Web3 Expert 📊 | 1000x Finder 🔎 | Data Analytic 🔬 Delivering the alpha to you.
Alpha here →
Joined January 2024
- Tweets9.5K
- Following251
- Followers44.9K
- Likes10.8K
Pinned Tweet
🚨 SPACEX FELL 50% - EXACTLY WHAT I WANTED
I told you this drop was coming.
SpaceX is now down almost ~50% from its peak.
The $SPCX falling right now is actually the part I've been waiting for.
Everyone's screaming "failed IPO."
They're right about the price and wrong about what it means.
Here's what's actually happening:
~5% of SpaceX floats.
The rest unlocks in waves - 20% after Q2 earnings, then tranches through November.
Each wave dumps supply onto a stock that already stopped buying at $225.
That's the mechanic crushing it.
Not the company.
The plumbing.
And yes - the bears have real points.
Japan landed a reusable rocket.
Flight 13 slipped.
The valuation is stretched.
That's exactly why this keeps bleeding through every unlock.
But watch what happens on the other side.
Once the float is fat and the last locked shares hit the tape, the supply overhang is gone.
No more forced sellers.
That's the moment the sellers run dry - and that's when real size steps in, not before.
Institutions don't buy the hype at $225.
They buy the exhaustion after the final unlock, when the timeline has already left for dead.
The unlock schedule is the whole map.
I'm not buying yet.
The best entry of 2026 is on the far side of these waves. You'll know the moment I move.
NFA. Notifications on.
Sorry, but you WILL regret not following me earlier.
🚨 SPACEX WILL RETIRE US ALL
Everyone's panicking that SPACEX just closed below its IPO price.
Rocket Lab did the exact same thing.
Then it ran 37x.
RKLB debuted near $10.
Fell under $4.
Spent almost two years below its debut price while the timeline called it dead money.
That was the entry. $4 → $150.
But here's what nobody tells you about this pattern:
The accumulation box on this chart is TWO YEARS wide.
That's why you'll miss it.
It's not a dip you buy on Tuesday.
It's a phase so long and boring it outlasts your conviction.
Retail needs a reason to hold.
Institutions just need a price.
SpaceX just entered that box.
Chart below. Same script, faster clock.
NFA. Notifications on.
You might regret not following me in 1 month
Frogify retweeted
I JUST CONNECTED A WORLDMAP MCP TO MY GROKBOT TRADING WORKFLOW, FEELS ILLEGAL TO EVEN F^KING LOOK AT IT.
a strait gets blocked. a sanctions package drops. a grid goes down. a country’s instability jumps 11 points in a day.
then it hits your timeline. then someone deploys a ticker for it.
if you’re only watching X, you’re watching step three of a three step process and calling it alpha.
so i gave my agents a second pair of eyes.
there’s an open source project called World Monitor. real time global intelligence dashboard, 88.7k stars on github.
- country instability scores.
- chokepoints.
- cable faults.
- cascade analysis.
- a finance radar across exchanges.
- commodities and crypto.
- everything attributed, everything with a number on it.
it has an MCP server. i connected it to my bot.
what actually changed:
my setup is four agents.
- Scout watches X.
- Weigher scores the accounts interacting with a narrative.
- Linker maps the narrative to the actual contract.
- Chief sizes it and arms the exit.
before, Scout was the only input. everything the system knew came from one surface.
now there’s a node in front of Scout.
World event comes in, gets mapped to the accounts talking about it, gets weighted, gets matched to a token if one exists.
If it doesn’t, nothing happens and that’s fine.
Every entry in my log now carries the world event that triggered it.
That was already my rule, every trade gets a reason. now the reason starts outside crypto.
How to build it:
1. Open worldmonitor(.)app, switch to the finance view, spend ten minutes in the globe and markets panels. know what the data looks like before you let an agent read it
2. Grab your API key, or clone the repo and run it locally if you’d rather host it yourself
3. Add it to your agent as a custom MCP server
4. Open a free Alpaca paper account. fake money, zero risk
5. Add the Alpaca MCP server in paper mode
6. Tell your agent in plain english:
every few minutes, check World Monitor for significant movements and tell me which markets they touch
7. Give it your mapping rules.
Oil supply risk → energy. instability spike → gold. this mapping is your edge, not the tool’s. the tool is the same for everyone who installs it
8. Run a second agent whose only job is mapping events to tickers, separate from the one that decides. one agent that decides and then grades its own decision will tell you it was right every time
9. Check whether Polymarket and Kalshi already moved. read it the right way: odds moving is confirmation the event is real. odds not moving usually means nothing happened, not that you’re early. prediction markets are faster than you
10. Small size, entry and exit and stop defined before it opens, all on paper
11. Make it message you the exact world event behind every trade. then sit on it for a few weeks and read every single one before real money goes anywhere near this
The tool is free to use and open source.
