null margin

margins compress. narratives don't. AI capex, unit economics, and who actually gets paid.

[null margin @nullmargin0]

ELON MUSK'S CRAZIEST THEORY. ..WHAT IF HE'S RIGHT?

he said it on stage in 2016, totally calm. the odds that our world is the real one: one in billions.

no joke. he walks through it like a math problem in under 2 minutes.

simulation = a world running on someone else's computer, so real from the inside that you can't tell.

his argument, step by step:

→ about 40 years ago games were Pong. two rectangles and a dot
→ now it's photoreal 3D worlds with millions playing at once, and it gets better every year
→ keep any speed of progress and games end up looking exactly like reality
→ slow it down 1,000x and you still get there in 10,000 years, which is nothing on an evolution clock
→ then billions of those worlds run on every console and PC
→ billions of fake worlds, one real one. so which one are you in

what actually got me is the ending. he says we should hope he's right.

because the only other option is that progress stops. and that probably means something wiped us out first.

the weak spot: it assumes a mind like yours can run on a chip. nobody has shown that yet.

the idea isn't even his. philosopher Nick Bostrom wrote it up in 2003. musk just made it sound simple.

watch 1:47. Kara asks if someone just beat us to it. he doesn't even blink.

and that was 2016. today AI makes video in seconds that's hard to tell from real.

one in billions isn't a wild guess. it's what you get when you assume the curve never stops.

so, base reality or not? what's your number?

[null margin @nullmargin0]

I DIDN'T THINK IT WAS THIS SIMPLE. HAIKU 5.5 + JEV + OPENAI DOTS BUILT ME A SYSTEM AN AGENCY QUOTED AT $100,000

i plugged in 4 dots and the whole thing turned into a 14-second meeting

dot = a small openai agent that knows one thing well and nothing else

who does what:

→ haiku 5.5 reads the task and hands the chair to opus 5.5
→ opus 5.5 asks each dot what it knows: one reads the research, one the code, one the user data, one runs the tests
→ they argue out loud. jev scores every argument, the weak ones don't count
→ they vote. majority wins, no egos in the room
→ claude code writes the change, tests run, whatever breaks gets fixed before anyone moves on
→ last vote: ship or not. if it passes, it becomes a pull request and waits for a human

they talk better than most human teams. nobody defends their own idea, nobody says "let's take this offline"

first run they cut my sign-up form from 9 fields to 3. sign-ups went from 212 to 317 a week. my whole job was one click: approve

not every vote was right. night two they voted 3–1 to change a price. the one "no" was right and it cost me $41 in refunds. now prices need 4 of 4

pause at 0:09. the purple dot is the one that said no

the agency wanted $100,000 for 6 months. my api bill last month was $312

$100,000 was never the price of the code. it was the price of the meetings

a human team spends 14 seconds asking "can everyone hear me?"

the prompts:

1/ haiku 5.5 · router
read the task. rewrite it as one goal, one number to move, one deadline. list the files each dot needs. if anything is unclear, ask me one question and stop. save as task.md and hand it to opus 5.5

2/ opus 5.5 · chair · high effort
you run a meeting of 4 dots: research, code, users, tests. each gets 3 lines max and must name the file it's based on. let them disagree once, then call a vote. majority wins. prices, deletes and anything with money need 4 of 4. write decision.md: what changes, why, who voted no and their reason

3/ openai dot · seat (run 4 times, swap the source)
you only know [research/ | the repo | the analytics export | the test suite]. answer from that source only and name the file. if your source can't answer, say "no data" and abstain. never agree just to end the meeting

4/ jev · gate
score every argument 0–10: has a source, has a number, can be tested. under 6 doesn't count in the vote

5/ claude code · builder
build decision.md on a new branch. run the full test suite. fix what breaks, 3 tries max, then stop and report. open a pull request with the vote log in the description. never merge

