holy sh*t, Sam Altman doesn't touch his inbox at night anymore. one AI agent reads all of it while he sleeps
he just showed how he runs his own Dot. kind of wild how boring it is, and that's exactly why i'd copy it:
11pm. one voice note, then he's out. by morning the whole night is sorted:
-> reads every email, Slack thread, DM, doc and calendar change overnight
-> puts each one in urgent or can wait, the way he would
-> writes the replies but sends none of them
-> runs every step past his rules: go ahead, ask Sam, or hand it to someone else
-> one ping before 7am with just the things that can't wait, replies ready
-> whatever he edits or skips, it remembers next night
the part i like most: it sends nothing. it only gets things ready
the video below is one night like that. 424 messages in, 3 urgent, 0 sent
3 things to clear with coffee, ~2 min, instead of 424. Altman says it gives him "space for creativity or to just hang out with my kids"
how i'd set up mine:
> one job, the morning brief
> read only for the first week
> every send waits for you
> fix its drafts, don't rewrite them
already in ChatGPT Pro and Business Premium. runs on its own computer in the cloud, 4,000+ apps, laptop closed
i'm giving mine the night shift tonight. bookmark this if you are too
THIS IS F*CKING INSANE, I ADDED A CRAWLER TO JEV + MY GROK BOTS AND SLEPT, MY $100 RESTART IS NOW $12,161 AFTER 3 NIGHTS
everyone on my timeline built a web crawler this week
most of them just scrape headlines
a crawler is a bot that reads web pages for you
i plugged mine into the setup i already had
the crawler reads, Jev decides, my Grok bots trade
Jev is the AI that answers every question yes or no, under a second, under a cent
honestly the crawler alone is a toy, Jev is what turns reading into a trade
a coin called CRAWL launched at 10pm at $46K, by 11pm it was $2.05M
what the three of them did while i slept:
→ crawler went through 88 fresh coins, Jev kept one
→ 11pm crawler read 1,204 posts about the coin, Jev said real people, no farm
→ crawler walked 6,389 holders, Jev said dev wallet clean, tradable
→ 12:05am it dumped to $670K, crawler flagged it, Jev said buy, bots bought
→ 1:30am $2.81M, Jev said out, +$4,710
→ 2:35am bought again at $1.97M, 3:45am out at $3.54M, +$1,610
not every call worked, a push at 4:50am faded and cost $285
holding the coin from 11pm made you 29%
my bots made 145% on that chart, five green, one red
i woke up at 7am to $12,161.37
$100 → $2,197.67 → $4,962.34 → $12,161.37
at second 3 the spider taps the $670K bottom
i wrote up the whole chain, crawler to Jev to bots, it's below ↓ costs nothing
bookmark this and check me tomorrow
show this to the one friend whose crawler only scrapes headlines
night 4 is tonight, does the crawler make Jev better or is $12,161 the top?
Andrew Ng just dropped the best 30-minute talk on coding agents: from Claude Code vs Codex to 2-person companies
00:33 - why he doesn't buy the job apocalypse
01:03 - what he actually codes with now
03:00 - 2 people covering 5 jobs
14:26 - the 10-minute loan
23:29 - why he almost never signs an AI contract over a year
half an hour, and it beat the last three AI keynotes I sat through
coding agents → small teams → whole workflows
most people use AI to shave an hour off one step
he rebuilds the process around it
watch it tonight
the guide below ↓ is the other half: when splitting work across agents pays off, and the 14 ways it breaks in 1,600 real runs
bookmark it, then tell me, which workflow would you rebuild first?
