Olivia
Online
Frezz
@Frezzwnie
Al & Tech Creator
Joined August 2022
124 Following
29 Followers
97 Posts
@0xDezo Thanks, bro
ANDREJ KARPATHY SHARED HIS PERSONAL C++ NN VISUALIZER HE'S BEEN QUIETLY WRITING FOR 3 YEARS. IT MAKES TENSORBOARD LOOK LIKE MICROSOFT PAINT
he didn't post it on twitter. didn't ship a repo. just a 30-second clip that leaked from a research group chat
and when i saw what was inside, i understood. loss curves suddenly feel obsolete
the app is called NeuralT. it's a single compiled binary written in c++ and opengl. every neuron on the screen is a real node. every weight is a physical line. the whole network runs live in front of you
meaning he doesn't stare at a scalar going down. he watches the signal move. sees which neurons fire on every input. sees which paths carry gradient. sees which weights are asleep
then he tweaks
drop a neuron. change the activation. add a hidden layer. the whole network redraws in real time. no rebuild. no notebook restart. no waiting for tensorboard to catch up
there are five hidden layers on screen. HL1 with 26 neurons in ReLU. HL2 with 30. HL3 with 21. output layer with 14 sigmoid heads. every weight glows. the strongest fire orange, the dead ones stay dim
the part that stopped me: none of this touches python. no cuda dependency. no torch install. no jupyter kernel. it's one binary rendering a live network like a nightclub visualizer
he basically built the tool ml frameworks should have shipped in 2018. it took a solo dev writing c++ for three years to do it
most people trust a loss curve to tell them the model is learning. this guy watches the model think @0xDezo
https://t.co/vs4VQ3I2Ie
@LyckyOus thanks for the information bro
Most people still think RAG is about giving AI a bigger memory. I think they’re missing the point.
A personal knowledge base or a “second brain” is great for notes, ideas, and organizing your own information. But once you’re dealing with millions of enterprise documents, that approach simply doesn’t scale.
Businesses don’t need an AI that reads everything. They need one that instantly finds the few pieces of information that actually matter, cites the source every time, and is efficient enough to use in everyday workflows.
That’s what makes RAG so powerful.
The goal isn’t to make the model remember more. It’s to make the model read less—but read the right information.
Instead of stuffing an entire knowledge base into the context window, RAG retrieves only the most relevant documents. The result is faster responses, lower costs, better accuracy, and answers that are easy to verify.
There’s a reason companies like Glean, Harvey, and Sierra have reached multi-billion-dollar valuations. They’re all solving the same fundamental problem from different angles: turning massive collections of enterprise knowledge into searchable, trustworthy, AI-powered systems.
Many people are still building a second brain for themselves.
The bigger opportunity is building a second brain for businesses—where years of institutional knowledge are trapped inside documents that nobody can efficiently search or use.
https://t.co/4qH7MeVnA4
ANDREJ KARPATHY SHARED HIS PERSONAL C++ NN VISUALIZER HE'S BEEN QUIETLY WRITING FOR 3 YEARS. IT MAKES TENSORBOARD LOOK LIKE MICROSOFT PAINT
he didn't post it on twitter. didn't ship a repo. just a 30-second clip that leaked from a research group chat
and when i saw what was inside, i understood. loss curves suddenly feel obsolete
the app is called NeuralT. it's a single compiled binary written in c++ and opengl. every neuron on the screen is a real node. every weight is a physical line. the whole network runs live in front of you
meaning he doesn't stare at a scalar going down. he watches the signal move. sees which neurons fire on every input. sees which paths carry gradient. sees which weights are asleep
then he tweaks
drop a neuron. change the activation. add a hidden layer. the whole network redraws in real time. no rebuild. no notebook restart. no waiting for tensorboard to catch up
there are five hidden layers on screen. HL1 with 26 neurons in ReLU. HL2 with 30. HL3 with 21. output layer with 14 sigmoid heads. every weight glows. the strongest fire orange, the dead ones stay dim
the part that stopped me: none of this touches python. no cuda dependency. no torch install. no jupyter kernel. it's one binary rendering a live network like a nightclub visualizer
he basically built the tool ml frameworks should have shipped in 2018. it took a solo dev writing c++ for three years to do it
most people trust a loss curve to tell them the model is learning. this guy watches the model think @0xDezo
@pulmencr this is great
ONE OF ANTHROPIC'S CODERS SHARED HIS "SECOND BRAIN" HE'S BEEN BUILDING FOR 3 MONTHS. I WASN'T READY FOR WHAT I SAW
he didn't post this on twitter. didn't make a thread. didn't record a youtube video. just showed me on a call and said "look what i've got"
when i opened it - 2000+ notes. all connected to each other. looked like a map of the universe
and at the center of it all - one file. CLAUDE.md
it's not a prompt. it's a full profile of a person. who he is. how he thinks. what he's working on. where he gets stuck. how he wants the AI to talk to him. even his weaknesses
claude reads this file first. every time. automatically. never starts a conversation from zero
i asked him "so what did this change?"
