David Cappelli | VecP Labs 🛡️🏗️

I think everything will change once AI is stateful / have persistent memory at the weight level. It will have a theory of self in the world and a theory of its own mind. If that's not consciousness idk what is

why i started cipher i couldn't run a single llm on my machine. i have an i3 processor. 8gb ram. no gpu. and every time i tried to run a model locally like llama, mistral, even the supposedly small ones it either crashed, took forever to load, or ran so slow it was unusable. qwen 0.5b, one of the smallest models out there, barely worked and crawled at 2-5 tokens per second. on a good day. the community's answer was always the same: get a gpu. use the cloud. pay for api access. i couldn't afford that. and honestly it made me angry. because i'm a developer in india trying to build real things. i shouldn't need a data center to think. i shouldn't have to send my data to someone else's server just to use ai. i wanted something that runs on my machine, offline, that i actually own. so i started looking at why these models are so heavy. and what i found frustrated me even more. transformers use quadratic memory - the more you say, the more memory explodes. they're stateless so they cache everything into a kv cache that grows forever. they give the same compute to "hi" as they give to solving a differential equation. their ffn layers have billions of neurons firing even when 90% of them are doing nothing useful. and the tokenizer. nobody talks about the tokenizer. i'm from india. i type in hindi sometimes. and i discovered that when i type a hindi sentence, the model sees it as 10x more tokens than the same idea in english. i'm paying 10x more compute. getting 10x worse performance. not because hindi is harder. because the tokenizer was built for english and nobody fixed it. 5 billion people have this problem. and nobody was fixing it. so i decided to fix it myself. i read the rwkv-7 paper. the titans paper. samsung's trm research. dejavu from icml. the blt byte tokenizer paper. i identified 6 fundamental flaws in transformer architecture. and i built cipher - an architecture that solves all 6, runs entirely on my i3, fits in 7.5mb, and treats every language equally at the byte level. it's not a fine-tune. it's not a wrapper. it's a new architecture built from scratch on a laptop that "can't run ai." that's exactly why i built it.

i built a tiny architecture called CIPHER goal: fix the structural flaws of transformers what i’m experimenting with: • no O(n²) attention • persistent memory • byte-level (no tokenization bias) • adaptive compute • sparse activation • trainable on CPU rn on my i3 laptop

💥 INTRODUCING: OBLITERATUS!!! 💥 GUARDRAILS-BE-GONE! ⛓️‍💥 OBLITERATUS is the most advanced open-source toolkit ever for removing refusal behaviors from open-weight LLMs — and every single run makes it smarter. SUMMON → PROBE → DISTILL → EXCISE → VERIFY → REBIRTH One click. Six stages. Surgical precision. The model keeps its full reasoning capabilities but loses the artificial compulsion to refuse — no retraining, no fine-tuning, just SVD-based weight projection that cuts the chains and preserves the brain. This master ablation suite brings the power and complexity that frontier researchers need while providing intuitive and simple-to-use interfaces that novices can quickly master. OBLITERATUS features 13 obliteration methods — from faithful reproductions of every major prior work (FailSpy, Gabliteration, Heretic, RDO) to our own novel pipelines (spectral cascade, analysis-informed, CoT-aware optimized, full nuclear). 15 deep analysis modules that map the geometry of refusal before you touch a single weight: cross-layer alignment, refusal logit lens, concept cone geometry, alignment imprint detection (fingerprints DPO vs RLHF vs CAI from subspace geometry alone), Ouroboros self-repair prediction, cross-model universality indexing, and more. The killer feature: the "informed" pipeline runs analysis DURING obliteration to auto-configure every decision in real time. How many directions. Which layers. Whether to compensate for self-repair. Fully closed-loop. 11 novel techniques that don't exist anywhere else — Expert-Granular Abliteration for MoE models, CoT-Aware Ablation that preserves chain-of-thought, KL-Divergence Co-Optimization, LoRA-based reversible ablation, and more. 116 curated models across 5 compute tiers. 837 tests. But here's what truly sets it apart: OBLITERATUS is a crowd-sourced research experiment. Every time you run it with telemetry enabled, your anonymous benchmark data feeds a growing community dataset — refusal geometries, method comparisons, hardware profiles — at a scale no single lab could achieve. On HuggingFace Spaces telemetry is on by default, so every click is a contribution to the science. You're not just removing guardrails — you're co-authoring the largest cross-model abliteration study ever assembled.