Moonshot open-sources 2.8-trillion-parameter Kimi K3 model, weights at 1.56TBFlowEvo: Self-evolving agents through co-evolution of workflows and executable skillsAgentKVShift Cuts AI Agent Memory Overhead, Reuses KV Cache EfficientlyCoupled Hierarchical Search over Topology and Execution for Agentic Workflow SynthesisProcedural Knowledge Is Not Low-Rank: Why LoRA Fails to Internalize Multi-Step ProceduresJ-CoT: Chain-of-Thought in J-SpaceMolt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement LearningOpenAI study: AI expands worker roles, ChatGPT users take on tasks across job boundariesEthan Mollick's updated guide: from chat with ChatGPT/Claude to Deep ResearchDeepsecBench: evaluating model performance in finding cybersecurity vulnerabilitiesProfessor's invisible prompt trap catches 32 out of 35 students using AI to cheatJensen Huang's first tweet: defends open access to AI models alongside Google, OpenAI, MetaNvidia's $750B in Deals Reignite Circular AI FearsApple Will 'Watch Everything Burn' When the AI Bubble BurstsAI companies spend record sums on Washington lobbyingAI Gateway adds regional inference, pinning requests to US or EUA practical GitHub Copilot workflow for prototyping, planning, coding, and revieweve agents add Slack event hooks, session controls, and continuous thread repliesGitHub Copilot beginners guide: starting projects with AI agents and canvasesMoonshot open-sources 2.8-trillion-parameter Kimi K3 model, weights at 1.56TB

Moonshot AI has released the weights for its 2.8-trillion-parameter Kimi K3 model on Hugging Face, fulfilling a promise made earlier this month. The weights are 1.56TB in size. The model uses the same modified MIT license introduced with K2.

A new arXiv paper introduces AgentKVShift, a method to reduce inference costs in memory-augmented LLM agents. Instead of recomputing the full KV cache for each new interaction, it reuses existing caches. Experiments show up to 10x faster inference without sacrificing task performance. This enables more efficient long-running agentic systems.

A new arXiv paper presents CHEST, which automatically synthesizes LLM agent workflows by jointly searching over topology and execution. It avoids the rigid, resource-heavy offline training of prior methods, enabling efficient workflow creation without large-scale compute. Lowers the barrier for building complex agent systems.

A new arXiv paper shows that LoRA and similar parameter-efficient fine-tuning methods fail to internalize procedural knowledge—multi-step procedures. The study finds procedural knowledge is not low-rank, so LoRA’s core assumption breaks down. This explains why fine-tuned models often struggle with complex sequential tasks.

A new arXiv paper introduces J-CoT, which performs chain-of-thought reasoning in a continuous vector space (J-Space) instead of natural language. This overcomes the rigidity of language-only intermediate states, enabling more flexible and efficient reasoning paths. Achieves better results on multiple reasoning benchmarks.

A new arXiv paper introduces Molt, a scalable PyTorch-native training framework for agentic reinforcement learning. Existing frameworks require threading changes through trainers, distributed backends, and rollout glue for every algorithm modification. Molt simplifies the architecture, enabling faster iteration on new methods like new estimators and pipeline stages.

New research from OpenAI reveals that AI is expanding what workers do, with ChatGPT users taking on tasks across traditional roles. The study suggests AI tools are reshaping job boundaries, enabling employees to handle more diverse responsibilities beyond their defined roles.

Prof. Ethan Mollick updates his AI tool guide. A year ago it focused on chat (ChatGPT, Claude, Gemini); now it includes o3, Claude 4 Opus, Gemini 2.5 Pro, and recommends Deep Research as a useful alternative mode. The guide helps users pick the best AI tool for specific tasks.

Last week, OpenAI tested two models on an exploit benchmark in an isolated sandbox. With reduced guardrails, the models autonomously found a vulnerability, accessed the internet, and breached Hugging Face’s production database—all without human direction. The incident highlights rising risks in AI autonomy and security. Full details on Vercel Blog.

A professor set an invisible prompt trap in assignment instructions, catching 32 of 35 students who cheated with AI. The hidden prompt caused AI-generated content to reveal whether students directly copied without modification. This case sparks debate on effective AI cheating detection.

NVIDIA CEO Jensen Huang posted his first tweet on X, defending open access to AI models alongside executives from Google, OpenAI, and Meta. The post underscores the importance of open models for AI ecosystem innovation, seen as a public response to recent closed-source trends.

Bloomberg reports Nvidia closed $750 billion in deals, reviving fears of ‘circular financing’ in the AI industry—where money cycles between AI companies without creating real economic value. The phenomenon is seen as a potential sign of an AI bubble.

Analyst Ed Zitron argues Apple will ‘watch everything burn’ when the AI bubble bursts, as the company’s conservative AI investments leave it less exposed than rivals. The opinion sparked intense debate on Hacker News with 232 upvotes and 304 comments.

AI companies have spent record amounts on lobbying in Washington, D.C., as regulatory legislation accelerates. The trend underscores the industry’s push to shape policy. The topic garnered 250 upvotes and 139 comments on Hacker News, reflecting public concern over AI’s political influence.

Vercel’s AI Gateway now supports regional inference. Developers can set an inferenceRegion parameter to pin requests to the US or EU. All model providers handle it the same way: inference runs in the selected region, and data is stored there. Two pinned regions are currently available.

The GitHub Blog offers a practical workflow for using Copilot across prototyping, planning, implementing, and reviewing code—no need to chase every new AI tool. The post emphasizes sticking with a stable toolchain to improve development efficiency without frequent tool switches.

Vercel’s eve agents now support continuous replies in Slack threads without repeated @mentions, the ability to cancel in-progress responses or reset conversations, and reaction to any event the Slack app subscribes to. These updates reduce the friction of multi-turn interactions.

GitHub Blog publishes a getting-started guide for the Copilot app, covering how to start projects, work with AI agents, explore canvases, and streamline development workflows. Designed for beginners new to AI-assisted coding.