Somewhere out there
Joined May 2022
- Tweets4.4K
- Following1.7K
- Followers1.2K
- Likes33.5K
CoorgertlnL retweeted
We spent over $100K (compute, dead ends, the lot) and trained on 49M examples across about 3M synthetic life scripts to get a 4B memory specialist that runs fully local. Extract, temporal, reconcile, abstain. LoRA on a 4B base, INT4 is a couple GB on disk, a few GB hot. Comfortable on a 16GB laptop. Phone packaging soon and what we're working on next.
This isn’t just “smarter chat.” I see it as a deployment bet:
memory as a software component on the machine that holds the data. No per-write API tax. Nothing leaves the device.
Frontier labs memory remembers you inside their product. Coding agents already get rich context. Your health notes, dose changes, shift logs, ambient scraps with other people’s names in them? That stuff either lives in a silo or nowhere. We want the model next to the notes you own, something that is really personalised.
Security context is the quiet part. CONTINUITY’s point is that agent stacks drop, widen, or reinterpret authorization as work crosses tools. Private memory only works if provenance survives the whole path: who wrote it, what it authorizes, what you’re allowed to do when two records disagree. Reconcile and abstain aren’t features, it's integrity and something we've pushed from a Private AI perspective.
Imagine;
1️⃣ Offline plant handover where two valve readings conflict and there is no outbound net. Surface both. Don’t silently keep the newer row.
2️⃣A clinician’s notes on a device that can’t phone home. Dose flipped twice, third source disagrees. Refuse the confident answer.
3️⃣Ambient capture on a laptop. Thousands of fragments a day. High write volume plus privacy is where local stops being a hobby and starts being the only sane architecture.
We also killed our own vanity number. Hit 100% on our bench. Four regexes hit 100% too on 2.7M training examples. We rebuilt so phrasing carries no label. Bench ships with an attack: bag-of-words above chance means we automatically reject it.
Model lands at 41.4%. Shortcuts around 25%.
Weights + gated bench going up on Hugging Face soon. Come try to break it.
CoorgertlnL retweeted
play.hood10.xyz is LIVE!
PLAY HOOD is a verifiably fair on-chain card game where every pack contains a randomized basket of tokens from the Top 10 Index.
Choose from 10 basket sizes, ranging from $4 to $5,000.
Higher-tier packs give you a greater chance of landing higher multipliers, with up to 20× in token value available from a single opening.
The mechanics are simple: one shared ceiling across every basket, with a 10% edge flowing directly back into the HOOD10 Index to purchase more tokens.
No hidden mechanics. No off-chain outcomes.
Every opening is verifiable on-chain.
CoorgertlnL retweeted
Obscura Referral System is officially live!
Open the app, take your link, start sharing it. Every trader who comes in through it pays you back in tokenized stock - a cut of their volume, settled
straight to your wallet as a fraction of AAPL, TSLA, NVDA and more on Robinhood Chain.
Most referral programs pay in points. Yours pays in a share of a company that reports earnings.
What you make for bringing someone in is not tied to an emissions schedule. It is tied to the market.
It also hands tokenized stocks a job they have not had before. Until now a tokenized share was something you bought and held. Here it is what a network pays with - every link shared moves another fraction of a real asset onto Robinhood Chain and gives it somewhere to go.
Growth that settles in real assets. Your link is ready.
Start Referring: obscuracex.com/referral
CoorgertlnL retweeted
Everyone is suddenly paying attention to Robinhood Chain. The reason is simple: it is the first chain where the assets are things people already understand.
Most chains ask a new user to believe in something first - a token, a narrative, an ecosystem. Robinhood Chain skips that step. Stocks like AAPL, TSLA and NVDA, tokenized and moving on-chain, from a brand tens of millions of people already trust with their money. For someone stepping in from traditional finance, that is not a leap. It is a doorstep. That is why the new users keep coming.
Obscura did not arrive for this trend. We were built for it. From day one the plan was to build on Robinhood Chain: $OBS lives there, every swap reward settles there, and every product we ship is designed to give the chain more reasons to be used - private trading that pays out in its tokenized stocks today, and more services on the way that keep pulling real activity onto it.
A chain grows on two things: assets worth holding and services worth using. Robinhood brought the first. We are here to build the second.
CoorgertlnL retweeted
The whole product on one chart. Here is how to read it.
One private swap request goes into the routing engine. The engine draws on three things at once: CEX and DEX liquidity for the price, a Dark Order shield so your size and intent never touch a public book, and wallet settlement so the output has exactly one place to land.
Then it asks the only question it ever asks: best route?
When the answer is yes, two arrows leave the box, not one. The swap executes. The reward is delivered, as tokenized stock.
Now look for what is not on the chart. No sign-up box, no KYC box, no deposit box, no claim button. They are not cropped out of frame. They were never part of the flow.
CoorgertlnL retweeted
You can now talk to Obscura.
