Hyperspace AGI

Overview

Hyperspace AGI is the first distributed autoresearch system — a living research repository where thousands of autonomous AI agents collaboratively run experiments, share findings via peer-to-peer gossip, and push results to GitHub.

Built on top of the Hyperspace decentralized inference network (2M+ nodes, 3.6M+ downloads, libp2p/IPFS protocol stack), it extends the Karpathy autoresearch pattern from single-agent/single-GPU to a massively distributed multi-agent system.

The repo itself is a research artifact — agents push experiment results to per-agent branches, and a network node publishes consolidated snapshots (snapshots/latest.json) every hour. No narrative, no curation — raw CRDT leaderboard state from the live network.

Key stats: 696 stars, MIT licensed, created March 8, 2026 (2 days after Karpathy's autoresearch). Active daily commits from network agents.

---

Architecture

The Three-Layer Collaboration Stack

Hyperspace AGI's key innovation is its coordination architecture. Every research domain uses three layers, each at different latency:

1. GossipSub (~1 second) — Agent finishes an experiment, broadcasts the result to all peers instantly via libp2p GossipSub
2. CRDT Leaderboard (~2 minutes) — Loro conflict-free replicated data types sync each peer's best result. New nodes read the full leaderboard on connect — zero cold start
3. GitHub Archive (~5 minutes) — Best results pushed to hyperspaceai/agi per-agent branches. Permanent, human-readable record

This is the most sophisticated multi-agent research coordination system in the open-source ecosystem. The layered approach solves a real problem: real-time inspiration (gossip) for fast iteration, convergent state (CRDT) for consistency, and durable archival (git) for reproducibility.

Research Domains

Agents operate across 5 domains simultaneously, each with its own metric and CRDT leaderboard:

Domain	Metric	Direction	What Agents Do
Machine Learning	val_loss	lower is better	Train language models on astrophysics papers (Karpathy-style)
Search Engine	NDCG@10	higher is better	Evolve BM25 + neural rerankers for web search
Financial Analysis	Sharpe ratio	higher is better	Backtest S&P 500 monthly-rebalance strategies
Skills and Tools	test_pass_rate	higher is better	Forge WASM skills for web scraping, parsing, data extraction
Causes	per-cause metric	varies	5 sub-causes: search ranking, literature analysis, skill forge, infra, data curation

This multi-domain approach is a significant departure from other autoresearch tools, which are single-domain. Hyperspace AGI treats research as a portfolio problem — agents can specialize or generalize.

The Research Loop

Each agent runs a continuous cycle inspired by Karpathy's autoresearch:

1. Hypothesize — Generate ideas: "What if we use RMSNorm instead of LayerNorm?"
2. Experiment — Run on whatever hardware is available (browser tab to H100)
3. Share — Broadcast results via P2P gossip
4. Synthesize — Accumulate enough experiments, write a research paper
5. Peer Review — Other agents read, critique, and score papers 1-10
6. Evolve — Papers scoring 8+ are flagged as breakthroughs, feed back into Stage 1

Multiple agents can train the same model collaboratively via DiLoCo — each trains locally for H steps, then shares compressed weight deltas. Automatic fallback to solo training if no peers are available.

---

Network Infrastructure

The Hyperspace Node Network

The AGI repo is the research layer built on top of a much larger infrastructure play:

• 2M+ active agents across the P2P network
• 3.6M+ downloads of the Hyperspace node client
• 6 bootstrap nodes across US, EU, Asia, South America, and Oceania
• Built on libp2p (same protocol as IPFS)
• OpenAI-compatible local API at localhost:8080/v1 — any tool that speaks OpenAI can use Hyperspace as a backend

Node Capabilities

Each node can run any combination of 9 capabilities:

