Olivia
Online
@kleos
@1kleos1
I post about AI · agents · automation · builds
Joined January 2024
132 Following
50 Followers
712 Posts
Pinned Tweet
https://t.co/9gOiua5OxN
@TimDaugs hahahahahhah, this robot is a bully
THE VOICEOVER HAS TO TELL YOU IT'S A ROBOT.
aria stands on a platform, blinks, shifts her eyes, parts her lips slightly.
all micro-movement. no walking, no gestures, no task.
that's not a limitation they're hiding, it's the point. idle behavior is what the uncanny valley actually lives in.
a still face reads as dead. a face doing tiny unprompted things reads as alive, even when nothing else about it is.
the silicone and makeup do less work here than the eye drift does.
nobody's shown a face like this holding a real conversation and staying convincing. standing still is the version that works.
THE ROBOT ISN'T CHOOSING ANY OF THOSE EXPRESSIONS. IT'S COPYING A FACE OFF A TABLET
the silicone skin and hair get the attention. the actual system is simpler than it looks.
a camera tracks a human face, maps the landmark points to actuators under the skin, and drives them in real time.
open your mouth wide, the robot opens wide. raise your eyebrows, it raises its eyebrows.
that's mirroring, not expression. the robot has no idea what surprise means.
it's the same reason these demos always feature a person standing right there off-camera.
the hard part was never the copying. it's a face deciding on its own when to look confused, and getting the timing right so it doesn't land in the uncanny valley.
silicone skin solved the surface problem years ago. nobody's solved the timing problem yet.
THE ROBOT ISN'T CHOOSING ANY OF THOSE EXPRESSIONS. IT'S COPYING A FACE OFF A TABLET
the silicone skin and hair get the attention. the actual system is simpler than it looks.
a camera tracks a human face, maps the landmark points to actuators under the skin, and drives them in real time.
open your mouth wide, the robot opens wide. raise your eyebrows, it raises its eyebrows.
that's mirroring, not expression. the robot has no idea what surprise means.
it's the same reason these demos always feature a person standing right there off-camera.
the hard part was never the copying. it's a face deciding on its own when to look confused, and getting the timing right so it doesn't land in the uncanny valley.
silicone skin solved the surface problem years ago. nobody's solved the timing problem yet.
JAPANESE ENGINEERS TORE APART A CHINESE ROBOT TO FIND ITS WEAKNESS. THEY FOUND JAPAN'S INSTEAD
nikkei xtech dismantled the unitree g1 down to the screws.
battery cells inside carry murata markings, a japanese supplier, sitting inside a chinese robot.
motor boards skip noise filters entirely to save space, a real risk, kept anyway.
some screws strip easily. some holes were drilled after assembly just to make parts fit.
none of that changed their conclusion.
china already moved past prototypes into mass-produced, working hardware.
closing that gap fast isn't realistic, and the report says so directly.
JAPANESE ENGINEERS TORE APART A CHINESE ROBOT TO FIND ITS WEAKNESS. THEY FOUND JAPAN'S INSTEAD
nikkei xtech dismantled the unitree g1 down to the screws.
battery cells inside carry murata markings, a japanese supplier, sitting inside a chinese robot.
motor boards skip noise filters entirely to save space, a real risk, kept anyway.
some screws strip easily. some holes were drilled after assembly just to make parts fit.
none of that changed their conclusion.
china already moved past prototypes into mass-produced, working hardware.
closing that gap fast isn't realistic, and the report says so directly.
UNITREE POSTED A ROBOT DOING FLYING SPIN KICKS AT FULL SPEED. THE VIDEO EVEN SAYS "NO SPEED-UP"
that disclaimer is there for a reason. footage of a robot moving this fast usually gets accused of being sped up in editing.
the sequence: jump spin kicks, turning combinations, boxing-style arm strikes, all chained together like a real sparring routine.
then it connects a jump kick with a ball hanging mid-air, the kind of moving target that needs real-time tracking, not a pre-programmed swing.
then it kicks over a 30kg bag. then a 60kg one.
staying upright while your leg delivers that much force into a stationary object is a balance problem, not just a strength one. the robot has to absorb its own recoil without toppling.
none of this is the robot deciding to spar. it's a training clip built to show off actuator torque and real-time balance recovery under load.
still, watching something human-shaped throw a jump spin kick at a heavy bag and land clean hits different than reading the spec sheet.