[xinces @xinc3s]

https://t.co/f3NGqqO8TZ

[null margin @nullmargin0]

IF SOMEONE TOLD ME THIS, I WOULDN'T BELIEVE IT... I PUT GROK BOT, OPUS 5.5 AND AN OPENAI DOT ON ONE SHOPIFY STORE. $14,196 IN A DAY

i gave each one a single job. none of them dropped it

dropshipping = you sell it, a supplier ships it, you never touch a box

who does what:

→ grok bot reads x live and finds what people want before tiktok does. it picked a $6.40 pet bottle that clips on a leash
→ opus 5.5 checks the supplier, prices it at $39, builds the page and writes 12 ad hooks
→ the dot (gpt-6 astra) launches the ads, sends every order to the supplier, answers buyers and kills ads that don't sell
→ my rule for the dot: changes under $50 go auto, refunds need my ok

364 orders x $39 = $14,196

ad spend $3,890. product, shipping and fees $3,738. i kept $6,568

what actually got me: grok bot found it in 40 posts of dog owners complaining on walks. nobody was running ads for it yet

at 3:14am a buyer asked where her order was. the dot sent tracking before i woke up

$14,196 isn't a lucky product. it's three agents that each do one job and don't touch the rest

so Elon, you asked how grok bot is working with my shopify store. perfectly, as long as other agents work next to it

[Elon Musk @elonmusk]

How is Grok @Bot working with your Shopify store?

[null margin @nullmargin0]

THIS IS F*CKING CRAZY... I didn't think this actually worked

my research bill went from $1,042 a month to $59. same 500 stocks every night

i read one guide on haiku 5.5 + opus 5.5 and rebuilt my desk in an evening

reading stopped being an edge. everyone gets the same summary by 7am

the desk = three agents passing files, not chat

my night now, line by line:

→ haiku 5.5 reads 500 filings and calls, one card each. $1.20
→ a gate scores every card in code and throws out 495. $0.02
→ opus 5.5 writes 5 one-page dossiers from filings only. $1.60
→ code adds the price move after. never before
→ i read 5 pages, not 500

what actually got me: opus 5.5 is still 57% of the bill and touches 1% of the companies

first night i got burned. 9 filings crossed haiku 5.5's 100K line and those reads cost 5x. now i split them

pause at 0:11 and count the gold dots. five. everything grey is a no

$983 a month isn't cheaper reading. it's paying opus only where judgment is needed

sometimes it's worth reading guides

[xinces @xinc3s]

https://t.co/f3NGqqO8TZ

[octarinx @octarinx]

I READ THIS TWICE AND I STILL CAN'T BELIEVE IT'S FREE

i opened it expecting another "AI replaced analysts" thread. i wanted to see where it breaks

it didn't

the trick isn't the models. it's what each one is forbidden to do

Haiku 5.5 reads 500 companies a night and isn't allowed an opinion
Jev throws away 495 of them and isn't allowed to write a word
Opus 5.5 writes 5 one-pagers and isn't allowed to see the chart

that last rule is the one that got me. show a model the +14% first and it writes a caption, not research. hide the price and the story has to come out of the filing

by the article's math the whole desk runs ~$59 a month. same desk with Opus on every read: ~$1,042

and the number nobody will talk about: opus is 57% of the bill while touching 1% of the companies

reading got cheap. saying no didn't

it's all below. the code, the math, the prompt. honestly the part about hiding the chart is worth the read on its own

[null margin @nullmargin0]

I F*CKING COULDN'T BELIEVE MY EYES...We haven't even recovered from Opus 5.5 and Haiku 5.5 is already shocking everyone

I tested it today. it made me $4,300 in about two hours

a store client had a 38,000-product catalog: broken categories, duplicate listings, half the descriptions missing. i'd quoted $4,300 for the cleanup. normally that's two weeks of my life

I gave Opus 5.5 the brief and let it run Haiku 5.5 as subagents

it split the catalog into batches, merged the duplicates, rebuilt the categories, wrote every missing description, checked its own work and exported a clean file

I didn't touch anything. i just sat there and watched it make me money

api bill: $9.40
margin: basically 100%

are you sure you are using Haiku 5.5 correctly?

