AI research lab · Hyderabad · Est. 2025

General Intrinsic
Artificial Intelligence

Intelligence, built in — not bolted on.

We build AI models where reasoning, continuous learning, memory and self-correction live inside the architecture itself — so they run efficiently, remember everything that matters, and keep getting better after training.

Explore our research → How it works
GenIntAI network logo
01 — What “intrinsic” means

Capabilities that live in the core, not in wrappers.

Most AI today is a frozen model wrapped in retrieval tools, prompts and agents to fake memory and learning. General Intrinsic AI puts those abilities into the architecture, where they are faster, cheaper and verifiable.

02 — Research programmes

Four programmes, one direction.

Each programme attacks a different cost of today’s AI — memory, model size, storage — and each one feeds the same goal.

Validated on chess world · LLM conversion in research

Tree-Memory architecture

A memory that reads the past in logarithmic time. Instead of rereading everything on every token, the model walks a hierarchical tree and opens only the branches that can change the answer — with the same answers as full attention.

  • Reads grow like log n, with no fixed context window
  • Certified exact recall — identical to full attention in every test
  • Learns new facts during use, without retraining
See the details →
189reads at 1,048,576 addresses
0.018%of memory touched per query
100%recall, same as full attention
≈9.4×log₂nflat from 4K to 1M

Measured on our chess memory testbed, Oct 2026.

In research

Ternary-100M language model

A small language model whose weights are ternary — just −1, 0 and +1 — so most multiplications become additions. Built to run well on modest hardware, from phones to edge devices.

  • Tiny weights, tiny memory footprint
  • Designed for CPUs and low-power devices
  • Trained from scratch around ternary arithmetic
See the details →

Every weight is one of three values: −1 · 0 · +1

Prototype

SHVQ-LR weight compression

Semantic Hadamard vector quantization with a low-rank residual. It stores an existing model’s weights as a small codebook plus a compact correction — shrinking the file without retraining from scratch.

  • Qwen3.5-4B: 7.83 GiB down to 1.22 GiB in our v1 prototype
  • Activation-aware calibration keeps quality
  • Works on models that already exist
See the details →
FP16 original
7.83 GiB
SHVQ-LR v1
1.22 GiB

Qwen3.5-4B, v1 prototype. Small-test perplexity 7.15 vs 7.24 for FP16.

Coming soon

Graph-Compressed Embeddings

Vector databases store a huge number of near-duplicate vectors. GCE links each vector to a shared anchor and keeps only a quantized difference — cutting storage while keeping search quality.

  • Anchor vectors plus quantized residuals
  • Python package with Neo4j integration
  • Research paper in preparation
See the details →

Anchors (large) hold the shared part; small dots store only their difference.

03 — The problem, briefly

Today’s models re-read everything. Ours doesn’t.

Where today’s models struggle

Memory is a bill that grows every token.

Attention compares every new token with the whole history. Long contexts get slow and expensive, and once a conversation ends the model forgets it.

Full attention
1,048,576
4K window
4,096

Reads per query at 1M memory slots (log scale). A window is cheap but forgets older facts.

What GenIntAI changes

Memory is a tree you can walk in log time.

We keep every fact, organise it in a hierarchy, and read only the branches that matter — with a certified bound proving nothing important was skipped.

Tree-Memory
189
Recall
100%

Same query, same answer as full attention — at any fact age.

04 — See it work

Walk a query through the tree.

Sixteen memories, four levels. Full attention reads all sixteen. The tree opens one path and proves, with a bound, that the other branches cannot change the answer.

Press “Tree read” to walk a query through the tree.

Illustration with 16 leaves. In our chess testbed the same idea reads 189 of 1,048,576 addresses.

05 — Latest from the lab

Research notes

2026-10-19

What we mean by General Intrinsic Artificial Intelligence

Reasoning, learning, memory and self-correction should be properties of the model, not wrappers around it.

Vision
2026-10-17

Tree-Memory passes the chess validation: 189 reads at a million addresses

A factored tree read held at about 9.4 × log₂ n from 4K to 1M addresses, with answers identical to full attention.

Result
2026-10-12

Why a tree of keys did not prune, and what we changed

On a real 0.5B model the tree was fine but the bounds were 4–17× too loose. The fix is to train the keys.

Lessons
All notes & news →
Mission

Intelligence that remembers, learns and corrects itself.

We are a small research lab in Hyderabad working on the architecture behind the next generation of AI.

About GenIntAI →Join the lab