https://ku5e.com/tags/ai/Recent content in AI on ku5e | Cybersecurity PortfolioHugo -- 0.162.1en-usTue, 12 May 2026 11:14:00 +0000https://ku5e.com/blog/frontier-models-are-overkill-for-most-production-workloads/Tue, 12 May 2026 11:14:00 +0000https://ku5e.com/blog/frontier-models-are-overkill-for-most-production-workloads/Open weight models are close enough to frontier performance for most production workloads. What changed this week, what the frontier models still win on, and what the practical question actually is.https://ku5e.com/blog/the-ethical-ai-company-billed-you-for-using-competitor-tools/Sun, 10 May 2026 22:30:00 +0000https://ku5e.com/blog/the-ethical-ai-company-billed-you-for-using-competitor-tools/Anthropic's billing system scanned user repositories for competitor keywords and charged extra or blocked access when it found them. The gap between stated values and operational behavior.https://ku5e.com/blog/claude-code-on-deepseek-17x-cheaper/Thu, 07 May 2026 10:27:00 +0000https://ku5e.com/blog/claude-code-on-deepseek-17x-cheaper/DeepClaude routes Claude Code's tool calls through DeepSeek instead of Anthropic's models. Same tool ecosystem, approximately 17x lower cost. How it works and where it falls short.https://ku5e.com/blog/ai-outperformed-er-doctors-in-a-harvard-trial/Wed, 06 May 2026 19:22:00 +0000https://ku5e.com/blog/ai-outperformed-er-doctors-in-a-harvard-trial/Harvard published a controlled trial showing AI outperformed emergency physicians in triage diagnoses. What the result means and which conversation it actually changes.https://ku5e.com/blog/the-47-percent-debugging-skill-drop/Tue, 05 May 2026 10:52:00 +0000https://ku5e.com/blog/the-47-percent-debugging-skill-drop/Anthropic's own research documented a 47% drop in debugging skills among developers who used AI coding agents aggressively. The supervision paradox built into these tools.https://ku5e.com/blog/deepseek-v4-broke-the-pricing-argument/Sun, 03 May 2026 21:17:00 +0000https://ku5e.com/blog/deepseek-v4-broke-the-pricing-argument/DeepSeek V4 is open weight, near-SOTA on benchmarks, and costs $1.74 per million input tokens. The pricing argument for closed frontier models just got harder to make.https://ku5e.com/blog/i-built-a-trading-bot-that-runs-its-llm-on-a-jetson-in-my-closet/Thu, 30 Apr 2026 21:03:00 +0000https://ku5e.com/blog/i-built-a-trading-bot-that-runs-its-llm-on-a-jetson-in-my-closet/How I built a paper trading bot that runs its LLM on a Jetson Orin Nano in my home lab — architecture, the bug that would have been catastrophic, and why the model is explicitly blocked from touching execution.https://ku5e.com/blog/i-built-ai-software-in-2009.-here-is-what-asu-s-new-learning-platform-gets-wrong./Tue, 28 Apr 2026 21:15:00 +0000https://ku5e.com/blog/i-built-ai-software-in-2009.-here-is-what-asu-s-new-learning-platform-gets-wrong./Arizona State University launched Atomic, a platform that converts faculty lectures into AI-generated learning modules without faculty consent. The modules are factually wrong in places. This is what happens when institutions treat AI as a credential to sell instead of a discipline to teach.https://ku5e.com/blog/gpt-5.5-images-2.0-claude-design-and-why-i-m-done-listening-to-ai-ceo-debates/Sun, 26 Apr 2026 20:51:00 +0000https://ku5e.com/blog/gpt-5.5-images-2.0-claude-design-and-why-i-m-done-listening-to-ai-ceo-debates/description: Three AI tools shipped this week that changed what I expect from the work. One public exchange between executives reminded me that capability and credibility are not the same thing.https://ku5e.com/blog/the-karpathy-loop/Wed, 22 Apr 2026 19:32:00 +0000https://ku5e.com/blog/the-karpathy-loop/ku5e.com blog article on the Karpathy Loop — the three-component auto-improvement pattern that lets an AI agent run hundreds of optimization experiments without human intervention, and what it means for small teams and business owners.https://ku5e.com/blog/run-a-private-ai-that-reads-your-documents-locally-with-no-internet-required/Wed, 22 Apr 2026 18:58:00 +0000https://ku5e.com/blog/run-a-private-ai-that-reads-your-documents-locally-with-no-internet-required/ku5e.com tutorial — run a private local AI with RAG using PrivateGPT and Ollama. No internet, no API keys, all local.https://ku5e.com/blog/193-applications-taught-me-that-hr-ai-agents-are-an-unmonitored-attack-surface/Mon, 13 Apr 2026 20:09:00 +0000https://ku5e.com/blog/193-applications-taught-me-that-hr-ai-agents-are-an-unmonitored-attack-surface/description: HR AI agents are running application screeners, confirmation senders, denial generators, and support chats. They read unstructured external input and route it into internal processes. That is an injection surface. Most companies did not buy them as security infrastructure.https://ku5e.com/blog/your-dlp-policy-does-not-know-what-your-employees-are-running/Sun, 12 Apr 2026 22:22:00 +0000https://ku5e.com/blog/your-dlp-policy-does-not-know-what-your-employees-are-running/76% of organizations now call shadow AI a definite or probable problem. The tools deployed against it have the same blind spot that plagiarism detectors have against a student who knows humanizer tools exist.https://ku5e.com/blog/how-to-use-ai-as-a-study-partner-without-using-it-as-a-cheat-sheet/Sun, 15 Mar 2026 19:52:00 +0000https://ku5e.com/blog/how-to-use-ai-as-a-study-partner-without-using-it-as-a-cheat-sheet/AI will give you the answer if you ask for it. Whether that answer does anything useful for you depends entirely on how you ask.https://ku5e.com/blog/building-a-cybersecurity-ai-lab-on-a-raspberry-pi-for-under-200/Sun, 15 Mar 2026 19:40:00 +0000https://ku5e.com/blog/building-a-cybersecurity-ai-lab-on-a-raspberry-pi-for-under-200/A Raspberry Pi 5 with the Hailo-8L AI Kit gives you a local, air-gapped inference platform for real network security work. No cloud. No subscription. Under $200.https://ku5e.com/blog/agi-is-good-for-humanity.-it-s-a-problem-for-your-network./Sat, 14 Mar 2026 20:44:00 +0000https://ku5e.com/blog/agi-is-good-for-humanity.-it-s-a-problem-for-your-network./AGI done right looks like Star Trek. The path there runs through a period where attackers get the tools first. That gap is where cybersecurity careers are made.