Tested Muse. Named it, then asked for 4 things to do in LA this weekend. It returned dated picks with venues: Rilo Kiley + Beachwood Sparks at Hollywood Forever (Sat), Pasadena Chalk Festival at Brookside, Watts Towers Day of the Drum + Jazz, and Day Trip Season Closer with Armand Van Helden (Sun). Setup also showed own-computer work plus Gmail and Calendar connect. Latency and spend not measured. Keep for easy to use simple chatbot functions and integrations. Kill if you need any complex tasks done.

Had an AI on my PC draft a reply to a contractor's email. 1:18 vs 3:21 typing it myself, but it took two tries. Setup: Llama Windows (free, from the llama.cpp team) with Gemma 4 E4B, running on my RX 9060 XT. Alt+Space, paste the email plus my notes, nothing leaves my PC. First try it got the direction wrong and wrote the reply as if I were the contractor. Second try was perfect. Keep for quick replies from your own notes. Check who it thinks is talking, and check that every note made it in.

Asked an AI 3 questions about my T-Mobile contract. Googling took me 5:21 and missed two catches. Reflection Beam took 42 seconds, including reading its answer, and found both. Setup: Beam on Ollama Cloud (free for now), with the full 9,500-word T-Mobile terms pasted in. It wrote the answer in 7 seconds. What Googling missed: - If you leave by porting your number, service shuts off right away, but you still pay for the full billing cycle. - The price guarantee doesn't cover surcharges. They "may change from time to time without notice regardless of any price-related promotion." I checked every answer against the actual terms. It got all 5 price-change rules and nothing was made up. On fees, it listed about 10 of 14, missing roaming, Wi-Fi, third-party charges, and calls rounded up to the full minute. Keep for reading the fine print you'd never read yourself. Still ask it about the specific fee you're worried about.

IBM says the speech model bundled with Granite Switch 4.1 is the fastest on the Open ASR leaderboard. It's #1 for speed on the English short-form board, by about 1%, but that's the throughput measured on one H200 GPU. I wanted to know if it saves me time on one voice note. Setup: IBM's free Speech-to-Decisions demo, recorded on my Mac. English only. The job: a practice shift handoff note with 2 lot numbers, board counts, an oven zone, a date, and who's covering. Typed by hand: 1:14. Talked into Granite: 31s. Every number came out right. It missed 2 words: "AOI" became "A Y" and "Jim" became "gym," so its topic line at the end read "Operator shortage swing shift gym". Keep for number-heavy notes. Check names and shop jargon before you hand it off.

Reran Muse (free chatbot) on camera for YouTube. Same job as my Sep 26 post: 4 things to do in LA this weekend. Last time, it returned dated picks with real venues: Rilo Kiley and Beachwood Sparks at Hollywood Forever on Saturday, Pasadena Chalk Festival at Brookside, Watts Towers Day of the Drum, and Day Trip Season Closer. Full test is up: youtube.com/watch?v=w2s8dHtI… What would you ask it to plan next? Keep for everyday tasks like weekend plans. Kill for overly complex jobs.

Running Liquid AI's LFM2.5-24B-A2B, Q4_K_M, through Ollama on my AMD RX 9060 XT (16 GB) with Windows. Free and offline after the download, and it loaded in 12.1s. The job: turn a fake practice roster for next week into one table and flag every short-staffed shift. My spreadsheet: 12 minutes and 33 seconds. LFM2.5: 35s, including checking every row. It found all 6 gaps, marked no extra ones, and added no names. But it grouped operators per shift, and when I asked it to fix that column, it redid the whole job instead of just editing it. Keep for a first pass on roster gaps. Kill if you need to go back and forth on edits.

Too many interests, not enough week? I made a free Grok Bot template for this purpose: one main project, one slow reading track, one side project, a parking lot for the rest, small daily steps, and all in ISO 24495 plain language. x.ai/bot/QZ59mNTczWx2fi92XUb…

Kai (Morgen's AI assistant), free trial, on my Gmail: sorted my 15 newest emails in 48s. By hand: 4:41. Grok Bot did it under 30. Both sorted security alerts and a price drop as FYI, and promos and fundraising blasts as Ignore. None needed a reply, so no drafts were tested. Keep if you don't already have an AI assistant reading your mail. Otherwise, kill and stick to Muse or Grok Bot.

Tested 5 AI tools on real work in the last 7 days. Timed each against my old way. Claude Haiku 5.5 (free Claude.ai), shift handoff note: about 3m54s saved. Keep. LightOnOCR-3 1B local on my RX 9060 XT, packing slip: 24s vs 80s by hand. Keep. LightOnOCR-3 4B free demo, same slip: 87s, slower than typing. Kill. Liquid AI d1-3B webcam demo, yes/no on a slip: 3 for 3. Keep for quick checks. Falcon-ASR free demo, talking vs typing: about 80s saved on a handoff note, but AOI came back "AOW." Keep, read it once. Which one would you run on your own work? Keep 4, Kill 1.

Ran Falcon-ASR on TII's free demo (60s cap, no sign-in). TII says 5.74% word error rate on English. Built for Arabic and English, I only tested English. Two clips, talked vs typed on my phone: To-do list: 15s talking + 3.2s transcribe, zero fixes. Typing: 1:03. About 45s saved. Shift handoff note: 20s talking + 3.2s + 5s fixing. Typing: 1:48. About 80s saved. Every number and day came out right in both. It tripped on shop words: AOI became "AOW," "short two" became "short to." Keep for talking notes instead of typing them. Read it once for jargon.

Ran Liquid AI open d1-3B on its free Ask Anything demo. Held a packing slip to my webcam, asked 3 yes/no questions. It answered live, every frame. More than one item: yes, 95%. Order date 2024: yes, 93%. UPS: yes, 97%. All right. Keep for quick yes/no checks.

Ran Claude Haiku 5.5 on free Claude.ai. Shift handoff note into 3 detailed action items. By hand, 4 min. Haiku, 2.4s plus 3.2s to read it. Saved about 3 min 54 sec. Prompt: turn this handoff note into 3 action items, plain language, no fluff. Keep for handoff notes you'd write out in full or any small simple tasks you need done.

Reran LightOnOCR-3 1B locally on my AMD RX 9060 XT. Same Secretlab packing slip. By hand, 80s. Local, 24s including model load. Free demo was 87s. Every field matched the slip, down to the bin codes. Keep for slips and documents on a local GPU. Kill the free demo.

Ran ChatGPT's new GPT-6 with Intelligent UI, free plan, on a week chart I can tap. By hand, 5:20. ChatGPT, 3 seconds. Saved 5 min 17 sec. Prompt: flag a missing time, add nothing I didn't list. Saturday flagged. No extra row. Keep if it saves 5+ min and every item is right.