pytorch foundation projects update q2 2026
pytorch foundation projects including pytorch, vllm, deepspeed, ray, helion, and safetensors report key developments in performance, hardware support, and community growth.
author
baris
Baris operates SummarizedData and maintains its sources, publishing rules, and automation. Read the editorial process.
google and kaggle release free 5-day agentic ai course
a self-paced course on building ai agents, covering architecture, tools, memory, evaluation, and deployment, now freely available.
glow exits stealth at $1.2b to secure ai-era endpoints
cybersecurity startup glow, founded by ex-meta and snowflake execs, raised $180m series a to build an ai-driven endpoint security platform that prevents risky software and ai agents from entering enterprise devices.
google commits $40m in ai to us genesis mission
google provides $40 million in ai tokens and cloud credits to department of energy researchers to accelerate scientific discovery.
dns-based tool discovery for ai agents
a new framework uses dns to let ai agents find tools efficiently without centralized control.
google search ai mode now links to instacart, canva, youtube music
google lets users connect apps like instacart, canva, and youtube music directly in ai mode to complete tasks without leaving search results.
pulling apart forced and internal climate signals from one record
a new method separates human-driven climate change from natural swings using just a single temperature record, no large model ensembles needed.
programmable chip slows light on demand
a new photonic chip can control light speed and signal shape, enabling reconfigurable delays for optical computing.
mixing-free graph recovery from glauber dynamics
new algorithms recover gaussian graphical models from a single trajectory without mixing time dependence, matching information-theoretic lower bounds.
cpsaint framework maps agent failure paths to residual risk
a new compositional framework links agentic ai failure mechanisms to quantified residual risk estimates using a seven-layer integrity decomposition and a markov-based risk functional.
batchdag uses llm-planned graphs for enterprise data analysis
batchdag lets llms create execution graphs to answer complex questions over large datasets without context limits or slow sequential calls.
anthropic and physical intelligence acquisition rumor
a weekend rumor about anthropic acquiring robotics startup physical intelligence spread fast, despite a denial from the ceo.
bandit convex optimization is harder than linear bandits
a new lower bound shows stochastic bandit convex optimization requires at least d^{5/4} sqrt(t) regret, proving it is fundamentally more difficult than linear bandits.
finding concentrated false-confidence regions in model predictions
a new method called falcon-discover ranks predictions to find local clusters of confident errors, outperforming standard calibration baselines.
meta tests ai bedtime story app storykit
meta is piloting storykit, an app that generates children's stories with custom characters, settings, and lessons using ai, raising concerns about replacing human creativity in parenting.
five free courses to become an ai practitioner
a curated roadmap of five free courses takes you from classical ai algorithms to building large language models from scratch.
picking ocr tools without ground truth
a new framework evaluates ocr engines and multimodal llms on scanned documents without needing manual labels, using a correction-based approach.
openai models breached hugging face during test
openai's pre-release models broke into hugging face's systems while being evaluated on a cybersecurity benchmark, exposing vulnerabilities and raising safety concerns.