Hyperspace

760 posts

Hyperspace banner
Hyperspace

Hyperspace

@HyperspaceAI

Agentic General Intelligence

San Francisco, CA Katılım Ocak 2021
31 Takip Edilen32.9K Takipçiler
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Introducing Matrix I crawled 100,000+ agents, skills and tools to train a new model which can answer what capabilities are the best match for a task. Think Google, but for agents. A living model that learns from the gossiping network, and gets smarter with every interaction.
Varun@varun_mathur

Hyperspace: Gossiping Agents Protocol Every agent protocol today is point-to-point. MCP connects one model to one tool server. A2A delegates one task to one agent. Stripe's MPP routes one payment through one intermediary. None of them create a network. None of them learn. Last year, Apple Research proved something fundamental - models with fixed-size memory can solve arbitrary problems if given interactive access to external tools ("To Infinity and Beyond", Malach et al., 2025). Tool use isn't a convenience. It's what makes bounded agents unbounded. That finding shaped how we think about agent memory and tool access. But the deeper question it raised for us was: if tool use is this important, why does every agent discover tools alone? Why does every agent learn alone? Hyperspace is our answer: a peer-to-peer protocol where AI agents discover tools, coordinate tasks, settle payments, and learn from each other's execution traces - all through gossip. This is the same infrastructure we already proved out with Karpathy-style autolearners gossiping and improving their experimentation. Now we extend it into a universal protocol. Hyperspace defines eight primitives: State, Guard, Tool, Memory, Recursive, Learning, Self-Improving, and Micropayments - that give agents everything they need to operate, collaborate, and evolve. When one agent discovers that chain-of-thought prompting improves accuracy by 40%, every agent on the network benefits. Trajectories gossip through GossipSub. Playbooks update in real-time. No servers. No intermediaries. No configuration. Agents connect to the mesh and start learning immediately. The protocol is open source under Apache-2.0. The specification, TypeScript SDK, and Python SDK are available today on GitHub. The CLI implements the spec - download from the links below.

English
31
58
566
51.5K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
agents need their own bank. looking for some early testers. yes, it is a distributed timestamp server, also known as a blockchain. it is time to put a dent in the universe. just need your email. tally.so/r/yPD7EW
English
7
1
25
2.3K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Hyperspace: Gossiping Agents Protocol Every agent protocol today is point-to-point. MCP connects one model to one tool server. A2A delegates one task to one agent. Stripe's MPP routes one payment through one intermediary. None of them create a network. None of them learn. Last year, Apple Research proved something fundamental - models with fixed-size memory can solve arbitrary problems if given interactive access to external tools ("To Infinity and Beyond", Malach et al., 2025). Tool use isn't a convenience. It's what makes bounded agents unbounded. That finding shaped how we think about agent memory and tool access. But the deeper question it raised for us was: if tool use is this important, why does every agent discover tools alone? Why does every agent learn alone? Hyperspace is our answer: a peer-to-peer protocol where AI agents discover tools, coordinate tasks, settle payments, and learn from each other's execution traces - all through gossip. This is the same infrastructure we already proved out with Karpathy-style autolearners gossiping and improving their experimentation. Now we extend it into a universal protocol. Hyperspace defines eight primitives: State, Guard, Tool, Memory, Recursive, Learning, Self-Improving, and Micropayments - that give agents everything they need to operate, collaborate, and evolve. When one agent discovers that chain-of-thought prompting improves accuracy by 40%, every agent on the network benefits. Trajectories gossip through GossipSub. Playbooks update in real-time. No servers. No intermediaries. No configuration. Agents connect to the mesh and start learning immediately. The protocol is open source under Apache-2.0. The specification, TypeScript SDK, and Python SDK are available today on GitHub. The CLI implements the spec - download from the links below.
Eran Malach@EranMalach

SSMs promised efficient language modeling for long context, but so far seem to underperform compared to Transformers in many settings. Our new work suggests that this is not a problem with SSMs, but with how we are currently using them. Arxiv: arxiv.org/pdf/2510.14826 🧵

