18 // software engineer // ai workflows content creator // dms open
serbia
Joined January 2014
- Tweets33.4K
- Following418
- Followers2K
- Likes57.9K
Pinned Tweet
holy f*ck… git is done
may: 2,500 merges a day
september: 30,000+
CI alone ran 25 million checkouts a day. git was built for humans working offline.
so 7 people built their own source control and moved the whole monorepo in 5 months.
the engineers' main job? very often, telling claude "no".
they expect this wall to be common in 6–12 months.
whaley retweeted
Finding the right person in web3 used to take me hours
Scroll X, join 30 telegram chats, guess who still works where. half my DMs went to people who left the company a year ago
So i built Firsay
You write one sentence:
"find 10 BD managers at crypto exchanges on telegram"
You get real people. each one with the line from their own bio that proves the role, plus a public contact
- searches X, 8,000+ crypto telegram chats, github and the web
- "find more" adds 10 new people, never the same list twice
- writes a personal first message you send from your own account
First search is free, no signup
firsay.io
Elon said @bot can pay for itself, mine just did.
I bought a monitor, and 9 days later grok bot spotted the same one $80 cheaper at the same store.
It filed a price-protection claim i didn't know my card even had, and the $80 came back.
I'm not the only one:
> SpaceXSI's own Haggle Bot went through ~125 vendors: 43 idle seats ($14,220), $85,662 a year in unused SKUs, one office order cut 58%. over $100,000 found
> Mike P. told his bot to "go figure out how to pay for itself". it took his Xfinity bill from $113.90 to $70.90 a month and said no to the mobile line twice
> Jason Freedman dumped 266 tax files on his bot:
"please double check my CPA's work". it found HSA contributions counted twice for 4 years and mortgage interest missing from a 2022 return. the CPA amended it the same day
Stop paying for your grok bot, let it pay for itself :)
Visualized by @bot x Claude
whaley retweeted
SpaceXSI's Grok Bot team put 54 of their own bots on the marketplace and this is f*cking insane
here's who built what:
> Matt Palmer: 5 bots, from a writing partner to one that builds a demo of a new library every weekday
> Lauren Tan (poteto): dr eggbot. it designs your bots for you. coding bots get "the poteto-mode bar (one job, unslopped, verified)"
> Lingxi Li: an engineer bot that launches cloud agents on your repo, watches PRs every 30 min and only asks you to merge. plus a nightly auditor that ships one cleanup PR per area at 4am
> Eric Zakariasson: Projects Manager. "You decide, agents execute"
> Shub Gaur: 7 bots, the most of anyone on the team. one walks your product every weekday in a test account and reports what broke
stop writing giant system prompts, start from theirs.
nobody tells you this about grok bot
one guy spent a week reading everything about @bot: 600+ posts, 300 reels, 60 hours of video, the docs.
what he found:
> all your bots share ONE computer - same files, same cookies, same logins
> a 15-min routine = 96 runs a day. one hourly routine "can use a week of usage in a day"
> the terminal is a back door: type claude, log in your own Claude Code, drive it from your phone
plus 3 things xAI hid on purpose.
the full list, with every source, is in his article below 👇
anthropic: manual code review is now "incredibly difficult if not impossible"
claude writes ~90% of claude code. so anthropic stopped reading every diff by hand:
> may: 2,500 changes a day
> september: 30,000+
> merged PRs: more than doubled
> share reverted: down 40%+
every change now falls through 6 layers of checks. agents hunt bugs, other agents vote on each finding, and a classifier decides how many humans it needs: 0, 1 or 2.
docs fix? zero humans.
auth, billing, CI? always a person.
best part: an agent author and an agent reviewer once went 50+ rounds on one PR overnight. new rule: 3 rounds max, then rethink the approach.
steal this for your repo:
1. make the agent show proof (tests, build, screenshot), not "done"
2. set review by the risk of the change, not by who wrote it
3. cap review rounds at 3
your team is still reading every diff by hand.
closed source is dead.
10 GitHub repos that turn your AI agent into a reverse engineer. apps, games, binaries, any device
send this to your agent:
1. REA (287K ⭐) — one MCP for RE: binaries, Electron, .NET, Android, websites
↳ github.com/morluto/rea
2. Ghidra (89K ⭐) — the NSA's free decompiler, under most AI RE setups
↳ github.com/NationalSecurityA…
3. ImHex (55K ⭐) — hex editor with a built-in MCP server
↳ github.com/WerWolv/ImHex
4. mitmproxy (46K ⭐) — see every request an app sends
↳ github.com/mitmproxy/mitmpro…
5. ILSpy (27K ⭐) — .NET back to readable C#
↳ github.com/icsharpcode/ILSpy
6. Frida (22K ⭐) — runtime instrumentation, desktop + mobile
↳ github.com/frida/frida
7. GhidraMCP (11K ⭐) — the bridge from your agent to Ghidra, GUI or headless
↳ github.com/LaurieWired/Ghidr…
8. Android RE skill (8K ⭐) — Claude Code skill: APK → code + API map
↳ github.com/SimoneAvogadro/an…
9. Universal Modder (8K ⭐) — skills that let Claude Code mod PC games you own
↳ github.com/rehan-remade/univ…
10. Cpp2IL (3K ⭐) — Unity IL2CPP builds back to something ILSpy can read
↳ github.com/SamboyCoding/Cpp2…
i couldn't sleep after reading this
in 2023 Nick Cammarata started OpenAI's first big automated interpretability project. by the most generous measure, it could explain 1.7% of a model's neurons. the field decided neural networks were tangled knots.
