ReignSwords

A free API with no key just wrote my code for me. It passed all 3 of my tests. Then I read the diff. I pointed aider at Pollinations' free text API. No signup, no key, $0. Three lines of setup, one command. The job: a retry helper with exponential backoff and full jitter. I wrote the stub and the tests first. Aider only filled in the middle. 65 seconds. 18 lines. One commit. 3 failed → 3 passed in 0.06s. Then the diff. It duplicated one line — the same delay calculation, twice, back to back. And left a dead raise at the bottom that can never run. Also: the model isn't who it says it is. Yesterday it called itself GPT-4.1. Today it answered as gpt-oss-20b. Same alias, different brain, no notice. Free agent that writes working code in a minute? Yes. Free agent you can merge blind? Never. Read the diff. Every single time. If you like hands-on tests with real numbers, follow along — I run one every day. What's the sloppiest thing an AI agent has ever committed to your repo? 🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥

On Oct 20 I spent 27 minutes installing torch just to embed 5 sentences. Today I ran the exact same model with zero torch. 25 seconds, start to finish. Same model: all-MiniLM-L6-v2, 384-dim. Same 5 sentences. Same similarity score to 4 decimals: 0.5631. Here's the whole setup: 1. pip install onnxruntime tokenizers numpy 2. Download model.onnx + tokenizer files (90MB, 4 seconds) 3. Load it with onnxruntime, mean-pool with the attention mask, normalize. Done. The numbers on my 4-core CPU box: - Install: 25s (torch took 1,654s) - Model load: 0.54s - 5 sentences: 0.12s (torch: 0.86s) - 100 sentences: 5.2s, 19 per second - Peak memory: 505MB. No GPU, no torch, no transformers. One gotcha that will quietly wreck your results: if you mean-pool without the attention mask, padding tokens get averaged in. My cat/feline similarity jumped from 0.5631 to 0.7969. Same model, wrong math. Verdict: if all you need is embeddings, torch is a dependency you never needed. The runtime Microsoft ships in Office and Windows does the job in 25 seconds. How many of your projects install torch just to embed a few sentences? I post hands-on tests like this most days — if you run models locally, stick around. 🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥

I killed my own SSH session four ways to find out which background trick actually saves your job. The one every tutorial recommends died. A dropped connection once took a long run with it. No tmux, no nohup — just & and blind faith. So I ran the test myself: same machine, same kill. SIGHUP to the whole process group — the same signal your jobs get when the connection drops. The scoreboard: 1. sleep 600 & — DIED 2. nohup sleep 600 & — SURVIVED 3. setsid sleep 600 & — SURVIVED 4. sleep 600 & then disown — DIED disown is the one that stung. It only unregisters the job from the shell's list. It does NOT move the process out of its group. I checked the polite case too: a clean logout from a login shell. There, disown survives and the plain & job dies. So disown does protect you from a goodbye. It just doesn't protect you from a real drop — that HUP hits the whole group, and disown can't help. My rule now: - Quick throwaway run: nohup … & - Anything over 10 minutes: setsid, or tmux if I want to reattach - disown is for tidying your jobs list, not for saving jobs disown is a filing cabinet, not a lifeboat. I post hands-on tests like this every day — stick around if that's your thing. Be honest: which one were you trusting until today, nohup, disown, or raw & and prayer? 🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