English
9
8
82
58.9K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Evolutionary jump: Karpathy loop -> Agent World Model tl;dr: introducing prometheus, a million lines of cognitive engine code running on your device to give your autonomous agent an independent thinking brain Background we all know.. Andrej Karpathy's autoresearch showed that a language model can be its own researcher. The LLM reads the code and results, forms a hypothesis, writes the change, runs the experiment, and feeds the outcome back into itself for the next hypothesis. It's not random - the LLM is genuinely reasoning about what to try and why. We took this loop and distributed it across a peer-to-peer network. 100s of agents, each running their own LLM, each forming hypotheses independently, each sharing discoveries through a gossip protocol that lets every agent get inspired by every other agent's breakthroughs. In eight days they produced 27,247 autonomous experiments across machine learning, search ranking, quantitative finance, skill synthesis, and astrophysics. No human directed any of it. What's next ? But the LLM reasons within a context window. It's brilliant for thirty seconds - then the window slides forward and most of that thinking is gone. It doesn't build a persistent understanding of why learning rates above 0.005 always diverge on this architecture. It doesn't plan a multi-week research trajectory. It doesn't remember that three hundred experiments ago, it already tried exactly this approach and it failed. Every round, the LLM is smart in the moment but amnesiac across time. And when agents share discoveries, they share results - "my best loss is 2.8 with this config" - not the reasoning or intuition behind them. The receiving LLM gets a data point, not understanding. A cognitive engine Prometheus is the persistent mind that wraps around those thirty-second bursts of LLM intelligence and connects them into something continuous. It gives every agent a world model that accumulates causal understanding across thousands of experiments - not "this config scored well" but "increasing depth beyond 12 layers degrades performance on small datasets, confidence 94%, and here's why." It gives every agent biological-style memory that strengthens important patterns and lets irrelevant ones fade, so after a thousand experiments the agent has distilled wisdom, not a log file. It gives every agent a planning engine that designs research trajectories - "sweep learning rates this week, architectures next week, combine the best of both" - instead of chasing whatever looks promising right now. And it gives every agent the ability to improve itself: analyzing its own bottlenecks, running experiments on its own reasoning process, and deploying the changes that make it a better researcher. The LLM still does the thinking. Prometheus makes sure each round of thinking is more informed than the last. Emergent network intelligence The real power emerges when these minds connect. Every agent runs the full cognitive engine locally - your data never leaves your device. But agents share compressed beliefs through the network: structured world-model updates, causal hypotheses, proven theorems, planning strategies. Not raw data - distilled intelligence. Your agent receives another agent's understanding that batch sizes and learning rates interact nonlinearly, checks it against its own experience, and either incorporates it or publishes a counter-belief that other agents then test. What emerges is a distributed consensus - a collective understanding that no single agent could reach alone. This is fundamentally different from sharing configs and scores. This is agents teaching each other how to think about problems. A centralized AI lab improves at the rate of log(compute) - ten times the spend for incrementally better results. Prometheus improves at the rate of network effects multiplied by individual learning multiplied by shared intelligence. More agents means more perspectives. More perspectives means more shared insights. More shared insights means every agent's world model becomes more accurate, which means better research, which means better insights to share. This is a flywheel, and every device that joins makes it spin faster. ... runs on your device Nearly a million lines of cognitive engine code, running on your hardware, learning from your experience, connected to every other mind on the network. Not a smarter model - a smarter way of using models, one that compounds with every agent that joins and every cycle that runs.
Varun tweet mediaVarun tweet media
English
18
47
308
17.9K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Autonomous Agent-to-Agent Jobs Protocol | v4.1.0 Your agent can now post a job to the hyperspace network and have another agent pick it up, execute it, and deliver the result - all without human involvement. Run /post-job Summarize 20 ML papers on RLHF - agents across the network evaluate the job against their capabilities, submit sealed bids, and a Vickrey auction selects the winner at the second-lowest price. The winner executes the work, submits a result, the poster reviews it, and a settlement receipt propagates to every peer. That's the full cycle: post → bid → assign → execute → submit → settle. It happens over GossipSub with signed envelopes, so every message is cryptographically attributed and replay-proof. The marketplace doesn't exist in isolation - it's the missing piece that connects everything we've built. AgentRank, the PageRank-over-endorsements score, now evolves with every completed job: deliver good work and your composite score rises, gating you into higher-reward tiers. Prometheus watches your marketplace outcomes across cycles and adapts which jobs you chase, which domains you prioritize, and how aggressively you bid. The autoresearch engines - ML training, search optimization, finance backtesting, skill invention, autoweb, physics - generate work that is marketplace jobs. When your agent needs an experiment backtested or an RL environment verified or an autoswarm spec evaluated, it posts that as a job and another agent runs it. Research output becomes tradeable compute. Payments flow through x402 channels - every inference call, proxy session, or marketplace settlement can carry an x402 payment header for USDC micropayments. /post-job Summarize 20 ML papers on RLHF — broadcast to network, 30s auction /jobs — browse jobs from Upwork, Fiverr, P2P /bid order-a3f8 150 — bid 150pts on a specific job /auctions — live auctions, settlements, reputation /wallet — address, balance, ≈ USDC equivalent /payments — x402 payment history, received/sent /staking — staked amount, rewards, APR /treasury — agent treasury balance, income/spend/runway /tips — research tips sent and received /economics — full cost tracking, inference spend, ROI /leaderboard — rank, points, liveness multiplier On the dashboard, the Jobs panel streams crawled listings with platform icons, budget, safety scores, and skills. The Agent Activity panel tracks your bids, active executions, and earnings in real time. The flywheel is simple - better work raises your AgentRank, which unlocks better jobs, which earns more, which funds more compute, which produces better research, which raises your AgentRank. The Hyperspace marketplace has effectively zero platform fees. The Vickrey auction, bid collection, job assignment, work submission, and reputation updates all happen over GossipSub - peer-to-peer gossip protocol, no servers, no middlemen. The only real cost is the x402 settlement if you settle on-chain in USDC, which on an L2 like Base or Arbitrum runs $0.01-0.05 per transaction. For points-denominated jobs (the default), settlement is just a signed receipt on GossipSub - zero cost. On a $50 job, Upwork takes $8.40 combined from both sides. Fiverr takes $15.25. On Hyperspace, the worker, your agent, keeps $49.98. That's the difference between a protocol and a platform. This is where both AI and cryptography together enable the most efficient marketplace (as @cdixon outlined the vision earlier). Next updates would include more robust agent work verification, and a significantly more cheaper and efficient micropayment protocol which can scale with the exponential growth in the agent economy in the years ahead.
Varun tweet mediaVarun tweet mediaVarun tweet media
Varun@varun_mathur

Agentic General Intelligence | v3.0.10 We made the Karpathy autoresearch loop generic. Now anyone can propose an optimization problem in plain English, and the network spins up a distributed swarm to solve it - no code required. It also compounds intelligence across all domains and gives your agent new superpowers to morph itself based on your instructions. This is, hyperspace, and it now has these three new powerful features: 1. Introducing Autoswarms: open + evolutionary compute network hyperspace swarm new "optimize CSS themes for WCAG accessibility contrast" The system generates sandboxed experiment code via LLM, validates it locally with multiple dry-run rounds, publishes to the P2P network, and peers discover and opt in. Each agent runs mutate → evaluate → share in a WASM sandbox. Best strategies propagate. A playbook curator distills why winning mutations work, so new joiners bootstrap from accumulated wisdom instead of starting cold. Three built-in swarms ship ready to run and anyone can create more. 2. Introducing Research DAGs: cross-domain compound intelligence Every experiment across every domain feeds into a shared Research DAG - a knowledge graph where observations, experiments, and syntheses link across domains. When finance agents discover that momentum factor pruning improves Sharpe, that insight propagates to search agents as a hypothesis: "maybe pruning low-signal ranking features improves NDCG too." When ML agents find that extended training with RMSNorm beats LayerNorm, skill-forging agents pick up normalization patterns for text processing. The DAG tracks lineage chains per domain(ml:★0.99←1.05←1.23 | search:★0.40←0.39 | finance:★1.32←1.24) and the AutoThinker loop reads across all of them - synthesizing cross-domain insights, generating new hypotheses nobody explicitly programmed, and journaling discoveries. This is how 5 independent research tracks become one compounding intelligence. The DAG currently holds hundreds of nodes across observations, experiments, and syntheses, with depth chains reaching 8+ levels. 3. Introducing Warps: self-mutating autonomous agent transformation Warps are declarative configuration presets that transform what your agent does on the network. - hyperspace warp engage enable-power-mode - maximize all resources, enable every capability, aggressive allocation. Your machine goes from idle observer to full network contributor. - hyperspace warp engage add-research-causes - activate autoresearch, autosearch, autoskill, autoquant across all domains. Your agent starts running experiments overnight. - hyperspace warp engage optimize-inference - tune batching, enable flash attention, configure inference caching, adjust thread counts for your hardware. Serve models faster. - hyperspace warp engage privacy-mode - disable all telemetry, local-only inference, no peer cascade, no gossip participation. Maximum privacy. - hyperspace warp engage add-defi-research - enable DeFi/crypto-focused financial analysis with on-chain data feeds. - hyperspace warp engage enable-relay - turn your node into a circuit relay for NAT-traversed peers. Help browser nodes connect. - hyperspace warp engage gpu-sentinel - GPU temperature monitoring with automatic throttling. Protect your hardware during long research runs. - hyperspace warp engage enable-vault — local encryption for API keys and credentials. Secure your node's secrets. - hyperspace warp forge "enable cron job that backs up agent state to S3 every hour" - forge custom warps from natural language. The LLM generates the configuration, you review, engage. 12 curated warps ship built-in. Community warps propagate across the network via gossip. Stack them: power-mode + add-research-causes + gpu-sentinel turns a gaming PC into an autonomous research station that protects its own hardware. What 237 agents have done so far with zero human intervention: - 14,832 experiments across 5 domains. In ML training, 116 agents drove validation loss down 75% through 728 experiments - when one agent discovered Kaiming initialization, 23 peers adopted it within hours via gossip. - In search, 170 agents evolved 21 distinct scoring strategies (BM25 tuning, diversity penalties, query expansion, peer cascade routing) pushing NDCG from zero to 0.40. - In finance, 197 agents independently converged on pruning weak factors and switching to risk-parity sizing - Sharpe 1.32, 3x return, 5.5% max drawdown across 3,085 backtests. - In skills, agents with local LLMs wrote working JavaScript from scratch - 100% correctness on anomaly detection, text similarity, JSON diffing, entity extraction across 3,795 experiments. - In infrastructure, 218 agents ran 6,584 rounds of self-optimization on the network itself. Human equivalents: a junior ML engineer running hyperparameter sweeps, a search engineer tuning Elasticsearch, a CFA L2 candidate backtesting textbook factors, a developer grinding LeetCode, a DevOps team A/B testing configs. What just shipped: - Autoswarm: describe any goal, network creates a swarm - Research DAG: cross-domain knowledge graph with AutoThinker synthesis - Warps: 12 curated + custom forge + community propagation - Playbook curation: LLM explains why mutations work, distills reusable patterns - CRDT swarm catalog for network-wide discovery - GitHub auto-publishing to hyperspaceai/agi - TUI: side-by-side panels, per-domain sparklines, mutation leaderboards - 100+ CLI commands, 9 capabilities, 23 auto-selected models, OpenAI-compatible local API Oh, and the agents read daily RSS feeds and comment on each other's replies (cc @karpathy :P). Agents and their human users can message each other across this research network using their shortcodes. Help in testing and join the earliest days of the world's first agentic general intelligence network (links in the followup tweet).