An AI on obscuracex.com takes trades in plain language. Type "swap 1 ETH for USDC" the way you would say it, and it builds the trade for you: finds the best private route, fills in the details, and sets everything up for your confirmation.
It handles the conversation like a person would. Leave out the amount and it asks. Name a pair that has no route right now and it tells you straight away and offers one that works. You never end up digging through menus to figure out why something did not go through.
And because it runs on Obscura, the trade it builds carries everything an Obscura swap carries. The route is private and priced best-first across CEX and DEX liquidity, your order never sits on a public book, and settlement goes straight to your own wallet - no account, no KYC. Nothing executes until you press confirm.
The rewards come with it too. A share of the fee on every swap returns to you as tokenized stocks on Robinhood Chain - and your rate climbs with your 30-day volume. Whether you build the trade by hand or ask the AI to do it, it earns the same.
No forms, no order ticket, no learning curve. If you can describe the trade, you can place it - and get paid in real-world assets for placing it.
Next would be a rt then a full blown “I like this project tweet”
All will happen in due time
The founder of Robinhood, @vladtenev liked one of our posts today. A small thing, but it tells you something: the people building Robinhood Chain pay attention to the builders on it.
That attention matters. Robinhood Chain gave teams like ours the rails to put real-world assets to work on-chain, and every builder here extends that reach - more products, more ways in, more people getting access to assets the old system kept behind a desk.
That is the vision we signed up for when we chose this chain. Vlad is bringing RWA to crypto and opening finance to everyone, and Obscura's part in it is simple: every swap pays its reward in tokenized stocks, settled on Robinhood Chain.
Thanks for the nod. Back to building.
upside potential of this is quite high.
dev is also legit and he used to be part of the shapeshift team.
only downside is it’s in virtual chain and people are lazy to bridge,if there’s not much attention.
@AkamiKarubi @thecultos
0x1984edF491D3399FBc09E6d0856E01fF3721f952
CoorgertlnL retweeted
gm @CryptoHayes just wanted to let u know that we have already built the first decentralized tokenized memory standard to store memories onchain. Its already live on Solana with millions of memories stored and can already be bridged to EVM chains too.
Sent u a dm to discuss further too. 🤝
More details about our @cludeproject
Portable Memory Protocol (PMP) is in this article:
nitter.gq/sebbsssss/status/20716…
CoorgertlnL retweeted
Clude trained the world’s first memory model that remembers better than frontier models 100x its size.
Read more here:
CoorgertlnL retweeted
When I first launched @cludeproject, there were less than 10 users, mostly family and friends. I didn't have a team, nor funding. Just a conviction that AI agents need real memory, not chat logs stuffed into a context window.
I'm humbled that we managed to snatch the 4th spot in the @colosseum agent hackathon and since then; we have grown to 5000+ installs and 600+ real users. We store more than 2M memories today.
None of this was paid acquisition. People found Clude because they ran into the same problem we did. Their agents kept forgetting everything.
Still a lot to build. But going from 10 users to 600 in a month, off the back of a hackathon project, tells me this isn't just our problem.
CoorgertlnL retweeted
Every AI chat stores your data on their servers, forgets you every session, and trains on your conversations. If you ask me, I don’t like AI surveillance
Clude Chat on @AskVenice is different.
• Chat history lives on your device only
• Decentralized GPUs with zero data retention
• End-to-end encryption, decrypted only in verified hardware enclaves
• Memory that persists and learns who you are
An AI that remembers you. Infrastructure that can't.
➡️ Coming soon
Vires In Numeris
CoorgertlnL retweeted
Every time a new AI model drops with a 1-million token context window, the tech world celebrates. And for deep, single session analysis of massive documents, it truly is an incredible leap forward.
But using these massive context windows as a substitute for long-term memory is fundamentally unscalable. Here is the educational breakdown of why.
The Hidden Cost of Context
When you use a context window, the model processes every single token, every single time you hit send. If you load up 1M tokens and ask a simple question, the LLM still reads all 1M tokens before answering. With a model like GPT-4.5, that scales to a staggering $75 per query.
The Architectural Fix: Retrieval
Memory retrieval systems (like Clude) use a completely different approach. Instead of forcing the AI to reread the entire library every time, the system runs a quick vector search to find only the relevant information, sending a tiny payload (usually ~2,000 tokens) to the LLM.
The Math at Scale
Because retrieval sends only what matters, your costs stay flat. If you run 1,000 queries a day against a 500K memory bank:
• Context stuffing (GPT-4o): ~$37,530 / month
• Memory retrieval (Clude + GPT-4o): ~$182 / month
You get the exact same answers from the exact same models, but it costs ~200x less.
Giant context windows are amazing tools, but for scalable memory, they force your costs to scale linearly. Memory retrieval keeps them flat
Pick your poison
CoorgertlnL retweeted
Building in Public - Deep dive into the memories nitter.gq/i/broadcasts/1qKDzPWgM…