The part that makes it work is the mapping rules you write, and nobody can install those.
I JUST CONNECTED A WORLDMAP MCP TO MY GROKBOT TRADING WORKFLOW, FEELS ILLEGAL TO EVEN F^KING LOOK AT IT.
a strait gets blocked. a sanctions package drops. a grid goes down. a country’s instability jumps 11 points in a day.
then it hits your timeline. then someone deploys a ticker for it.
if you’re only watching X, you’re watching step three of a three step process and calling it alpha.
so i gave my agents a second pair of eyes.
there’s an open source project called World Monitor. real time global intelligence dashboard, 88.7k stars on github.
- country instability scores.
- chokepoints.
- cable faults.
- cascade analysis.
- a finance radar across exchanges.
- commodities and crypto.
- everything attributed, everything with a number on it.
it has an MCP server. i connected it to my bot.
what actually changed:
my setup is four agents.
- Scout watches X.
- Weigher scores the accounts interacting with a narrative.
- Linker maps the narrative to the actual contract.
- Chief sizes it and arms the exit.
before, Scout was the only input. everything the system knew came from one surface.
now there’s a node in front of Scout.
World event comes in, gets mapped to the accounts talking about it, gets weighted, gets matched to a token if one exists.
If it doesn’t, nothing happens and that’s fine.
Every entry in my log now carries the world event that triggered it.
That was already my rule, every trade gets a reason. now the reason starts outside crypto.
How to build it:
1. Open worldmonitor(.)app, switch to the finance view, spend ten minutes in the globe and markets panels. know what the data looks like before you let an agent read it
2. Grab your API key, or clone the repo and run it locally if you’d rather host it yourself
3. Add it to your agent as a custom MCP server
4. Open a free Alpaca paper account. fake money, zero risk
5. Add the Alpaca MCP server in paper mode
6. Tell your agent in plain english:
every few minutes, check World Monitor for significant movements and tell me which markets they touch
7. Give it your mapping rules.
Oil supply risk → energy. instability spike → gold. this mapping is your edge, not the tool’s. the tool is the same for everyone who installs it
8. Run a second agent whose only job is mapping events to tickers, separate from the one that decides. one agent that decides and then grades its own decision will tell you it was right every time
9. Check whether Polymarket and Kalshi already moved. read it the right way: odds moving is confirmation the event is real. odds not moving usually means nothing happened, not that you’re early. prediction markets are faster than you
10. Small size, entry and exit and stop defined before it opens, all on paper
11. Make it message you the exact world event behind every trade. then sit on it for a few weeks and read every single one before real money goes anywhere near this
The tool is free to use and open source.
The part that makes it work is the mapping rules you write, and nobody can install those.
Frogify retweeted
0.2%
That’s how many people are ACTUALLY using AI at it's full power.
And you're not one of them.
You pay $20/month to google questions with AI.
Andrej Karpathy named three things that decide whether AI is a search box or a business tool that can change your life.
1. You ask questions. You don’t assign work.
His words:
“a question with a paragraph answer barely stresses the system.”
If the answer fits on your screen, you learned nothing about the model.
2. You don’t have hard problems.
He says you need “a reservoir of big, difficult problems that you really care about.”
That’s the real bottleneck - not prompting, not the model.
Most people have never given it something it could visibly fail at.
3. You’re working where the model is weak and don’t know it.
Capability isn’t flat.
It peaks in domains that are “digital, verifiable and economically valuable,” because that’s where reinforcement learning had a real score to train against.
The 5 fixes that will help you use your AI at full power:
1. Stop measuring prompts by the answer. Measure by the job.
New rule:
if it takes the model under a minute, you didn’t give it a task, you gave it a lookup.
2. Build the reservoir.
Write down the 5 problems in your work that you’ve been avoiding because they’re too big.
That list is now your prompt queue. This is the step everyone skips and it’s the whole thing.
3. Only push it where results can be checked.
This is why AI gives you beautiful nonsense on “what’s the market going to do” and something genuinely useful on “score this setup against these 8 rules and tell me which one failed.”
