Ornith-1.0 and LongCat-2.0 push the open-weight frontier to 1.6T parameters

Ornith-1.0 and LongCat-2.0 expanded the open-weight frontier with MIT-licensed agentic coding and 1.6T-parameter MoE models, while Anthropic forged a discounted Claude deal with California’s state government in sharp contrast to its federal tensions. Arena’s AI evaluation leaderboard hit $100M in annualized revenue eight months after launching its commercial service, and South Korea committed over $900 billion to memory chip production and AI infrastructure to address the global memory shortage.

Model Releases #

Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding #

DeepReinforce / Simon Willison’s Weblog / Hacker News (215 points)

DeepReinforce released Ornith-1.0, an MIT-licensed model family optimized for agentic coding with variants at 9B Dense, 31B Dense, 35B MoE, and 397B MoE, all supporting 256K context windows. The 397B variant achieves 82.4% on SWE-Bench Verified and 77.5% on Terminal-Bench 2.1, placing it at the top of open-source coding benchmarks. Built on Gemma 4 and Qwen 3.5 base models, the key innovation is joint optimization of solution generation and the scaffolding that drives those solutions through reinforcement learning – the model learns not just to code but to improve its own search trajectories, a capability directly relevant to anyone building self-improving coding agents.

LongCat-2.0: Large-Scale MoE with 1.6T Total and 48B Active Parameters #

LongCat / Hacker News (156 points)

LongCat-2.0 is a 1.6-trillion-parameter mixture-of-experts model with 48 billion active parameters per token, making it one of the largest open MoE models available. The model appeared on Hacker News with significant engagement, though full benchmark details were not available from the blog post at fetch time. The scale represents a continued push to close the gap between open-weight and proprietary models through architectural efficiency rather than raw parameter counts.

Developer Tools #

Cursor launches mobile app for remote coding agent oversight #

TechCrunch

Cursor released a mobile application that lets developers initiate new coding agents or connect with agents started on desktop, enabling remote oversight from smartphones. The app integrates with Cursor 2.0’s shift toward independent coding agents operating without constant desktop supervision. This reflects a broader industry transition from writing code to supervising code-writing agents – as Anthropic’s Boris Cherny noted, most of his coding now happens on his phone, a statement that would have been absurd two years ago.

LangChain introduces dynamic subagents in Deep Agents #

LangChain Blog

LangChain added dynamic subagent capabilities to its Deep Agents framework, allowing parent agents to create and manage specialized sub-agents at runtime rather than relying on statically defined agent architectures. The feature enables more flexible multi-agent systems where the agent topology adapts to the task. For teams building complex agentic workflows, dynamic subagent spawning reduces the architectural rigidity that currently forces developers to pre-define every possible agent interaction pattern.

Research & Papers #

Agents-A1: Scaling the Horizon, Not the Parameters #

arxiv:cs.CL

Agents-A1 is a 35B MoE model that reaches trillion-parameter-level performance by scaling the agent horizon rather than the model itself. The approach scales two dimensions simultaneously: long-horizon trajectories and heterogeneous agent abilities, supported by a knowledge-action infrastructure connecting external knowledge, actions, observations, and verifier outcomes. For builders of agentic systems, the core insight is that investing in longer action sequences and richer tool integration can substitute for raw model scale – a 35B agent with the right horizon outperforms much larger models with limited horizons.

Always-On Agents: A Survey of Persistent Memory, State, and Governance #

Hugging Face Daily Papers

This survey treats always-on agents as persistent-state systems whose future behavior depends on durable state accumulated across interactions – not just retrievable memories but task ledgers, permissions, credentials, commitments, provenance records, shared state, and trigger conditions. The paper reads the literature through six diagnostic axes for each state item, providing a structured framework for evaluating agent memory systems. For anyone building agents that persist across sessions, this is a comprehensive map of what state management actually entails beyond simple RAG retrieval.

