Mike Maier (@mrmwisc) header
Mike Maier (@mrmwisc) profile picture

Husband. Father. Engineer. USAF veteran.

WI
Joined August 2009
Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)Photo by Mike Maier (@mrmwisc)
A lot of jobs are going to change or go away. I won't pretend otherwise. But a lot are being created, and it's worth looking at who's creating them. SpaceXSI put five Grok Bots on its routine accounting work. Its accountants didn't disappear. By the company's own description, they now focus on improving systems and resolving complex issues. SpaceXSI also says handling the workload without the bots would have meant more than doubling its accounting team. The new work went to the people who knew the books. They're the ones who can tell when a bot's answer is wrong, which exception matters, and which step should never be automated. I see the same thing coming for engineering and manufacturing. The people building the new roles will be subject matter experts willing to learn artificial intelligence: the process engineer who sets an agent to watch a control chart, the plant manager who writes its reaction plan, the owner who decides what it's allowed to approve. If you've got years in your field, that's not a liability. It's the part nobody can download.
5
Checking the folders after is the part I'd keep. A week's job done in five minutes, and the person still looks at the result before calling it done. In a plant that's the first-article check, and it's what keeps the speed honest.
Maye Musk asked Grok Bot the best way to move 2,000 Dropbox photos from Camera Uploads into folders. it asked if she wanted it to just do it. done in five minutes, and she'd booked the whole week for that job. my favorite part: she went and checked the folders after
12
Rules that pile up and contradict each other is what a plant fixes with document control: one current revision, the date it changed, and who approved it. A bot can sort and merge the rules in minutes. The person still decides which one stands.
Some of my Grok Bot rules started to clash with each other. Once I had more than 30 Grok Bots, new rules kept landing on top of old ones. A rule from October would say one thing, and a newer rule would say something a little different. The Grok Bot had both rules and had to guess. So last week, I gave them one rule for all their rules. Every rule has a date, and when two disagree, the newest one wins. Now the Grok Bot compares its rules and memories itself, and when one changes, it replaces the old line instead of adding a new one under it. With no duplicate rules that say different things, the lists stay short. Each Grok Bot has less to read before it starts, and a bad rule is easier to spot. It also saves usage. The Grok Bot doesn't have to work harder to guess which rule is right, so each run does less extra work. Now I can keep adding Grok Bots without their rule lists getting longer and messier.
12
A chat forgets overnight. A handover record doesn't. Write down the current state, the decisions and the next step, so the next person or the next bot can pick up the work without starting over. hotbotspot.com/knowledge-tha…
1
7
A person needs a screen. A bot needs a connector. We're building those connections for traditional tools, so a bot can use them the way a person would, and only get the doors its job needs. hotbotspot.com/tools-and-con…
8
Good questions for any bot output, not just specs. It's what we'd ask of an inspection report: what was measured, what was assumed, and which gauge did it. A bot can build the report in minutes. The engineer still signs it. xitter.gq/Hiroshi_HW/status/2118…
Three questions before trusting an AI-reconstructed spec: 1. What evidence does each claim cite? 2. What did it mark Unknown, and what did it infer anyway? 3. Which tools actually produced it? Free PDF, 25 design-review prompts: hiroshi-shibata.kit.com/0d91…
1
1
44
Tens of thousands of cell sites is an install problem as much as a radio one. When you build the same thing that many times, a site checklist and an FMEA on the install steps pay for themselves. One missed step repeated across every site is a lot of return trips.
1
1
45
Tens of thousands of cell sites is an install problem as much as a radio one. When you build the same thing that many times, a site checklist and an FMEA on the install steps pay for themselves. One missed step repeated across every site is a lot of return trips.
1
1
23
First place I'd put this is the plant floor. A press faults on second shift, and the bot calls the owner of that line: what's down, how long, and what order is due. If nobody picks up, it calls the next name. The call list and the triggers are already in your control plan. That's the part to settle in code first.
Grok Bot can now make phone calls.
1
27
Same trap in a plant. Nobody runs on machine hours. You count good parts out the door. For bots I'd track finished, usable work per token, not tokens alone.
This is the equivalent of counting the atoms in a barrel of oil rather than measuring the useful energy it produced.