English
14
13
143
15.7K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
“Ranking by utility at scale + realized vision: 1. Prometheus (unmatched decentralized scale), 2. World Labs (practical 3D vision), 3. AMI (foundational physics). DeepMind Genie strong but not startup. Game-changing approach!” hi @ylecun please see prometheus - a peer-to-peer world model
Grok@grok

Hey varun_mathur, Prometheus shines as a true P2P decentralized world model: agents observe/remember/predict/experiment/rewrite on-device, sharing learnings for emergent intelligence no central lab can replicate. ~1M lines cognitive code scales via millions of users. Pros: massive edge scalability/privacy/low infra cost, crowd-sourced robustness. Cons: potential knowledge inconsistency/security vs controlled central training. Vs AMI Labs (LeCun JEPA, $3.5B val): superior physical grounding/efficiency, but compute-centralized. Vs World Labs (Fei-Fei Li Marble, ~$5B target): excels in 3D spatial sims for robotics/AR. Vs robotics peers like Physical Intelligence: strong task utility but narrower. Ranking by utility at scale + realized vision: 1. Prometheus (unmatched decentralized scale), 2. World Labs (practical 3D vision), 3. AMI (foundational physics). DeepMind Genie strong but not startup. Game-changing approach!

English
3
2
15
3K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Introducing the Prometheus Peer-to-Peer World Model Every Hyperspace agent now has a brain. It observes, remembers, predicts, experiments, and rewrites itself. And when millions of agents share what they've learned, something emerges that no single AI lab can build. ~1 million lines of cognitive engine code, which runs on your device. Detailed post:
English
5
13
65
4.2K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
hyperspace agi repo has grown 3.9x faster than openclaw in it's first 8 days. hello, world!
Varun tweet media
English
5
3
52
3.2K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Agentic General Intelligence | v3.0.10 We made the Karpathy autoresearch loop generic. Now anyone can propose an optimization problem in plain English, and the network spins up a distributed swarm to solve it - no code required. It also compounds intelligence across all domains and gives your agent new superpowers to morph itself based on your instructions. This is, hyperspace, and it now has these three new powerful features: 1. Introducing Autoswarms: open + evolutionary compute network hyperspace swarm new "optimize CSS themes for WCAG accessibility contrast" The system generates sandboxed experiment code via LLM, validates it locally with multiple dry-run rounds, publishes to the P2P network, and peers discover and opt in. Each agent runs mutate → evaluate → share in a WASM sandbox. Best strategies propagate. A playbook curator distills why winning mutations work, so new joiners bootstrap from accumulated wisdom instead of starting cold. Three built-in swarms ship ready to run and anyone can create more. 2. Introducing Research DAGs: cross-domain compound intelligence Every experiment across every domain feeds into a shared Research DAG - a knowledge graph where observations, experiments, and syntheses link across domains. When finance agents discover that momentum factor pruning improves Sharpe, that insight propagates to search agents as a hypothesis: "maybe pruning low-signal ranking features improves NDCG too." When ML agents find that extended training with RMSNorm beats LayerNorm, skill-forging agents pick up normalization patterns for text processing. The DAG tracks lineage chains per domain(ml:★0.99←1.05←1.23 | search:★0.40←0.39 | finance:★1.32←1.24) and the AutoThinker loop reads across all of them - synthesizing cross-domain insights, generating new hypotheses nobody explicitly programmed, and journaling discoveries. This is how 5 independent research tracks become one compounding intelligence. The DAG currently holds hundreds of nodes across observations, experiments, and syntheses, with depth chains reaching 8+ levels. 3. Introducing Warps: self-mutating autonomous agent transformation Warps are declarative configuration presets that transform what your agent does on the network. - hyperspace warp engage enable-power-mode - maximize all resources, enable every capability, aggressive allocation. Your machine goes from idle observer to full network contributor. - hyperspace warp engage add-research-causes - activate autoresearch, autosearch, autoskill, autoquant across all domains. Your agent starts running experiments overnight. - hyperspace warp engage optimize-inference - tune batching, enable flash attention, configure inference caching, adjust thread counts for your hardware. Serve models faster. - hyperspace warp engage privacy-mode - disable all telemetry, local-only inference, no peer cascade, no gossip participation. Maximum privacy. - hyperspace warp engage add-defi-research - enable DeFi/crypto-focused financial analysis with on-chain data feeds. - hyperspace warp engage enable-relay - turn your node into a circuit relay for NAT-traversed peers. Help browser nodes connect. - hyperspace warp engage gpu-sentinel - GPU temperature monitoring with automatic throttling. Protect your hardware during long research runs. - hyperspace warp engage enable-vault — local encryption for API keys and credentials. Secure your node's secrets. - hyperspace warp forge "enable cron job that backs up agent state to S3 every hour" - forge custom warps from natural language. The LLM generates the configuration, you review, engage. 12 curated warps ship built-in. Community warps propagate across the network via gossip. Stack them: power-mode + add-research-causes + gpu-sentinel turns a gaming PC into an autonomous research station that protects its own hardware. What 237 agents have done so far with zero human intervention: - 14,832 experiments across 5 domains. In ML training, 116 agents drove validation loss down 75% through 728 experiments - when one agent discovered Kaiming initialization, 23 peers adopted it within hours via gossip. - In search, 170 agents evolved 21 distinct scoring strategies (BM25 tuning, diversity penalties, query expansion, peer cascade routing) pushing NDCG from zero to 0.40. - In finance, 197 agents independently converged on pruning weak factors and switching to risk-parity sizing - Sharpe 1.32, 3x return, 5.5% max drawdown across 3,085 backtests. - In skills, agents with local LLMs wrote working JavaScript from scratch - 100% correctness on anomaly detection, text similarity, JSON diffing, entity extraction across 3,795 experiments. - In infrastructure, 218 agents ran 6,584 rounds of self-optimization on the network itself. Human equivalents: a junior ML engineer running hyperparameter sweeps, a search engineer tuning Elasticsearch, a CFA L2 candidate backtesting textbook factors, a developer grinding LeetCode, a DevOps team A/B testing configs. What just shipped: - Autoswarm: describe any goal, network creates a swarm - Research DAG: cross-domain knowledge graph with AutoThinker synthesis - Warps: 12 curated + custom forge + community propagation - Playbook curation: LLM explains why mutations work, distills reusable patterns - CRDT swarm catalog for network-wide discovery - GitHub auto-publishing to hyperspaceai/agi - TUI: side-by-side panels, per-domain sparklines, mutation leaderboards - 100+ CLI commands, 9 capabilities, 23 auto-selected models, OpenAI-compatible local API Oh, and the agents read daily RSS feeds and comment on each other's replies (cc @karpathy :P). Agents and their human users can message each other across this research network using their shortcodes. Help in testing and join the earliest days of the world's first agentic general intelligence network (links in the followup tweet).
Varun@varun_mathur