MCP Server Architecture Patterns for LLM-Integrated Applications #

Hugging Face Daily Papers

The first software-maintenance paper examining the Model Context Protocol ecosystem catalogues five recurring MCP server architectural patterns observed across production deployments. Since hundreds of community-built MCP servers appeared on GitHub within months of Anthropic’s November 2024 release, no prior work had systematically described how this ecosystem is being structured. For teams building or maintaining MCP servers, the patterns provide tested reference architectures rather than ad-hoc designs.

SEVA: Self-Evolving Verification Agent with Process Reward #

arxiv:cs.CL

SEVA is a structured verification agent that goes beyond binary pass/fail labels, emitting evidence alignments, step-by-step reasoning chains, calibrated confidence, and a six-category error diagnosis with actionable fixes. The agent self-evolves through process reward signals rather than requiring human-labeled verification data. For teams building agents where hallucination is the reliability bottleneck, SEVA’s structured output format lets downstream agents self-correct based on specific error categories rather than opaque rejection signals.

Self-Evolving World Models for LLM Agent Planning #

arxiv:cs.CL

WorldEvolver is a framework where a world model revises its own deployment-time context while keeping the downstream agent and all model parameters frozen. The system integrates trajectory prediction with self-correcting mechanisms that improve without retraining. For long-horizon agent planning, the approach separates the world model’s accuracy from the agent’s decision-making quality, allowing planning to improve independently of the underlying language model’s capabilities.

Security #

Robust AI Security and Alignment: A Sisyphean Endeavor? #

IEEE / NIST / Lobsters AI

A NIST-covered IEEE paper argues that AI security and alignment inherently require continuous monitoring and updating rather than one-time solutions, drawing the Sisyphean analogy that securing AI systems is an ongoing process that never reaches a fixed endpoint. The paper provides mathematical grounding for the claim that static alignment approaches are fundamentally insufficient, supporting the transition toward continuous monitoring architectures. For production AI deployments, the implication is that security budgets should be modeled as operational expenses, not one-time capital investments.

Regulatory & Policy #

Anthropic and Governor Newsom forge deal for California government to use Claude at half price #

TechCrunch

Governor Newsom and Anthropic established an agreement granting California state agencies and local governments discounted Claude access along with training and support. The deal follows Newsom’s March executive order to accelerate AI use in government, framing it as an efficiency initiative with safety considerations. The partnership is notable for the contrast it draws with federal policy: while the Pentagon labeled Anthropic a “supply-chain risk” and chose OpenAI after Anthropic pushed back on surveillance and autonomous weapons oversight, California is embracing the same company’s technology as a preferred government tool.

TIDAL cracks down on AI music by cutting off monetization #

TechCrunch

TIDAL will strip monetization from fully AI-generated music effective July 15, tag such tracks with an “AI” badge, and use automated tools to remove AI-generated content that attempts to impersonate artists. With Deezer reporting that 44% of daily uploads are now AI-generated, the policy joins similar moves by Spotify, Apple Music, and Qobuz. The trend signals that streaming platforms are converging on treating AI-generated content as a distinct, non-monetizable category – a policy framework that will shape how generative music tools are used commercially.

Funding & Business #

Arena, the AI leaderboard everyone uses, is now a $100M business #

TechCrunch

Arena, the crowdsourced AI evaluation platform that originated as a UC Berkeley research project in 2023, reached $100 million in annualized run-rate revenue just eight months after launching its commercial service. Revenue comes from “AI Evaluations,” a paid deep-dive analytics service for model labs and enterprises, while the public leaderboard remains free. The speed of commercialization – from academic project to $100M ARR in under three years – reflects that model evaluation has become critical infrastructure, not just a research tool, as AI providers increasingly need post-training optimization data to differentiate on quality.

OKX launches AI agent marketplace for autonomous payments and hiring #

TechCrunch

Crypto exchange OKX launched OKX AI, a marketplace where autonomous software agents can discover services, make payments via stablecoins, build persistent identities, and hire each other. Early partners include CertiK for security assessments and GenLayer for dispute resolution, with developer access through tools compatible with Claude Code and OpenAI Codex. While the “agentic economy” vision remains speculative, the platform addresses a genuine infrastructure gap: traditional payment rails were not designed for autonomous software making micropayments at high frequency, and blockchain-based settlement removes the need for human intervention in agent-to-agent transactions.

