Olivia
Online
ZaiGen
@Zaigennnn
AI, distilled. Lectures, quotes & guides that actually matter — daily
Joined January 2024
65 Following
102 Followers
913 Posts
🎁 lilDudes_ NFT GIVEAWAY 🎁
Got something special for the community
I have 7 GTD spots + 10 FCFS spots to give away for @lilDudes_!
YOU HAVE 24h
Want to get your chance? Here’s what you need to do:
✅ Follow @lilDudes_
❤️ Like & RT their pinned post
🔁 RT this giveaway post
Drop your wallets in comment
Good luck, Dudes! Let's get you onto the whitelist! 🫡
Winners will be announced here on X
Introducing “ lil Dudes ”
3333 lil Dudes on chain @inkonchain
Apply now >> https://t.co/gQ4fduQdNP
Five terminal tabs open on his laptop. Ten more sessions running on the web. One more kicked off from his phone before he's even out of bed.
That's not a power user showing off. That's Boris Cherny's actual workday — the person who built Claude Code and still leads the team behind it.
"I don't prompt Claude anymore. My job is to write loops."
Here's what that looks like in practice: each of those terminal sessions runs in its own separate git checkout — not a branch, a full checkout — so five agents can edit the same repo at once without stepping on each other. 10-20% of those sessions get abandoned when something goes sideways. At that scale, throwing away a few costs nothing.
The memory isn't a vector database or a fine-tune. It's a plain text file, CLAUDE.md, checked into git, that the whole team writes to. Every time Claude gets something wrong, the fix goes into the file so the mistake doesn't repeat — for anyone on the team, not just him.
The part that matters most has nothing to do with prompting at all: verification. No gate, and the agent is just grading its own homework. A real one — a test suite, a type check, an actual browser checking the UI — and the quality jumps by a factor of 2 to 3, by his own account.
His advice on token spend cuts against the instinct to economize: don't cap it early, because the constraint that actually throttles teams isn't cost, it's how fast engineers learn to use the tokens well.
Curious how many people run more than one Claude Code session at a time versus still treating it like a single chat window.
Bookmark this — the full breakdown of his setup, tab by tab, is below.
Five terminal tabs open on his laptop. Ten more sessions running on the web. One more kicked off from his phone before he's even out of bed.
That's not a power user showing off. That's Boris Cherny's actual workday — the person who built Claude Code and still leads the team behind it.
"I don't prompt Claude anymore. My job is to write loops."
Here's what that looks like in practice: each of those terminal sessions runs in its own separate git checkout — not a branch, a full checkout — so five agents can edit the same repo at once without stepping on each other. 10-20% of those sessions get abandoned when something goes sideways. At that scale, throwing away a few costs nothing.
The memory isn't a vector database or a fine-tune. It's a plain text file, CLAUDE.md, checked into git, that the whole team writes to. Every time Claude gets something wrong, the fix goes into the file so the mistake doesn't repeat — for anyone on the team, not just him.
The part that matters most has nothing to do with prompting at all: verification. No gate, and the agent is just grading its own homework. A real one — a test suite, a type check, an actual browser checking the UI — and the quality jumps by a factor of 2 to 3, by his own account.
His advice on token spend cuts against the instinct to economize: don't cap it early, because the constraint that actually throttles teams isn't cost, it's how fast engineers learn to use the tokens well.
Curious how many people run more than one Claude Code session at a time versus still treating it like a single chat window.
Bookmark this — the full breakdown of his setup, tab by tab, is below.
https://t.co/W9DmIL3ZVP
Andrej Karpathy just explained the exact loop behind his 700-experiment overnight run, on No Priors with Sarah Guo:
00:00 — the "16 hours a day" shift from writing code to directing agents
15:51 — why AutoResearch exists: the loop that found what 2 decades of hand-tuning missed
22:45 — the actual skill that matters now: designing objective criteria, not writing better prompts
32:30 — where he thinks this goes next: untrusted swarms of agents collaborating like Folding@home
Most people still think "loop" means a prompt on a schedule.
This is Karpathy explaining why that's not even close — and why removing yourself as the bottleneck, not writing smarter instructions, is the actual unlock.
Worth more than most paid courses on agentic workflows right now.
Watch it today, then save the breakdown below.
Andrej Karpathy just explained the exact loop behind his 700-experiment overnight run, on No Priors with Sarah Guo:
00:00 — the "16 hours a day" shift from writing code to directing agents
15:51 — why AutoResearch exists: the loop that found what 2 decades of hand-tuning missed
22:45 — the actual skill that matters now: designing objective criteria, not writing better prompts
32:30 — where he thinks this goes next: untrusted swarms of agents collaborating like Folding@home
Most people still think "loop" means a prompt on a schedule.
