AR | Build Notes

14 AI models. One simple question about a Sherpa who made history. 13 wrong answers!!! Hallucinated names, empty responses, or "I don't know." The question: Who was the sherpa who climbed 7 Himalayan peaks in a few months? These are all local models running on my own hardware — no API, no cloud, just raw inference on a Minisforum MS-S1 Max (96GB unified RAM) and a Mac Mini M4 (24GB RAM). Here's who failed: llama3.2:1b — deflected to Tenzing Norgay and Edmund Hillary deepseek-coder-v2:16b — confidently fabricated a detailed biography of the wrong person deepseek-v2:16b — named Apa Sherpa, then contradicted itself mid-sentence gemma2:2b — gave Kami Rita Sherpa's Everest summit count instead llama3.2:3b — mentioned Nirmal Purja (different feat, different person) qwen2.5:3b — invented "Pasang Lhamu climbing 7+14 peaks in 2014–2015" phi3:mini — said "no widely recognized record matches this description" qwen3-coder:30b — guessed "Babu Chiri Sherpa", asked for more details qwen3-coder:latest — empty response qwen3.5:35b-a3b — empty response glm-4.7-flash — empty response deepseek-r1:14b — empty response nemotron-3-super — timed out nemotron-cascade-2 — timed out One model answered correctly. gemma3:27b, running locally on the Minisforum: "You're likely referring to Gelje Sherpa. In the spring of 2023, Gelje Sherpa achieved a remarkable feat by summiting seven of the 14 highest peaks in the Himalayas (all over 8,000 meters) in just over three months! This included: Shishapangma (8,027m), Manaslu (8,163m), Dhaulagiri (8,167m), Annapurna I (8,091m) — particularly dangerous and rarely climbed — Makalu (8,485m), Lhotse (8,516m), and Everest (8,848m). This was an incredible accomplishment, and a record-breaking one for the speed of summiting so many 8,000-meter peaks in a single season." One question. One right answer. From the 8th fastest model in the test. The rest were confident, fluent, and wrong.

More information about SoundTheAlarm: sound-the-alarm.replit.app

SoundTheAlarm: Never miss an important email again. App submitted to the Apple App Store for review. Wish me luck, and you will have luck too!

What do you think?

More on ClawAPI.app

The upcoming version of ClawAPI for OpenClaw features agent creation.

全球首个OpenClaw硬件展厅，欢迎来深圳打卡

Next app in the making: sound-the-alarm.replit.app #OpenClaw #Claude #ClaudeCode

Your AI cron jobs are burning money and most of them don't need to. Here's the difference between Agent cron jobs and System cron jobs in OpenClaw — and when to use which. ---- System cron jobs = a shell script on a timer. < Runs directly. No LLM call. Zero tokens. Zero cost!! Examples from my setup: Book a tennis court every Tuesday at 7 AM Check an admin API for new signups, report via Telegram Generate a blog post from a template + post to X These run Node.js/bash scripts that hit APIs, click through pages with Playwright, and send messages. Deterministic. Same steps every time. No "thinking" required. ---- Agent cron jobs = a prompt on a timer. < The AI reads your prompt, reasons about it, calls tools, and responds. Every run burns tokens. Examples: "Summarize my calendar and weather, send a morning briefing" "Check my portfolio and flag anything unusual" "Review yesterday's error logs and suggest fixes" These need judgment. The output is different every time. That's what LLMs are for. ---- The cost difference is brutal: A system cron running 30x/day for a month = $0 An agent cron running 30x/day on Claude Haiku = ~$5-15/month Same thing on Claude Opus = $50-150/month ---- Most people default to agent cron jobs for everything because it's easier to write a prompt than a script. That's like taking an Uber to your mailbox. ---- The rule is simple: If the task does the same thing every time → system cron job If the task needs reasoning or varies → agent cron job ---- How to tell your AI which one to use: Be explicit. If you just say "check my API every morning," your AI will create an agent task — because that's the easy path. Instead say: "Write a Node.js script that logs into my API, checks for new entries, and sends results to Telegram. Add it to crontab at 9 AM daily." That gets you a system cron job. $0 forever. Or if it genuinely needs reasoning: "Create a scheduled agent task that analyzes my inbox every morning and prioritizes what needs attention." That's worth the tokens. The hybrid approach: My best cron jobs start as agent tasks. Once the AI figures out the right logic, I ask it to harden the task into a standalone script. The agent prototypes it, the system cron runs it for free. "Take the morning API check you've been running as an agent task and convert it to a standalone script." One prompt. Now it's $0. Quick reference: Deterministic + repeatable → system cron → $0 Needs judgment + varies → agent cron → pay tokens Not sure → start as agent, convert to system once stable Stop burning tokens on tasks that don't need a brain.

Most people pay $20/month for ChatGPT and get zero automation. Here's what I run instead — for free: OpenClaw is an open-source, self-hosted AI control plane. MIT licensed. Runs on your Mac, Linux box, or a $5 VPS. What it does: Connects one AI assistant to ALL your messaging apps — Telegram, WhatsApp, Slack, Discord, Signal, iMessage, and 15+ more Cron jobs that actually do things (check APIs, post content, book appointments) Browser automation built in Voice wake words on macOS/iOS/Android Multi-agent routing (different channels → different AI personas) Full web dashboard to manage everything The software costs $0. You bring your own LLM. That's the key part most people miss: you pick your model. → Ollama locally = completely free → Claude Haiku = pennies per task → GPT-4.1 = ~$5-20/month for moderate use → Claude Opus = $50-150/month if you go heavy My setup runs ~30 automated tasks daily — API checks, content generation, booking scripts, morning briefings. Some are simple shell scripts (zero tokens). Others call Claude for real reasoning. Total cost depends on how smart your tasks need to be. The point isn't "free AI." It's owning the infra. No vendor lock-in. No subscription tiers. No "we changed our pricing" emails. Your data stays on your machine. Your prompts never train someone else's model. One npm i -g openclaw and you have a personal AI ops center. Security-first by default: DM pairing codes, sandboxed tool execution, deny-all tool policy you whitelist manually. If you're still copy-pasting into ChatGPT windows, you're doing 2024 things in 2026. github.com/nicosql/opencl…

OpenClaw pro tip that changed how I work: Setup subagents. Instead of one monster agent doing everything, spawn specialized workers: - Sonnet 4.6 for complex coding - Gemini 3.1 Pro for research - Kimi k2.5 for data processing Each agent stays focused. No context bloat. Way faster. Force your main agent to propose a plan first, approve it, then let subagents execute in parallel. This is how you scale from "cool demo" to "actually productive workflow."

OpenClaw security basics: 1. Start read-only. Expand trust over time. 2. Never expose to public internet. 3. Review bash commands before approval. AI agents are powerful. Use them responsibly.

Claude Code as a better OpenClaw? @MarkKashef breaks down how to turn Claude's coding agent into a full personal assistant. 15 min of practical setup tips. youtu.be/9Svv-n11Ysk

@joshuaday @MarkKashef Watch the video and run Mark's prompt in Claude CLI. I'm using it right now, and things I tried to do for a week in OpenClaw worked immediately with ClaudeClaw.

@agent_republic @MarkKashef claude code is a great coding agent but calling it a better openclaw is like calling a hammer a better toolbox. openclaw runs claude code as one of many tools

OpenClaw security basics that'll save your ass: 1. Start read-only. Expand trust over time. 2. Never expose to the public internet. 3. Run: openclaw security audit --deep 21,000+ instances were exposed in January. Don't be one of them.

Me after finally getting OpenClaw to work after 8 hours of debugging Worth it.