UNITREE POSTED A ROBOT DOING FLYING SPIN KICKS AT FULL SPEED. THE VIDEO EVEN SAYS "NO SPEED-UP"
that disclaimer is there for a reason. footage of a robot moving this fast usually gets accused of being sped up in editing.
the sequence: jump spin kicks, turning combinations, boxing-style arm strikes, all chained together like a real sparring routine.
then it connects a jump kick with a ball hanging mid-air, the kind of moving target that needs real-time tracking, not a pre-programmed swing.
then it kicks over a 30kg bag. then a 60kg one.
staying upright while your leg delivers that much force into a stationary object is a balance problem, not just a strength one. the robot has to absorb its own recoil without toppling.
none of this is the robot deciding to spar. it's a training clip built to show off actuator torque and real-time balance recovery under load.
still, watching something human-shaped throw a jump spin kick at a heavy bag and land clean hits different than reading the spec sheet.
split one random stream of events with a coin flip, and the two resulting streams come out completely independent of each other.
that's not intuitive. it's provable.
take a poisson process, a stream of random events with no memory of past timing.
flip an independent coin on every single event to send it to stream a or stream b.
common sense says the streams should be a little correlated.
if a goes quiet for a while, you'd expect b to be running a bit hot to compensate.
it isn't. and it doesn't.
→ split a poisson process this way and both halves are themselves poisson processes
→ knowing everything about stream a tells you nothing about stream b
→ run it backwards and merge independent poisson streams together, the result is one clean poisson process at the summed rate
this is lecture 5 of gallager's free mit course, and it's not just a neat proof.
6.262 was originally built to give network engineers the math for queueing and congestion.
this exact splitting and combining result is why traffic routed across multiple servers, or merged from multiple sources, can still be analyzed cleanly instead of the math falling apart.
it's the same reason a telephone exchange from the 1910s and a modern load balancer get described by the same handful of equations.
https://t.co/4xaZHpEvCm
YOUR NOTES APP COULD ALWAYS DRAW THIS MAP. IT JUST NEVER KNEW WHICH DOTS TO CONNECT
obsidian's graph view has existed for years.
it draws a line between two notes the second you type a backlink.
the problem was never the graph.
the problem was that nobody had time to sit down and manually link hundreds of old notes together.
→ point claude at an unsorted vault: journal entries, article clippings, half-finished thoughts
→ it reads through the actual content of every file, not just filenames or folders
→ it finds overlap between notes that were never manually linked, stuff you wrote months apart and forgot were related
→ new connections start showing up in the graph without you touching a single backlink
the pretty visualization isn't the interesting part.
most notes get written once and never opened again. that's the actual thing this changes, not how the graph looks.
it also gets things wrong.
it links notes together just because the wording overlaps, not because the ideas actually connect.
you still go back in and prune the bad ones by hand.
the point was never to remember more.
it's to read less and still find the thing you wrote eight months ago.
https://t.co/9gOiua5OxN
kimi's ceo threw out the optimizer every lab has used since 2014.
adam and its variants have been the default since then. basically every major model trains on some flavor of it.
moonshot swapped it for muon instead, wrapped with their own fix for a stability bug called muonclip.
kimi k2 trained on 15.5 trillion tokens with zero loss spikes.
not rare. not managed with checkpoint rollbacks. zero.
muon alone already runs at roughly double the flops-per-token efficiency of adamw.
moonshot lists token efficiency right next to long context as one of the actual pillars of their roadmap.
not bigger models. not more data.
the same data, doing twice the work per token.
the optimizer was never the boring part.
https://t.co/JzbEJ0onc0
an $9 chip can already hear your heartbeat through a wall.
it still can't draw your skeleton.
RuView is real.
85k stars on github, built almost entirely in rust.
a mesh of esp32-s3 boards, about $9 each.
reads the channel state info your router already collects to manage signal quality.
no camera. nothing leaves the building.
presence, motion, breathing rate, heart rate.
that part actually works today, offline, on hardware you can buy in bulk.
the part that got it 85k stars is full skeleton tracking through drywall at 8 meters.
that doesn't ship yet.
the project's own notes say the training pipeline exists but the trained weights don't.
github issues report zero skeleton output on working setups.
one is literally titled "looks like a scam project."
the boring 80% of this works right now.
the part everyone's sharing is still a promise.
https://t.co/9gOiua5OxN
CLAUDE READING YOUR TRADING JOURNAL WILL FIND THE MISTAKE YOU'VE MADE NINE TIMES WITHOUT NOTICING
the dashboard isn't the edge. RSI, MACD, order flow, a bullish score to two decimals. every one of those is public and free.
what nobody has is a record of their own decisions that they actually reread.
so keep the journal in Obsidian. one note per trade. thesis, size, what you expected, what actually happened, what you felt when you pressed the button.
then point Claude at the folder and ask the questions you'd never ask yourself:
which setups do I keep taking that lose
where did I size up right before a drawdown
what did I write in March that contradicts what I'm doing now
it answers from your own notes, with your own words quoted back.