This is Karpathy explaining why that's not even close — and why removing yourself as the bottleneck, not writing smarter instructions, is the actual unlock.
Worth more than most paid courses on agentic workflows right now.
Watch it today, then save the breakdown below.
https://t.co/BnNGOai7km
This is Barry Zhang and Mahesh Murag — the two Anthropic engineers who actually built the Skills system the post below is about.
Their core point: agents don't need to be smarter, they need domain expertise packaged as something reusable. That's literally what every skill in the list below is doing.
Anthropic actually maintains an official skills repo — https://t.co/zeYFtfxSkN — and a couple of these solve a genuinely specific problem.
Take frontend-design, installed 667K times per public usage trackers. Its actual documented mechanism: pick an aesthetic direction first — brutalist, minimalist, retro-futurist — before writing a single line of code. That's the real fix for the thing every Claude-generated site starts to look the same: purple gradients, Inter font, everywhere.
theme-factory ships with 10 preset professional themes, one command to apply to any artifact — slides, docs, landing pages — plus the option to generate a custom one on the fly.
Worth knowing before you install all ten: five of these (frontend-design, canvas-design, theme-factory, algorithmic-art, skill-creator) are official, straight from Anthropic's repo. The other five — systematic debugging, superpowers, file search, context optimization — are community projects that plug into the same skill system, including Jesse Vincent's well-known superpowers framework. Good tools, just a different trust level than the ones shipped by Anthropic itself.
The "3-4x faster iteration" and revenue numbers in the post are one person's experience, not a benchmark — worth testing on your own workflow before repeating the figure.
Curious which of these ten people are actually running daily versus just starred on GitHub and forgot about.
Bookmark this — the install links are below, easier to grab them now than dig through a thread later.
This is Barry Zhang and Mahesh Murag — the two Anthropic engineers who actually built the Skills system the post below is about.
Their core point: agents don't need to be smarter, they need domain expertise packaged as something reusable. That's literally what every skill in the list below is doing.
Anthropic actually maintains an official skills repo — https://t.co/zeYFtfxSkN — and a couple of these solve a genuinely specific problem.
Take frontend-design, installed 667K times per public usage trackers. Its actual documented mechanism: pick an aesthetic direction first — brutalist, minimalist, retro-futurist — before writing a single line of code. That's the real fix for the thing every Claude-generated site starts to look the same: purple gradients, Inter font, everywhere.
theme-factory ships with 10 preset professional themes, one command to apply to any artifact — slides, docs, landing pages — plus the option to generate a custom one on the fly.
Worth knowing before you install all ten: five of these (frontend-design, canvas-design, theme-factory, algorithmic-art, skill-creator) are official, straight from Anthropic's repo. The other five — systematic debugging, superpowers, file search, context optimization — are community projects that plug into the same skill system, including Jesse Vincent's well-known superpowers framework. Good tools, just a different trust level than the ones shipped by Anthropic itself.
The "3-4x faster iteration" and revenue numbers in the post are one person's experience, not a benchmark — worth testing on your own workflow before repeating the figure.
Curious which of these ten people are actually running daily versus just starred on GitHub and forgot about.
Bookmark this — the install links are below, easier to grab them now than dig through a thread later.
https://t.co/p4teS18nnJ
@crimson_moti That's the whole shift. Curious if you've run into a case where even a hook wasn't enough — where you needed to go a layer deeper?
https://t.co/3cSzN0f1QU
https://t.co/3cSzN0f1QU
$50 per million output tokens. 1 million token context window. 128,000 tokens of output per request. And on SWE-Bench Pro, it scores 80.3% against GPT-5.5's 58.6%.
That's Claude Fable 5 — Anthropic's first publicly available Mythos-class model, launched June 9, 2026, priced exactly double Opus 4.8 per token.
The clip above is Adel El Hallak, VP of Agentic AI at NVIDIA, making the point that applies just as much here, different vendor: an agent isn't just a model — it's a model plus the system engineered around it. The system is what compounds, not the model itself.
Here's the part the numbers don't capture: most people are using Fable 5 like a bigger Sonnet. Prompt it, get an answer, close the tab. That's paying Mythos-tier pricing for zero compound effect.
Anthropic's own Claude Code team put it plainly: "a verifier sub-agent tends to outperform self-critique with Fable 5." The model that wrote the code should not be the model grading it — that's the entire mechanism behind why it works for hours instead of minutes.
