A new open leaderboard for Relational Transformers, modded-nanoRT, speeds up relational pretraining by ranking wall-clock time to a quality target on a single GPU. It uses calibrated holdout sets, typed trilinear attention, Muon, and SIGReg, achieving 305.63 seconds on an NVIDIA H200 and promising public release of benchmark, artifacts, and code.
SK Telecom’s “AI for Everyone” campaign promotes wider use of AI in South Korea and outlines plans for data centres, its A.X K2 model, and a personal AI agent. The company says it aims to make AI an accessible everyday service while expanding its AI business.
The post explains that BigQuery AI functions allow regression on two pandas dataframes via a single call. It describes how the BigQuery accessor translates pandas operations into Google Standard SQL, executes a foundation model on BigQuery, and returns results as a pandas DataFrame, with setup steps and code examples using pandas, bigframes, and a Google Cloud project.
Translated title. Original: Warum das Cern mit Mistral ein eigenes KI-Modell entwickelt
CERN and French AI vendor Mistral are collaborating to create a pre-trained Large Physics Model for particle physics, trained on collision data rather than text, described as the first foundation model for collision physics. The article also includes journalistic background on a physics journalist and standard community guidelines for engagement and sharing.
Former Anthropic researcher Jacob Coxon says a chart in the company’s IPO roadshow helped prompt his resignation. Anthropic presents accelerating pretraining efficiency as a business advantage, while Coxon sees it as a warning that self-improving AI may arrive sooner than expected.
Companies increasingly customize open-weight models via continued pre-training and post-training to fit specific domains; the approach often blends multiple models rather than replacing all systems, with data engineering and realistic training environments playing key roles.
The document provides a concise overview of leading LLM companies, model labs, and emerging developers in 2026, detailing dominant foundation and enterprise models, primary technologies, and key markets.
A look at how firms tailor open-models via continued pre-training and post-training to fit corporate tasks, highlighting lessons on when fine-tuning helps, the value of domain data, and the role of specialized models alongside general ones.
ABHBench tests 12 foundation models’ moral choices in first-person, multimodal scenarios adapted from Detroit: Become Human. Most models changed their behavior between text-only and situated scenarios, and reasoning enhancements did not consistently improve their results.
Amulya Vishwanath, ex-NVIDIA robotics executive and founder of Techable VC, describes robotics as still in a basic building-block stage, contrasting with views that a single general foundation model will trigger a rapid, widespread shift in the field. The interview covers her early-stage investing focus in physical AI, and is part of TechCrunch’s Equity podcast ecosystem.