I timed four popular AI models writing the same 1,200-word memo. The slowest wrote 60 words a second. The fastest wrote 93.

So even the slow one writes about 15x faster than you can read. Scale that up (slowest AI, and you reading nonstop):

An hour: you read 48 pages. It writes 720.

A workday: you read one long novel. It writes all seven Harry Potter books, with three hours to spare.

A work week: you read 1,900 pages. It writes 28,800.

A work year: you read 95,000 pages. It writes 1.4 million.

(Writing isn't working, of course. Plenty of those pages would be junk. But it tells you who the slow one is, and it isn't the AI.)

Now the part that bugs me.

A year has 8,760 hours. A full-time job is about 2,000 of them. Most people only use AI while they're sitting in front of it, so for the other 6,760 hours it just... waits.

That's 77% of the year. Nights, weekends, the commute, the dentist.

And the 23% when you are there? You're the bottleneck anyway. It writes 15 pages in the time you read one.

So the cheapest upgrade most people can make isn't a smarter model. It's giving the one you already have a night shift.

That's all a "software factory" is. People throw the phrase around like everyone already knows how to run one. Most don't. Here's the simplest version:

• Before you leave, hand it one job that would take a person a full day. Read the last 200 support emails and sort the complaints. Draft three versions of the pricing page. Check 20 competitors' prices.
• Drafts only. Nothing gets sent, bought or deleted overnight. (Rule from my last post: make each mistake small.)
• Ask it to check its own work and put three sentences at the top.
• In the morning, read the three sentences. Open the rest only if something looks off.

Do that every weeknight and the 77% starts working for you.

Showing how to build these, step by step, is most of what we write about at @theunbiasedco.

I co-founded @theunbiasedco, which builds all sorts of stuff with (and for) AI agents.

💫 Welcome to the https://t.co/OKwVXQYsdB era https://t.co/Dmf40TBDkN pic.twitter.com/LxJnCjzTA6

Grok Bot's recent issues stem from you suddenly having to care which model was powering it. You couldn't see it, and you couldn't bring plans you already pay for. (We wrote about it on November 9.)

OpenBot fixes all three. Here's why I'm so excited.

The setup screen, in one line: "Sign in with a plan you already have, or use an API key." The big labs are there. As is Unbiased AI.

On November 9 I said a bot should work like Google Flights: Google doesn't own planes, it puts all the airlines on one screen. Same idea.

No landlord: it's open source and runs on your own laptop. (Bring your own cloud 🤫?!?) Nobody can change the terms overnight, or evict your bot. Grok Bot runs on a cloud computer, and its own help page says it doesn't show which model answered you.

Why this team: CopilotKit says the majority of the Fortune 500 use it, and it has 38,000+ GitHub stars. OpenBot itself has 10,000+ more. A friend introduced us (hi @aviel!), and we quickly realized we see the future in the same way.

Why Pareto fits: we make blended models, so several models work on every request. You pick it once and stop caring which one is on. Think auto updates. Like your bot read a bunch of textbooks, the news, learned some recipes and workout routines, all while you slept.

That was Grok Bot's promise, before the bill came.

With us under the hood, $10 a month gets 50M tokens each week. Run far more bots, or feed your main one.

My bet from November 9 stands: by November 30, Grok Bot follows their lead and lets you plug in another company's plan. They should!

If you'd rather not wait and pray, here's how to move your bot over. You can hand it straight to your agent.

p.s. Remember the Sphere at the top? OpenBot did that. Did it buy the ad space, or just generate an image? Who knows! I bet if you ask it, you just might find out.

I co-founded Circuit & Chisel, which makes Pareto. We're partners with OpenBot. Drafted with agents, edited by me.

Make me money. Bet $100 at a time, but only when the odds are at least 51–49 in your favor.

Seems pretty safe. A 51–49 bet doesn't win every time, but do enough of them and each one is worth about $2 to you (51 wins × $100 – 49 losses × $100 = $200 per 100 bets).

Casino empires hinge on this elementary math.

But AI is (currently) a little different. Every so often the AI misreads one. It thinks it's looking at 51–49, and it's actually 40–60 against. Same $100 on the line, but the return is suddenly –$20.

One mistake erases ten good bets.

(10 good bets × $2) + (1 mistake × –$20) = $0

That's an AI that's right 10 times out of 11. 91%.

In school, that was an A- and a high five from my dad. Here, it's a waste of electricity.

"Cool...but my AI doesn't bet." Psh, that's what you think!

Every time it books, buys, or sends something for you, it's betting that it understood what you meant. Most of the time it did, and you get a few minutes back. Every so often it didn't, and now you own a nonrefundable flight for the wrong Tuesday.

I ran this experiment on myself pre-AI. I bought Bitcoin at about $520 in 2014, felt like a genius selling two of them at $2,700 and $4,200, then bought back in at 15x what I first paid.

Being early? Cheap (albeit lucky).

The expensive part was one wimpy sell. In addition to being lazy, I'm also a wimp!

To recap, Warren Buffett has two rules. Rule No. 1: never lose money. Rule No. 2: never forget Rule No. 1.

For AI, it's going to be awhile before it never loses money. So for now, Rule No. 2 might have to be:

• Make each mistake small: a draft instead of a send, or a spending limit
• Watch the mistakes for a few weeks before you hand it the card

I co-founded @theunbiasedco, which builds all sorts of stuff with (and for) AI agents.