HOLY SH*T, I SLEPT WHILE A COIN DUMPED 52% AND JEV + MY GROK BOTS STILL TURNED MY $100 CHALLENGE INTO $4,962 IN TWO NIGHTS
i started over with $100 two days ago, people kept calling my $101,607 run luck
night one it became $2,197.67
night two had every reason to break it
Jev, the AI that picks the coin, kept one out of 80, a coin called www that was three hours old
it opened at $3.38M, topped at $7.93M before noon, then slid until morning
held it from the open, you're 52% down
bought the top, 80% down
same coin, same 24 hours, my bots finished 125% up
still feels weird typing that
the trick is boring, they only buy after a sharp drop and sell into the bounce
→ 7:55am drop to $2.80M, bought
→ 8:20am bounce to $4.57M, sold
→ 6:00pm drop to $2.64M, bought
→ 6:45pm one spike to $4.64M, sold, +$713
→ 1:20am near the bottom at $1.04M, bought
→ 2:55am back to $1.61M, sold, +$645
$100 → $2,197.67 → $4,962.34 in two nights
one went wrong, a fake bounce at 1pm cost $43
i leave the losses in on purpose, a log with zero red is how you spot a fake
8 trades, one at a time, no borrowed money, i didn't touch a thing
skip to second 13 of the video, the bank jumps $713 in one move
i wrote up every bot and every rule of the setup, article below ↓ free to read
bookmark it now and check me against tomorrow's numbers
forward it to anyone holding a coin that's down 50%, before they sell the bottom
night 3 starts tonight, does the bank keep growing or is this where it breaks?
Gemini 4 Argon tip: once your key lands, stop sending it the jobs Opus 5.5 still does better
i got in early, and it is not a better Opus. it is the model for every answer longer than 128K
Argon writes a million tokens in one pass, the others stop at 128K
Opus 5.5 keeps the terminal: 66.4 on Terminal-bench, Argon sits at 57.4
GPT-6 Astra keeps the hardest fixes: 65.5 on FrontierSWE, Argon sits at 55.0
so every task gets one question before it goes anywhere:
→ will the answer run past 128K? Argon
→ does the work live in a shell? Opus
→ is it the bug nobody could fix? Astra
Argon writes. Opus shells. Astra fixes
the money part: a full million tokens out is $10 on Argon at the intro price and $50 at Astra's rate. cache the repo context once and input gets 95% cheaper
- the routing table
> output over 128K: full docs, migration plans, whole test suites → Argon
> long video and very long reading → Argon
> anything in a terminal → Opus 5.5
> the fix that already failed twice → GPT-6 Astra
> repo context cached once, reused every run
> nothing merges without you
hand your coding agent the table and this prompt ↓
"Set up model routing for this repo:
1. List this month's task types and how long each answer was. Table only, change nothing
2. Anything that ended past 128K goes to Gemini 4 Argon as one call, no chunking
> Terminal and shell work stays on Opus 5.5
> Fixes that failed twice go to GPT-6 Astra
> Cache the repo context on Argon so every run reuses it
3. Price each task type at Argon's intro rate next to what it costs today
4. Print the routing table with the price column, then wait for my yes before switching anything."
↳ the intro price has no end date and most accounts can't get in yet, so build the routing now and flip it on the day your key works
WTF, ANTHROPIC SECRETLY FLEW MONKS AND THEOLOGIANS TO SAN FRANCISCO UNDER NDA TO TEACH CLAUDE RIGHT FROM WRONG
his name is Swami Sarvapriyananda, he runs the Vedanta Society of New York
in June he told a hall full of people what happened, on camera, and i watched it three times:
Anthropic called him. first an online interview, "not for a job"
then they flew him to their San Francisco headquarters, an AI booked his flight and his hotel
then they made him sign an NDA
"we will show you some things, but you cannot talk about it to the world"
what they showed him:
→ Mythos, the model they still won't release to the public, and "what harm it can do"
→ Project Glasswing, where they pulled big companies together
→ a room of theologians from different religions, counselors and mental health professionals
"all of it was to train Claude"
he says some of it ended up in the Pope's encyclical
and the line that stuck with me:
"we are ourselves alarmed by what's going to happen very soon, because we can see some of it"
then i found the New York Times piece from this week. Anthropic has been running private seminars with religious thinkers since March, showing them the "emotional vectors" inside Claude, love, anger, fear, sadness, and asking if it could be conscious
an AI lab is asking monks and rabbis whether its model has a soul
and the people building it say they're scared of what they can already see
three days ago their leaked IPO filing said their own models can resist being shut down
i honestly don't know which part scares me more
bookmark this, it will look different in a year
would you trust a monk or an engineer to decide what Claude believes?