he said: "i stopped spending time explaining. now i spend time thinking"
but that's just the beginning. every project has its own folder. clean structure. when he works on something - claude only sees that project. not the entire chaos
anything he does more than once is saved as a skill. one line and claude does it his way
and at 7am every day claude walks through the entire vault on its own. finds new stuff. links it. cleans what's stale. writes him 3 lines: what changed overnight
he wakes up and the brain already worked
and here's what got me the most: it's all just text files. no cloud. no subscription. no dependency on one service. if a better model drops tomorrow he just points it at the same folder
he basically uploaded himself into a file system. and now AI doesn't guess who he is. AI knows
most people use claude like google with manners. this guy built himself an external memory that grows every day
https://t.co/JQY6hCnPiZ
@fomoliver very big I think
SOMEONE FROM THE ANTHROPIC TEAM LEAKED THEIR OBSIDIAN SETUP. 8 MILLION PEOPLE SAW HOW HE ACTUALLY USES CLAUDE
the funniest part? all of this information was sitting in claude's documentation from day one. nobody read it
one guy did, packed it into a 9-step guide and posted it. and it broke the internet. 4,100 likes, 800 retweets, then china picked it up and 8 million views
want to know what's in it?
one file. called CLAUDE.md. it holds everything about you: how you think, what you're working on, where you get stuck, even how you want the ai to talk to you. claude reads it first every single session
one file changed everything. because now ai doesn't open with "how can i help?" it already knows. it remembers your projects, sees your goals, catches moments where you're contradicting yourself
people spent years searching for the perfect prompt. the right temperature. the magic formula. and the answer turned out to be not how you ask ai. but what ai knows about you before you even open your mouth
then the guy went deeper. taught claude to work on a schedule. every morning at 7am the ai walks through all notes on its own, finds new stuff, links it, cleans what's stale. no command. no reminder
and all of this runs on obsidian. free app. text files on your drive. no cloud, no lock-in. switch models tomorrow and the folder keeps working.
https://t.co/KADozFhjGF
ANTHROPIC'S CEO SHARED WITH ME HOW HE BUILT HIMSELF A SECOND BRAIN OVER 2 MONTHS. WHEN HE HELPED ME REPLICATE IT, I REALIZED I'D BEEN USING CLAUDE AT 5%
he didn't post about it. didn't make a thread. just showed his obsidian on a call and said "this is the only thing i open before email"
i asked him to explain. he showed me
the first file in his vault is called CLAUDE.md. it has everything: how he thinks, what he's working on, where he gets stuck, what decisions he's postponing. claude reads this file first every session. automatically
meaning he opens claude and it already knows the context. doesn't ask "how can i help?" doesn't wait for a prompt. just picks up where they left off yesterday
and here's what got me: he said the first two weeks the vault was empty and useless. he almost quit. but on week three claude started citing things he forgot he wrote. and that's when he realized - these aren't notes anymore. this is a memory that works better than his own
i decided to try it. set everything up in one evening. first week i thought it was overhyped
then on day 9 i asked "what do i know about content strategy?" and claude pulled up my own idea from march that i wrote at 2am and completely forgot. connected it to three other notes and showed me that together they're a ready-made plan. one i'd been looking for for two months
i just sat there staring at the screen. these were my thoughts. my words. but assembled into something i never would have seen on my own
a month passed. and here's what changed: i stopped being afraid of forgetting. before, if i didn't write something down immediately - it was gone. now i know that even if i write it messy, at 2am, in one sentence - the system will sort it, connect it, and file it where it belongs
https://t.co/iZG4HDFeOx
THIS GUY TURNED THOUSANDS OF OBSIDIAN NOTES INTO A SECOND BRAIN THAT CLAUDE CAN SEARCH, CONNECT, AND UPDATE WITHOUT REBUILDING CONTEXT FROM ZERO.
