Created 3 notebooks
從 Memex 到長青筆記· 10/27
親手拆開 AI 時代的軟體· 10/27
AI agent 怎麼管理脈絡· 10/27
Added 10 to 3 notebooks
親手拆開 AI 時代的軟體· 4 added · 10/27
Martin Gratzer explains why he updated his agentic workflow to allow automated PR merging once sufficient guardrails and fresh-context reviews were in place. He notes that human attention has shifted from reading diffs to defining problems and shaping the surrounding system as automated checks catch fewer human-noticed errors.
Over a few months between June and September 2026, OpenAI and Anthropic substantially reduced their large language model prices by up to 80 percent, driven by intense competition with each other and the market pressure of low-cost open-weight alternatives.
This post introduces an eleven-hour video course consisting of eight coding sessions that teach how to build reasoning models from scratch using Python and PyTorch. The materials cover evaluation, inference-time scaling, reinforcement learning, and distillation, acting as a companion to the author's book on building reasoning models.
Recent advancements in AI and developer tools enable software builders to tackle much more ambitious projects, such as complex rewrites, performance optimizations, and domain-specific tools, in a fraction of the time it previously took.
從 Memex 到長青筆記· 3 added · 10/27
Bi-directional links are gaining renewed attention through personal digital gardens and tools like Roam Research, despite historical challenges with global web implementation. Tracing back to Vannevar Bush's Memex and Ted Nelson's Project Xanadu, these links provide social awareness between connected pages while avoiding global web moderation issues when used on single-author sites.
This blog post explores how to develop transformative tools and mediums for thought, arguing that modern computing has not yet realized its visionary potential to deeply augment human intelligence. The authors introduce their experimental mnemonic medium, Quantum Country, designed to make memorization effortless and address the overwhelming cognitive load of learning complex subjects.
AI agent 怎麼管理脈絡· 3 added · 10/27
In this Dwarkesh Podcast episode, OpenAI researcher Noam Brown discusses multi-agent systems, reasoning models, and the implications of scaling test-time compute. They also explore math progress, AI research automation, recursive self-improvement, and alignment.
The text discusses the need for richer, thicker interfaces and better boundary objects to improve collaborative planning and decision-making with agents. It critiques current planning systems as computationally rudimentary while exploring potential solutions for human-agent collaboration.
In a talk on the state of context engineering, Drew Breunig examines how expanding context windows have introduced new challenges rather than solving AI harness issues. He details four primary failure modes—context poisoning, distraction, confusion, and clash—and highlights why precise, deliberate instruction management remains essential.
