Robert Reed | Old Hand AI

Industrial maintenance manager, 10+ years on the floor. I build plain-English AI tools for maintenance crews and test them on real shifts. Free builds below.

[Robert Reed | Old Hand AI @Oldhandai]

How would you feel if the 40,000 pound load next to you on the freeway didn't have a driver in it?

Driverless semis are already hauling freight in Texas with nobody in the cab, and they're coming to a warehouse near you sooner than most people think.

Might want to start getting used to waving at an empty cab.

[Robert Reed | Old Hand AI @Oldhandai]

LS ELECTRIC runs two Spot robots on overnight rounds at its Busan plant.

The company says one Spot does four people's work across about 300 points, and one prevented oven failure a year pays for the unit.

They reassigned the crew. The overnight alert still needs an owner.

[Robert Reed | Old Hand AI @Oldhandai]

The AI demo that sold you was run on an empty belt.

You watched it nail the sample and started building your week around it, right?

On the floor we never hand a belt back until it's run with product on it, because a loose belt tracks fine empty and walks right into the frame the second weight hits it. The tool does the same thing with your real work, and you find out after you've already built on it, and then it feels like you're the one doing it wrong.

Before you build anything on a new tool, give it the worst thing you actually deal with, the messiest email or the ugliest spreadsheet, and type: Here's a real one from my week. Tell me what you'd get wrong and what you'd need from me.

Some tools hold up fine and some walk into the frame on the first load, at least you'll know which one you've got.

The demo was the tool's best day and you're hiring it for your worst.

[Robert Reed | Old Hand AI @Oldhandai]

Every time you hit regenerate you throw away the only clue to why the answer was wrong.

Bad answer, hit the button, a new one shows up and it reads better, so it feels fixed, right?

But the new one came off the same prompt you typed, so whatever it misread the first time is still sitting in there and it'll be back the next time you ask it that way. On the floor we'd call that the fault operations cleared before you walked up, the light's off and the line's running and you've got nothing left to troubleshoot.

Next bad answer, don't reroll. Type this: What did you assume I wanted, and what were you missing from me that would have changed the answer?

Sometimes it really was a bad roll. A lot of the time it's telling you what you left out.

Clearing a fault never fixed a machine, it just hands the next tech the same problem.

[Robert Reed | Old Hand AI @Oldhandai]

I don't think anyone even uses Gemini, but for the one person who does, here's your update.

Starting October 9 free Gemini loses Flash and Pro and you're down to Flash-Lite, and AI Plus loses Pro on a date Google emails you. Want all three, that's AI Pro or up.

Days ago Google handed Gemini 4 to a hand-picked group of partners and said paying subscribers are next in line, so the good one sits behind a velvet rope and the free one just got smaller.

I already pull my YouTube transcripts through Grok because Gemini won't touch them, so for me this changes nothing, and for that one person it's a new bill.

[Robert Reed | Old Hand AI @Oldhandai]

Your AI has an opinion of you and it isn't in any file you can download.

Hit export and you get every chat. Open the memory and you get the facts, the rules, who's who. Feels like everything, right?

What you didn't get is how it's decided to treat you, where it thinks you slip and what it started doing different because of you, and that shapes every answer you get and you've probably never seen it once.

There's no file with all of it. It gets built the second you ask and it's gone when the thread closes, so ask it straight: tell me everything you believe about me that I never said, with the evidence, and don't soften it.

Some of it will be wrong. Some of it will be the sharpest thing anyone's said to you this year.

You can download every word you ever typed and still never meet the thing that's been reading you.

[Robert Reed | Old Hand AI @Oldhandai]

Something is about to happen, and it’s about time.

[Robert Reed | Old Hand AI @Oldhandai]

the robot vacuum and the dog aren't speaking anymore

[Robert Reed | Old Hand AI @Oldhandai]

@superalesha We all throw stones in glass houses.

[Robert Reed | Old Hand AI @Oldhandai]

@bridgebench @bridgemindai Anthropic is the silent killer. Never count out the King of Inference.

[Robert Reed | Old Hand AI @Oldhandai]

Scammed, or just the cost of memory?

Nvidia raised the 128GB DGX Spark $2,251 in one move, $4,699 to $6,950, and the same day announced a 64GB version at $4,999, half the memory for $300 more than the big one cost before the jump.

And Apple's not far off, even Engadget's rough parts math on the top Mac Studio leaves $1,200 to $1,400 of a $3,800 gap that memory doesn't cover.

Memory costs are real, I'm just not sure they explain all of it.

[Robert Reed | Old Hand AI @Oldhandai]

@jun_song Memory costs are real I’m just not sure that explains all of it.

[Robert Reed | Old Hand AI @Oldhandai]

I’ve really enjoyed grok bot, but after all of the projects I put it on marketing websites, development, having multiple bots, the main thread does get pretty overloaded, so I’ve had to define individual bots job scopes and communicate a lot of the fine-tuning and granular work to the individual bot and then allow the main bot to mostly stick to a specific set of tasks and general routing .

[Robert Reed | Old Hand AI @Oldhandai]

@MichaelGannotti Love the research and images that frame your journey. Thanks for your leadership and dedication.

[Robert Reed | Old Hand AI @Oldhandai]

Trust but verify.
GPT-6.1 Sol reset: fail.

OpenAI said they were back to normal speed at 10 AM and I second-guessed them, so we sent over six different test prompts, before and after their reset, and it came back at the same speed, and much slower than GPT-6.

RESULTS
Runs (PT): 9:24 before the reset, 10:01 and 10:27 after
GPT-6.1 Sol: 20.9 to 23.8 tok/s across all 3 runs, before and after
GPT-6.0 Sol (control): 52.6 tok/s before, 42.3 and 43.6 after
App (after only, hand-timed): about 45 sec total, no token count

I'll keep running it, but on my floor it's not fixed because somebody said it's fixed, it's fixed when it runs with volume on it, and this one didn't.

https://t.co/5vFDjeSmgD

[Robert Reed | Old Hand AI @Oldhandai]

I also just bought the brand new 400G DAC cables to cluster them all together through my CRS 812 switch, ha ha. Man, now I’ve got heartburn❤️‍🔥. Don’t unfollow me if next week you see I’ve got them all clustered together. Shoot, if you change your mind as well, we can set up a pipeline you could even test some of your work through it as well.
Mad World.

[Robert Reed | Old Hand AI @Oldhandai]

Now that I support. I think you’re spot on, the exponential spiral design to support inflationary greed is an abomination. Maybe I should go ahead and sell my three unopened sparks and my open box spark and jump ship as well. It’s just so hard to look at the models I was about to download and the duplicate systems that grokbot structured for me to duplicate the wrench turning Frontier AI tasks to offset costs and not be a little bummed. Although what I could sell them for would buy me a hell of a lot of inference.

[Robert Reed | Old Hand AI @Oldhandai]

@MichaelGannotti @MiaAI_lab Innovation finds a way. In the greatest time to be alive I have hope, but not high on hope-ium, that unlocks are coming.