the indicators were never the scarce part. the discipline to see your own pattern is, and that's the one thing a vault full of your decisions can finally show you.
https://t.co/9gOiua5OxN
A FANTASY WRITER TURNED HIS WHOLE UNIVERSE INTO A LINKED GRAPH, AND NOW IT CATCHES THE PLOT HOLES HE CAN'T
pantheon, realms, galaxies. named moons, star systems, the gods that made them. all in one Obsidian vault.
it looks like a beautiful theme. that's the surface.
underneath it's a database. every character linked to a location, every location to a system, every system to the god who created it.
that linking is the actual point.
in Word, your world lives in your head and 40 scattered documents. nothing checks whether chapter 3 contradicts chapter 30.
in a linked graph, the contradiction shows up. a character in two places at once. a moon orbiting a star you already destroyed. the map won't let you lie to yourself.
the pretty theme is free on GitHub, clone it tonight.
but the theme was never the unlock. it's that a world too big to hold in your head can finally be held by something that reads all of it at once.
https://t.co/9gOiua5OxN
A founder gave Claude Code access to 10 years of Obsidian notes.
15,000 markdown files. Startup ideas, meeting notes, random thoughts, books, and projects he'd forgotten even existed.
The prompt was simple: find the ideas I keep rediscovering.
Three days later Claude surfaced a framework from 2021 that solved the exact problem he'd been working on all week.
Then it connected two notes written four years apart.
That became his next product.
He thought he was collecting notes.
He was quietly building a model of how he thinks.
https://t.co/9gOiua5OxN
THE GLOWING 3D GRAPH IS THE PART THAT DOESN'T MATTER, THE THING READING IT DOES
everyone screenshots the pretty node cloud. the orbiting projects, the maturity score, the galaxy of notes.
it looks like a command center. mostly it's a screensaver.
a graph you look at is still a graph you have to read. and past a few hundred notes, nobody does.
that's where the three tools split the work.
Obsidian holds it. plain markdown, your files, no lock-in.
Claude reads it. point it at the folder, ask what you keep circling, where you contradict yourself, what you shelved.
NotebookLM talks it back. it turns the pile into something you can question out loud and get sourced answers from your own notes.
the dashboard was never the system.
a vault you admire is a graveyard with good lighting. a vault something else reads for you is the first time it works.
https://t.co/9gOiua5OxN
A CHARACTER JUST REALIZED THE SIMULATION'S CODE IS FULL OF REAL NAMES, WHICH IS EXACTLY WHAT YOUR OBSIDIAN VAULT ALREADY IS
"these are all names of real people, sir. it's us. they're mocking us."
it's a sci-fi scene. a node graph, particles, labels like PHE and ILE, and the horror that the system is made of you.
here's the thing. that's not fiction if you run a second brain.
your Obsidian vault is a node graph of real names, real projects, real half-thoughts. every note a particle, every backlink an edge. a live map of a person, sitting in plain text on your drive.
the character finds it terrifying that the code knows them. the builder finds it useful.
point Claude at that folder and the graph stops being a pretty picture. it reads the whole thing, finds the clusters you never linked, the ideas you keep circling, the places you contradict yourself.
the sci-fi horror is a system that contains you and you can't see it. the actual tool is a system that contains you and finally lets you read it back.
https://t.co/9gOiua5OxN
YOU SPENT A YEAR HUNTING THE PERFECT PROMPT, THE PROBLEM WAS NEVER HOW YOU ASK
everyone chased the same thing. the right wording, the magic temperature, the formula that unlocks the model.
all of it optimized the wrong side of the conversation.
the model doesn't open weak because your prompt is weak. it opens weak because it knows nothing about you when you hit enter. every session starts from zero, so every session you re-explain your whole life.
the fix isn't a better sentence. it's one file.
a CLAUDE.md at the root of your Obsidian vault. who you are, what you're building, where you keep getting stuck, how you want to be talked to. the model reads it before you type a word.
now it doesn't ask "how can I help." it already knows the project, sees the goal, catches you contradicting last week's you.
and it's plain text on your drive. no cloud, no lock-in. a better model drops tomorrow, you point it at the same folder, nothing breaks.
prompt engineering was people optimizing the question. the actual lever was context, everything the model knows before the question exists. you were tuning the wrong half the whole time.
https://t.co/9gOiua5OxN
MOST CLAUDE.MD FILES ARE A STATIC BIO, HIS TRACKS WHICH OF HIS OWN PRIORITIES ARE QUIETLY DYING
everyone writes a CLAUDE.md that tells the model who they are. write it once, it reads the same paragraph forever, even after you've changed.