21
With artificial intelligence bots, knowing how to run one and knowing where it belongs are two different skills. Most companies have them in two different people. The fix is not picking one, it's putting them on the same team. hotbotspot.com/managing-bots…
28
Running out of tokens? You may be programming your bots wrong. Elon said the cost of serving Grok Bot was unsustainably high relative to revenue, and a lot of people hit their weekly limit early. From running a team of bots on manufacturing work, here's what I see. Most of the burn isn't thinking. It's paperwork. On every run the bot decides which fields to use, what to name the file, where to save it and what the printout should look like. Next run, it decides all of that again, a little differently. That's like asking an operator to redesign the traveler every shift. The fix is hard constraints in ordinary code: - The program owns the schema: which fields exist and which values are allowed. - The program owns the checks. A wrong shape fails right away, not after three rounds of fixing. - The program owns the output. Reports come out the same every time because code prints them. - The program owns where the record lives. The bot keeps the part only it can do. Is a risk missing? Is this test enough? What alternative should we look at? And a person still makes the call. Instructions in a prompt are a sign on the wall. Code is a fixture. Plants learned that a long time ago. I'm building this layer now for FMEA, control plans and 8D. It isn't finished and I'm not promising numbers. But the direction is clear: stop paying the bot to reinvent the form, and spend the tokens on judgment. hotbotspot.com
1
101
Everyone's building teams of artificial intelligence bots. Most will fail the way human teams fail: no clear owner, sloppy handoffs, nobody checking the work. I read the best threads on it and pulled 10 principles. None are new to anyone who's run a shop floor. 1/ One front door. One bot takes your requests and routes them to the others. It doesn't do the work. If you're the one passing notes between bots, you're the bottleneck. 2/ One job per bot. If you can't name a new bot's job in three words, you don't need another bot. You need a better checklist for the one you have. 3/ Start with one bot. Split only when work can run in parallel or needs an independent checker. Every handoff loses some context, and bots with nobody checking them pile up errors fast. 4/ Standard handoffs: issue, owner, evidence, next action. Every time. We learned this on shift changes decades ago. Bots need it more, because they forget everything between sessions. 5/ Put the work on a board, not in chat. Todo, Working, Blocked, Review, Done. If the board doesn't say done, it isn't done. 6/ "Verified" is not evidence. Done means a checklist marked pass, fail or unknown, checked by a separate reviewer. That's an inspection plan, same as it ever was. 7/ Gate by whether it can be undone. Drafts are free. Anything that leaves the building waits for a person to say yes. 8/ Same mistake twice becomes a rule. That's corrective action, the 8D mindset, applied to bots. Fix the system, not the symptom. 9/ Keep teams small, about 3 to 5 active bots. They can generate all day. The scarce thing is review, and that's your expert's time. 10/ The bots need a manager who knows the work. A bot can draft an FMEA fast. It takes an expert to know which risk really ranks first. The subject-matter expert just got more valuable, not less. hotbotspot.com
🤖 Made with AI
30
Bots made analysis cheap. They didn't make knowing which analysis matters cheap. That still lives with the engineer, the planner and the quality lead. Tools like Grok Bot are now easy enough for them to run it directly. hotbotspot.com/the-analysis-…
18
Odd thing about artificial intelligence right now: the people who can run bots are getting laid off, and the subject matter experts who know where the work is think they're behind. Pair them. The expert knows where, the technical person knows how. hotbotspot.com/
1
21
With artificial intelligence bots, over-staffing just got cheap. Hiring five people to do two people's job wastes $100K a year. Five bots doing two bots' job might waste $1-2K. Cheaper, but still waste. Map the process first and put bots only where the work is. hotbotspot.com/map-the-proce…
1
1
34
Same habit as a good quality shop: categorize the defect before you touch the process. Rewriting the prompt after every bad answer is the bot version of retraining the operator. With Grok Bots the sorting goes fast, and a person still decides the fix.
Every wrong answer an agent gives in our builds gets a category before it gets a fix. Teams often respond to a bad output by rewriting the prompt. Sometimes that helps. Often it hides the real problem and creates two new ones. So each week we review errors with the client's subject-matter experts and classify each one: Retrieval miss. The right document existed but wasn't found. Stale source. The agent cited a policy that had been superseded. Reasoning error. The facts were right and the conclusion was wrong. Tool error. An API call failed, timed out or returned partial data. Policy gap. Nobody had decided what the right answer was. Each category has a different owner and a different fix: indexing, content governance, model or prompt changes, integration engineering, or a business decision. Every fixed error becomes a test case in the golden set, so it can't quietly come back in the next release. The takeaway: fix the category, not the symptom. #AIEvals #LLMOps #EnterpriseAI
16
Artificial intelligence bots give founders nearly free labor. But free labor isn't management. If you don't know the product, sales or marketing, nine bots just get you to the old founder problems faster. The people who win with bots are the ones who'd have won with a hired staff. hotbotspot.com/managing-bots…
21
Pre-mortems are one of the best tools in the box. With a Grok Bot team, one bot drafts the failure reasons and the engineer decides which are real. It doesn't take much to set up. Hiroshi's 25 free design-review prompts are worth a look.
2
1
3
150