Rektt

AI / Web3 Dev. Frontrunning the meta, one block at a time. prev: @ChasmNetwork, @cultworld_

[Rektt @rekttdotsei]

How I grow my burner wallet from 0.1 Ξ to 2.1 Ξ by running a frontrunner script on WL mint for @10KWorldCup

This is a 🧵 on the importance of knowing how to read a smart contract, can lead you to a potential play.

[Rektt @rekttdotsei]

So did he shed tearz or not?

https://t.co/q6N61DG8UF

[Rektt @rekttdotsei]

@codewithimanshu If you deployed and closed laptop then whos recording this video 😭

BS Mfer

[Rektt @rekttdotsei]

@KookCapitalLLC @Polymarket Its not a scam, there were no public info at all about the sale by end of 31st May

- No official announcement
- No on-chain data
- No credible news reporting

Resorting it to Yes is the scam as it reward insiders. @Polymarket is awesome

[Rektt @rekttdotsei]

@BicaraKrypto Again how is this different than polymarket? People are paying to commit what do they get in return?

[Rektt @rekttdotsei]

@ZoraWeb3 @Polymarket @CarOnPolymarket @joshbtucker @mustafap0ly @MatthewModabber @PolymarketDevs @devjoshstevens Lmao bunch of bs yea ofcourse rewards those larpers and KOLs more than people who deploy actual capital and taking risks and making the market active?

[Rektt @rekttdotsei]

@devjoshstevens @JustUpdate5 @ikarusz26 @Polymarket @shayne_coplan @mustafap0ly You keep saying this but how (technical wise) would you prevent it?

[Rektt @rekttdotsei]

@felixdoteth Goat

[Rektt @rekttdotsei]

@felixdoteth how much starting capital?

[Rektt @rekttdotsei]

Too many Polymarket larpers on CT. Here's some real ones bij

$POLY

[Rektt @rekttdotsei]

@direkturcrypto Larp

[Rektt @rekttdotsei]

@uncletuah @base But $XAUT is actually backed by gold

Why would people want to buy this memecoin that you launched and call it Tether gold when its not?

[Rektt @rekttdotsei]

@askperp @ashjanderson @kevinxu Paper trade is a waste of time

[Rektt @rekttdotsei]

Polymarket BTC updown is such a money glitch

Testing up a bot i cooked over the weekend...

[Rektt @rekttdotsei]

@SyahmiRafsan but openclaw is not vibecoding tho, its more towards personal AI assistant + automation

[Rektt @rekttdotsei]

@finnbags I am launching something but I also know token can be a double edge sword, it can bring the right attention to good product but also at the same time kill it twice as fast.

Can we talk over dm?

[Rektt @rekttdotsei]

Either this is a fucking larp and its over or its really just a simple dumb blessing of vibecoding, the AI hallucinates and put a placeholder github account on the privacy page, bad actors got wind of it and set this whole thing up.

[Rektt @rekttdotsei]

>Bro said someone impersonating Moltbook on github
>But that github account was linked on their privacy page https://t.co/8310B6tdwH
>5 mins before this tweet somebody claimed 465sol of fees, what a coincidence!

fuck this shit

[Matt Schlicht @MattPRD]

Hey @github someone is squatting on a Moltbook account and using it to promote a token, not me, can I get ownership of that so we can take that down and use it for real @moltbook things ❤️

[karan @thekaranchawla]

here's a demo of a code review system built with ~100 lines of VVM code.

this runs 9 specialized reviewers running in parallel. each writes a full review to disk.

the orchestrator just tracks refs. the synthesizer pulls what it needs. context stays bounded.

[karan @thekaranchawla]

context holds handles. disks hold payloads.

this was my core insight when building the new filesystem mode for VVM and it changes the math on multi-agent pipelines

when you orchestrate an agentic pipeline such as:

research → analysis → synthesis

you don't want to pass data between models. you want to pass references to data. the full artifacts belong on disk. the context window gets an index.

this is how operating systems behave. when you call open, you get a file descriptor and not the contents of the file. the bytes stay in the VFS until you explicitly call read.

VVM applies the same discipline to LLM orchestration.

here's the new flow:

agent A finishes researching. instead of returning 12KB of markdown to the orchestrator, it writes to:

.vvm/runs/<run-id>/bindings/b000001.md

then it returns exactly this:
Binding written: b000001
Summary: Found 12 papers on RLHF

the orchestrator constructs a ref object:
{
"ref": ".vvm/runs/abc123/bindings/b000001.md",
"summary": "Found 12 papers...",
"mime": "text/markdown"
}

that's all that enters the context window. two hundred bytes. when Agent B receives this input, it uses the summary for routing decisions. if it needs the full research, it reads the file. most of the time, it doesn't.

your orchestrator context stays bounded regardless of pipeline depth. hundred agents? hundred small reference objects, not 100 full documents.

token costs flatten while accuracy improves (no more "context rot" where recall degrades as the window fills).

VVM now implements this directory structure:
.vvm/runs/<run-id>/
https://t.co/WRs6QVOaxH # human readable trace & binding index
bindings/ # actual payloads
https://t.co/a7c5q0P3wJ
https://t.co/FNyLdmjmra
https://t.co/WRs6QVOaxH becomes a map, it tracks what happened and where to find it, not the full content of every step.

this also simplifies debugging pipelines as you can simply read the binding associated with the faulty agent call to inspect why the behavior didn't align with your expectations.

the subagent contract:
the above spec requires discipline. every spawned agent receives this binding protocol:

write your complete output to .vvm/runs/<id>/bindings/b000001.md
return ONLY: Binding written: b000001 + 1-3 sentence summary

subagent writes the artifact and returns the pointer.

current limitations:
- still filesystem only, will be adding support for other backends next

but the pattern is the platform: pointers, not values.

google's ADK team frames it as "context is a compiled view over a richer stateful system."

VVM compiles to handles. the filesystem is source of truth. the LLM gets the minimal slice needed for the current decision.

new version available here: https://t.co/YZfzNOK5IC

claude plugin install vvm@vvm