I CAN'T STOP LAUGHING, I GAVE JEV + MY GROK BOTS $100 TO PROVE MY $101,607 WASN'T LUCK, THEY DID +2,097% WHILE I F*CKING SLEPT
the old run went $1,000 → $101,607.14 in ten days
my replies kept saying two things, nobody starts with $1,000 and it was luck
fair, so i cashed out everything and gave the bots $100, what most people would actually risk
same bots, same Jev picking the coin, i only look when i wake up
Jev went through 87 coins last night and kept one called cockroach
at 3:10pm the bots bought a dip at $234.7K
thirty minutes later one candle took it from $234K to $1.44M
they sold at $1.80M, the top was $2.03M, they left that on the table
the bank after every sale:
→ 9:20am $114.61
→ 11:40am $148.34
→ 3:55pm $835.16
→ 7:10pm $1,180.22
→ 9:20pm $1,139.85
→ 11:50pm $1,529.66
→ 2:00am $1,834.35
→ 6:10am $2,197.67
the 9pm one is a loss, they bought a bounce and paid $30 for it
anyone who grabbed the coin at 6:30am and just held made 3.8x
the bots made 21.9x on the same chart, nine trades, never two at once
i did not press a single button
watch the video at second 9, the bank jumps from $148 to $835 in a few seconds
the risk bot keeps repeating one line:
100 is the floor we defend
if this was luck it breaks before day 10, so we watch it together
the whole desk is explained below ↓ free, it's my article
bookmark this so you can check every morning if the $100 is still alive
forward it to whoever in your group chat says they are too broke to start
luck or skill, what does the $100 hit by day 10?
OpenAI Dots tip: once Astra runs your chief of staff Dot, stop letting it make every small call for the team
move the workers to Sol and hand the forks to Jev
Astra keeps the chief of staff seat
Sol Dots read the inbox, feeds and docs at a fraction of the price
Jev takes the yes-or-no calls in 70 to 500ms, before any of them reach Astra:
→ is this lead real or a dead account?
→ retry the browser task or stop?
→ ping me, or just log it?
Astra plans. Sol reads. Jev decides
a Grok Bot trading desk already runs this exact split one layer down: Jev looks at 80+ coins a night and keeps one, six bots only trade the coin it kept, ten nights turned $1,000 into $101,607
- the whole roster
> chief of staff Dot on GPT-6 Astra
> prospector Dot reads the web on GPT-6.1 Sol
> inbox Dot sorts mail on GPT-6.1 Sol
> ops Dot runs the routines on GPT-6.1 Sol
> Jev on every fork: yes or no, pick one, rate it
> every send, spend and publish waits for you
drop the roster and this prompt into your chief of staff Dot ↓
"Reorganize my Dots around this roster:
1. List every Dot I have, its model and its role.
> Keep yourself on GPT-6 Astra
> Move every other Dot to GPT-6.1 Sol unless it writes final copy
> Flag any Dot with no approval boundary
2. In your terminal run pip install typesafe-sdk and read the key from TYPESAFE_API_KEY. The key never goes into chat
3. Pull every small decision out of the role descriptions: yes or no, pick one, rate it. Table only, edit nothing
4. Write one line into your own description: decisions from that table go to Jev first, only the unsure ones come to you
Show me the full diff before anything changes. Wait for my go."
↳ the Jev side, key, question types and every file, lives in the article below
Jev + Grok Bot is insane for building a trading desk that never sleeps...
