AI Infrastructure Pulse

BDH-CQ presents in-context learning with recurrent latent reasoning, pointing to more efficient test-time approaches for smaller models.

Anthropic's largest acquisition aims to close two critical weaknesses at once.

• Video leap: Decart's Lucy 2.5 and Oasis 3 world models deliver...

• ToolHazard scales adversarial environments via simulators to discover injection points and generate payloads for agent evaluation.
• It produces...

Jeff Dean is in talks for a $10B valuation on his new AI startup, underscoring strong investor interest in high-profile AI ventures leaving big tech.

A new peer-reviewed study finds that AI tools helping oil and gas companies could generate 3.3 to 13.3 times more emissions than powering AI data...

MiniMax H3 delivers free multimodal video generation via text, image, and frame control, plus voice transfer and precise editing.

Key capabilities...

The shift toward local AI execution is gaining clear traction, moving from browser demos to practical on-device agents.

• Transformers.js hit 10M...

Cloudera survey highlights real-world barriers to AI scaling

• Cost surge: 84% report higher infrastructure costs tied to AI workloads
• Governance...

Qwen3.8-Max launched as a 2.4T-parameter open-weight model with 95B active params, 1M context, and agentic coding capabilities.

Unsloth then enabled...

Genesis makes the software project itself the persistent entity, letting finite-lived agents propose changes through recursive delegation while only...

Multiple approaches now target cost-quality Pareto frontiers in reasoning systems.

• Reasoning modes outperform non-reasoning on agentic tasks but...

• Capy v2 beats Claude Code, Codex, Devin and Cursor on DeepSWE while costing 50% less
• VMs deliver 4x lower cost and 10x faster performance with...

AI models are already supporting high-stakes national security and foreign policy decisions, yet evaluation methods remain far too weak for these...

HireRoad slashed an 18-month product rewrite to 15 weeks by redesigning team structures and daily work around AI tools instead of layering them onto...

NVIDIA enters efficient open-weight agents with Nemotron 3.5 Lightning, a 30B MoE model (3B active params) delivering up to 4× faster output for...

Two advances push large-scale human behavior simulation for interactive AI:

• MatrAIx simulates the world with 8.3 billion persona agents
• UserIDA...

• Open positioning: Meta releases Muse Glimmer (30B local agentic model) under Apache 2.0 and plans to open Muse Spark 1.2 weights, directly...

AI investment is accelerating at unprecedented scale, with capital flows reshaping both private markets and future fundraising.

• UK AI firms raised...

SPOT tackles key limits in standard on-policy distillation by using an acquisition-exploration-exploitation loop: it allocates probing budget via...

Databricks' new FILE column type stores unstructured data like images, video, and audio natively in lakehouse tables, making it AI-ready with unified...