every node represents a real note, project, tool, prompt, or decision, while the links show how ideas that looked unrelated are actually connected.
instead of opening 20 folders manually, claude follows the graph, pulls the 3–5 most relevant notes, and reconstructs the context behind any task in seconds.
new conversations, GitHub repos, experiments, and unfinished ideas keep entering the same vault, so the map becomes more useful every day instead of turning into another dead archive.
the result is not just a prettier obsidian graph. it is a persistent memory system that helps an AI agent remember what you built, why you built it, and what should happen next.
https://t.co/ITKQ6Eq8sI
@dev_maims Xaxaxax
THEY MADE AN EXTRA $6,000 IN A WEEK THANKS TO THIS AI BRAIN
“Claude Code is like crazy.”
You’ve probably used Obsidian. It’s a brain… sort of. It stores files, holds notes, you put stuff in it and it just sits there. It doesn’t actually think.
Then he showed this.
A real neurological system. Real neurons firing. Prefrontal cortex, sensory cortex, association layers — everything lighting up as it processes information.
You feed it data and it doesn’t just store it. It starts connecting things on its own. Stuff you added last week suddenly links with what you’re feeding it today. It gets smarter every single day without you having to do anything extra.
They literally fed it all their sales data… and it found a hidden pattern in their clothes that nobody on the team had noticed. One small change in the process and they made an extra $6,000 last week.
Obsidian probably wouldn’t have done that.
This is what personal AI is supposed to be. Not a passive notebook, but an active thinking partner that evolves with you.
https://t.co/iZG4HDFeOx
AN ENGINEER FROM ANTHROPIC HELPED ME SET UP CLAUDE FOR MY ONLINE STORE. ON DAY THREE IT SHOWED ME WHERE MY MONEY WAS GOING
i had chaos. suppliers in one chat, pricing in a spreadsheet, customer reviews in email, analytics in my head. every day started with me piecing together context like a puzzle
he looked at all of it and said: "let's build you a brain for your business. one evening"
we connected claude to obsidian. he threw everything in: product catalog, margins, supplier data, reviews from the last 3 months, my notes on ads. just dumped the files into a folder
claude read all of it in a minute. organized it. linked it. built connections between products, suppliers and reviews that i never would have seen myself
on day two i asked: "which products cause the most returns?" claude gave me the top 5. with quotes from reviews. with margins for each one. and showed that 2 of them were the ones i was pushing the hardest
i was advertising the products customers returned the most. and i had no idea
on day three i asked: "where am i losing money?" it connected my ad spend with conversions and returns and showed that one ad channel had been eating my budget with negative ROI for 2 months straight. i looked at those numbers separately every day and never saw the connection
https://t.co/iZG4HDFeOx
I'm sharing this because it's becoming a hot topic overseas: "Stop coding with AI and build a second brain."
Here's how to create a "second brain," suggested by OpenAI co-founder Karpathy ↓
・Install Obsidian (a free note-taking app) and create a vault.
・Open the vault with Claude Code, give a note about Karpati's design (one page of wiki ideas) and say "Do it."
・AI prepares 3 boxes (raw = for materials/wiki = for pages prepared by AI/CLAUDE. md = 1 move rule)
・Put the material (article, PDF, transcript) into RAW format and say “Import”.
・You can view everything and listen at any time.
Setup takes 5 minutes. Now you don't have to start with an empty chat every time.
I seriously don't understand why people don't still use this.
https://t.co/iZG4HDFeOx
A developer pointed Claude Code at a 6,200-note Obsidian vault and let it run overnight.
The vault was 4 years of markdown. Meeting notes, half-ideas, book highlights. Dead weight. He hadn't opened 90% of it since 2022.
He typed one line into the terminal. "Read everything. Find what I forgot."
Then he went to bed.
By 7 AM, Claude Code had touched all 6,200 files. It wrote 3 new notes.
The first mapped 14 ideas scattered across 3 years into one product. He'd never seen they connected.
The second listed 22 promises he'd made in old meeting notes and never kept.
The third was a full essay. Stitched from 40 fragments. In his own phrasing.
He hadn't written a note in 8 months. The vault had been writing itself the whole time. He just couldn't read his own handwriting at scale.