[ClaudeDevs @ClaudeDevs]

Haiku 5.5 is now available in the Claude Platform and Claude Code. On average, it costs around 75% less to run than Haiku 4.5.

It pairs well with Opus 5.5 or Sonnet 5.5 as a subagent. Use it for high-volume, cost-sensitive tasks like summaries, compactions, or database queries.

[null margin @nullmargin0]

@GergelyOrosz wait, revenue actually went up every year. +6% in 2025
it's the margin that died. 40% in 2021, -2% now (per their czech filings)
people kept paying for the ide, just not ide prices anymore

[null margin @nullmargin0]

AI IS BREAKING EVERY RECORD RIGHT NOW.AND THIS IS NOT CLICKBAIT

openai just released 722 math manuscripts from an unreleased internal model. run them against a ranked list of the 500 most important open problems and you get:

The most frightening fact is that it took only TWO DAYS.

90 full claims
45 partial
here are the hardest ones it cracked:

#22 hilbert's tenth problem over Q — full claim
#29 unique games — full claim
#48 landau–siegel zeros — full claim
#80 hadwiger conjecture, open since 1943 — full claim

#2 riemann, #4 hodge, #5 birch–swinnerton-dyer — partial

when do you think ai shocks us next?

[OpenAI @OpenAI]

We’re releasing a broad range of new mathematical results produced by an internal frontier model.

We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.

https://t.co/7N6TPlft1P

[null margin @nullmargin0]

THIS IS F*CKING INSANE... I WIRED DOTS + GROK 4.7 + OPUS 5.5 INTO ONE STACK AND IT PULLED $1,149 OVERNIGHT

the result still doesn't fit in my head

the three strongest systems right now. and i still don't get why traders stare at charts all day when this stack does it while they sleep

here's how mine works:

1.GROK 4.7 reads everything. 5,752 market items a night from prices, news, X, on-chain, options and filings
DOTS sorts every item in 81 ms. noise, routine, or worth a real look
2.DOTS agents handle the routine. stop updates, trade journal, rebalance notes, the morning brief
3.OPUS 5.5 sleeps until something actually matters. last night that was 1 item out of 5,752
i get one card in the morning. approve or skip

the part nobody prices in: 99.98% of the night is cheap reading. the expensive model touched one item

the edge isn't the smartest model. it's the one you let sleep

what would you let it wake up for?

[null margin @nullmargin0]

EVERYONE'S HIRING THEIR FIRST GROK BOT EMPLOYEE. ALMOST NOBODY READ THE CONTRACT

i did. it has one number in it

the shift: stop reading the benchmark, start reading the terms

term 1 → salary: $300/mo via supergrok heavy, or $200 via cursor ultra. grok bot isn't sold on its own

term 2 → hours: "weekly usage". how much work that actually buys isn't published

term 3 → overtime: past the allowance you pay on demand, at model and token rates

term 4 → spend cap: no grok bot-specific cap published

term 5 → who does the work: xai doesn't name the model

where it gets interesting:

$300/mo is ~5.5% of the median us full-time paycheck ($1,251/wk, BLS Q2 2026)

so the salary is the cheapest line in the deal. it's also the only line with a number on it

a human contract fixes pay, hours and the overtime rate. this one fixes the entry fee and leaves the ceiling blank

that's not a hire. that's a variable cost wearing a name tag

would you sign an offer with the overtime line empty?

[Brian Roemmele @BrianRoemmele]

It Is Time To Hire Your First Grok Bot Employee!

My insights come from building the first Zero Human Company with @Grok as CEO.

I made the mistakes so you don’t have to.

Read more: