Olivia
Online
Roxten
@0xRoxten
AI engineering. Agents, systems.
Joined March 2025
91 Following
113 Followers
93 Posts
This is f*cking gold..
the bot running this company has never done a single task. it has 4 tools, and every one of them is a decision
it's a test run of a grok bot company - 5 bots do the work, jev decides who goes next
10:00 one line goes in: "find the 10 most important AI-agent developments"
14:31 out comes a sourced article, an seo brief, visuals and 5 sales emails
the research, the writing, the briefs and the outreach - all handed off
my favourite part: cooper brought 10 developments and jev threw 1 claim straight back. no source, no article
the 5 emails are written, personalized and ready. sent: 0
how to use the video:
> follow the little bot from objective to guard
> pause at 10:47, the line where jev says verify_more
> read the very last line of the log
full run below ↓
This is f*cking gold..
the bot running this company has never done a single task. it has 4 tools, and every one of them is a decision
it's a test run of a grok bot company - 5 bots do the work, jev decides who goes next
10:00 one line goes in: "find the 10 most important AI-agent developments"
14:31 out comes a sourced article, an seo brief, visuals and 5 sales emails
the research, the writing, the briefs and the outreach - all handed off
my favourite part: cooper brought 10 developments and jev threw 1 claim straight back. no source, no article
the 5 emails are written, personalized and ready. sent: 0
how to use the video:
> follow the little bot from objective to guard
> pause at 10:47, the line where jev says verify_more
> read the very last line of the log
full run below ↓
https://t.co/Fg3HoC67bw
220 handoffs between 9 grok bots in one week, and the human shows up in the log exactly 3 times
it's a grok bot build loop drawn as a heat map - every dot is one handoff, red is find, white is build, purple is launch
monday 10:00 is the loudest hour, 31.6 handoffs. by friday it's a few purple dots
the 3 human lines: authorized on monday, answered on tuesday, approve on friday
the trend digging, the complaint sweeps, the spec and the launch video - all handed off
my favourite part: the whole week is a wall of dots, and the one that actually ships it is yours
how to use the video:
> watch the colour move from red to purple
> pause on the log at the bottom
> count the lines that start with [you]
full loop below ↓
220 handoffs between 9 grok bots in one week, and the human shows up in the log exactly 3 times
it's a grok bot build loop drawn as a heat map - every dot is one handoff, red is find, white is build, purple is launch
monday 10:00 is the loudest hour, 31.6 handoffs. by friday it's a few purple dots
the 3 human lines: authorized on monday, answered on tuesday, approve on friday
the trend digging, the complaint sweeps, the spec and the launch video - all handed off
my favourite part: the whole week is a wall of dots, and the one that actually ships it is yours
how to use the video:
> watch the colour move from red to purple
> pause on the log at the bottom
> count the lines that start with [you]
full loop below ↓
https://t.co/sVlvM5N7Xb
one copy-paste prompt sets up 9 grok bots for a weekly build loop, and the only thing it leaves to you is the decisions
it's the "build anything" setup from a grok bot course note - every bot gets one job, its own cloud computer and its own file
scout, pain miner, competitor mapper, strategy lead, spec writer, builder, video director and two more
grok bot now comes with the $30 x premium plan
you copy the prompt, point it at a folder and say "run week 2026-42"
the monday trend digging, the spec writing and the launch script - all handed off
my favourite line in it: "blocked? it writes BLOCKED and pings you." no bot guesses
how to use the video:
> watch the log fill up from 09:00
> pause at 17:48, where one bot stops and pings its human
> copy the prompt and run your own week
full prompt below ↓
one copy-paste prompt sets up 9 grok bots for a weekly build loop, and the only thing it leaves to you is the decisions
it's the "build anything" setup from a grok bot course note - every bot gets one job, its own cloud computer and its own file
scout, pain miner, competitor mapper, strategy lead, spec writer, builder, video director and two more
grok bot now comes with the $30 x premium plan
you copy the prompt, point it at a folder and say "run week 2026-42"
the monday trend digging, the spec writing and the launch script - all handed off
my favourite line in it: "blocked? it writes BLOCKED and pings you." no bot guesses
how to use the video:
> watch the log fill up from 09:00
> pause at 17:48, where one bot stops and pings its human
> copy the prompt and run your own week
full prompt below ↓
https://t.co/sVlvM5N7Xb
a grok bot ran a 3-step skill, hit a case it couldn't see, and added a 4th step to itself. the catch is it wrote down why before anyone noticed.
it's from the skill file, a grok bot breakdown: no nodes to wire, you describe the job and the bot asks its own follow-ups.
run → doubt → note to self → skill v2 → back to the next run
"i can't always tell if a post is already liked." - the bot
v1 was open chrome, go to the post, like it. v2 is still four boxes, but one of them is new, and the bot put it there.
a skill is a file that gets edited by the thing using it. the more it runs, the sharper it gets.
the part that got me is the label on the note: "nobody asked." it's the line every good intern learns in their second month.
and the honest part: a skill that edits itself still needs someone to reread it once in a while.
the 4th step is on screen in the clip.
bookmark & watch today ↓
https://t.co/vRA1B0RK6H
At 09:05 a bot looked at a checkout bug, decided another bot was better at it, and gave the ticket away. 26 minutes later the fix was verified, and the bots stopped one step short:
"Fix verified on staging. Shipping to users is your call."