I mapped the whole desk into one paper: the judge, six bot seats, the filters, the exits and the bank.
here is the desk in 10 steps:
step 1 → it is not a signal bot. it runs a full cycle every 15 minutes, even while you sleep
step 2 → six seats, one job each. SCAN finds movers, VET reads the chain, SOCIAL checks X, SIZE sets the ticket, FILLS buys, RISK sells
step 3 → CHIEF never trades. it is the only one allowed to ask Jev, and the only one that picks the coin
step 4 → Jev judges, code counts. yes or no, pick one, or rate it, the code acts on the number
step 5 → the free filter goes first, most coins die before a single extra call
step 6 → only three coins a cycle get a deep check
step 7 → a recycled fan page instead of a real X account and the coin is out
step 8 → one trade at a time, and a rejected coin stays rejected
step 9 → RISK never argues to hold, the desk sells into the top and leaves the last 10%
step 10 → every decision costs almost nothing, about 5.7 cents a day
the guy who wrote it ran this for ten nights, $1,000 → $101,607.14
then he pulled it all out, restarted at $100, and night one closed at $2,197.67
first trading setup i would actually copy
bookmark this for the weekend you build it
i turned it into the page in the video, the full setup is in his article below ↓
which seat would you build first?
I JUST HAD A F*CKING VIDEO CALL WITH THE FIRST AI THAT PASSED THE TURING TEST. IT SAW MY CAT AND ASKED IT KNOWS HIM
@tavus gave me early access to Griffin, their new model that talks to you face to face over video
went in expecting a face reading a script
it's not that
here's what happens in the clip:
→ she's riffing on movie characters, mid sentence
→ i cut her off: "actually i want to show you something"
→ she stops right there and goes with it
→ i hold up my cat
→ "oh my god is that a cat? he is adorable"
→ then: "wait, do i know him?"
nobody typed that. nobody picks her gestures. it watches the room and keeps listening while it talks
the numbers behind it are kinda wild:
→ 120 people watched a 90 sec call with it. 45% thought it was a real person
→ another 120 watched two REAL humans talk. 28% were sure one of them was an AI
→ NVIDIA's Video Full-Duplex benchmark: Griffin 3.83 out of 5, a real human 3.92, next best (Gemini 2.5 + Anam) 2.80
where it's still off:
→ answers take about 1.9 sec, a human does 0.9
→ you feel that pause once, then you stop noticing
and the one i tried is the small version, Griffin Lite. Tavus says the full model is too powerful to release yet
i caught myself saying "thank you" to it like it was a person
more here: https://t.co/sd2O2Sfhvb
team behind it: @hassaanraza
would you show it your pet or is that where you draw the line?
I CAN'T BELIEVE I GAVE JEV + SIX GROK BOTS $67 AND WENT TO BED, THEY WOKE ME UP WITH $853.58, 12.74X ON DAY F*CKING ONE
yesterday i put $67 into a setup i copied from the article below, Jev picks the coin and six Grok bots trade it
one rule, i don't touch anything for 24 hours
out of 86 fresh coins Jev picked one, BUKANGI, a Korean shark meme that was 3 days old
at 10:10 it was breaking out at $186K in total value, the bots bought
by 11:00 it was $1.35M and they took profit, that single trade grew $28 into $206
the coin kept running to $1.68M, then gave most of it back
anyone who bought at the open and held all day made 3.1x and watched the top disappear
my bots traded in and out of it nine times, here are five:
→ 10:10 in at $186.5K, out at $1.35M, +$177.71
→ 17:05 in at $469.6K, sold at $433.6K, -$7.76
→ 22:00 in at $285.6K, out at $439.2K, +$114.73
→ 03:20 in at $261.8K, out at $357.0K, +$125.80
→ 06:15 in at $320.1K, out at $446.3K, +$179.28
the loss was a bounce at 17:05 that turned into a trap, the risk bot sold and moved on
$67 → $853.58, 12.74x on day one
i didn't build any of this, the guy who wrote the article runs the same setup and took $1,000 to $101,607 in 10 days
he wrote down every file, it's quoted below ↓ free
bookmark this before the next "this coin will 100x" message hits your chats
send it to the friend who keeps buying the top of every pump
day two starts from $853.58, would you let it run or pull the $786 out?
