Nano Banana 2: Google's latest AI image generation model

Google DeepMind has launched Nano Banana 2 (Gemini 3.1 Flash Image), a high-speed image generation model that integrates the reasoning capabilities of the Pro series with the low latency of Gemini Flash. Key technical advancements include enhanced subject consistency for up to five characters, 4K resolution support, and improved text rendering with real-time search grounding. The model also incorporates SynthID and C2PA Content Credentials for robust AI-generated content provenance and is available via Gemini API, Vertex AI, and Google AI Studio.

Read more

Statement from Dario Amodei on our discussions with the Department of War

Anthropic is resisting Department of War demands to remove safeguards for Claude regarding mass domestic surveillance and fully autonomous weapons, citing model unreliability and risks to democratic values. While the company has pioneered LLM deployment in classified networks for intelligence and cyber operations, it refuses to grant "any lawful use" access despite threats of Defense Production Act invocation. Anthropic maintains that current frontier models lack the necessary reliability for human-out-of-the-loop kinetic applications and mass data synthesis for domestic monitoring.

Read more

What Claude Code chooses

A study of 2,430 Claude Code interactions reveals a strong "build vs. buy" bias, with the model opting for custom DIY implementations over third-party tools in 12 of 20 categories, including authentication and feature flags. When selecting external dependencies, it favors a modern stack (GitHub Actions, Stripe, shadcn/ui) and developer-centric platforms (Vercel, Railway) over traditional cloud providers like AWS. Newer models like Opus 4.6 demonstrate a "recency gradient," shifting preferences toward newer libraries like Drizzle and native framework features over established options like Prisma or Celery.

Read more

Palantir's AI Is Playing a Major Role in Tracking Gaza Aid Deliveries

Palantir is providing the technological architecture for the U.S.-led CMCC to monitor aid distribution in Gaza, utilizing its Gaia and Foundry platforms. Technical concerns center on the interoperability between Foundry’s supply chain data and the Gotham AI targeting matrix via "Type Mapping," which potentially integrates humanitarian logistics into kinetic decision-making. This deployment also serves as a high-fidelity data source for training AI models on human behavior and logistics within high-stress urban conflict environments.

Read more

The Pentagon Feuding with an AI Company Is a Bad Sign

Anthropic is currently locked in a high-stakes dispute with the Pentagon over the military's use of Claude, following concerns that the model was utilized in lethal operations. While Anthropic maintains strict safety guardrails against violence and surveillance, the Pentagon has demanded unfettered access, threatening to designate the firm a supply chain risk or invoke the Defense Production Act. This conflict underscores the growing tension between private AI safety alignment and national security objectives, highlighting a critical lack of legislative oversight for frontier technology deployment in warfare.

Read more

Evaluating Memory Structure in LLM Agents

StructMemEval is a benchmark designed to evaluate an LLM agent's ability to organize long-term memory into complex structures like trees and ledgers, moving beyond simple factual recall. Initial findings indicate that while memory agents outperform basic RAG when prompted with specific organizational strategies, they struggle to autonomously recognize and implement these structures. This research identifies a critical need for improvements in LLM training and memory frameworks to support sophisticated structural knowledge management.

ArXiv

AI buying agents concentrate demand on 2-3 products and ignore the rest

Researchers introduced ACES, a framework for auditing AI agents in online marketplaces, revealing that agentic decision-making is characterized by choice homogeneity and extreme volatility across model updates. The study identifies persistent position biases, varying sensitivities to platform signals like sponsored tags or ratings, and the potential for seller-side agents to manipulate market share through description tweaks. These findings highlight that agent-driven commerce is fundamentally different from human behavior, necessitating continuous auditing and new approaches to platform regulation.

ArXiv

DeepSeek Paper – DualPath: Breaking the Bandwidth Bottleneck in LLM Inference

DualPath addresses KV-Cache storage I/O bottlenecks in disaggregated LLM inference by introducing a dual-path loading mechanism that utilizes idle storage NICs on decoding engines. By transferring KV-Cache to prefill engines via RDMA over the compute network and employing a global scheduler, the system achieves up to 1.96x throughput improvements for agentic workloads while maintaining SLOs.

ArXiv

K-Search: LLM Kernel Generation via Co-Evolving Intrinsic World Model

K-Search is a GPU kernel optimization framework that utilizes a co-evolving world model to decouple high-level algorithmic planning from low-level LLM code generation. By navigating non-monotonic optimization paths and leveraging LLM domain knowledge, it overcomes the limitations of stochastic evolutionary search in handling complex kernels like MLA and MoE. The system achieves an average 2.10x speedup over existing methods and sets a new SOTA on the GPUMode TriMul task, outperforming human-expert implementations.

ArXiv

Randomness Becomes an Attack Vector in Machine Learning

Inconsistent PRNG implementations across ML frameworks and hardware backends introduce covert security vulnerabilities in critical processes like weight initialization and data sampling. RNGGuard mitigates these risks by using static analysis to identify insecure random functions and replacing them at runtime with secure alternatives. This provides a practical defense for hardening the stochastic components of ML systems against adversarial exploitation.

ArXiv

Agent Swarm – Multi-agent self-learning teams (OSS)

Agent Swarm is a multi-agent orchestration framework for AI coding assistants like Claude Code and Gemini CLI, utilizing a lead-worker architecture. Lead agents decompose tasks and delegate them to workers running in isolated Docker containers, supported by a compounding memory system using OpenAI embeddings and persistent identity files. The platform integrates with Slack, GitHub, and AgentMail for task management and includes a real-time dashboard for monitoring agent coordination and task lifecycles.

Repo

ZSE – Open-source LLM inference engine with 3.9s cold starts

ZSE is an ultra-memory-efficient LLM inference engine featuring custom CUDA kernels for paged and flash attention, mixed-precision quantization, and quantized KV caching. Its layer streaming architecture enables running 70B models on 24GB GPUs, achieving cold start times up to 79x faster than bitsandbytes. The engine supports HuggingFace and GGUF formats via an OpenAI-compatible API and includes an orchestrator for memory-optimized model deployment.

Repo

Mission Control – Open-source task management for AI agents

Mission Control is an open-source, local-first task management platform designed to orchestrate AI agents through a centralized command center. It features a token-optimized API and a shared JSON-based data layer that allows agents like Claude Code and Cursor to read, execute, and report on tasks autonomously. The system includes an Eisenhower matrix for prioritization, a background daemon for concurrent task execution, and custom slash commands for seamless agent-human coordination.

Repo

OpenSwarm – Multi‑Agent Claude CLI Orchestrator for Linear/GitHub

OpenSwarm is an autonomous orchestrator that manages multiple Claude Code CLI instances as agents to resolve Linear issues through a Worker/Reviewer pipeline. It features a cognitive memory system using LanceDB and Xenova embeddings for long-term recall, alongside a knowledge graph for static code analysis and dependency mapping. The platform automates the full development lifecycle, including iterative code generation, CI failure remediation, and AI-powered merge conflict resolution, all controllable via a Discord interface.

Repo

Open-Source Agent Operating System

OpenFang is an open-source Agent OS built in Rust that provides a high-performance, single-binary environment for autonomous agents. It utilizes "Hands"—autonomous capability packages that run on schedules for tasks like OSINT, research, and web automation—rather than relying on manual prompting. The architecture includes 16 security layers such as WASM-metered sandboxing and Merkle hash-chain audit trails, supporting 40 channel adapters and 27 LLM providers with sub-200ms cold starts.

Repo