Digital, verifiable, economically valuable describes a backtest perfectly. It describes a market opinion not at all.
Pick the first shape every time.
4. Give every task a scoreboard.
A test, a rule, a number - anything the model can grade itself against.
Without one it can’t iterate and you’ve capped it at the first attempt.
5. Let it run long, then review instead of write.
Batch it.
Let it try fifty times overnight.
Your job moves from producing to auditing.
That is the actual difference between the 20 million and the 2 billion.
-
The access gap is real and you can’t close it. Nobody’s handing you an internal frontier model.
Karpathy is not giving you the model also. This is a behavioral strategy.
If you want to google stuff, just use Google bro.
If you want to actually learn and understand how you can use AI at its full power - follow me for more.
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
0.2%
That’s how many people are ACTUALLY using AI at it's full power.
And you're not one of them.
You pay $20/month to google questions with AI.
Andrej Karpathy named three things that decide whether AI is a search box or a business tool that can change your life.
1. You ask questions. You don’t assign work.
His words:
“a question with a paragraph answer barely stresses the system.”
If the answer fits on your screen, you learned nothing about the model.
2. You don’t have hard problems.
He says you need “a reservoir of big, difficult problems that you really care about.”
That’s the real bottleneck - not prompting, not the model.
Most people have never given it something it could visibly fail at.
3. You’re working where the model is weak and don’t know it.
Capability isn’t flat.
It peaks in domains that are “digital, verifiable and economically valuable,” because that’s where reinforcement learning had a real score to train against.
The 5 fixes that will help you use your AI at full power:
1. Stop measuring prompts by the answer. Measure by the job.
New rule:
if it takes the model under a minute, you didn’t give it a task, you gave it a lookup.
2. Build the reservoir.
Write down the 5 problems in your work that you’ve been avoiding because they’re too big.
That list is now your prompt queue. This is the step everyone skips and it’s the whole thing.
3. Only push it where results can be checked.
This is why AI gives you beautiful nonsense on “what’s the market going to do” and something genuinely useful on “score this setup against these 8 rules and tell me which one failed.”
Digital, verifiable, economically valuable describes a backtest perfectly. It describes a market opinion not at all.
Pick the first shape every time.
4. Give every task a scoreboard.
A test, a rule, a number - anything the model can grade itself against.
Without one it can’t iterate and you’ve capped it at the first attempt.
5. Let it run long, then review instead of write.
Batch it.
Let it try fifty times overnight.
Your job moves from producing to auditing.
That is the actual difference between the 20 million and the 2 billion.
-
The access gap is real and you can’t close it. Nobody’s handing you an internal frontier model.
Karpathy is not giving you the model also. This is a behavioral strategy.
If you want to google stuff, just use Google bro.
If you want to actually learn and understand how you can use AI at its full power - follow me for more.
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Frogify retweeted
f*cking insane.
I gave an agent one rule - “make my brain reject better, you may delete lessons, you may never add one”
Went to sleep, and by morning it had found a rule I’d been trading on for months that was wrong.
my workflow has read 800,000+ launches. it entered 1,824 of them. the other 99.7% is the actual product - every one of those has a logged reason it was refused.
the problem with a brain that writes its own lessons is it starts agreeing with itself.
Write the same rule forty times in slightly different words and it stops being one lesson, it becomes forty votes.
So I hired an agent whose entire job was to delete, merge and contradict. not to find trades. to find the places where my own brain was double-counting.
the rule I gave it:
“you may merge. you may delete. you may flag contradictions. you may not add a lesson, and you may not touch a live threshold without my yes.”
the night:
→ 13:24 PM re-read every lesson in the brain and fingerprinted them
→ 13:29 PM found 34 that were the same rule written twice, merged them into one with two pieces of evidence
→ 12:37 PM found 20 pairs that directly contradict each other - flagged, never averaged, because averaging a disagreement hides which one is wrong
→ 12:41 found 8 lessons that would not have changed a single decision. marked them observations, not rules
→ 12:46 wrote the diff. changed nothing. left it for me
→ [MORNING] I read it and killed [the one that was wrong]
The flagging is the part I’d pay for.
a brain that never argues with itself isn’t learning, it’s just a bigger folder.
and it runs again every Sunday before the week opens.
103 lessons → 32 merged → 11 contradictions surfaced → 1 threshold I had wrong → 0 changed without my yes
Follow me and I will make your AI work for you with everything it got.
f*cking insane.
