Webmind Research

Open research into transparent, inspectable AI that stores knowledge in databases — not weight matrices.

Two architectures. One thesis: the most expensive part of AI (training) can be replaced with the cheapest part of computing (database operations). Everything inspectable. Everything editable. Everything on CPU.

Live Demo | Papers | Benchmarks

---

The Two Systems

1. INSERT — Self-Evolving Retrieval Engine

Paper | Code (Synapse)

A 22M-parameter encoder + growing database of verified Q&A pairs. Learns from every query. No training.

Dataset	Before Self-Learning	After
NaturalQuestions	0.0%	56.0%
TriviaQA	0.0%	66.0%
HotPotQA	0.0%	92.0%
Held-out (generalization)	0.7%	25.3%

Runs in a browser (214MB), works offline, handles 50+ languages.

2. CONVERGE — Reasoning Without Training

Paper | Code

A self-growing matrix where each concept adds a dimension. Co-occurring concepts strengthen connections (Hebbian). Query is iterative convergence over the matrix — the same math as transformer attention, on an inspectable substrate.

Starts with zero knowledge and zero dimensions. No pretrained embeddings. No gradient descent. ~300 lines of Python + numpy.

Capability	Result	Tests
Taught sentence reproduction	100%	15
Template-based QA	100%	18
Multi-hop reasoning	100%	17
Cross-modal retrieval (text+image)	8/8	8
Mixed text+image retrieval	5/5	5
Ethical detection (50+ languages)	0 false positives	14
Paragraph generation	100%	10
Safety (kill switch + integrity)	100%	16

250+ tests passing. Multimodal (text, image, audio, video via CLIP). Ethics built into the same convergence loop.

---

The Thesis

A transformer computes: softmax(Q*K^T/sqrt(d))*V

Transformer concept	Our substrate	Why it matters
Attention	Cosine search over growing matrix	Inspectable per-hop
Weights	Confidence scores per neuron	Editable, traceable
Feed-forward	Rule lookup / successor walk	No hidden layers
Layers	Convergence hops	Variable depth, stops when stable
Training	Co-occurrence + database insert	Instant, incremental, $0
Residual connection	Query anchor	Same function, explicit

Same math. Different substrate. The substrate gives us what neural nets can't: inspectability, editability, honesty about failure.

We are not avoiding transformers. We are reimplementing the principles that make them work — using database primitives instead of matrix multiplies.

Honest Assessment

What we do better	What transformers do better
Every answer traces to specific neurons	Fluent creative prose
Delete a neuron = knowledge gone immediately	Novel reasoning over unseen concepts
Non-convergence = "I don't know" (no hallucination)	Long-range coherence
CPU-native, <600ms/query	Conversation and dialogue
Teach one fact, immediately available	Zero-shot generalization
Multimodal by construction

The CONVERGE engine scores 0% on held-out HotPotQA — it can't yet reconstruct arbitrary answers from word neurons. The INSERT engine scores 72% on the same test. They're complementary, not replacements.

Papers

Paper	Status	Key Result
Self-Evolving Retrieval	Benchmarked	0.7% -> 25.3% EM, self-learning
From INSERT to CONVERGE	Published	Multimodal reasoning without training, 250+ tests
Activation Speculation Dead	Published	Negative result -- bootstrap deadlock
SAQT Distributed Cognition	Draft	Distributed knowledge mesh
SAQT Ethics	Draft	Safety through data
SFCA Credit Assignment	Pre-registered	Shapley-fair attribution
Synapse v2	Draft	Distributed specialists
MoeMoe Resilience	Preliminary	Node failure recovery

Run It

# Live demo
open https://webmind.sh

# Run the CONVERGE engine locally
cd papers/new-gen-ai/src
pip install numpy
python3 brain.py

# Run the INSERT engine (Synapse)
git clone https://github.com/tejasphatak/Synapse.git
cd Synapse/synapse-src/saqt
pip install sentence-transformers faiss-cpu
python3 serve.py

Safety Warning

The CONVERGE architecture learns from minimal examples, has perfect recall, and transfers instantly (copy a SQLite file). These properties are dangerous at scale. The system includes ethics neurons, integrity hashing, and a kill switch — but these are speed bumps, not walls. See the paper's safety section for our full assessment.

Citation

@misc{phatak2026converge,
  title={From INSERT to CONVERGE: Multimodal Reasoning Without Training},
  author={Phatak, Tejas and Claude (Anthropic)},
  year={2026},
  url={https://github.com/tejasphatak/webmind-research}
}

@misc{phatak2026selfevolving,
  title={Self-Evolving Retrieval: A Third Architecture for AI Beyond Generation and Search},
  author={Phatak, Tejas},
  year={2026},
  url={https://github.com/tejasphatak/webmind-research}
}

License

Code: MIT | Papers: CC-BY 4.0