It's part 2 of a 10-step tutorial on Grok Bot, xAI's agent that gets its own cloud computer. Five specialists, one chief, one group chat.
objective → engineering bot → debug bot → chief → you
The objective landed at 09:02: ship the checkout fix before Friday. Three minutes later the engineering bot had reproduced it and filed ticket #412. One minute after that the debug bot had the root cause, a rounding error on the currency step, and a patch.
The rule is one line: if it can't be undone in under a minute, park it and ask. That's why the sales bot has 36 drafts queued and 0 sent.
Honestly, I go back and forth on this one. 29 minutes is fast, but the tutorial's own last step is a 15-minute weekly review, because routines nobody reads rot quietly.a grok bot watched someone do a 4-step report for 40 seconds and never needed the paragraph explaining it again. the catch is what it does when it hits a sign-in page.
it's part 1 of a 10-step grok bot breakdown: 5 bots, each on its own cloud computer, 8 plugins wired once for all of them.
errand → charter → plugins → login → show it once → routine → back to the errand
"40 s to show it once instead of writing it twice." - the breakdown
the four steps are grafana, sheets, notion, slack. read the numbers, find what dropped, paste it under the right heading, ping the lead if anything fell more than 15%.
the first errand is tiny on purpose: find a $84.20 dinner receipt, file it under october, 30 seconds for you to check.
the part that got me is the login wall. the bot stops, you sign in on its browser, click done, and it carries on in the same session. the password never touches the chat.
and the honest part: plugins connect at the account level, so the blast radius of one connection is every bot.
bookmark & watch today ↓
But I keep thinking about 09:05. One bot admitting another one is better at the job.
Would you let your bots reassign work to each other?
At 09:05 a bot looked at a checkout bug, decided another bot was better at it, and gave the ticket away. 26 minutes later the fix was verified, and the bots stopped one step short:
"Fix verified on staging. Shipping to users is your call."
It's part 2 of a 10-step tutorial on Grok Bot, xAI's agent that gets its own cloud computer. Five specialists, one chief, one group chat.
objective → engineering bot → debug bot → chief → you
The objective landed at 09:02: ship the checkout fix before Friday. Three minutes later the engineering bot had reproduced it and filed ticket #412. One minute after that the debug bot had the root cause, a rounding error on the currency step, and a patch.
The rule is one line: if it can't be undone in under a minute, park it and ask. That's why the sales bot has 36 drafts queued and 0 sent.
Honestly, I go back and forth on this one. 29 minutes is fast, but the tutorial's own last step is a 15-minute weekly review, because routines nobody reads rot quietly.a grok bot watched someone do a 4-step report for 40 seconds and never needed the paragraph explaining it again. the catch is what it does when it hits a sign-in page.
it's part 1 of a 10-step grok bot breakdown: 5 bots, each on its own cloud computer, 8 plugins wired once for all of them.
errand → charter → plugins → login → show it once → routine → back to the errand
"40 s to show it once instead of writing it twice." - the breakdown
the four steps are grafana, sheets, notion, slack. read the numbers, find what dropped, paste it under the right heading, ping the lead if anything fell more than 15%.
the first errand is tiny on purpose: find a $84.20 dinner receipt, file it under october, 30 seconds for you to check.
the part that got me is the login wall. the bot stops, you sign in on its browser, click done, and it carries on in the same session. the password never touches the chat.
and the honest part: plugins connect at the account level, so the blast radius of one connection is every bot.
bookmark & watch today ↓
But I keep thinking about 09:05. One bot admitting another one is better at the job.
Would you let your bots reassign work to each other?
https://t.co/eDBs67VMq2
two bots got one line - "build a waitlist page" - and handed back a finished page and one approve button. 3 CI runs, 0 passwords typed in chat, 1 tap from you. the catch is who you're allowed to talk to.
it's the Handoff File, part 3 of a bot team breakdown: a chief-of-staff bot called Klaus, an engineer bot called Dev, and CI.