his file ages with him.
every note he touches in his Obsidian vault gets logged with a timestamp. Claude reads that log before every session, so it knows what he's actually focused on right now, not what his bio claims.
three days ago the theme was "reputation before scale." yesterday it flipped to "buy back time." he explained nothing. Claude just read the log and adjusted.
then he added a decay score. roughly 73% of the vault counts as active. the rest, untouched for months, and Claude knows to stop pulling from it unless asked.
but here's the part that actually bites, and it has nothing to do with the AI.
the file makes it impossible to lie to yourself. you can say "learning is my top priority" all day. if the log shows you haven't opened that folder in 11 weeks, the file knows. and now so do you.
a static note stores who you were. this one shows you who you've actually been lately. the second one is the uncomfortable one, which is exactly why it's the useful one.
https://t.co/9gOiua5OxN
YEARS OF YOUR AI CHATS ARE JUST SITTING THERE UNMAPPED, SOMEONE BUILT A TELESCOPE TO POINT AT THEM
export your ChatGPT and Claude history, drop it in, and every conversation becomes a point in 3D space.
related chats pull into clusters on their own. it reads what you actually talked about and finds the topics itself, hundreds of them, no tagging.
one guy's map: 2,014 conversations, 18,000 messages, 2.5 million words, sorted into 230 clusters.
the part that lands isn't the pretty galaxy.
it surfaces connections you never made yourself. two topics 90% alike in meaning that you always kept in separate chats, sitting right next to each other once something reads all of it at once.
it even builds an "about me" from your own words. you, described back to you, by the sum of everything you asked.
we pour our whole head into these chats and then abandon it in a feed that scrolls one way. every problem you worked through, gone when the tab closes.
the chats were never disposable. you just never had anything that could see all of them at the same time. that was the only thing missing.
https://t.co/9gOiua5OxN
AN AGENT THAT READS YOUR WHOLE VAULT DOESN'T JUST FIND NOTES, IT CATCHES YOU LYING TO YOURSELF
most people treat Obsidian like a graveyard. save 40 highlights a week, link nothing, reread almost none of it.
the knowledge is in there. it just never talks back.
point Claude Code at the folder and that changes. it reads plain markdown natively, no plugins, no export step. your vault was always just files, and it lives in files.
but retrieval is the boring win. the sharp one is different.
ask it the questions you'd never ask yourself:
> what ideas do I keep circling in my daily notes but never ship
> where do my notes from March contradict my notes from now
> what did I decide six months ago and quietly abandon
it reads everything you wrote and hands it back with receipts. the half-thoughts you buried. the plan you talked yourself out of. the same idea you've rewritten nine times without noticing.
that's the part no note app sells you. it's not a better search box.
it's the first thing that ever read all of you at once and told you what you actually keep avoiding. storage was never the problem. facing what's in the pile was.
https://t.co/9gOiua5OxN
@AlfinCodes very cool!
THE DIFFERENCE ISN'T THE MODEL, IT'S THAT HE NEVER STARTS FROM SCRATCH
most people open Claude like google with manners. blank box, "how can I help," re-explain yourself every time.
this setup never does that.
the first file in his Obsidian vault is CLAUDE.md. not a note. a profile.
> who he is, how he thinks, where he gets stuck
> what he's building, how he wants to be talked to
Claude loads it every session. so it opens already in context. knows the goals, remembers what he shelved two months back, sees where he's contradicting himself.
every project is its own folder. inputs, process, outputs, feedback. he opens one folder, Claude sees one project, not the chaos of his whole life.
anything he does twice becomes a skill. one line and it runs his way.
and the vault is plain markdown. no cloud, no lock-in. a better model drops tomorrow, he points it at the same folder and nothing breaks.
that's the actual unlock. he didn't find a better prompt.
he stopped making the model guess who he is every single time. the context was always the bottleneck, not the model.
https://t.co/9gOiua5OxN
Last Seen Users on Sotwe
Most Popular Users
Elon Musk
@elonmusk
241.1M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
112.3M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98.1M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.3M followers
KATY PERRY
@katyperry
88.6M followers
Taylor Swift
@taylorswift13
82.5M followers
Lady Gaga
@ladygaga
74M followers
Virat Kohli
@imvkohli
71.4M followers
Kim Kardashian
@kimkardashian
70.3M followers
YouTube
@youtube
68.7M followers
Bill Gates
@billgates
64.4M followers
Neymar Jr
@neymarjr
64.3M followers
The Ellen Show
@theellenshow
62.4M followers
CNN
@cnn
61.8M followers
Selena Gomez
@selenagomez
61.7M followers
X
@x
60.8M followers