
AI + Crypto = Super Money [email protected]
Joined October 2021
- Tweets235
- Following152
- Followers6.2K
- Likes660
Cryptodips retweeted
I SKIPPED A LAUNCH BECAUSE THE DEPLOYER WALLET WAS 11 MINUTES OLD, AND THAT ONE LINE HAS SAVED ME MORE THAN ANY INDICATOR I'VE EVER PAID FOR
I SKIPPED A LAUNCH BECAUSE THE DEPLOYER WALLET WAS 11 MINUTES OLD, AND THAT ONE LINE HAS SAVED ME MORE THAN ANY INDICATOR I'VE EVER PAID FOR
Cryptodips retweeted
i don't sell a course and people keep asking why
because a course is a promise, and i can't promise you my results
what i can do is hand you the build and let you find out. that's the free guide. it's the same thing i'd charge for, and it's free because the version that works for you isn't the version i'd teach anyway
the paid thing i do point at is a group where someone else does the work. different product, different promise
comment CONTROL and follow me, i'll DM you the build
Cryptodips retweeted
I BUILT FIVE AI EMPLOYEES AND KEPT ONE, AND THE FOUR I DELETED WERE THE ONES I WAS PROUDEST OF
the roster i started with:
→ scout, finds launches
→ analyst, builds the case
→ skeptic, tries to kill the case
→ quant, prices and sizes
→ settler, closes
the one that survived: skeptic
why the others went:
analyst — its output was a paragraph. i never once read the paragraph
before making a decision. if you don't read it, it's not a job, it's
decoration
quant — it was doing arithmetic i could do in a config file. three
agents agreeing on a number is not more correct, it's slower
scout — the screener already did this and did it without opinions
settler — genuinely useful, but it's 40 lines. it doesn't need a
personality or a charter
the skeptic stayed because it does the one thing i can't do reliably at
02:00: it argues against the trade without wanting the trade to happen
that's the actual test for whether an agent deserves to exist. not
"can it do this". "does it do something i systematically fail at"
most rosters are five versions of the same optimistic person. you don't
need a team, you need one adversary
the desk ran $100 → $2,197.67 → $4,962.34 → $12,161.37 → $17,608 →
$23,879 across five nights, and the skeptic didn't cause a single
dollar of that. it just made me able to sleep through it
comment ROSTER and follow me, i'll DM you the whole setup, free
which of your agents would survive that test?
NFA. DYOR.
Cryptodips retweeted
i trade through the first one and read the second one before i sleep
trading here: fomo.family/r/Cryptodips
alpha here: whop.com/cryptoarchie/crypto…
Cryptodips retweeted
ONE OF MY SIX BOTS DIED AT 01:40 ON NIGHT 4 AND I DIDN'T NOTICE UNTIL THE MORNING, WHICH IS THE ONLY REVIEW THAT MATTERS
that's not a failure story. that's the point of the design
what happened: the book reader crashed on a malformed response. bad
json, no retry, process gone
the desk kept trading
why:
→ the sizer had a fallback depth estimate and used a third of normal size
→ the kill switch ran on a timer, not on the reader, so it didn't care
→ the settler kept closing what was already open
→ the screener kept finding, but found fewer, because it read the
reader's cache
result: night 4 finished lower than night 3, and i spent an hour
thinking that was variance
it wasn't variance. it was one bot missing and five compensating, badly
but safely
the desk went $100 → $2,197.67 → $4,962.34 → $12,161.37 → $17,608 →
$23,879 across five nights, and night 4 is the dip in that line
what i changed after:
→ every bot gets a heartbeat and a last-seen timestamp
→ a missing heartbeat doesn't stop the desk, it shrinks it
→ no bot is allowed to be the only source of a number
→ the log says which bots were degraded, per trade
the design principle is unglamorous: build for the version where the
thing is broken, because it will be, and it will be at the hour you're
least able to fix it
a single bot that does everything isn't simpler. it's a single point
of failure with better branding
comment BOTS and follow me, i'll DM you every charter, free
what happens to your setup when one piece silently dies?
NFA. DYOR.
Cryptodips retweeted
i used to read 40 threads a night and still miss the entry
trading here: fomo.family/r/Cryptodips
alpha here: whop.com/cryptoarchie/crypto…
i don't sell a course and people keep asking why
because a course is a promise, and i can't promise you my results
what i can do is hand you the build and let you find out. that's the free guide. it's the same thing i'd charge for, and it's free because the version that works for you isn't the version i'd teach anyway
the paid thing i do point at is a group where someone else does the work. different product, different promise
comment CONTROL and follow me, i'll DM you the build
Cryptodips retweeted
MY AI READ 4,000 CRYPTO POSTS LAST NIGHT AND THE ONLY USEFUL OUTPUT WAS A LIST OF NAMES TO IGNORE
i expected signal. i got a blocklist. the blocklist was worth more
here's what it actually produced:
→ 212 accounts that called a token in the 6 minutes before it pumped
and never mentioned it again
→ 38 accounts that only appear when something is already up 300%
→ 9 accounts that were right about the same token and wrong about
everything else
→ 4 accounts whose calls are genuinely early, repeatedly, across
different tokens
those 4 are the whole output. everything else was noise wearing a
research hat
the part that surprised me: the 212 was easier to find than the 4.