Safety classifiers decline under 5% of sessions — cybersecurity, biology, chemistry, model-distillation — and automatically fall back to Opus 4.8 when they trigger. That's the one number in this space I'd call fully confirmed against official docs; treat the more specific benchmark claims elsewhere as worth double-checking before you repeat them.
Curious how many people running Fable 5 right now have actually built a verifier loop versus just prompting it straight.
Bookmark this — the full 14-step system is what turns that $50/million into hours of unattended work instead of five minutes and a closed tab.
$50 per million output tokens. 1 million token context window. 128,000 tokens of output per request. And on SWE-Bench Pro, it scores 80.3% against GPT-5.5's 58.6%.
That's Claude Fable 5 — Anthropic's first publicly available Mythos-class model, launched June 9, 2026, priced exactly double Opus 4.8 per token.
The clip above is Adel El Hallak, VP of Agentic AI at NVIDIA, making the point that applies just as much here, different vendor: an agent isn't just a model — it's a model plus the system engineered around it. The system is what compounds, not the model itself.
Here's the part the numbers don't capture: most people are using Fable 5 like a bigger Sonnet. Prompt it, get an answer, close the tab. That's paying Mythos-tier pricing for zero compound effect.
Anthropic's own Claude Code team put it plainly: "a verifier sub-agent tends to outperform self-critique with Fable 5." The model that wrote the code should not be the model grading it — that's the entire mechanism behind why it works for hours instead of minutes.
Safety classifiers decline under 5% of sessions — cybersecurity, biology, chemistry, model-distillation — and automatically fall back to Opus 4.8 when they trigger. That's the one number in this space I'd call fully confirmed against official docs; treat the more specific benchmark claims elsewhere as worth double-checking before you repeat them.
Curious how many people running Fable 5 right now have actually built a verifier loop versus just prompting it straight.
Bookmark this — the full 14-step system is what turns that $50/million into hours of unattended work instead of five minutes and a closed tab.
https://t.co/hmpJtXuXCq
Last post covered what a loop actually is — goal, verify, iterate. Here's what happens when someone builds one that never stops running.
Hermes is a cloud agent that lives in your messenger instead of a chat tab you open and close. The part worth understanding isn't that it runs 24/7 — plenty of agents do that now. It's the self-improving loop underneath: a background process that watches your conversations and turns what it learns into permanent upgrades to its own memory and skills.
It doesn't analyze every message, because that would burn tokens for nothing. Instead it waits on two counters: a memory trigger every 10 prompts, a skill trigger every 10 tool-call iterations in a single turn — the idea being if the agent just fought through something for that many steps, that fight is worth turning into a reusable skill.
The part most people would skip building: a curator. Left alone, a system like this eventually produces hundreds of skills, some redundant, some just wrong. Once every 7 days, once the agent's been idle 2+ hours, the curator reviews the whole library — keeps what's accurate, fixes what's stale, merges duplicates, archives the rest. Anything untouched for 90 days gets moved out automatically.
The one line worth remembering if you build anything like this yourself: the agent that did the work is a bad judge of the work. Every serious version of this splits the builder from the checker.
Curious whether people think self-improving memory is actually a step forward, or just a more expensive way to accumulate agent hallucinations nobody reviews.
Bookmark this if you're mid-build on something similar — the trigger thresholds and curator cadence here are worth stealing outright.
Last post covered what a loop actually is — goal, verify, iterate. Here's what happens when someone builds one that never stops running.
Hermes is a cloud agent that lives in your messenger instead of a chat tab you open and close. The part worth understanding isn't that it runs 24/7 — plenty of agents do that now. It's the self-improving loop underneath: a background process that watches your conversations and turns what it learns into permanent upgrades to its own memory and skills.
It doesn't analyze every message, because that would burn tokens for nothing. Instead it waits on two counters: a memory trigger every 10 prompts, a skill trigger every 10 tool-call iterations in a single turn — the idea being if the agent just fought through something for that many steps, that fight is worth turning into a reusable skill.
The part most people would skip building: a curator. Left alone, a system like this eventually produces hundreds of skills, some redundant, some just wrong. Once every 7 days, once the agent's been idle 2+ hours, the curator reviews the whole library — keeps what's accurate, fixes what's stale, merges duplicates, archives the rest. Anything untouched for 90 days gets moved out automatically.
The one line worth remembering if you build anything like this yourself: the agent that did the work is a bad judge of the work. Every serious version of this splits the builder from the checker.
Curious whether people think self-improving memory is actually a step forward, or just a more expensive way to accumulate agent hallucinations nobody reviews.
