← All episodes
Daily AI Brief

How Together AI built the world’s fastest speech-to-text stack

Thursday, May 28, 2026 · 25 stories · 5:43 · 665 words

On Thursday, May 28, 2026, a mixed bag of AI and data engineering news emerged, with notable developments in speech-to-text, trustworthy AI evaluations, and data lakehouse advancements. Together AI built the world's fastest speech-to-text stack on Artificial Analysis by treating ASR as a full-path systems problem, not just a GPU inference problem, achieving a significant milestone in the field.

OpenAI published a case study sharing guidance on third-party AI evaluations, covering how to assess model capabilities, safeguards, and validity for frontier systems, providing valuable insights for developers and researchers. This guidance is particularly relevant for evaluating complex AI models, with 10 key considerations outlined in the study.

DuckDB posted a update on their Iceberg integration, with new features in v1.5.3 supporting Iceberg Tables and Iceberg REST Catalogs, including `MERGE INTO`, `ALTER TABLE`, partition transforms, V3 support, and others, making it easier to manage and analyze data. Hugging Face also released a beginner's guide to PyTorch profiling, providing a step-by-step introduction to using `torch.profiler` for optimizing model performance.

On the application side, Confluent introduced the Customer Intelligence Hub, a single pane of glass for customer insight and action, helping GTM teams detect risk, uncover growth, and act faster with real-time context, leveraging AI-powered insights to drive business decisions. ClickHouse recapped Open House 2026 Day 2, covering 10 years of open source, the House Mates partner program launch, and ecosystem announcements from Fivetran, Sigma, dbt, Notion, Replit, and others, highlighting the growing importance of open-source data management.

Meanwhile at AWS, Azercell Telecom LLC built an Azerbaijani large language model on Amazon SageMaker AI for telecom use cases and customer-facing applications, demonstrating the potential of AI in regional languages. AWS also published a post on building a custom portal with embedded Amazon SageMaker AI MLflow Apps, providing a detailed guide on using a React front end paired with a Flask reverse proxy to create a seamless user experience.

In funding news, a number of significant investments were announced, with Together AI and OpenAI making notable strides in their respective areas. On the infrastructure side, Google Research highlighted advancements in Dataflow, a platform for processing massive datasets for machine learning, which has been used to solve scaling problems in data processing for over 20 years.

A quieter note from Databricks: they posted a benchmark on scaling AI agents across organizations, featuring insights from enterprise leaders, including Dee Fitzgerald, CDO of Danone, and Prem Natarajan, EVP and Chief Scientist at Capital One. Databricks also announced the general availability of Iceberg v3, Open Sharing, and Unified Governance, marking a significant milestone in the development of the open lakehouse.

OpenAI published a case study on how Endava uses Codex to build an agentic organization, accelerating software delivery and reducing requirements analysis from weeks to hours, demonstrating the potential of AI in software development. Microsoft Research introduced Data Formulator 0.7, a platform for AI-powered data analytics, enabling data teams to easily bring enterprise data into an AI-ready workspace for exploration, analysis, and visualization.

Google Cloud AI announced the general availability of Nano Banana 2 and Nano Banana Pro, which are already powering creative workflows across industries, enabling organizations to unlock new ways to use image generation and editing. NVIDIA Research advanced robotics from simulation to the real world, presenting new developments at the International Conference on Robotics, and GeForce NOW launched 007 First Light, a cloud gaming experience that lets members play from almost any device.

Snowflake announced Claude Opus 4.8 on Snowflake Cortex AI, bringing advanced reasoning and agentic capabilities to their secure governed AI platform, and also published a post on retail predictive analytics for like-for-like growth, exploring how retail and consumer goods leaders use AI, store intelligence, and supply chain analytics to drive growth.

And one for the security folks: Vercel AI published a post on building a real-time power outage map with Next.js on Vercel, demonstrating the potential of AI in monitoring and responding to critical infrastructure events. That's the Daily AI Brief for Thursday, May twenty-eighth. More tomorrow.

Sources covered

1 story · 25 sources (0 news, 25 company, 0 voices)

  1. How Together AI built the world’s fastest speech-to-text stack