You don’t amplify a model by giving it more. You amplify it by giving it less, better. Human brains and LLMs fail the same way. Here’s how a trader I follow lost money on US equities last week — three mistakes, in his own words: Read these
Claude’s auto mode — and every other AI agent safety system today — only sees what the agent declares it wants to do. Not what actually happens. For simple tools this is fine. Agent says “run git status”, safety layer checks git status, tha…
Today Anthropic announced auto mode for Claude Code — instead of approving every file write and bash command, Claude now makes permission decisions on your behalf. Safeguards check each action before it runs. 3.1 million views in hours. Gre…
You want to fine-tune a small LLM on your own data. A safety classifier, a code reviewer, a domain expert. The plan is obvious: take an open-weight model, quantize it to INT4, LoRA it, deploy locally. No cloud. No API costs. Then you try to…
Big Tech layoffs aren’t the end of programmers — they’re the end of programmer concentration. AI will scatter us everywhere. In a previous post, I described how GuardClaw secures multi-agent workflows on OpenClaw. Today I want to focus on a…
Another round of layoff rumors from Meta — 25–35% cuts targeting infra, RL, and M10, with E3s re-evaluated at E4 standards. Unverified, but the pattern across Big Tech is unmistakable. Are programmers going extinct? No. But the profession i…
In a previous post, I described how GuardClaw secures multi-agent workflows on OpenClaw. Today I want to focus on a simpler but equally common use case: Claude Code. If you use Claude Code, you have probably faced a familiar dilemma: the de…
In a recent post, I argued that agentic safety is the next frontier — that we need to reason about what AI agents do, not just what they say. GuardClaw is our open-source answer to that challenge: a safety layer that evaluates every tool ca…
As large language models evolve from passive question-answerers into autonomous agents that execute code, call APIs, and manage real-world infrastructure, the safety landscape shifts fundamentally. Traditional AI safety focuses on what mode…
With the rapid progress of models like GPT-4, Claude, and Llama 2, large language models have demonstrated remarkable capability across a wide range of tasks. Yet they remain fundamentally generic — they give the same answer to the same que…
I’m Yingqiang Ge, an Applied Scientist at Amazon. I earned my Ph.D. degree in the Computer Science Department at Rutgers University, New Brunswick, NJ in 2023, under the supervision of Prof. Yongfeng Zhang. Before that, I received M.S. degr…
I’m Yingqiang Ge, an Applied Scientist at Amazon. I got my Ph.D. from Rutgers University (2023, advised by Prof. Yongfeng Zhang), where I worked on Trustworthy AI, fairness, and explainability in recommender systems. During my PhD, I publis…