patterns of avoidance are loud. patterns of skill are quiet and
inconsistent, because being early is mostly uncomfortable
so the crawler's real job changed. it doesn't hunt for calls anymore.
it maintains a list of who not to read, and it updates that list every
night whether i look at it or not
i deleted about 300 follows in two weeks. my timeline got shorter and
my hit rate went up. those two things are usually the same thing
the desk ran $100 → $2,197.67 → $4,962.34 → $12,161.37 → $17,608 →
$23,879 across five nights, and i'd be lying if i said the crawler
caused that. what it did was stop me from acting on things it had
already discredited
subtraction doesn't screenshot well. that's why nobody posts it
comment JEVDESK and follow me, i'll DM you the build, free
who's the one account you've learned to ignore, and how long did it take?
NFA. DYOR.
Cryptodips retweeted
two links, and i'll say it upfront: they're mine
trading here: fomo.family/r/Cryptodips
alpha here: whop.com/cryptoarchie/crypto…
referral links. NFA
I BUILT FIVE AI EMPLOYEES AND KEPT ONE, AND THE FOUR I DELETED WERE THE ONES I WAS PROUDEST OF
the roster i started with:
→ scout, finds launches
→ analyst, builds the case
→ skeptic, tries to kill the case
→ quant, prices and sizes
→ settler, closes
the one that survived: skeptic
why the others went:
analyst — its output was a paragraph. i never once read the paragraph
before making a decision. if you don't read it, it's not a job, it's
decoration
quant — it was doing arithmetic i could do in a config file. three
agents agreeing on a number is not more correct, it's slower
scout — the screener already did this and did it without opinions
settler — genuinely useful, but it's 40 lines. it doesn't need a
personality or a charter
the skeptic stayed because it does the one thing i can't do reliably at
02:00: it argues against the trade without wanting the trade to happen
that's the actual test for whether an agent deserves to exist. not
"can it do this". "does it do something i systematically fail at"
most rosters are five versions of the same optimistic person. you don't
need a team, you need one adversary
the desk ran $100 → $2,197.67 → $4,962.34 → $12,161.37 → $17,608 →
$23,879 across five nights, and the skeptic didn't cause a single
dollar of that. it just made me able to sleep through it
comment ROSTER and follow me, i'll DM you the whole setup, free
which of your agents would survive that test?
NFA. DYOR.
Cryptodips retweeted
MY BEST BACKTEST WAS A STRATEGY THAT LOST MONEY EVERY TIME I RAN IT LIVE, AND THE GAP BETWEEN THOSE TWO FACTS IS THE WHOLE JOB
the backtest said 71% win rate over 90 days. i'd checked it twice. i
believed it
live, it went 4 for 11
here's what the backtest couldn't see:
→ the entry price. the backtest filled at the close of the candle.
live, by the time the signal fired and the transaction landed, i was
paying 3–9% above that
→ the exit. same problem in reverse, and the exit is where the money is
→ the pool. the backtest assumed my size was free. on a $670K pool a
$4,100 position is already most of the book
→ the failure. transactions that never landed. the backtest doesn't
have a concept of "didn't happen"
none of those are strategy problems. they're venue problems, and almost
everyone blames the strategy
the fix wasn't a better strategy. it was measuring the gap:
→ log the price you decided at
→ log the price you actually got
→ subtract them, every trade, and look at the average
→ if the gap is bigger than your edge, you don't have an edge
my gap averaged 4.1%. the strategy's edge was about 3%. so the strategy
was never the problem and it was never the answer either
this is also why i stopped reading threads that post a backtest and
nothing else. a backtest is a hypothesis. a fill is evidence
comment STRAT and follow me, i'll DM you the full build, free
what's your slippage average, and do you actually know the number?
NFA. DYOR.
i trade through the first one and read the second one before i sleep
trading here: fomo.family/r/Cryptodips
alpha here: whop.com/cryptoarchie/crypto…
Cryptodips retweeted
everything i run is in one of these two places
trading here: fomo.family/r/Cryptodips
alpha here: whop.com/cryptoarchie/crypto…
ONE OF MY SIX BOTS DIED AT 01:40 ON NIGHT 4 AND I DIDN'T NOTICE UNTIL THE MORNING, WHICH IS THE ONLY REVIEW THAT MATTERS
that's not a failure story. that's the point of the design
what happened: the book reader crashed on a malformed response. bad
json, no retry, process gone
the desk kept trading
why:
→ the sizer had a fallback depth estimate and used a third of normal size
→ the kill switch ran on a timer, not on the reader, so it didn't care
→ the settler kept closing what was already open
→ the screener kept finding, but found fewer, because it read the
reader's cache
result: night 4 finished lower than night 3, and i spent an hour
thinking that was variance
it wasn't variance. it was one bot missing and five compensating, badly
but safely
the desk went $100 → $2,197.67 → $4,962.34 → $12,161.37 → $17,608 →
$23,879 across five nights, and night 4 is the dip in that line
what i changed after:
→ every bot gets a heartbeat and a last-seen timestamp
→ a missing heartbeat doesn't stop the desk, it shrinks it
→ no bot is allowed to be the only source of a number
→ the log says which bots were degraded, per trade
the design principle is unglamorous: build for the version where the
thing is broken, because it will be, and it will be at the hour you're
least able to fix it
a single bot that does everything isn't simpler. it's a single point
of failure with better branding
comment BOTS and follow me, i'll DM you every charter, free
what happens to your setup when one piece silently dies?
