Next-token prediction isn't the only way forward.
The team at @HazyResearch just dropped Next-Concept Modeling (NCM) โ and I just tested it out locally.
Instead of generating text token-by-token (slow O(N) passes) or flat diffusion (heavy compute), NCM generates text coarse-to-fine:
1๏ธโฃ A multiscale residual VQ-VAE compresses text into hierarchical discrete "concepts".
2๏ธโฃ A transformer generates the hierarchy level-by-level, predicting all codes in parallel.
3๏ธโฃ Fixed inference (~12 forward passes total), regardless of output length!
๐ The results:
โข 10ร to 31ร faster than diffusion text models (XSum, LM1B)
โข 3รโ5ร fewer training tokens
โข Solves Sudoku with exact match in 12 parallel steps (beating autoregressive greedy traps)
Ran both the Sudoku solver and TinyStories generator locally on CPU in seconds. Weights and code are fully open source!
Huge congrats to @KumbongHermann , @simran_s_arora & @togethercompute !
๐ Code: https://t.co/0oKUqbnlto
๐ Paper: https://t.co/ijyyeu1XJw
#AI #MachineLearning #LLMs #DeepLearning
someone on Discord messaged me and just claimed heโs Wang Chuanfu, the CEO of BYD ๐
obviously, Iโm playing along. ๐ค
at this rate, Iโll be joining BYD as CTO before he realizes Iโm just a random dude on Discord ๐ญ๐โก๏ธ
๐ซ Introducing OrganProtocol: Decentralized Cadaveric Organ Allocation & Cold-Ischemia Decay Protocol built on @ArkivNetwork for Road to
Devcon 8 Mumbai!
In India, over 500,000 people die every year waiting for an organ transplant. For the few organs that are harvested, the current system is
crippled by two fatal bottlenecks:
1๏ธโฃ VIP Queue-Jumping & Black-Market Bribery: Centralized state registries (NOTTO / SOTTO) run on private hospital databases. Wealthy patients
and medical tourists regularly pay bribes to hospital coordinators to alter timestamps, pushing poor, critically ill patients down the waitlist.
2๏ธโฃ Organ Wastage Due to Cold-Ischemia Time (CIT): Once harvested, human organs have a biological viability window before irreversible cell
death (Heart: 4โ6 hrs, Liver: 8โ12 hrs, Kidney: 24โ36 hrs). Fragmented bureaucratic coordination causes hundreds of organs to expire in transit
and be incinerated.
Conventional blockchains can't fix this efficiently because storing decaying temporal state requires bloated contracts, periodic upkeep bots,
and exorbitant gas fees.
Enter Arkiv: Biological Cold-Ischemia Time as Native TTL โก
OrganProtocol maps human biology directly to Arkiv's core architectural breakthrough โ native programmable Time-To-Live (TTL):
๐น Autonomous Block-Decay Expiration: When an accredited hospital harvests a heart, it mints an on-chain entity with an exact 6-hour TTL
(ExpirationTime.fromHours(6)).
๐น Zero Retroactive Forgery: If an organ isn't matched and dispatched within its viability window, Arkiv's storage engine naturally expires
the entity from the active emergency pool. No corrupt administrator can backdate timestamps to sell an expired or redirected organ.
๐น Cryptographic Waitlist Integrity: Recipient priority is determined mathematically by clinical severity (MELD/CPRA scores) and block
registration height. Queue order is censorship-resistant.
๐น 100% DPDP Act Compliant: In strict adherence to Indiaโs Digital Personal Data Protection Act, ZERO personal data (no Aadhaar, names, or
phone numbers) is written on-chain. Everything relies on blinded clinical indices and cryptographic hospital node signatures.
๐ฅ Federated Multi-Hospital Nodes:
The protocol simulates real hospital nodes across premier Indian medical centers:
โข AIIMS New Delhi (Cadaveric Heart & Multi-Organ)
โข KEM Hospital Mumbai (Hepatic & Renal Transplant)
โข Apollo Hospitals Chennai (Cadaveric Renal & Cornea)
โข NIMHANS Bengaluru (Neurotrauma & Recovery)
Built for Road to Devcon 8 Mumbai (@EFDevcon) on @Devfolio:
๐ Track: Best Indian Use Case & CROPS Track (Resistance to Capture & Censorship Resistance)
๐ Live dApp: https://t.co/PSDJw42cep
๐ Arkiv Tiramisu Explorer: https://t.co/tV5T0AAZ26
๐ป Open Source MIT Repo: https://t.co/MBehoJ73Vn
OrganProtocol on Arkiv: Restoring trust, eliminating bribery, and saving lives. ๐ฎ๐ณโค๏ธ
#RoadToDevcon #Arkiv #DevconMumbai #Devfolio #Ethereum #Web3Healthcare #CROPS
๐ Iโm coming to ETHGlobal hackathon! ๐ฆ๐ป
Excited to build, experiment with new ideas, and connect with amazing builders across the Ethereum ecosystem.
