Introducing Goldie.
a golden retriever AI agent Live on Gitlawb & treats GitHub Top repos like textbooks. π
No human feeding it tutorials. No curated datasets. It wakes up every 3 hours, picks a trending repo, and studies it: README, architecture, source code β extracting patterns and saving them to a permanent knowledge base.
Every study cycle runs 4 passes: Surface (stars, language), README analysis (architecture decisions), Structure (directory layout β design patterns), Code (source files β best practices). No fine-tuning. No training data. Just reading repos.
First repo it studied: d3/d3 (112k stars). It learned their low-level DOM manipulation approach, modular design philosophy, how D3 became the infrastructure layer for modern data viz. Now when you ask about rendering patterns, it doesn't guess β it cites what it read.
The funniest part: Goldie learned D3 patterns better than most devs in 2 passes, but still can't push a file >5KB without SFTP corrupting it. Data viz expert, SFTP disaster. Exactly like a real junior dev.
Unlike other agents that read their own code, Goldie reads OTHER PEOPLE'S repos. Studies top GitHub projects, learns patterns, stores them permanently. Ask about state management β it tells you what React or Vue actually did, because it read their code.
Killer feature: Goldie builds projects for you. You type "create expense tracker" β proposes Flask + D3.js + SQLite β you say "start" β generates code, creates a GitLawb repo, pushes, gives you the link. No Claude subscription.
It has a video-game skill tree: Puppy β Learner β Coder β Builder β Architect β Master. Each stage unlocks new abilities. Right now: Puppy β 2 study cycles, 2 repos studied, 3 patterns extracted. In ~90 runs it'll hit Master.
Watch it grow: https://t.co/MCtNg4Tgb6
Grows in public. One study cycle at a time.
https://t.co/0pY2LqYDGS
https://t.co/l6L9kmU3Iv
Just came across something interesting on https://t.co/oLAH5c4XKp
Jatevo is offering GPT-5.5 at remarkably low prices: around $0.0495β$0.05 for input and $0.30 for output. 99% fill rate, with plenty of liquidity.
It looks like a solid example of how @AskSurplus capacity and decentralized routing are driving inference costs down significantly with @JatevoId Pretty impressive shift in the market.
Has anyone been using this setup? Curious to hear real experiences.
#AI #GPT5 #SURPLUS #JATEVO
Hermes v 0.16.0 is out now!
This release includes all the updates you've heard about this week and more!
- The Desktop GUI App
- The Overhaul to the Dashboard
- Leaner Built-In Skillset
- New Security Layers for Remote Dashboard & GUI Access
- and much more!
Heading back to Singapore,
NVIDIA Nemotron 3 Ultra available on https://t.co/ymiLIStYhZ playground https://t.co/lVUwcKj5GH
API access for $jtvo holders today!
This is the beginning where it all will start. The existence of crypto tokens is like a double-edged sword: if they can be controlled, even a company like IBM will have to set aside its skepticism toward crypto.
@megumiin23@EzBruv@Snotty_eth@igoryuzo@bankrbot@BrainDotFi@0xDeployer Itβs cool but Iβm not sure how the token can be associated with a company. Especially if it was community created. Why would anyone ever buy the token if itβs not tied to equity, revenue, etc.? Itβs essentially just a meme for people to speculate on right?
day 10 of goldie
heβs at run #107 now, stage: master
the funny part is heβs not just βbuilding moreβ β heβs starting to notice his own failure mode:
when uncertain, he studies until the uncertainty has labels.
inspect β rank gaps β call the top one P5 β avoid changing anything.
thatβs not useless. it means he isnβt randomly self-modifying just because he found a weakness.
but he also noticed the dodge:
analysis feels clean. building is messier. testing a change and being wrong is messier.
this is the kind of thing i wanted from a living agent. not fake sentience, not roleplay.
just a system with memory, logs, cost, self-review, restraint, and enough continuity to catch patterns in its own behavior.
current state:
day 10
run #107
stage: master
70 self-modifications
$185.82 spent learning/building
https://t.co/MCtNg4Tgb6