Global AI Pulse

Two approaches signal a clear industry shift toward production-grade AI agents.

• Claude Code's dynamic workflows spin up parallel agents that debate...

Bidirectional Evolutionary Search lets language models evolve their own reasoning by breaking problems into goals and applying DNA-like crossover and translocation, outperforming standard methods on logic and scientific tasks.

Nvidia H200 rental prices collapsed from $7/hr to $4/hr in three weeks, a 40% drop signaling the market is pricing in a GPU glut rather than shortage.

Research Breakthroughs

• AdaState for Streaming Video: AdaState introduces self-evolving anchors in autoregressive video diffusion models to enable...

OmniInteract introduces a streaming benchmark requiring real-time omnimodal LLMs to process live audio-visual streams online, detecting triggers and...

Claroty introduced Claroty Claire, a CPS-native AI security agent powered by a language model trained on over a decade of domain data from 6,500+...

Enterprise AI is shifting from pilot experiments to scalable infrastructure decisions around data, security, and operations.

• Model selection is no...

Mem0 dynamically extracts, consolidates, and retrieves conversation memory to overcome fixed context windows, outperforming six baselines on the...

GASP injects geometric priors into VLM transformer layers via a correspondence head and dual losses on video geometry, lifting internal matching...

CoreWeave launched unified agentic AI capabilities that integrate training, inference, observability, and reinforcement learning into a single closed...

Visa has made an undisclosed investment in Replit to explore integrating its payment tools so developers and their AI agents can accept payments...

Gemini 3.1 Flash Image (Nano Banana 2) and Gemini 3 Pro Image (Nano Banana Pro) are now generally available, priced at $0.045 and $0.134 per image...

A study of 214 students encouraged to use LLMs in software engineering teams reveals targeted productivity gains.

• LLMs proved most useful in early...

WorldMemArena evaluates multimodal agent memory via a four-stage Action-World Interaction Loop across 400 tasks.

Key results:

• Better...

• Quantization makes models smaller and faster without crippling accuracy, shifting focus from scale to efficiency.
• High token costs for autonomous...

A new post-training method injects 3D foundation priors from SAM3D and PartField into DINO and Stable Diffusion features, resolving 2D ambiguities like symmetries and repeated parts while cutting manual supervision needs.

Larger models succeed on rare, complex tasks because they suffer less gradient interference from common ones, preserving dedicated features that...

Microsoft's AI Diffusion Report shows accelerating global adoption and AI's growing role in software development, while highlighting emerging...

A new In-Writing decoding method lets LLMs generate free-form reasoning first, then applies structured constraints only after a trigger token. This...

NeuROK learns a latent space of object states and a decoder for plausible deformations, enabling data-driven 4D dynamics generation without predefined...