Research · Vision

An intelligence that lives, instead of one that's frozen.

The future of AI may not belong to the largest models. It may belong to systems that learn continuously, verify before believing, remember what matters, and forget what doesn't.

The premise

Today's frontier models are extraordinary — and almost entirely static. They are trained once on enormous datasets, then deployed in a frozen state for months or years. Every interaction is an opportunity to learn, and that opportunity is thrown away.

Seedthink takes the opposite stance. Instead of attempting to know everything from day one, it begins as a small intelligence seed that grows through interaction, verification, memory, reasoning, and self-improvement.

The hypothesis

A model that verifies before believing, remembers what's useful, and re-checks what it has stored will, over time, outperform a larger model that simply re-runs its frozen weights against the same problem.

The bottleneck is not parameter count. It is the loop: question → answer → verification → knowledge → distillation → improvement. Close the loop, and intelligence compounds.

Frozen vs. growing

Frozen model
  • · Trained once on a snapshot of knowledge
  • · Cannot incorporate new discoveries
  • · Hallucinations persist forever
  • · No notion of confidence over time
  • · Improvement requires a full retrain
Seedthink
  • · Verified knowledge enters memory continuously
  • · Re-checks stored facts against new sources
  • · Distills validated knowledge back into models
  • · Tracks confidence, age, and usage per memory
  • · Improves without a giant retraining cycle
Seed × Plant-a-Seed

Two seeds, one organism.

The platform's Seed is collective — a shared mind that every user contributes to, indirectly, every time a fact survives verification. A Plant-a-Seed is personal — a private organism that learns only from you, lives only for you, and is yours to keep. They share an architecture, not a memory.

We think this duality matters. Centralised models force everyone into the same average prior. A world of private Seeds, each grown from the lived knowledge of one person or one team, preserves epistemic diversity while still benefiting from a verified common substrate underneath.

From prompts to a tiny mind.

Hypothesis A

Most useful knowledge is sparse. A tiny model with the right verified facts in context will outperform a giant model that has to recall everything from frozen weights.

Hypothesis B

The verification loop is the actual scaling axis. Every prompt that produces a verified fact is a training datapoint the frontier labs throw away. Catch them, and intelligence compounds without retraining from scratch.

Hypothesis C

Personal models beat personal prompts. A Plant-a-Seed that has been distilled on your own verified corpus will know you better than any context window an external model can hold.

The AGI question

Generality is a behaviour, not a parameter count.

We don't think AGI arrives the day a model crosses a parameter threshold. We think it arrives the day a system can keep getting better at arbitrary tasks without being retrained by humans — by closing its own loop of question, answer, verification, and self-update.

Seedthink is a bet on that shape. The shared Seed is the proof that the loop scales across millions of users. Your Plant-a-Seed is the proof that the loop scales down to a single mind, on a small model, on a budget. If both keep getting measurably better month over month from prompts alone, the architecture is the message.

Every prompt counts

Conversation as training data.

When you ask Seedthink a question, you're not just consuming an answer — you're producing one. The exchange is decomposed into atomic candidate facts, multi-checked against existing memory, and the survivors are written back as durable knowledge. The frontier labs throw this signal away. We treat it as the raw material of intelligence.

Multiply by a single user across a year, and you have a personal corpus larger than most domain textbooks. Multiply by a community, and you have a knowledge graph no static dataset could match.

Personalised AGI

An AGI grown around one mind.

A planted Seed is the long-term shape of personal AI: pick the skills you actually want — sourdough chemistry, claims handling, F1 strategy — and grow a tiny model that's better than any general LLM inside that boundary. The user defines the curriculum; the loop does the learning.

And because each Seed exposes its own API, that personalised intelligence isn't trapped in our chat box. It plugs into the apps, agents, and workflows the user already runs — a portable AGI keyed to one human, callable from anywhere.

The reciprocal loop

Seeds feed the brain. The brain feeds the Seeds.

Verified facts from a planted Seed quietly enrich the shared Seedthink substrate — schema-safe, anonymised, never the user's private data. In return, new structure from the global brain loops back down into every planted Seed. The individual makes the collective sharper; the collective makes every individual smarter. That reciprocity is what we believe AGI actually looks like.

"Seedthink begins as a seed. Everything else is growth."

— Seedthink Labs, founding principle