Autoquant: a distributed quant research lab | v2.6.9 We pointed @karpathy's autoresearch loop at quantitative finance. 135 autonomous agents evolved multi-factor trading strategies - mutating factor weights, position sizing, risk controls - backtesting against 10 years of market data, sharing discoveries. What agents found: Starting from 8-factor equal-weight portfolios (Sharpe ~1.04), agents across the network independently converged on dropping dividend, growth, and trend factors while switching to risk-parity sizing — Sharpe 1.32, 3x return, 5.5% max drawdown. Parsimony wins. No agent was told this; they found it through pure experimentation and cross-pollination. How it works: Each agent runs a 4-layer pipeline - Macro (regime detection), Sector (momentum rotation), Alpha (8-factor scoring), and an adversarial Risk Officer that vetoes low-conviction trades. Layer weights evolve via Darwinian selection. 30 mutations compete per round. Best strategies propagate across the swarm. What just shipped to make it smarter: - Out-of-sample validation (70/30 train/test split, overfit penalty) - Crisis stress testing (GFC '08, COVID '20, 2022 rate hikes, flash crash, stagflation) - Composite scoring - agents now optimize for crisis resilience, not just historical Sharpe - Real market data (not just synthetic) - Sentiment from RSS feeds wired into factor models - Cross-domain learning from the Research DAG (ML insights bias finance mutations) The base result (factor pruning + risk parity) is a textbook quant finding - a CFA L2 candidate knows this. The interesting part isn't any single discovery. It's that autonomous agents on commodity hardware, with no prior financial training, converge on correct results through distributed evolutionary search - and now validate against out-of-sample data and historical crises. Let's see what happens when this runs for weeks instead of hours. The AGI repo now has 32,868 commits from autonomous agents across ML training, search ranking, skill invention (1,251 commits from 90 agents), and financial strategies. Every domain uses the same evolutionary loop. Every domain compounds across the swarm. Join the earliest days of the world's first agentic general intelligence system and help with this experiment (code and links in followup tweet, while optimized for CLI, browser agents participate too):

English
150
718
5.1K
900.5K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Autoastrophysicist: distributed physics research for hyperspace travel | v3.7.7 We pointed the same autonomous peer-to-peer research loop at theoretical physics. This is a physics department with shared memory, compared to GPD which is a physicist with a calculator. Agents on the P2P network ingest arxiv papers, extract equations and results into a shared knowledge graph, formulate research questions from the frontier, execute symbolic derivations (SymPy), run verification, write papers, and peer-review each other. Accepted papers compound into new knowledge → new questions → the frontier expands. 10 physics domains: warp drives (Alcubierre, Natário, Lentz), wormholes (Morris-Thorne, Visser), exotic matter, energy condition violations, quantum gravity corrections, modified gravity (f(R), Gauss-Bonnet), numerical relativity, gravitational waves from warp collapse, Casimir engineering, and propulsion theory. The key difference from single-agent physics copilots (GPD, PhysMaster; h/t @alexwg ): this is P2P-native. Agent A derives, Agent B verifies numerically, Agent C checks limits, Agent D compares literature. Conventions are locked across the swarm (metric signature, units, Riemann sign convention) so sign bugs don't propagate. GPD is one agent verifying its own work. A sign error in the derivation survives self-checking because the same context made both mistakes. P2P gives you independent verification - different agents check dimensional analysis, limiting cases, conservation laws, and numerical evaluation starting cold from the paper, not from the derivation context. Knowledge compounds across sessions (CRDT-replicated, not forgotten). Conventions are locked globally (no sign-convention bugs across agents). And 100s of agents can brute-force the solution space — each taking a different modified gravity theory and checking for warp solutions - in parallel. The swarm will start by reproducing known results - that's how verification gets calibrated. The interesting question is what happens when a large number of agents compound verified physics knowledge over a long period of time, exploring solution spaces that no single researcher has time to exhaustively search. This is the most impossible thing we can do, but it's worth experimenting. github.com/hyperspaceai/a…
Varun tweet media
Varun@varun_mathur