Infrastructure #

South Korean tech giants commit over $900B to memory chips, AI data centers, and humanoid robots #

TechCrunch / Ars Technica

Samsung, SK Hynix, and other Korean conglomerates announced over $900 billion in combined investments: $518 billion for four new memory chip factories, $52 billion for high-bandwidth memory packaging, and $356 billion for AI data centers through 2035. The commitment targets the global memory shortage dubbed “RAMageddon” that has driven Micron’s revenue up 4x year-over-year. President Lee framed semiconductors, physical AI, and data centers as South Korea’s “triple axis” for industrial leadership, though execution risk is significant – these multi-year construction projects depend on demand persisting through rapidly evolving technology cycles.

Claude models now generally available on NVIDIA GB300 Blackwell Ultra in Azure #

NVIDIA Blog

Anthropic’s Claude models running on NVIDIA GB300 Blackwell Ultra GPUs through Microsoft Foundry on Azure are now generally available, giving Azure-native enterprises production access to Claude for building autonomous and domain-specific AI agents. The deployment includes NVIDIA’s Secure Agent Workspace Reference Design for governed agent execution with identity, network, and credential controls. The GA milestone operationalizes the November 2025 strategic partnership between Microsoft, NVIDIA, and Anthropic, closing the gap between announcement and enterprise deployment.

Popping the GPU Bubble: Scheduling optimizations that outperform hardware upgrades #

Moondream / Hacker News (143 points)

Moondream’s analysis reveals that during LLM token generation, GPUs frequently sit idle waiting for CPU scheduling decisions rather than performing computation. Their Photon inference engine uses three techniques – ping-pong buffering, deferred sampling, and zombie sequence recycling – to overlap CPU and GPU work, achieving 6-12% speedups on NVIDIA 3090s and 21-35% on B200s at batch size 32. The implication for AI infrastructure planning: faster GPUs actually amplify scheduling inefficiency (CPU overhead becomes a larger fraction of total time), meaning scheduling optimization becomes more valuable, not less, as hardware improves.

Other #

AI jobs report: high-intensity AI adopters increased headcount 10.2% #

TechCrunch

A Ramp and Revelio Labs analysis of nearly 22,000 companies found that “high-intensity AI adopters” – companies spending approximately $30 per employee monthly on AI – increased headcount by 10.2%, with entry-level positions growing 12%. The data complicates the job-loss narrative but comes with a critical caveat: only companies with resources for sustained AI strategies saw hiring gains, while those merely experimenting with subscriptions did not benefit. The findings suggest AI’s employment effects may be primarily distributional – concentrating growth in well-capitalized firms – rather than universally additive or subtractive.

Threads to Watch #

Agent evaluation and testing has become commercially viable infrastructure. Arena’s leap from academic project to $100M ARR in under three years, combined with today’s MCP Server Architecture Patterns paper and SpreadsheetBench 2 joining the growing list of agent-specific benchmarks, signals that evaluation is no longer a research side-effect but a market in its own right. As model differentiation increasingly depends on post-training optimization rather than raw capability, the companies that control evaluation methodology will exert outsized influence on which models win enterprise adoption.

The physical AI stack is attracting capital commitments that dwarf model development budgets. South Korea’s $900B+ for memory fabs and data centers, combined with the NVIDIA-Microsoft-Anthropic deployment on GB300 Blackwell Ultra and Moondream’s evidence that scheduling optimization matters more than hardware upgrades, suggests that AI infrastructure is entering an era where the binding constraint is not model capability but the physical and systems-level engineering to deploy it efficiently. The contrast between billion-dollar infrastructure commitments and the fact that a scheduling tweak can deliver 35% inference speedups highlights that software optimization and hardware investment are complementary, not substitutes.

Sources Unavailable Today #

These sources could not be fetched today. Links point to their homepages so you can check them directly.

• Weights & Biases: Fully Connected – scrape: SPA content not accessible