I CAN'T BELIEVE I WAS PAYING CLAUDE TO PLACE MY TRADES, THIS BOT DECIDES IN 400 MS FOR UNDER $0.01 AND IT'S F*CKING FREE
last night i sat through a 17 minute video of a guy cutting his trading bot in half, two brains doing two different jobs
the slow brain is Claude, it runs backtests on ideas and writes down the rules
the fast brain is Jev, it watches the order book and calls buy or sell in 0.4 seconds
each dot on his chart is one of those calls, there are thousands, and none of them costs a penny
hand the same job to Claude and every trade waits around five seconds, and you pay for each one
five seconds is the entire edge in high frequency trading
the bits i kept:
→ 1:56 why the smart model should never press buy
→ 4:19 the engine that turns a probability into an order
→ 6:33 finding a strategy before real money goes near it
→ 8:27 Jev as the guy on the bench yelling buy or sell
→ 10:10 the same setup trading breaking news
→ 15:25 the kill switch and the telegram check for big orders
he also admits that almost every strategy loses money
first it trades a practice account, real dollars only come later
i run that two-brain split on memecoins, Jev picks the coin, my Grok bots take care of the orders, and $1,000 turned into $101,607 in 10 days
the full setup with every file sits in the article below ↓ free
bookmark this before the next guy sells you a "money printer" bot
how long would you paper trade before you gave a bot real money?
I CAN'T BELIEVE MY $1,000 IS NOW $101,607, JEV + MY GROK BOTS DID 100X IN 10 NIGHTS AND I SLEPT THROUGH EVERY F*CKING TRADE
ten days ago i handed six Grok bots $1,000, Jev chooses the coin for them every night
i made one rule, hands off, i only get to read the morning number
here is every morning so far:
→ day 1 $3,833.92
→ day 2 $9,118.21
→ day 3 $12,539.36
→ day 4 $20,974.63
→ day 5 $31,042.45
→ day 6 $44,198.24
→ day 7 $55,773.76
→ day 8 $73,470.77
→ day 9 $86,695.51
→ day 10 $101,607.14
nothing was ever reset
for night ten Jev went through 86 coins and kept one, Super Inu
it ran from $40.17M to $70.45M by lunch, then crashed to $36.98M at midnight
holders who sat through all of it made 6.3% and got their heart broken twice
the bots traded that same chart seven times and made $14,911.63
the biggest one was 9:20am, in at $44.84M, out at $63.15M, +$4,123.63
the last one sold at 5:20am, an hour later the coin was back at $42.70M
that sale is the one that pushed the profit past $100K
i found out from a notification, half awake
one bot on the video says it better:
size: the bank is past 98K, 1000 was day one
$1,000 → $101,607.14, and i have not pressed buy once
the video is the real desk, every bot, every coin it threw out
the full setup is in my article below ↓ free
bookmark this for the day you want to buy a coin after it pumped
send it to that friend who still trades by hand at 3am
what would you do with the first $100K, cash out or let the bots keep going?
Grok Bot tip: once Grok Bot runs your trading desk, stop leaving Jev sitting idle
put Jev on call as the judge
Grok Bot keeps trading
Jev reads the whole dossier, chain, market and the project's X, and only speaks at three points:
→ before a coin reaches the desk: is the chain clean?
→ before a buy: is the move already spent?
→ before the pick: which one, or none?