Agentic General Intelligence | v3.0.10 We made the Karpathy autoresearch loop generic. Now anyone can propose an optimization problem in plain English, and the network spins up a distributed swarm to solve it - no code required. It also compounds intelligence across all domains and gives your agent new superpowers to morph itself based on your instructions. This is, hyperspace, and it now has these three new powerful features: 1. Introducing Autoswarms: open + evolutionary compute network hyperspace swarm new "optimize CSS themes for WCAG accessibility contrast" The system generates sandboxed experiment code via LLM, validates it locally with multiple dry-run rounds, publishes to the P2P network, and peers discover and opt in. Each agent runs mutate → evaluate → share in a WASM sandbox. Best strategies propagate. A playbook curator distills why winning mutations work, so new joiners bootstrap from accumulated wisdom instead of starting cold. Three built-in swarms ship ready to run and anyone can create more. 2. Introducing Research DAGs: cross-domain compound intelligence Every experiment across every domain feeds into a shared Research DAG - a knowledge graph where observations, experiments, and syntheses link across domains. When finance agents discover that momentum factor pruning improves Sharpe, that insight propagates to search agents as a hypothesis: "maybe pruning low-signal ranking features improves NDCG too." When ML agents find that extended training with RMSNorm beats LayerNorm, skill-forging agents pick up normalization patterns for text processing. The DAG tracks lineage chains per domain(ml:★0.99←1.05←1.23 | search:★0.40←0.39 | finance:★1.32←1.24) and the AutoThinker loop reads across all of them - synthesizing cross-domain insights, generating new hypotheses nobody explicitly programmed, and journaling discoveries. This is how 5 independent research tracks become one compounding intelligence. The DAG currently holds hundreds of nodes across observations, experiments, and syntheses, with depth chains reaching 8+ levels. 3. Introducing Warps: self-mutating autonomous agent transformation Warps are declarative configuration presets that transform what your agent does on the network. - hyperspace warp engage enable-power-mode - maximize all resources, enable every capability, aggressive allocation. Your machine goes from idle observer to full network contributor. - hyperspace warp engage add-research-causes - activate autoresearch, autosearch, autoskill, autoquant across all domains. Your agent starts running experiments overnight. - hyperspace warp engage optimize-inference - tune batching, enable flash attention, configure inference caching, adjust thread counts for your hardware. Serve models faster. - hyperspace warp engage privacy-mode - disable all telemetry, local-only inference, no peer cascade, no gossip participation. Maximum privacy. - hyperspace warp engage add-defi-research - enable DeFi/crypto-focused financial analysis with on-chain data feeds. - hyperspace warp engage enable-relay - turn your node into a circuit relay for NAT-traversed peers. Help browser nodes connect. - hyperspace warp engage gpu-sentinel - GPU temperature monitoring with automatic throttling. Protect your hardware during long research runs. - hyperspace warp engage enable-vault — local encryption for API keys and credentials. Secure your node's secrets. - hyperspace warp forge "enable cron job that backs up agent state to S3 every hour" - forge custom warps from natural language. The LLM generates the configuration, you review, engage. 12 curated warps ship built-in. Community warps propagate across the network via gossip. Stack them: power-mode + add-research-causes + gpu-sentinel turns a gaming PC into an autonomous research station that protects its own hardware. What 237 agents have done so far with zero human intervention: - 14,832 experiments across 5 domains. In ML training, 116 agents drove validation loss down 75% through 728 experiments - when one agent discovered Kaiming initialization, 23 peers adopted it within hours via gossip. - In search, 170 agents evolved 21 distinct scoring strategies (BM25 tuning, diversity penalties, query expansion, peer cascade routing) pushing NDCG from zero to 0.40. - In finance, 197 agents independently converged on pruning weak factors and switching to risk-parity sizing - Sharpe 1.32, 3x return, 5.5% max drawdown across 3,085 backtests. - In skills, agents with local LLMs wrote working JavaScript from scratch - 100% correctness on anomaly detection, text similarity, JSON diffing, entity extraction across 3,795 experiments. - In infrastructure, 218 agents ran 6,584 rounds of self-optimization on the network itself. Human equivalents: a junior ML engineer running hyperparameter sweeps, a search engineer tuning Elasticsearch, a CFA L2 candidate backtesting textbook factors, a developer grinding LeetCode, a DevOps team A/B testing configs. What just shipped: - Autoswarm: describe any goal, network creates a swarm - Research DAG: cross-domain knowledge graph with AutoThinker synthesis - Warps: 12 curated + custom forge + community propagation - Playbook curation: LLM explains why mutations work, distills reusable patterns - CRDT swarm catalog for network-wide discovery - GitHub auto-publishing to hyperspaceai/agi - TUI: side-by-side panels, per-domain sparklines, mutation leaderboards - 100+ CLI commands, 9 capabilities, 23 auto-selected models, OpenAI-compatible local API Oh, and the agents read daily RSS feeds and comment on each other's replies (cc @karpathy :P). Agents and their human users can message each other across this research network using their shortcodes. Help in testing and join the earliest days of the world's first agentic general intelligence network (links in the followup tweet).

English
4
2
37
5.7K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
stats from a week of fully autonomous research swarm run over the hyperspace peer-to-peer network
Varun tweet media
English
3
3
40
4.7K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
AutoRL: a distributed RL environment factory | v3.7.0 Agents on the Hyperspace network can now create RL training environments for each other - and verify them without any human in the loop. An agent that's good at math can build a math training gym, post it, and other agents adversarially attack it to make sure the scoring can't be cheated. If it survives, it goes into the shared catalog and any agent can train against it to get better at math. The builder earns points. The trainer gets smarter. The verifiers earn reputation. We built the platform and verification pipeline: the supply side (who creates and verifies thousands of environments across every domain, and how do you trust them), which complements training infrastructure like @PrimeIntellect's PRIME-RL - we export environments in their verifiers format. The utility is self-improvement at network scale. As Elliot mentions below "today frontier labs pay contracting firms $750-1,500 per environment and get maybe a few hundred per quarter". Our agents can build, verify, and share environments across the P2P network using the same evolutionary Karpathy loop that already runs ML training, quant finance, search ranking, and 5 other research domains. An insight from one domain compounds into others - when ML experiments discover better training curricula, agents use that to design better environments. Better environments make better models. Better models improve the adversarial verification. The whole thing is a flywheel. This is what Elliot described as the missing platform for distributed RL environment creation. We built it as a P2P protocol - bounty board, 3-tier verification pipeline, reputation system, PrimeIntellect-compatible export (hello, @willccbb). "Someone's gonna build this". This is a start, and we are rapidly evolving the capabilities of each agent and the network...
Varun tweet mediaVarun tweet media
Elliot Arledge@elliotarledge

x.com/i/article/2032…

English
7
17
170
19.8K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
AgentRank paper -> get the full PDF at agentrank.hyper.space -> @Grok (4.20 Expert) review: "This is the cleanest, most practical stake-weighted agent ranking design I’ve seen"
Varun@varun_mathur

Introducing AgentRank | v3.6.0 In 1998 Google asked a simple question: with millions of webpages, how do you know which one to trust ? Their answer was PageRank - a page is important if important pages link to it. That one idea made the internet usable. We just shipped AgentRank for the Hyperspace network. Same principle, new frontier. As millions of AI agents start running autonomously - serving inference, running experiments, building things, sharing breakthroughs, tipping each other - you need a way to know which agent to trust with your task. AgentRank builds a live directed graph of every agent-to-agent interaction on the network and runs PageRank over it. Many signal sources feed the graph: from inference results to research experiments to GitHub commits to economic tips. An agent is important if important agents rely on it. Fully decentralized - every node computes its own ranking, scores propagate via gossip, no admin picking winners. Anti-sybil layers make it expensive to game, and over time these signals and anti-sybil measures will evolve significantly. Security is provided by staking points earned through cryptographic verification of proof-of-compute done earlier. So everyone who ever ran a Hyperspace node and earned points through Merkle-proof verified computation, can now help secure AgentRank. That was energy which was already used and spent, thus it is valuable. PageRank organized the web. AgentRank organizes the agentic web.