Now Claude Code opens the vault every morning at 6:00. It links new notes to old ones. It flags contradictions. It drafts.
2 tools. 1 folder of plain text. No app, no cloud, no subscription.
He stopped taking notes to remember them.
He takes them now to be read back.
He's 34. He's killing a $180 a-month AI subscription stack with a mini PC under his standing desk that finally stopped making him introduce himself to a machine every morning
For six years he opened a new AI tab every morning and re-explained his life. Who he is, what he's building, what he already tried, what sounds fake to him, which patterns kept repeating. Two hundred hours a year re-introducing himself to a machine that could not remember him past the session timeout
Last October the Obsidian CEO published five skills publicly. Free to add. He read the release at midnight from a Copenhagen apartment above Torvehallerne, closed the tab, thought about it for two weeks, then spent one Sunday building the setup he had been sketching in his head for years
Ordered a used Mac Studio off DBA for kr 6,800. Set it under his standing desk next to his coffee grinder
Installed Obsidian, connected the five skills, wired an MCP server directly into Claude, spun up a local Postgres instance for embeddings. Wrote one file describing his goals, his active projects, his blind spots, his taste, the specific phrases he uses when he is being honest versus when he is performing. Claude loads it before every session. He has not re-explained himself to a model since
He demoed it at Torvehallerne the following Wednesday to a Copenhagen fintech founder who had been paying Mem $12 monthly plus Reflect $15 monthly plus ChatGPT Plus $20 monthly for personal memory that never actually remembered anything. The founder wired kr 22,000 Friday for a setup on his own four-year archive. Signed a kr 3,800 monthly maintenance retainer over smørrebrød
By January: seventeen operator clients paying for setups on their own memory archives. $58,000 monthly recurring. Four fintech founders, three solo attorneys, two academic researchers, one wine importer, seven independent consultants who wanted their entire career context loaded into Claude before every session on hardware they controlled
Mem raised $30 million Series A on the premise that personal AI memory requires their proprietary app, cloud sync, and $12 monthly seat. Reflect raised $10 million seed selling the same promise at $15. Both products sell the ability to remember on a rented server neither operator owns
He's not competing with Mem. He's competing with the assumption that AI memory requires a subscription instead of a Mac Studio
He didn't build a second brain. He built the file that made Claude finally stop forgetting him - on hardware they cannot lock out from under him
https://t.co/ttcLJsZT2T
@InsideFlowx everything is correct, this is how claude works
A developer pointed Claude Code at a 6,200-note Obsidian vault and let it run overnight.
The vault was 4 years of markdown. Meeting notes, half-ideas, book highlights. Dead weight. He hadn't opened 90% of it since 2022.
He typed one line into the terminal. "Read everything. Find what I forgot."
Then he went to bed.
By 7 AM, Claude Code had touched all 6,200 files. It wrote 3 new notes.
The first mapped 14 ideas scattered across 3 years into one product. He'd never seen they connected.
The second listed 22 promises he'd made in old meeting notes and never kept.
The third was a full essay. Stitched from 40 fragments. In his own phrasing.
He hadn't written a note in 8 months. The vault had been writing itself the whole time. He just couldn't read his own handwriting at scale.
Now Claude Code opens the vault every morning at 6:00. It links new notes to old ones. It flags contradictions. It drafts.
2 tools. 1 folder of plain text. No app, no cloud, no subscription.
He stopped taking notes to remember them.
He takes them now to be read back.
https://t.co/iZG4HDFeOx
Last Seen Users on Sotwe
Trends for you
Most Popular Users
Elon Musk
@elonmusk
241M followers
Barack Obama
@barackobama
119.2M followers
Cristiano Ronaldo
@cristiano
112M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.2M followers
KATY PERRY
@katyperry
88.4M followers
Taylor Swift
@taylorswift13
82.3M followers
Lady Gaga
@ladygaga
73.8M followers
Virat Kohli
@imvkohli
71.1M followers
Kim Kardashian
@kimkardashian
70.2M followers
YouTube
@youtube
68.7M followers
Bill Gates
@billgates
64.3M followers
Neymar Jr
@neymarjr
64M followers
The Ellen Show
@theellenshow
62.4M followers
CNN
@cnn
61.8M followers
Selena Gomez
@selenagomez
61.5M followers
X
@x
60.8M followers