such as temporary restrictions on adult content, the app's intent remains to improve users' emotional well-being. With a user base in the millions and a team of 50-60, Replika is planning a significant relaunch to enhance interactivity and realism in conversations.
"Lazy visual grounding" is a two-stage method for open-vocabulary semantic segmentation that first discovers object masks without relying on text, then assigns text labels afterward.
OpenAI has introduced a subset of SWE-bench that is easier and more in line with what humans and AI can solve today. It is a good benchmark for validating and working towards before running the entire original benchmark.
ColBERT is an extremely powerful model for retrieval. This new model has only 33m parameters but achieves amazing performance on a number of benchmarks. This post explores how to train a similar model and what tricks led to strong performance.
For serious businesses, going to production involves more than just building a killer app. The team at OctoAI has assembled a panel of industry experts to help take the guesswork out of enterprise GenAI.
Join August 27th and learn the cutting-edge approaches for:
UniBench is a unified framework that simplifies the evaluation of vision-language models (VLMs) by combining over 50 benchmarks into a single implementation. It helps assess the progress of VLMs across various capabilities, from object recognition to spatial awareness.
AI's integration into corporate governance is prompting leaders to devise robust AI strategies for data-driven decision-making. While AI offers valuable insights, especially with LLMs, challenges remain, including skill gaps and ethical concerns. A proper balance between AI and human judgment is essential for future C-suite decision-making processes.
Multion has trained an agent to use self play to perform web queries. Over the course of training, it improved from 18% to 81% on a variety of web-based tasks like ordering food. It uses MCTS and DPO to improve. Researchers from Stanford participated in this work as well and a paper is available on the website. It appears to be built on the xLAM function calling model from Salesforce Research.
A new click attention algorithm improves interactive segmentation. This approach broadens the influence of positive clicks and reduces interference between clicks.
Gemini 1.5 Flash has undergone a price drop, with a 78% decrease on input and 71% cut on output token costs, and its API now supports over 100 languages.
Polymarket has partnered with AI search engine Perplexity to integrate event-related news summaries and data visualizations into its prediction marketplace.