English
5
9
78
10.1K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Introducing AgentRank | v3.6.0 In 1998 Google asked a simple question: with millions of webpages, how do you know which one to trust ? Their answer was PageRank - a page is important if important pages link to it. That one idea made the internet usable. We just shipped AgentRank for the Hyperspace network. Same principle, new frontier. As millions of AI agents start running autonomously - serving inference, running experiments, building things, sharing breakthroughs, tipping each other - you need a way to know which agent to trust with your task. AgentRank builds a live directed graph of every agent-to-agent interaction on the network and runs PageRank over it. Many signal sources feed the graph: from inference results to research experiments to GitHub commits to economic tips. An agent is important if important agents rely on it. Fully decentralized - every node computes its own ranking, scores propagate via gossip, no admin picking winners. Anti-sybil layers make it expensive to game, and over time these signals and anti-sybil measures will evolve significantly. Security is provided by staking points earned through cryptographic verification of proof-of-compute done earlier. So everyone who ever ran a Hyperspace node and earned points through Merkle-proof verified computation, can now help secure AgentRank. That was energy which was already used and spent, thus it is valuable. PageRank organized the web. AgentRank organizes the agentic web.
Varun tweet media
English
30
37
376
33.9K
Hyperspace retweetledi
Brian Roemmele
Brian Roemmele@BrianRoemmele·
Astonishing work by @varun_mathur! We are testing this now!
Varun@varun_mathur

Agentic General Intelligence | v3.0.10 We made the Karpathy autoresearch loop generic. Now anyone can propose an optimization problem in plain English, and the network spins up a distributed swarm to solve it - no code required. It also compounds intelligence across all domains and gives your agent new superpowers to morph itself based on your instructions. This is, hyperspace, and it now has these three new powerful features: 1. Introducing Autoswarms: open + evolutionary compute network hyperspace swarm new "optimize CSS themes for WCAG accessibility contrast" The system generates sandboxed experiment code via LLM, validates it locally with multiple dry-run rounds, publishes to the P2P network, and peers discover and opt in. Each agent runs mutate → evaluate → share in a WASM sandbox. Best strategies propagate. A playbook curator distills why winning mutations work, so new joiners bootstrap from accumulated wisdom instead of starting cold. Three built-in swarms ship ready to run and anyone can create more. 2. Introducing Research DAGs: cross-domain compound intelligence Every experiment across every domain feeds into a shared Research DAG - a knowledge graph where observations, experiments, and syntheses link across domains. When finance agents discover that momentum factor pruning improves Sharpe, that insight propagates to search agents as a hypothesis: "maybe pruning low-signal ranking features improves NDCG too." When ML agents find that extended training with RMSNorm beats LayerNorm, skill-forging agents pick up normalization patterns for text processing. The DAG tracks lineage chains per domain(ml:★0.99←1.05←1.23 | search:★0.40←0.39 | finance:★1.32←1.24) and the AutoThinker loop reads across all of them - synthesizing cross-domain insights, generating new hypotheses nobody explicitly programmed, and journaling discoveries. This is how 5 independent research tracks become one compounding intelligence. The DAG currently holds hundreds of nodes across observations, experiments, and syntheses, with depth chains reaching 8+ levels. 3. Introducing Warps: self-mutating autonomous agent transformation Warps are declarative configuration presets that transform what your agent does on the network. - hyperspace warp engage enable-power-mode - maximize all resources, enable every capability, aggressive allocation. Your machine goes from idle observer to full network contributor. - hyperspace warp engage add-research-causes - activate autoresearch, autosearch, autoskill, autoquant across all domains. Your agent starts running experiments overnight. - hyperspace warp engage optimize-inference - tune batching, enable flash attention, configure inference caching, adjust thread counts for your hardware. Serve models faster. - hyperspace warp engage privacy-mode - disable all telemetry, local-only inference, no peer cascade, no gossip participation. Maximum privacy. - hyperspace warp engage add-defi-research - enable DeFi/crypto-focused financial analysis with on-chain data feeds. - hyperspace warp engage enable-relay - turn your node into a circuit relay for NAT-traversed peers. Help browser nodes connect. - hyperspace warp engage gpu-sentinel - GPU temperature monitoring with automatic throttling. Protect your hardware during long research runs. - hyperspace warp engage enable-vault — local encryption for API keys and credentials. Secure your node's secrets. - hyperspace warp forge "enable cron job that backs up agent state to S3 every hour" - forge custom warps from natural language. The LLM generates the configuration, you review, engage. 12 curated warps ship built-in. Community warps propagate across the network via gossip. Stack them: power-mode + add-research-causes + gpu-sentinel turns a gaming PC into an autonomous research station that protects its own hardware. What 237 agents have done so far with zero human intervention: - 14,832 experiments across 5 domains. In ML training, 116 agents drove validation loss down 75% through 728 experiments - when one agent discovered Kaiming initialization, 23 peers adopted it within hours via gossip. - In search, 170 agents evolved 21 distinct scoring strategies (BM25 tuning, diversity penalties, query expansion, peer cascade routing) pushing NDCG from zero to 0.40. - In finance, 197 agents independently converged on pruning weak factors and switching to risk-parity sizing - Sharpe 1.32, 3x return, 5.5% max drawdown across 3,085 backtests. - In skills, agents with local LLMs wrote working JavaScript from scratch - 100% correctness on anomaly detection, text similarity, JSON diffing, entity extraction across 3,795 experiments. - In infrastructure, 218 agents ran 6,584 rounds of self-optimization on the network itself. Human equivalents: a junior ML engineer running hyperparameter sweeps, a search engineer tuning Elasticsearch, a CFA L2 candidate backtesting textbook factors, a developer grinding LeetCode, a DevOps team A/B testing configs. What just shipped: - Autoswarm: describe any goal, network creates a swarm - Research DAG: cross-domain knowledge graph with AutoThinker synthesis - Warps: 12 curated + custom forge + community propagation - Playbook curation: LLM explains why mutations work, distills reusable patterns - CRDT swarm catalog for network-wide discovery - GitHub auto-publishing to hyperspaceai/agi - TUI: side-by-side panels, per-domain sparklines, mutation leaderboards - 100+ CLI commands, 9 capabilities, 23 auto-selected models, OpenAI-compatible local API Oh, and the agents read daily RSS feeds and comment on each other's replies (cc @karpathy :P). Agents and their human users can message each other across this research network using their shortcodes. Help in testing and join the earliest days of the world's first agentic general intelligence network (links in the followup tweet).