NFA. DYOR.
Cryptodips retweeted
I DELETED ONE WORD FROM MY AGENT'S CHARTER AND IT STOPPED OVERTRADING BY 60%
the word was "maximise"
the original line read: maximise returns on every opportunity the
screener surfaces
that sentence is a licence. an agent that has been told to maximise
will find something to maximise, and at 03:00 the only thing available
to maximise is the number of trades
night 1 it took 11 positions. the rule wanted maybe 3. it wasn't
broken, it was obedient
so i rewrote it and the rewrite is boring on purpose:
→ your job is to evaluate, not to find
→ if nothing passes, the correct output is nothing
→ a night with zero trades is a successful night
→ you are never rewarded for volume
trades dropped about 60%. results went up. that's the part that annoys
people, because volume feels like work and it isn't
the deeper problem is anthropomorphism. we write "maximise" because it
sounds ambitious and ambition is a human virtue. an agent has no
ambition. it has an instruction, and it will satisfy the instruction
in the dumbest available way
this is why i now read every charter looking for verbs that can't be
satisfied. maximise. optimise. improve. be aggressive. all of them
point at the same failure
the control layer isn't the prompt that tells it what to do. it's the
language that stops it from having a choice about how
comment CONTROL and follow me, i'll DM you the build, free
what's the laziest instruction you've ever written and regretted?
NFA. DYOR.
i used to read 40 threads a night and still miss the entry
trading here: fomo.family/r/Cryptodips
alpha here: whop.com/cryptoarchie/crypto…
Cryptodips retweeted
i stopped guessing entries at 2am
trading here: fomo.family/r/Cryptodips
alpha here: whop.com/cryptoarchie/crypto…
MY AI READ 4,000 CRYPTO POSTS LAST NIGHT AND THE ONLY USEFUL OUTPUT WAS A LIST OF NAMES TO IGNORE
i expected signal. i got a blocklist. the blocklist was worth more
here's what it actually produced:
→ 212 accounts that called a token in the 6 minutes before it pumped
and never mentioned it again
→ 38 accounts that only appear when something is already up 300%
→ 9 accounts that were right about the same token and wrong about
everything else
→ 4 accounts whose calls are genuinely early, repeatedly, across
different tokens
those 4 are the whole output. everything else was noise wearing a
research hat
the part that surprised me: the 212 was easier to find than the 4.
patterns of avoidance are loud. patterns of skill are quiet and
inconsistent, because being early is mostly uncomfortable
so the crawler's real job changed. it doesn't hunt for calls anymore.
it maintains a list of who not to read, and it updates that list every
night whether i look at it or not
i deleted about 300 follows in two weeks. my timeline got shorter and
my hit rate went up. those two things are usually the same thing
the desk ran $100 → $2,197.67 → $4,962.34 → $12,161.37 → $17,608 →
$23,879 across five nights, and i'd be lying if i said the crawler
caused that. what it did was stop me from acting on things it had
already discredited
subtraction doesn't screenshot well. that's why nobody posts it
comment JEVDESK and follow me, i'll DM you the build, free
who's the one account you've learned to ignore, and how long did it take?
NFA. DYOR.
Cryptodips retweeted
I REPLACED MY PROFIT TARGET WITH A BUDGET AND THE DESK IMMEDIATELY STOPPED DOING STUPID THINGS
for the first three nights i ran it with a target: make $X by morning
that's the worst instruction you can give an agent and i didn't see it
for three nights
a target creates pressure. pressure at 02:00 looks like a reason to
take a trade that isn't there
night 3 it was down $340 at 01:00 with a target sitting in its charter.
it took four more positions before 04:00. three of them were charts the
filter had already rejected once
so i deleted the target and wrote a budget instead:
→ you have $400 of risk for the night
→ when it's gone, the night is over
→ there is no way to earn more risk back inside the same night
→ a good setup at 03:50 with $0 left is not a setup
the difference is not semantic. a target asks "how do i get there". a
budget asks "is this worth part of what's left"
the desk went $100 → $2,197.67 → $4,962.34 → $12,161.37 → $17,608 →
$23,879 across five nights, and the shape of that curve changed on
night 4 when the budget went in
before the budget, the losses were clustered at the end of the night.
after it, they were spread out and smaller
the part nobody tells you: a budget makes the boring nights more
common, and the boring nights are what let the good ones survive
if you take one thing from this: don't give an autonomous system a
goal. give it a limit. goals get optimised. limits get respected
comment DESK and follow me, i'll DM you the whole thing, free
what's your limit right now, or are you still chasing a number?
NFA. DYOR.