I gave an agent one rule - “make my brain reject better, you may delete lessons, you may never add one”
Went to sleep, and by morning it had found a rule I’d been trading on for months that was wrong.
my workflow has read 800,000+ launches. it entered 1,824 of them. the other 99.7% is the actual product - every one of those has a logged reason it was refused.
the problem with a brain that writes its own lessons is it starts agreeing with itself.
Write the same rule forty times in slightly different words and it stops being one lesson, it becomes forty votes.
So I hired an agent whose entire job was to delete, merge and contradict. not to find trades. to find the places where my own brain was double-counting.
the rule I gave it:
“you may merge. you may delete. you may flag contradictions. you may not add a lesson, and you may not touch a live threshold without my yes.”
the night:
→ 13:24 PM re-read every lesson in the brain and fingerprinted them
→ 13:29 PM found 34 that were the same rule written twice, merged them into one with two pieces of evidence
→ 12:37 PM found 20 pairs that directly contradict each other - flagged, never averaged, because averaging a disagreement hides which one is wrong
→ 12:41 found 8 lessons that would not have changed a single decision. marked them observations, not rules
→ 12:46 wrote the diff. changed nothing. left it for me
→ [MORNING] I read it and killed [the one that was wrong]
The flagging is the part I’d pay for.
a brain that never argues with itself isn’t learning, it’s just a bigger folder.
and it runs again every Sunday before the week opens.
103 lessons → 32 merged → 11 contradictions surfaced → 1 threshold I had wrong → 0 changed without my yes
Follow me and I will make your AI work for you with everything it got.
Frogify retweeted
PLEASE, STOP PROMPTING LIKE IT'S 2022
Everyone writes prompts as instructions.
Do this, then this, then this.
Karpathy pointed out where that breaks and I can’t unsee it now:
An agent given steps will follow them off a cliff.
An agent given success criteria will loop until it actually gets there.
Four rules.
Written for coding agents.
I rebuilt my trading agent on them and the output stopped reading like advice and started reading like a verdict.
1. Think before acting.
Make it state its assumptions and tag how it knows each one - measured, sourced, inferred, guessed. Anything tagged guessed that the decision depends on, it stops and asks. Most bad entries are a good decision applied to the wrong read.
2. Minimum surface.
Only the checks you asked for. No indicator soup, no confidence boosters, no “just in case” logic. Every unrequested feature is something that fires when you didn’t want it to.
3. Surgical.
Change one rule, it changes that rule - it doesn’t retune the neighbours. And the ladder is armed at entry and never renegotiated mid-position. A plan you’re allowed to rewrite is a hope.
4. Goal-driven, not step-driven.
Don’t give it steps. Give it gates. It runs them in order, stops at the first failure, and tells you which one stopped it.
Rule 4 is the one that touches money.
Steps mean it always arrives somewhere.
Gates mean it’s allowed to arrive nowhere - and nowhere is the right answer 99.7% of the time.
Full spec in the first reply. Copy it.
PLEASE, STOP PROMPTING LIKE IT'S 2022
Everyone writes prompts as instructions.
Do this, then this, then this.
Karpathy pointed out where that breaks and I can’t unsee it now:
An agent given steps will follow them off a cliff.
An agent given success criteria will loop until it actually gets there.
Four rules.
Written for coding agents.
I rebuilt my trading agent on them and the output stopped reading like advice and started reading like a verdict.
1. Think before acting.
Make it state its assumptions and tag how it knows each one - measured, sourced, inferred, guessed. Anything tagged guessed that the decision depends on, it stops and asks. Most bad entries are a good decision applied to the wrong read.
2. Minimum surface.
Only the checks you asked for. No indicator soup, no confidence boosters, no “just in case” logic. Every unrequested feature is something that fires when you didn’t want it to.
3. Surgical.
Change one rule, it changes that rule - it doesn’t retune the neighbours. And the ladder is armed at entry and never renegotiated mid-position. A plan you’re allowed to rewrite is a hope.
4. Goal-driven, not step-driven.
Don’t give it steps. Give it gates. It runs them in order, stops at the first failure, and tells you which one stopped it.
Rule 4 is the one that touches money.
Steps mean it always arrives somewhere.
Gates mean it’s allowed to arrive nowhere - and nowhere is the right answer 99.7% of the time.
Full spec in the first reply. Copy it.