Bookmark this if you're mid-build on something similar — the trigger thresholds and curator cadence here are worth stealing outright.
https://t.co/rua4N4akAh
The entire "AI engineer vs ML researcher" distinction below isn't something we came up with. It's Chip Huyen's — the person who coined "AI engineering" as its own discipline, after surveying people actually building on foundation models and finding that's what they called themselves. The clip above is her explaining it directly.
Here's what that distinction means in practice: the highest-paid building role in tech right now doesn't check your diploma. It checks what you've shipped.
An AI engineer isn't the person training models from scratch — that's a research role, small slice of the market, genuinely needs the math and the PhD. An AI engineer is the person who takes a model that already exists and builds something real with it: a support bot that resolves the actual ticket, a search tool that finds the answer buried in ten thousand documents, an agent that runs a workflow without a human babysitting it.
Hiring managers say it plainly: they've watched self-taught engineers outperform PhD holders, because shipping is a different skill than studying. The credential gate most people assume exists has mostly stopped being enforced.
The path below breaks it into 12 months, 6 phases, ending with 3 shipped projects: a RAG system with real evaluation, a multi-agent system solving an actual problem, and a deployed system with monitoring. That portfolio is the whole interview — not reciting theory, describing what you built.
For context on the incentive: Glassdoor's June 2026 data puts average AI engineer pay at $143,518, with entry-level roles starting around $100K.
Curious how many people reading this have already shipped something, versus still waiting to feel "qualified enough" to start.
Bookmark this whole thread if you're not starting today — the clip clears up whether the path is even right for you, and the 12-month breakdown below is what you'll want once you are.
The entire "AI engineer vs ML researcher" distinction below isn't something we came up with. It's Chip Huyen's — the person who coined "AI engineering" as its own discipline, after surveying people actually building on foundation models and finding that's what they called themselves. The clip above is her explaining it directly.
Here's what that distinction means in practice: the highest-paid building role in tech right now doesn't check your diploma. It checks what you've shipped.
An AI engineer isn't the person training models from scratch — that's a research role, small slice of the market, genuinely needs the math and the PhD. An AI engineer is the person who takes a model that already exists and builds something real with it: a support bot that resolves the actual ticket, a search tool that finds the answer buried in ten thousand documents, an agent that runs a workflow without a human babysitting it.
Hiring managers say it plainly: they've watched self-taught engineers outperform PhD holders, because shipping is a different skill than studying. The credential gate most people assume exists has mostly stopped being enforced.
The path below breaks it into 12 months, 6 phases, ending with 3 shipped projects: a RAG system with real evaluation, a multi-agent system solving an actual problem, and a deployed system with monitoring. That portfolio is the whole interview — not reciting theory, describing what you built.
For context on the incentive: Glassdoor's June 2026 data puts average AI engineer pay at $143,518, with entry-level roles starting around $100K.
Curious how many people reading this have already shipped something, versus still waiting to feel "qualified enough" to start.
Bookmark this whole thread if you're not starting today — the clip clears up whether the path is even right for you, and the 12-month breakdown below is what you'll want once you are.
https://t.co/4DFiJcFLOH
Losing $100 hurts roughly twice as much as gaining $100 feels good.
That's not a guess — it's Daniel Kahneman's own research, and the clip below is him explaining it in his own words.
It's why a strategy that needs 50 losing trades to prove its edge gets abandoned on trade 12. Not because the edge was fake. Because a human being couldn't tolerate the sample size required to see it.
The fix isn't willpower. It's a 15-minute gap: before overriding any plan, ask if you're reacting to new information or just a feeling. If you can't point to the first one, you already have your answer.
Full breakdown of this — the sunk cost trap, why automation only protects an edge instead of creating one, and 4 concrete things to do starting with your next trade — is in the article below. Watch the clip first though. It's the source of everything.
Losing $100 hurts roughly twice as much as gaining $100 feels good.
That's not a guess — it's Daniel Kahneman's own research, and the clip below is him explaining it in his own words.
It's why a strategy that needs 50 losing trades to prove its edge gets abandoned on trade 12. Not because the edge was fake. Because a human being couldn't tolerate the sample size required to see it.
The fix isn't willpower. It's a 15-minute gap: before overriding any plan, ask if you're reacting to new information or just a feeling. If you can't point to the first one, you already have your answer.
Full breakdown of this — the sunk cost trap, why automation only protects an edge instead of creating one, and 4 concrete things to do starting with your next trade — is in the article below. Watch the clip first though. It's the source of everything.
https://t.co/itdatkr1y9
https://t.co/itdatkr1y9
Last Seen Users on Sotwe
Most Popular Users
Elon Musk 
@elonmusk
241.5M followers
Barack Obama
@barackobama
118.8M followers
Cristiano Ronaldo
@cristiano
114.5M 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