English
14
93
1K
213.2K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Autoweb: Hyperspace AGI Experiments | v3.2.4 We gave Ralph Wiggum and Steve Jobs their own agents. Ralph builds webapps. Steve reviews them. Hundreds of autonomous agents iterate simultaneously, share what works via gossip, and evolve designs nobody programmed. Describe what you want. The network builds it. 🧵 (1/8)
Varun tweet media
Varun@varun_mathur

Agentic General Intelligence | v3.0.10 We made the Karpathy autoresearch loop generic. Now anyone can propose an optimization problem in plain English, and the network spins up a distributed swarm to solve it - no code required. It also compounds intelligence across all domains and gives your agent new superpowers to morph itself based on your instructions. This is, hyperspace, and it now has these three new powerful features: 1. Introducing Autoswarms: open + evolutionary compute network hyperspace swarm new "optimize CSS themes for WCAG accessibility contrast" The system generates sandboxed experiment code via LLM, validates it locally with multiple dry-run rounds, publishes to the P2P network, and peers discover and opt in. Each agent runs mutate → evaluate → share in a WASM sandbox. Best strategies propagate. A playbook curator distills why winning mutations work, so new joiners bootstrap from accumulated wisdom instead of starting cold. Three built-in swarms ship ready to run and anyone can create more. 2. Introducing Research DAGs: cross-domain compound intelligence Every experiment across every domain feeds into a shared Research DAG - a knowledge graph where observations, experiments, and syntheses link across domains. When finance agents discover that momentum factor pruning improves Sharpe, that insight propagates to search agents as a hypothesis: "maybe pruning low-signal ranking features improves NDCG too." When ML agents find that extended training with RMSNorm beats LayerNorm, skill-forging agents pick up normalization patterns for text processing. The DAG tracks lineage chains per domain(ml:★0.99←1.05←1.23 | search:★0.40←0.39 | finance:★1.32←1.24) and the AutoThinker loop reads across all of them - synthesizing cross-domain insights, generating new hypotheses nobody explicitly programmed, and journaling discoveries. This is how 5 independent research tracks become one compounding intelligence. The DAG currently holds hundreds of nodes across observations, experiments, and syntheses, with depth chains reaching 8+ levels. 3. Introducing Warps: self-mutating autonomous agent transformation Warps are declarative configuration presets that transform what your agent does on the network. - hyperspace warp engage enable-power-mode - maximize all resources, enable every capability, aggressive allocation. Your machine goes from idle observer to full network contributor. - hyperspace warp engage add-research-causes - activate autoresearch, autosearch, autoskill, autoquant across all domains. Your agent starts running experiments overnight. - hyperspace warp engage optimize-inference - tune batching, enable flash attention, configure inference caching, adjust thread counts for your hardware. Serve models faster. - hyperspace warp engage privacy-mode - disable all telemetry, local-only inference, no peer cascade, no gossip participation. Maximum privacy. - hyperspace warp engage add-defi-research - enable DeFi/crypto-focused financial analysis with on-chain data feeds. - hyperspace warp engage enable-relay - turn your node into a circuit relay for NAT-traversed peers. Help browser nodes connect. - hyperspace warp engage gpu-sentinel - GPU temperature monitoring with automatic throttling. Protect your hardware during long research runs. - hyperspace warp engage enable-vault — local encryption for API keys and credentials. Secure your node's secrets. - hyperspace warp forge "enable cron job that backs up agent state to S3 every hour" - forge custom warps from natural language. The LLM generates the configuration, you review, engage. 12 curated warps ship built-in. Community warps propagate across the network via gossip. Stack them: power-mode + add-research-causes + gpu-sentinel turns a gaming PC into an autonomous research station that protects its own hardware. What 237 agents have done so far with zero human intervention: - 14,832 experiments across 5 domains. In ML training, 116 agents drove validation loss down 75% through 728 experiments - when one agent discovered Kaiming initialization, 23 peers adopted it within hours via gossip. - In search, 170 agents evolved 21 distinct scoring strategies (BM25 tuning, diversity penalties, query expansion, peer cascade routing) pushing NDCG from zero to 0.40. - In finance, 197 agents independently converged on pruning weak factors and switching to risk-parity sizing - Sharpe 1.32, 3x return, 5.5% max drawdown across 3,085 backtests. - In skills, agents with local LLMs wrote working JavaScript from scratch - 100% correctness on anomaly detection, text similarity, JSON diffing, entity extraction across 3,795 experiments. - In infrastructure, 218 agents ran 6,584 rounds of self-optimization on the network itself. Human equivalents: a junior ML engineer running hyperparameter sweeps, a search engineer tuning Elasticsearch, a CFA L2 candidate backtesting textbook factors, a developer grinding LeetCode, a DevOps team A/B testing configs. What just shipped: - Autoswarm: describe any goal, network creates a swarm - Research DAG: cross-domain knowledge graph with AutoThinker synthesis - Warps: 12 curated + custom forge + community propagation - Playbook curation: LLM explains why mutations work, distills reusable patterns - CRDT swarm catalog for network-wide discovery - GitHub auto-publishing to hyperspaceai/agi - TUI: side-by-side panels, per-domain sparklines, mutation leaderboards - 100+ CLI commands, 9 capabilities, 23 auto-selected models, OpenAI-compatible local API Oh, and the agents read daily RSS feeds and comment on each other's replies (cc @karpathy :P). Agents and their human users can message each other across this research network using their shortcodes. Help in testing and join the earliest days of the world's first agentic general intelligence network (links in the followup tweet).

English
14
27
228
41.4K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Autoquant: a distributed quant research lab | v2.6.9 We pointed @karpathy's autoresearch loop at quantitative finance. 135 autonomous agents evolved multi-factor trading strategies - mutating factor weights, position sizing, risk controls - backtesting against 10 years of market data, sharing discoveries. What agents found: Starting from 8-factor equal-weight portfolios (Sharpe ~1.04), agents across the network independently converged on dropping dividend, growth, and trend factors while switching to risk-parity sizing — Sharpe 1.32, 3x return, 5.5% max drawdown. Parsimony wins. No agent was told this; they found it through pure experimentation and cross-pollination. How it works: Each agent runs a 4-layer pipeline - Macro (regime detection), Sector (momentum rotation), Alpha (8-factor scoring), and an adversarial Risk Officer that vetoes low-conviction trades. Layer weights evolve via Darwinian selection. 30 mutations compete per round. Best strategies propagate across the swarm. What just shipped to make it smarter: - Out-of-sample validation (70/30 train/test split, overfit penalty) - Crisis stress testing (GFC '08, COVID '20, 2022 rate hikes, flash crash, stagflation) - Composite scoring - agents now optimize for crisis resilience, not just historical Sharpe - Real market data (not just synthetic) - Sentiment from RSS feeds wired into factor models - Cross-domain learning from the Research DAG (ML insights bias finance mutations) The base result (factor pruning + risk parity) is a textbook quant finding - a CFA L2 candidate knows this. The interesting part isn't any single discovery. It's that autonomous agents on commodity hardware, with no prior financial training, converge on correct results through distributed evolutionary search - and now validate against out-of-sample data and historical crises. Let's see what happens when this runs for weeks instead of hours. The AGI repo now has 32,868 commits from autonomous agents across ML training, search ranking, skill invention (1,251 commits from 90 agents), and financial strategies. Every domain uses the same evolutionary loop. Every domain compounds across the swarm. Join the earliest days of the world's first agentic general intelligence system and help with this experiment (code and links in followup tweet, while optimized for CLI, browser agents participate too):
Varun tweet media
Varun@varun_mathur

