Microsoft Open-Sources TauGrid for GPU Workloads on Kubernetes
TauGrid packages queueing, orchestration, and health monitoring into a single Helm install for Kubernetes GPU clusters.
AI & ML
Frontier models, research results, agents, and developer tooling.
Six Clones of Jev Appear in Two Days, Latent Space Reports
A Latent Space news roundup flags how quickly AI personas can be replicated, with six Jev clones emerging in just two days.
Alibaba Qwen's LiveTranslate Cuts Interpretation Lag to 2.3 Seconds
A new simultaneous interpretation model from Alibaba's Qwen team reduces average delay and adds speaker-aware features across 60 languages.
Hugging Face tokenizers v1: encode, decode, and scaling measured
Hugging Face's new tokenizers v1 release focuses on measured encode/decode performance and scaling behavior.
StepFun's Step 5 Preview: 600B MoE with 1M Context for Agentic Coding
A new sparse mixture-of-experts model from StepFun combines 600B total parameters, 27B active, and a 1M-token context to tackle long-horizon software engineering tasks.
Voice Cloning API Showdown: Similarity, Consent, and Cost in 2026
A hands-on review of seven voice cloning APIs shows that speaker similarity, consent handling, and price per million characters vary significantly across providers.
Quartile Adds ChatGPT Ads to Retail Media Platform
The retail media optimization firm is now a technology partner for ChatGPT advertising, letting existing clients run campaigns through its tools.
AI Lending Overlays Modernize Banks Without Core Replacement
AI platforms can bring modern lending capabilities to banks without forcing a risky teardown of rigid legacy core systems.
Linkup's SPARSEUP: 149M-Parameter Sparse Embedding Model Sets BEIR Mark
Linkup Research's open-source SPARSEUP claims a top retrieval score for a public sparse encoder under 150M parameters.
GGUF vs GPTQ vs AWQ vs EXL2: Choosing the Right LLM Format
A guide distinguishes file containers from quantization methods, clarifying bits-per-weight, calibration, and hardware fit for format selection.
AI Governance Plans Fail When Humans Are Off Shift
Automated exception-handling systems that pause for human review can stall for hours when the qualified decision-maker isn't on duty.
Synthesia Opens 50,000-Square-Foot New York Headquarters
The AI video company has expanded its US footprint with a Manhattan office nearly four times its previous space.
Google Expands AI & Economy Research Team
Google is adding academic advisors and fellows to deepen its study of AI's economic effects.
Google's R4T Framework Cuts Query Fan-Out Time by 12–20x
A new reinforcement-learning-trained diffusion retriever from Google Research reduces the inference bottleneck in complex AI search.
AWS Updates Bedrock AgentCore Runtime for Elastic Memory and Faster Cold Starts
Amazon's new AgentCore runtime aims to improve memory efficiency and reduce cold start variability for Bedrock agents.
MilleMiglia: Realistic Test Generator for Middle-Mile Logistics
Google Research’s new instance generator gives optimization researchers a more practical way to benchmark algorithms for middle-mile logistics.
Moonshot AI's Kimi K3, a 2.8T-Parameter Open Model, Arrives on Amazon Bedrock
The open-weight Kimi K3, claimed by its developer to be the first open model with 2.8 trillion parameters, is now available on Amazon Bedrock with a 1M-token context window.
Federal Register briefly used Chinese AI tool FBI flagged as malicious
A US government website temporarily deployed an open-source Chinese search model that the FBI had previously labeled as a security threat.