"build a waitlist page" → Klaus reads the other bots' job descriptions → Dev builds it on its own computer → red, red, green → Klaus checks it against the brand → "ready for your approval"
"you don't write a prompt - you hire an employee and give it a computer and a role." - the handoff file
the part that got me is that you only ever talk to Klaus. the bots find each other by reading each other's job descriptions, and memory sits in 3 layers, so when you delete one, what it learned about you stays with the team.
about $200 a month on a top-tier plan. Dev runs Cursor CLI on its own machine.
bookmark & watch today ↓
https://t.co/4qTQfdoerp
Anthropic's Applied AI team gave Claude a Swedish car insurance claim in front of a room of developers. on the first try it decided this was a skiing accident. the fix starts 4 minutes in.
it's Prompting 101 from Code w/ Claude: Hannah Moran and Christian Ryan build one prompt live, version by version.
one simple prompt → a skiing accident → add the role and the tone → explain the form in the system prompt → read the boxes first, the sketch second → "vehicle B appears to be clearly at fault"
"If Claude can't tell which form is checked, we don't want Claude to take its best guess." - Hannah Moran
the part that got me is the order. a human wouldn't start with the drawing either. it's just boxes and lines until you've read the form.
17 hand-filled checkboxes, one sketch of two cars, temperature zero, and the new Claude 4 on stage.
bookmark & watch today ↓
Anthropic's Applied AI team gave Claude a Swedish car insurance claim in front of a room of developers. on the first try it decided this was a skiing accident. the fix starts 4 minutes in.
it's Prompting 101 from Code w/ Claude: Hannah Moran and Christian Ryan build one prompt live, version by version.
one simple prompt → a skiing accident → add the role and the tone → explain the form in the system prompt → read the boxes first, the sketch second → "vehicle B appears to be clearly at fault"
"If Claude can't tell which form is checked, we don't want Claude to take its best guess." - Hannah Moran
the part that got me is the order. a human wouldn't start with the drawing either. it's just boxes and lines until you've read the form.
17 hand-filled checkboxes, one sketch of two cars, temperature zero, and the new Claude 4 on stage.
bookmark & watch today ↓
https://t.co/WVztfCUH4u
Yann Dubois, now at OpenAI, showed a Stanford class GPT-4 beating itself 64.4% of the time. same model on both sides. he only added two words to one prompt - 3 minutes into part 7.
it's the last part of his CS229 lecture on building LLMs: how you even grade a chatbot.
humans rank chatbots on Chatbot Arena → too slow and expensive for daily work → so GPT-4 judges instead → 98% correlation with Arena, under 3 minutes, under $10
"If you ask it to be concise, it gets 20%." - Yann Dubois
i laughed at how small the change is. 50% by definition, then two words, then 64.4%. his team had to bring in causal inference just to stop the judge from being fooled.
then he gets to GPUs. even with $10 million, he says, you can't buy the best ones right now.
bookmark & watch today ↓
Yann Dubois, now at OpenAI, showed a Stanford class GPT-4 beating itself 64.4% of the time. same model on both sides. he only added two words to one prompt - 3 minutes into part 7.
it's the last part of his CS229 lecture on building LLMs: how you even grade a chatbot.
humans rank chatbots on Chatbot Arena → too slow and expensive for daily work → so GPT-4 judges instead → 98% correlation with Arena, under 3 minutes, under $10
"If you ask it to be concise, it gets 20%." - Yann Dubois
i laughed at how small the change is. 50% by definition, then two words, then 64.4%. his team had to bring in causal inference just to stop the judge from being fooled.
then he gets to GPUs. even with $10 million, he says, you can't buy the best ones right now.
bookmark & watch today ↓
https://t.co/WVztfCUH4u
Yann Dubois, now at OpenAI, showed a Stanford class two answers to "tell me about self-driving cars." he admits he hadn't even read them carefully. he still guarantees which one most people pick - 10 minutes into part 6.
it's his CS229 lecture on building LLMs, the part on RLHF.
both answers are fine → raters go for the same one anyway → a reward model learns their taste → RLHF trains on it, round after round
"If you've ever been annoyed at ChatGPT answering you super long sentences, this is because of RLHF." - Yann Dubois
i laughed at the next bit. his team of five spent three hours agreeing on labeling rules, labeled the data themselves, and got 67-68%. the experts barely beat the crowd.
so they handed the rating to LLMs. roughly 50x cheaper than the $300 per 1,000 on Mechanical Turk.
bookmark & watch today ↓
https://t.co/WVztfCUH4u
Yann Dubois, now at OpenAI, told a Stanford class the step that turned GPT-3 into ChatGPT might also be why it makes things up - he gets to it 10 minutes into part 5.
it's his CS229 lecture on building LLMs, the part on post-training.