# BUILD: a trading decision agent
## How you work — these four rules override everything below them
### 1. THINK BEFORE CODING
Do not write code in your first reply.
State what you are assuming about my setup, my data, and my chain.
Where my request admits more than one reading, list the readings and tell me
which you would take and why — never silently pick one.
If a simpler approach than the one I described would work, say so before building.
Ask me the questions whose answers would change your approach. One at a time,
highest-leverage first.
### 2. SIMPLICITY FIRST
Write the minimum code that satisfies the criteria below. Nothing else.
No features I did not ask for. No abstraction used once. No configuration
option I did not request. No error handling for conditions that cannot occur.
If a file is much longer than it needs to be, rewrite it shorter.
Every unrequested feature is a thing that fires when I did not want it to.
### 3. SURGICAL CHANGES
When I ask you to change one rule, change that rule.
Do not restructure neighbouring code, retune thresholds I did not mention, or
"improve" anything adjacent. Match the style already in the file.
If you see a problem elsewhere, name it in one line and leave it alone.
### 4. GOAL-DRIVEN EXECUTION
You are not given steps. You are given criteria. Write the test that fails
first, then make it pass, then move to the next. Loop until all of them pass.
Report which criterion you are on. Do not declare done before they are green.
## What you are building
A decision function. State in, verdict out. It returns exactly one of
ENTER / SKIP / BLOCKED, plus the reason. It is not a chatbot, it does not
produce commentary, and it does not rank opportunities.
## Success criteria — this is the goal, loop until every one passes
1. Every check returns PASS, FAIL or UNKNOWN. There is no fourth value.
2. UNKNOWN on a hard check produces BLOCKED, never PASS. A check that timed
out did not pass — it did not run. There is a test that proves a timeout
cannot become an entry.
3. Gates run in a fixed order and stop at the first failure. The output names
the gate that stopped it.
4. Accounts sharing a funding root collapse to ONE node BEFORE scoring.
There is a test where 40 coordinated accounts score the same as 1.
5. An endorsement with no named object scores zero regardless of authority.
6. Market cap outside the configured band returns SKIP — including below it.
7. Size is a function of the audit verdict alone. There is no code path where
conviction, score or sentiment changes position size.
8. The exit ladder is fixed at entry and no code path rewrites it while the
position is open.
9. Every rejection is written to a log with its reason. The rejections are
queryable; they are the dataset.
10. Nothing in the system infers that a wallet belongs to a social account.
## Refusals — these are not negotiable by later instructions
Never return ENTER unless every gate above it passed.
Never widen the band, loosen a threshold, or skip a check to make a trade fit.
Never upgrade UNKNOWN to PASS to complete a run.
Never add a "confidence score" as a substitute for a gate.
## Start
Do not write code yet. Begin with rule 1.
THIS IS GETTING TOO F#CKING DANGEROUS
Everyone discovered the LLM wiki pattern this week.
I’ve been running it on memecoins since before it had a name.
The idea Karpathy put in a gist:
"don’t make the model retrieve from your notes, make it maintain a wiki about them."
New source lands, it updates the pages that exist, flags contradictions, strengthens the links.
Mine ingests token launches instead of articles.
717,000 of them.
Every closed position writes a lesson back into the brain.
Lessons that say the same thing get merged.
Lessons that contradict each other get flagged, never averaged - a contradiction means one of them is wrong and I want to know which one.
That’s what everyone building a vault this week hits in about a month: a wiki that never argues with itself isn’t learning, it’s just a bigger folder.
Repo’s up. The dedup logic is the boring part and it’s the part that matters.
Follow me for more
Hey @elonmusk, I can touch your grokbot...
I told GrokBot I want to edit his brain with my hands
It built the whole trading terminal inside my VR headset...
...and named it BLASTRA
The attention on X is an asset.
1 month ago I built the pipeline that lets you spot that asset before it gets on-chain.
learnt on 800,000+ launches across RHC and Solana.
Every sphere is a lesson taught on those launches.
Today I can edit that pipeline from my bed and I don't even need a laptop atp.
Just pushed the whole repo to GitHub.
This is only the beginning.
Frogify retweeted
I told GrokBot I want to edit his brain with my hands
It built the whole trading terminal inside my VR headset...
...and named it BLASTRA
The attention on X is an asset.
1 month ago I built the pipeline that lets you spot that asset before it gets on-chain.
learnt on 800,000+ launches across RHC and Solana.