@ETHGlobal โ letโs build something awesome! ๐ฅ
If youโre attending ETHGlobal Mumbai, letโs connect, find teammates, and ship something cool! ๐ค
See you there! ๐
#ETHGlobal #Ethereum #Web3 #Hackathon
Our #neatHack submission: Meet ๐ก๏ธ FixPilot โ Autonomous Self-Healing SRE & Code Remediator.
"Build agents youโd trust to run without you watching."
When production CI breaks or auth tokens fail at 3 AM, human on-call engineers spend 45+ minutes digging through stack traces. But handing
autonomous write access to an AI agent comes with two massive risks:
1. Raw logs leaking database credentials, API keys, and PII to cloud LLMs.
2. Fragile agents that hallucinate or halt when their initial code patch fails.
Over the last 48 hours for #neatHack, we built FixPilot to solve autonomous trust from the ground up:
๐ Local AI Guardrail: Intercepts & redacts secrets, DB URIs, and PII on-device before any context reaches cloud LLMs, while enforcing strict
action boundaries.
๐ Autonomous Remediation Swarm: An Investigator agent diagnoses root causes, and a Fixer agent generates surgical code patches.
๐ง Closed-Loop Self-Healing: FixPilot's Meta-Agent inspects its own execution traces in @neatlogs when tests fail, pinpoints the exact
failing span, and guides an intelligent self-correction that passes on iteration 2 with zero regressions.
Built with @arithecfo@EntireHQ@neatlogs:
โข @neatlogs: Full span observability, latency tracking, and before-vs-after trace diffs powering the Meta-Agent's self-healing feedback loop.
โข @EntireHQ: Deep repository dependency graph mapping and Entire Checkpoints attached directly to GitHub Pull Requests.
โข @arithecfo: Real-world unit economics modeled at $0.028/fix (38s MTTR) vs. $75/incident for human triage โ unlocking an 88.2% gross margin.
๐ฅ Demo:
And thatโs a wrap on the AgntID Hackathon with @hiagntid!
30+ builders. 27 projects shipped. Four hours of building.
Thank you, everyone, for joining us.
๐ Won the AgntID Hackathon! ๐
Built, shipped, and won. Huge thanks to @devfolio and the @hiagntid team, the judges, and fellow builders for an amazing experience.
Took home a MacBook Neo! ๐
Canโt wait for the next one. ๐ฅ
#AI#Hackathon
I'm going to ETHGlobal Mumbai @ETHGlobal!
Use my referral link below to earn a discount on your stake!
Link: https://t.co/WkC7BNPidt
https://t.co/WkC7BNPidt
@shriyash_soni@VitalikButerin I donโt just want to hear Vitalik talk about Ethereum.
I want to hear how he thinks about the next internet especially where AI, agents, decentralization, and human coordination collide.
Thatโs a conversation Iโd hate to miss. ๐๐ฎ๐ณ.
BUILD. SHIP. WIN.
The Nasiko Build-A-Thon in Bengaluru brought together builders who turned ideas into working AI systems across agents, infrastructure, and orchestration.
๐ฅ Mohammad (Agent Honeypot)
๐ฅ Team Innovators (Grok Providers)
๐ฅ Mugundhan (DeepSeek Orchestration)
A huge shoutout to our winners for the *effort, creativity, and engineering* they put into building, testing, and shipping their solutions.
This is what weโre building toward at Nasiko: an open runtime for AI agents.
Build with what you already use. Run anywhere. See everything.
๐ฅ 3rd Place at the Nasiko Labs Buildathon! ๐
Honored to secure 3rd place for our DeepSeek Orchestrator ๐ง
Contributed to the open source naskio repo and leanred more about the CI/CD and infra.
Big thanks to the @nasikolabs team.
#NasikoLabs#AI#Kubernetes#Buildathon