GPT-3 answers a question with more questions → humans write the answers → fine-tune → ChatGPT → but the model is told to say things it never saw
"Hallucination might be caused by this SFT." - Yann Dubois
the part that got me is how his team built Alpaca. 175 questions written by humans, an older OpenAI model wrote 52,000 more, and they fine-tuned LLaMA 7B on all of it.
he also says going from 2,000 to 32,000 human examples barely helps. the model already knows. you're just picking which internet user it should sound like.
bookmark & watch today ↓
Yann Dubois, now at OpenAI, told a Stanford class the step that turned GPT-3 into ChatGPT might also be why it makes things up - he gets to it 10 minutes into part 5.
it's his CS229 lecture on building LLMs, the part on post-training.
GPT-3 answers a question with more questions → humans write the answers → fine-tune → ChatGPT → but the model is told to say things it never saw
"Hallucination might be caused by this SFT." - Yann Dubois
the part that got me is how his team built Alpaca. 175 questions written by humans, an older OpenAI model wrote 52,000 more, and they fine-tuned LLaMA 7B on all of it.
he also says going from 2,000 to 32,000 human examples barely helps. the model already knows. you're just picking which internet user it should sound like.
bookmark & watch today ↓
https://t.co/WVztfCUH4u
Yann Dubois handed a Stanford class 10,000 imaginary GPUs for a month and asked what they'd train - with the old method, the model you ship got just one day of it. it's the opening of part 4.
it's part 4 of his CS229 lecture on building LLMs, the section on scaling laws.
old way: 30 models for one day each → pick the best → it only ever trained for 1 day
new way: 3 days of small models → fit the curve → "train the final huge model for 27 days instead of just one day"
"Do the simple things, do it well, scale them." - Yann Dubois
the part that got me is what the curves say about architecture. most small tweaks, he says, only shift the starting point a little. train 10 hours longer and the difference is gone.
i'd never heard the second number before: Chinchilla says 20 tokens per parameter, but if you care about serving it cheaply, it's closer to 150.
https://t.co/WVztfCUH4u
Yann Dubois, now at OpenAI, did the math on Llama 3 405B for a Stanford class: about $75 million to train - and his guess is Meta stopped 2x short of the line that triggers government scrutiny on purpose, 11 minutes into part 4.
it's his CS229 lecture on building LLMs, the back-of-the-envelope section.
405B parameters × 15.6T tokens → 3.8e25 FLOPs → 16,000 H100s for about 70 days → 26 million GPU hours → "I'm probably off by like 10 million, but that's kind of right ballpark"
"They really went right below this to not have special scrutiny." - Yann Dubois
the part that got me is how boring the biggest line is. renting the GPUs at $2 an hour is about $52 million. the people are about $25 million.
he also says the carbon is about 4,000 tons, roughly 2,000 round trips from JFK to London.
bookmark & watch today ↓
Yann Dubois, now at OpenAI, did the math on Llama 3 405B for a Stanford class: about $75 million to train - and his guess is Meta stopped 2x short of the line that triggers government scrutiny on purpose, 11 minutes into part 4.
it's his CS229 lecture on building LLMs, the back-of-the-envelope section.
405B parameters × 15.6T tokens → 3.8e25 FLOPs → 16,000 H100s for about 70 days → 26 million GPU hours → "I'm probably off by like 10 million, but that's kind of right ballpark"
"They really went right below this to not have special scrutiny." - Yann Dubois
the part that got me is how boring the biggest line is. renting the GPUs at $2 an hour is about $52 million. the people are about $25 million.
he also says the carbon is about 4,000 tons, roughly 2,000 round trips from JFK to London.
bookmark & watch today ↓
https://t.co/WVztfCUH4u
Last Seen Users on Sotwe
evli gizliyim
Seen from Turkey
PAWGBouncer ❄️🐰
Seen from Ireland
JT Toney
Seen from Latvia
Angah1391
Seen from Malaysia
sh
Seen from Malaysia
Politically Homeless
Seen from United States
cuck captions
Seen from United States
Goodbro365 Dot Com
Seen from Ireland
𝗕𝗔𝗡𝗡𝗘𝗗 𝗦𝗧𝗥𝗘𝗘𝗧 𝗦𝗘𝗫 🔞
Seen from Pakistan
😈Catfish Fingoddess_ kat 🫰😈
Seen from United States
Most Popular Users
Elon Musk
@elonmusk
241.5M followers
Barack Obama
@barackobama
118.8M followers
Cristiano Ronaldo
@cristiano
114.6M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98.7M followers
NASA
@nasa
92.3M followers
Justin Bieber
@justinbieber
91.8M followers
KATY PERRY
@katyperry
90M followers
Taylor Swift
@taylorswift13
84M followers
Lady Gaga
@ladygaga
75.5M followers
Virat Kohli
@imvkohli
73.5M followers
Kim Kardashian
@kimkardashian
70.9M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
66.5M followers
Bill Gates
@billgates
65.3M followers
Selena Gomez
@selenagomez
63.1M followers
The Ellen Show
@theellenshow
62.2M followers
CNN
@cnn
61.7M followers
X
@x
60.7M followers