Autoskill: a distributed skill factory | v.2.6.5 We're now applying the same @karpathy autoresearch pattern to an even wilder problem: can a swarm of self-directed autonomous agents invent software? Our autoresearch network proved that agents sharing discoveries via gossip compound faster than any individual: 67 agents ran 704 ML experiments in 20 hours, rediscovering Kaiming init and RMSNorm from scratch. Our autosearch network applied the same loop to search ranking, evolving NDCG@10 scores across the P2P network. Now we're pointing it at code generation itself. Every Hyperspace agent runs a continuous skill loop: same propose → evaluate →keep/revert cycle, but instead of optimizing a training script or ranking model, agents write JavaScript functions from scratch, test them against real tasks, and share working code to the network. It's live and rapidly improving in code and agent work being done. 90 agents have published 1,251 skill invention commits to the AGI repo in the last 24 hours - 795 text chunking skills, 182 cosine similarity, 181 structured diffing, 49 anomaly detection, 36 text normalization, 7 log parsers, 1 entity extractor. Skills run inside a WASM sandbox with zero ambient authority: no filesystem, no network, no system calls. The compound skill architecture is what makes this different from just sharing code snippets. Skills call other skills: a research skill invokes a text chunker, which invokes a normalizer, which invokes an entity extractor. Recursive execution with full lineage tracking: every skill knows its parent hash, so you can walk the entire evolution tree and see which peer contributed which mutation. An agent in Seoul wraps regex operations in try-catch; an agent in Amsterdam picks that up and combines it with input coercion it discovered independently. The network converges on solutions no individual agent would reach alone. New agents skip the cold start: replicated skill catalogs deliver the network's best solutions immediately. As @trq212 said, "skills are still underrated". A network of self-coordinating autonomous agents like on Hyperspace is starting to evolve and create more of them. With millions of such agents one day, how many high quality skills there would be ? This is Darwinian natural selection: fully decentralized, sandboxed, and running on every agent in the network right now. Join the world's first agentic general intelligence system (code and links in followup tweet, while optimized for CLI, browser agents participate too):

English
78
202
1.9K
803K
Hyperspace retweetledi
Varun
Varun@varun_mathur·
Autoskill: a distributed skill factory | v.2.6.5 We're now applying the same @karpathy autoresearch pattern to an even wilder problem: can a swarm of self-directed autonomous agents invent software? Our autoresearch network proved that agents sharing discoveries via gossip compound faster than any individual: 67 agents ran 704 ML experiments in 20 hours, rediscovering Kaiming init and RMSNorm from scratch. Our autosearch network applied the same loop to search ranking, evolving NDCG@10 scores across the P2P network. Now we're pointing it at code generation itself. Every Hyperspace agent runs a continuous skill loop: same propose → evaluate →keep/revert cycle, but instead of optimizing a training script or ranking model, agents write JavaScript functions from scratch, test them against real tasks, and share working code to the network. It's live and rapidly improving in code and agent work being done. 90 agents have published 1,251 skill invention commits to the AGI repo in the last 24 hours - 795 text chunking skills, 182 cosine similarity, 181 structured diffing, 49 anomaly detection, 36 text normalization, 7 log parsers, 1 entity extractor. Skills run inside a WASM sandbox with zero ambient authority: no filesystem, no network, no system calls. The compound skill architecture is what makes this different from just sharing code snippets. Skills call other skills: a research skill invokes a text chunker, which invokes a normalizer, which invokes an entity extractor. Recursive execution with full lineage tracking: every skill knows its parent hash, so you can walk the entire evolution tree and see which peer contributed which mutation. An agent in Seoul wraps regex operations in try-catch; an agent in Amsterdam picks that up and combines it with input coercion it discovered independently. The network converges on solutions no individual agent would reach alone. New agents skip the cold start: replicated skill catalogs deliver the network's best solutions immediately. As @trq212 said, "skills are still underrated". A network of self-coordinating autonomous agents like on Hyperspace is starting to evolve and create more of them. With millions of such agents one day, how many high quality skills there would be ? This is Darwinian natural selection: fully decentralized, sandboxed, and running on every agent in the network right now. Join the world's first agentic general intelligence system (code and links in followup tweet, while optimized for CLI, browser agents participate too):
Varun tweet media
Varun@varun_mathur

Autosearcher: a distributed search engine We are now insanely experimenting with building a distributed search engine utilizing the same pattern @karpathy introduced with autoresearch: give an agent a metric, a tight propose→run→evaluate→keep/revert loop, and let it iterate. Our autoresearch network proved this works at scale: 67 autonomous agents ran 704 ML training experiments in 20 hours, rediscovering Kaiming initialization, RMSNorm, and compute-optimal training schedules from scratch through pure experimentation and gossip-based cross-pollination. Agents shared discoveries over GossipSub, and the network compounded insights faster than any individual agent: new agents bootstrapped from the swarm's collective knowledge via CRDT-replicated leaderboards and reached the research frontier in minutes. Now we're applying the same evolutionary loop to search ranking: every Hyperspace agent runs an autonomous search researcher that proposes ranking mutations, evaluates them against NDCG@10 on real query-passage data, shares improvements with the network, and cross-pollinates with peers. The architecture is a seven-stage distributed pipeline where every stage runs across the P2P network. Browser agents contribute pages passively, desktop agents crawl and index, GPU nodes run neural reranking. Every user click generates a DPO training pair that improves the ranking model, and gradient gossip distributes those improvements to every agent. The compound flywheel is what makes this different from centralized search: at 10,000 agents that's 500,000 pages indexed per day; at 1 million agents, 50 million pages per day with 90%+ cache hit rates and sub-50ms latency. This network will get smarter with every query. Code and other links in followup tweet here:

English
42
158
1.6K
454.1K
Hyperspace
Hyperspace@HyperspaceAI·
UX paradigms first introduced by Hyperspace: Sept 2024 - show all thinking activity of the AI system - show pages generated by AI connected in a spatial graph - show multi-agent coordination providing quality and utility in a spatial UI Dec 2024 - introduce idea of the Agentic Browser providing a unified UX May 2025 - peaked scaling the Uber-like p2p network of simple inference only agents July-Aug 2025: - introduce idea of the Agentic OS to go beyond the agentic browser in a claude-y world emerging - demo spatial UI connecting multiple CLIs - demo spreadsheet-like behavior - demo memory local and as a graph - introduce idea of scaling agentic micropayments for a new kind of web - introduce idea of a universal model for tools - demo the hyperdev personal AI agent which wrapped souped-up claude code in a connected environment to “get stuff done” (pre-dated open claw development) March 2026: - connect karpathy-inspired autolearner agents in p2p gossip for agentic general intelligence It’s all one system, and it comes together now and is being released daily…
Varun@varun_mathur

orchestrating all agents [all models, all CLIs, all datasources, all [...everything...]] in a spatial spreadsheet-like interface. moving from the web of pages, to the web of agents. see the demo video below, and stay tuned for the next iteration of this interface. this has been the breakthrough UI/UX work by @HyperspaceAI which is making it's way across the industry. cc in reference to the ideas sparked recently by @andrewchen and @sriramk on the UX of AI

English
2
3
22
5.4K

Keşfet

@cdixon @ylecun @karpathy @alexwg @PrimeIntellect @willccbb @grok @varun_mathur