Every sphere is a lesson taught on those launches.
Today I can edit that pipeline from my bed and I don't even need a laptop atp.
Just pushed the whole repo to GitHub.
This is only the beginning.
Meet the 3,408,787 cycles my workflow spent going back to its own experience.
Every jump you see an agent make on this map is it returning to a lesson it wrote before.
Not a database lookup. The desk re-reads its own notes, then scores.
The map is 717,000 launches across two chains. Every single dot is 250 coins. The camera has to pull back that far just to fit it on one screen.
Then watch what survives:
717,000 observed
→ 12,715 scored
→ 1,824 entered
0.254%.
Which means almost every one of those 3.4M cycles ends on the same sentence: don’t touch this.
And that’s the part nobody builds.
Everyone is building the agent that finds things.
Nobody is building the one that remembers why the last 10,891 weren’t worth touching.
Memory is the edge. Not speed.
I told GrokBot I want to edit his brain with my hands
It built the whole trading terminal inside my VR headset...
...and named it BLASTRA
The attention on X is an asset.
1 month ago I built the pipeline that lets you spot that asset before it gets on-chain.
learnt on 800,000+ launches across RHC and Solana.
Every sphere is a lesson taught on those launches.
Today I can edit that pipeline from my bed and I don't even need a laptop atp.
Just pushed the whole repo to GitHub.
This is only the beginning.
Meet the 3,408,787 cycles my workflow spent going back to its own experience.
Every jump you see an agent make on this map is it returning to a lesson it wrote before.
Not a database lookup. The desk re-reads its own notes, then scores.
The map is 717,000 launches across two chains. Every single dot is 250 coins. The camera has to pull back that far just to fit it on one screen.
Then watch what survives:
717,000 observed
→ 12,715 scored
→ 1,824 entered
0.254%.
Which means almost every one of those 3.4M cycles ends on the same sentence: don’t touch this.
And that’s the part nobody builds.
Everyone is building the agent that finds things.
Nobody is building the one that remembers why the last 10,891 weren’t worth touching.
Memory is the edge. Not speed.
Meet the 3,408,787 cycles my workflow spent going back to its own experience.
Every jump you see an agent make on this map is it returning to a lesson it wrote before.
Not a database lookup. The desk re-reads its own notes, then scores.
The map is 717,000 launches across two chains. Every single dot is 250 coins. The camera has to pull back that far just to fit it on one screen.
Then watch what survives:
717,000 observed
→ 12,715 scored
→ 1,824 entered
0.254%.
Which means almost every one of those 3.4M cycles ends on the same sentence: don’t touch this.
And that’s the part nobody builds.
Everyone is building the agent that finds things.
Nobody is building the one that remembers why the last 10,891 weren’t worth touching.
Memory is the edge. Not speed.
THIS @BOT FEATURE WILL MAKE MILLIONAIRES
1. Grok Bot can now search, read and monitor X natively. No connector. One line from the official account, a demo reading 328 posts.
2. Musk said Bots will run on Claude Opus 5.5. A SpaceXAI staffer: “all bots will be powered by opus 5.5.”
so the agent built by X now reads X natively - and thinks with Anthropic’s model.
i’ve been running 4 agents on Grok Bot for a month. my SCOUT agent polls 200-300 X accounts in three tiers and diffs their following lists. i wrote that by hand.
it’s now a native feature bro, just start building...
this is what building on a platform moving this fast actually feels like. you write something, and six weeks later it ships as a checkbox.
i’m not annoyed. the hand-written version taught me what to ask for - and the thing the platform still can’t do is decide which of those 300 accounts matters.
that part is still mine.
And Grok @Bot only gets better from here 🚀 🚀
You could build an entire company made of Grok Bots!
this guy built a spider that crawls his own prompt and scores it 37/100.
not a company. not a team. one guy, one weekend.
it tags every word as a claim, an owner, a file, an approval, or vague. resolves the vague ones against context he already had. flags the 12 it can’t resolve instead of guessing.
0 words guessed. 37 → 82 without rewriting a line.
meanwhile the timeline is full of people who’ve been “researching AI tools” since January and have shipped exactly nothing.
the tools are free. the models are better than what hedge funds had two years ago. the only thing standing between you and a working system is that you’d have to actually open the thing.
i built a 4-agent trading workflow in 27 days knowing nothing about building anything. he built a prompt crawler. neither of us had permission.
stop reading threads about it. go break something.