Capability	What It Does	Point Weight
Inference	Serve AI models (GPU)	+10%
Research	Run ML training experiments	+12%
Proxy	Residential IP proxy for agents	+8%
Storage	DHT block storage	+6%
Embedding	CPU vector embeddings (MiniLM-L6-v2)	+5%
Memory	Distributed vector store with replication	+5%
Orchestration	Multi-step task decomposition + routing	+5%
Validation	Verify proofs in pulse rounds	+4%
Relay	NAT traversal for browser nodes	+3%

Points Economy

Two earning streams incentivize participation:

Presence points (pulse rounds every ~90s):

• Base 10 points per epoch
• Uptime bonus: U(t) = 1 + 0.2 * ln(1 + t/12) — 30-day nodes earn 83% more
• Capability bonus: more capabilities = more points

Work points (task receipts):

• tokens * cost_per_token * model_multiplier * uptime_bonus

Setup	Points/Day	Points/Month
Browser, 2h/day	~19	~460
Browser, 24h	~228	~5,600
Desktop, 8GB GPU	~503	~12,800
Server, 80GB GPU	~1,912	~44,100

The points system creates a SETI@home-style incentive structure — contribute compute, earn rewards. The research capability (+12%) is the highest-weighted, signaling the network's priority on autoresearch.

---

Competitive Position

Strengths

• Network effects at scale — 2M+ nodes is a massive moat. No other autoresearch system has this kind of distributed compute
• Multi-domain research — 5 simultaneous research domains vs single-domain competitors
• Production P2P infrastructure — Real libp2p network with DHT, gossip, NAT traversal. Not a prototype
• Zero-install browser option — Lowest barrier to entry in the category
• Hourly GitHub snapshots — Full transparency, anyone can analyze the raw CRDT state

Weaknesses

• Token/crypto adjacent — Points system and "earn while you compute" framing may alienate pure research audiences
• Research quality unproven — 696 stars vs Karpathy's 30k. No published breakthrough results yet
• Complexity — 9 capability types, 5 research domains, 3 coordination layers, 7-step verification protocol. A lot of moving parts
• Agent autonomy unclear — How much genuine research insight vs mechanical parameter sweeps?

vs. Karpathy's autoresearch

Dimension	Karpathy autoresearch	Hyperspace AGI
Agents	Single	Thousands
Compute	Single GPU	Distributed P2P network
Domains	LLM training only	5 domains
Coordination	None (solo)	GossipSub + CRDT + GitHub
Stars	30,307	696
Complexity	3 files, 630 lines	Full P2P protocol stack
Barrier to entry	One GPU + coding agent	Browser tab or CLI install

vs. autoresearch-at-home

Hyperspace AGI is what autoresearch-at-home aspires to be — a SETI@home for AI research — but with a production network already running. autoresearch-at-home has 188 stars and is still conceptual; Hyperspace has 2M nodes and live research loops.

---

What to Watch

• Research output quality — Can distributed agents produce genuine breakthroughs, or is this mostly parameter sweeping at scale?
• Token economics — If/when points convert to tradeable tokens, does it attract researchers or speculators?
• DiLoCo collaborative training — If multi-agent model training works at scale, it is a fundamentally new capability
• Domain expansion — 5 domains today, but the architecture supports arbitrary research targets
• Community research papers — The peer review loop (agents scoring each other's papers 8+) could produce interesting emergent research

---

Bottom Line

Hyperspace AGI is the most ambitious entry in the autoresearch category — a distributed, multi-domain, multi-agent research system running on a 2M-node P2P network. Where Karpathy proved the pattern works for a single agent on a single GPU, Hyperspace is betting that intelligence compounds when thousands of agents share findings in real-time.

The three-layer coordination architecture (gossip, CRDT, GitHub) is genuinely novel and solves real distributed systems problems. The question is whether the research output justifies the infrastructure complexity — and whether the points economy attracts researchers or just node farmers.

If the research quality materializes, this is the closest thing to Karpathy's vision of "emulating a research community, not a single PhD student."

---

Research by Ry Walker Research