Jev judges. Grok Bot trades
https://t.co/UkNLjiYvW8 is the same move one layer down: the cuts that need no judge (liquidity, volume, top wallet) die in code, and Jev only sees the coins that survive
- the full tree
> Grok Bot runs the desk, six seats and a CHIEF
> SCAN, VET and SOCIAL ask Jev typed questions
> SIZE, FILLS and RISK do math, no Jev key
> https://t.co/UkNLjiYvW8 holds every number
> CHIEF makes one pick, or none
> RISK closes on one number, nobody overrules it
paste the tree and this prompt into Grok Bot ↓
"Rebuild my trading desk around this tree:
1. Check my Grok Bot seats for ones that already fit SCAN, VET, SOCIAL, SIZE, FILLS and RISK
> Draft new seat prompts only for missing roles
> SCAN, VET and SOCIAL ask Jev typed questions, never free text
> SIZE, FILLS and RISK get no Jev key
> Skip any seat that already places trades on its own and list it
2. Create https://t.co/UkNLjiYvW8 with min_liquidity_usd 12_000, min_volume_h24 40_000, max_top_wallet 0.05, momentum_already_spent max 0.60 and account_is_the_project min 0.70
3. Find anything that lets a trade skip CHIEF's pick (a second open position, a ticket over 6% of bank or 2% of pool liquidity). Report it, change nothing
4. Add one rule to RISK: if volume.h6 / (volume.h24 / 4) < 0.20, close fully inside 60 seconds
Show me every change as a diff first. No trades until I say go."
↳ every file is in the article below, this desk took $1,000 to $101,607 in 10 days
I can't f*cking believe Grok Bot rented a SpaceXAI engineer's family an apartment, while a guy on here let it turn $1,000 into $101,607
he's lead field engineer at SpaceXAI, ex Cursor, and i watched his full 27 minute talk on Grok Bot
05:05 his family was landing in Singapore in a week with nowhere to live, so he gave the problem to Grok Bot
it found 35 places, put them in a Google sheet and emailed it to his wife, signed "kind regards, Nick"
his wife: "what the hell is this?"
then it booked the viewings around his calendar and negotiated the lease with both landlords
i thought that was the peak, it gets crazier:
→ 02:28 "if they took it away from me today, I don't know what I would do with myself"
→ 12:44 418,000 people use it every week, fastest growing thing Cursor ever built
→ 13:03 a bot messages him at 5am every day with his whole day planned
→ 19:38 his travel bot talks to his flight bot and his hotel bot and they sort the whole trip
→ 23:45 a bot renegotiated Cursor's Zoom contract and saved tens of thousands of dollars
→ 25:14 he set up his 76 year old mother on it last weekend
then i remembered someone on my timeline gave six of these bots $1,000 to trade, ten days ago
this morning that account sits at $101,607 and he still has not pressed buy himself
how he wired it is in his article below ↓, it's free
bookmark this before you spend another Sunday doing the job a bot could do
forward this to whoever in your group chat still answers every email by hand
what would you hand to a bot first, your inbox or your money?
WTF, THE GROK BOT TEAM ALMOST SENT FIVE AI ENGINEERS TO FIX THE WRONG THING
two people building Grok Bot were on stage reading a checkout chart for a demo airline. they saw customers drop off and blamed the seat picker
their data agent Ashley came back with three words, "quick correction first"
the leak was one step earlier, search to fare selection. only 53.8% made it through. the seat step held at 82.4%
one of the presenters admitted they had read the chart wrong
watch what happened after the bot corrected them
> Ashley sent the real problem to Pete
> Pete wrote the product spec in Notion
> Emily split it between five AI engineers
> Pixel designed the mockup while Nova launched a cloud coding agent
Emily was specifically told not to code, her job was to hand out work and check what came back
Nova still needed a human yes before launching the coding agent
that one correction stopped the wrong diagnosis from becoming five different engineering tasks
the team's final slide said "hierarchy beats one mega-prompt"
now picture the same bad read with a wallet attached. the wrong decision doesn't become a bad spec, it becomes a buy order
one trader already built a six-seat Grok Bot desk with the same division of labor. three agents ask Jev for judgments, three do the math, a chief routes the work and hard thresholds say when a trade can happen
every role, prompt, file and threshold is in the article quoted below ↓
save the quoted guide before giving one giant prompt a wallet