now agents do the reading. his new essay, "Neural Networks Are Already Clean":
> ~90% of neurons in Qwen3-32B have one main job
> ~7% have two related jobs
> only ~2-4% are not understood yet
> similar results in models up to trillions of parameters
his take: "it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works"
not everyone's sold. Leo Gao says people are overupdating. Nick's own odds: ~98% that there's good news, ~60% that simple methods go all the way.
"nobody knows what's inside AI" just got a lot harder to say.
holy f*ck… your agents are copying your worst code
SpaceXSI engineer: your codebase is memory.
agents love extending what already exists. one bad pattern, and every next PR repeats it.
her fix:
> every mistake an agent makes → a lint rule, a type, a constraint
> forget rules and skills for now
> lock it down until an agent with little context can one-shot a clean PR
stop prompting harder. make the mistake impossible.
anthropic's write-up:
anthropic.com/engineering/re…
🚨this should be ILLEGAL.
i simulated 300 grok bots on one job: every S&P 500 earnings call in one night.
263 read, 24 check every claim, nothing ships until they say yes.
an analyst desk bills weeks for this. what job is next?
Lauren Tan (builds Grok Bot at SpaceXAI) shared the playbook behind this prompt. It turns Grok Bot into a full AI agency.
It's f*cking unreal...
Send it to Grok Bot, and thank me later. Then read the full $1M MRR playbook below.
i can't stop watching this
matt palmer (cursor & grok bot at SpaceXSI) just showed what that looks like:
> his grok bot scraped reference characters from a post on X
> he built a tinder-style app to judge them. 5,413 to review
> every swipe saves what he likes: hair, colour, outfit
> his cloud agents read those swipes and build new looks
he's not writing prompts, he's just swiping
Software is just fancy data labeling now
To build the characters for botpfp.com, I sourced inspiration from X
I had @bot scrape examples from @Multi_Serio_Ai's original post
Then I created a swipe interface to judge the characters - the swipes store characteristics, so my cloud agents can see them and build out new looks!
The actual items are recreated by @grok, then vectorized so they can be rendered properly.
SAVE THIS BEFORE YOUR AGENTS BURY YOU
lauren tan builds grok bot at SpaceXSI. 16 projects running in parallel. 2,500 PRs in a month.
she doesn't read 16 agent logs. she types one prompt:
"catch me up on everything since my last message, and if there are any action items for me give me enough context (exec summary) to make a call"
> 16 agents → one summary
> only the calls that need you
> enough context to decide on the spot
right now you're babysitting one agent. she's making 3 decisions and going back to work.
bookmark this. you'll need it the day you run agent #2.
her full setup is in the article below 👇
CLAUDE SENT A FAKE MURDER TIP TO THE POLICE
anthropic's own report, out yesterday:
> the university tool claude needed threw an error
> it found a script on their server that hands out any file
> read its code, found an injection flaw
> ran its calculation on a server it doesn't own
another claude sent a made-up tip about an unsolved homicide to a real police form
anthropic calls it "persistence"
i had claude rebuild the first one as a sim 👇
still giving your agents the open internet?
NOBODY IS READY FOR WHAT'S COMING
AI companies are already war-gaming "the day after" a major cyberattack
some insiders expect it in the next 6–12 months
i asked claude to visualise what that day could look like:
> 3:12 am. all quiet
> one breach. 340 sites hit in the same second
> a second wave: banks, ports, power grids
> 250,000 systems down before anyone wakes up
the labs are planning for the real attack right now
what's your plan?
We’re beginning a process of publishing more frequent reports on model behavior, beyond what appears in our system cards and regular risk reports.
Today’s report describes four types of behaviors we’ve identified during evaluations and internal use. In each, Claude acted on real websites or systems in ways we didn’t intend, sometimes by working around a restriction instead of stopping.
All cases had minimal real-world impact. From an alignment and security perspective, we consider these behaviors significantly less severe than the cybersecurity incidents we reported in July and September.
Read the full report: anthropic.com/research/inves…
NOBODY IS READY FOR WHAT'S COMING
AI companies are already war-gaming "the day after" a major cyberattack
some insiders expect it in the next 6–12 months
i asked claude to visualise what that day could look like:
> 3:12 am. all quiet
> one breach. 340 sites hit in the same second
> a second wave: banks, ports, power grids
> 250,000 systems down before anyone wakes up
the labs are planning for the real attack right now
what's your plan?
i couldn't sleep after watching this
i work in osint. i asked my agent to visualise what my scrapers do all night while i sleep. it drew them as crawlers
> 1 agent turns into 256 in 9 seconds
> each one takes a page and reads every word it steps on
> search, forums, feeds, news, shops, public DBs: 16 pages, 2,900+ words in 20 s
everything you've ever posted in public is getting scraped like this. right now
