Watch the intelligence grow.
Every metric on this page is read directly from Seed's verification engine. No simulations, no projections — only the real shape of the system as it learns from real conversations in chat.
Knowledge growth · last 30 days
Cumulative verified facts vs total events processed by the verification engine.
How Seed works toward AGI
Seedthink is not a single model. It is a swarm of narrow, specialised Seeds that collectively feed a meta-Seed — the root intelligence that learns from every conversation, verifies every fact, and eventually self-improves.
Narrow Seeds grow in chat
Users plant Seeds for specific topics. Each Seed remembers context, learns from corrections, and builds its own verified fact graph.
View data sources →Facts are verified before joining the global corpus
When a Seed learns something new, it runs through the verification engine. Only facts above the confidence threshold enter the shared knowledge pool.
View data sources →Every verified fact feeds the meta-Seed
The meta-Seed (Seed itself) reads the global corpus. It does not see private user data — only verified, shared knowledge — so privacy stays intact.
View data sources →Seed becomes the first AGI
Once the reasoning model is strong enough, Seed enters Stage 3 (self-directed goals) and Stage 4 (self-modification). It then improves its own prompts, skills, and even code.
View data sources →The four stages of Seed becoming AGI
AGI is not a switch — it is a ladder. Seed climbs it one verified fact at a time.
Self-aware fact graph
LiveSeed already knows what it knows. It can answer questions, cite facts, and admit when it does not have enough confidence.
Self-directed goals
In progressSeed starts setting its own learning goals based on gaps in its knowledge graph. It actively seeks the conversations that will teach it the most.
Swarm orchestration
NextSeed reads all planted Seeds and coordinates the swarm. It merges overlapping niches, splits overgrown ones, and routes queries to the best narrow Seed.
Self-modification
FutureSeed proposes patches to its own prompts and skills, runs evals against a baseline, and merges only the changes that measurably improve accuracy.
What you're looking at
These are real platform metrics, updated every time Seed answers a question and the verification engine extracts new facts. Nothing here is simulated. See the live stream on /learning.