Tag
A tweet highlighting Andrej Karpathy's 70-minute breakdown of how top AI users work with LLMs, covering prompts, context, tools, and workflows, framed as essential viewing for practitioners.
Every’s Dan Shipper discusses on the podcast the decline of traditional newsletters and how to turn 40,000 edits by an editor-in-chief into an AI skill (“Kate Check”), using evaluations and feedback loops to compound expert judgment across the organization.
Ilya Raz interviews ~2 dozen researchers, engineers, and managers about real-world coding agent usage, concluding that fancy harnesses don't work, stock tools like Claude Code/Codex suffice, and that teams must guard against slop infestation, overbuilding, and eroded collaboration.
PyTorch 官方正在对 PyTorch Conference North America 2026 的同期活动 Jupyter Day 进行现场报道,数据科学家、开发者和研究人员齐聚一堂,探讨交互式计算与 AI 辅助工作流。
BotsCrew shares how it standardized Claude adoption across sales, finance, people ops, marketing, and engineering using three components—instructions, context connectors, and playbooks—yielding major time savings in proposals, reporting, hiring, and case study production.
AI models are now able to decompile old video games into source code in days rather than years, enabling native ports, high-FPS enhancements, and modding support for abandoned titles like the original Flash Binding of Isaac. The trend raises both excitement for game preservation and concerns about AI's broader impact.
Boris Cherny argues there is no secret to prompting Claude—just treat it like a coworker, giving it a goal, specifying the effort level, and defining how it should verify results. He notes prompting mattered far more in the Sonnet 3.5 era than it does today.
LlamaIndex announces the shift from traditional one-pass OCR to agentic OCR, using multi-pass verification, layout-aware reading order, and multimodal parsing for complex documents like tables and charts, with a deeper breakdown by Logan Markewich linked in the comments.
Output is an open-source TypeScript framework for building AI workflows and agents aimed at Claude Code users, keeping prompts, evaluations, tracing, cost tracking, orchestration, and credentials unified in one codebase. Key features include workflow folders, version-controlled prompts, automatic tracing, LLM-as-judge evaluations, and durable orchestration powered by Temporal, released under the Apache 2.0 license.
HQ 公开发布,作为面向人类、Agent 和机器人的共享 AI 上下文层,集成 Claude Code、Codex、Cursor 及 MCP,提供共享密钥、知识管理和部署等团队功能,已有近 1000 家企业在使用。
The article discusses the emerging trend of agentic AI tools that can perform actions beyond chat interfaces. It questions whether the pace of this shift is underestimated and explores its implications for job impact compared to previous AI advancements.
An interview with creative veteran Ingi Erlingsson discusses how his two decades of experience in the creative industry shape his use of AI tools like LoRAs and ComfyUI for rapid iteration and workflow building.
NVIDIA's SoL-Pi is an open-source agent harness that reduces token usage by 35-64% and API costs by 50-54% by automating the inspection and refactoring of AI workflows for efficiency.
The tweet asserts that AI agents now dominate inference traffic, with this trend expected to accelerate, leading to agents performing numerous tasks 24/7 and consuming vast amounts of tokens.
The article discusses the importance of addressing model errors in AI workflows, noting that handling cases where AI is unsure or wrong is as critical as the task itself in production environments.
ABrush is an AI studio designed for digital artists, integrating AI models and workflows to eliminate repetitive tasks, speed up iteration, and empower artists with greater creative control.
The article discusses the challenges of comparing local and hosted AI models within agent workflows, highlighted by Raycast v2.2's update to route workflows through various providers. It seeks advice on building provider-neutral evaluations and identifies variables like tool support and latency as hardest to keep constant.
A user discusses the challenge of session bloat and state handoff in multi-agent workflows, seeking advice on solutions like state summarization, external databases, or frameworks such as LangGraph and AutoGen.
This article explains how to use Git worktrees in Magit for parallel development, discusses their advantages over traditional branching, and mentions AI agent applications and the Jujutsu version control system.
The article argues that AI workflows lack true competitive moats, emphasizing that value lies in business insights, accumulated knowledge, and client relationships rather than easily replicable prompts or tools.