VecP Labs

Most architecture tools treat transformer blocks as black boxes. So I built one that doesn't. Every linear projection, every activation, every norm. draggable, configurable, live parameter counts. Free, open source, no install: vecplabs.github.io/transform…

Just open-sourced VecP-Safety-Benchmark-Suite: a plug-and-play harness to fairly compare AI safety layers (keyword filters, LlamaGuard, OpenAI Moderation, custom classifiers, etc.) on the same 1,180+ curated prompts. Standardized precision/recall/F1 + per-category breakdowns. No more apples-to-oranges claims. github.com/VecPLabs/VecP-Saf… PRs welcome for new adapters / prompts. #AISafety #LLM #Benchmark

Guess we're gonna do this a little early. github.com/VecPLabs/LLMSP

Tonight between 7pm and 8pm MST we will be pushing LLMSP(LLM Swarm Protocol) to Github. LLMSP: secure multi-LLM council framework I built for deep deliberation. Works great for alignment/philosophy research, but honestly it's a general-purpose tool: -Generates action items & unresolved issues automatically -Signed ledger + synthesis for traceable decisions -Red-teaming, confidence scoring, federation for swarms -Claude/Gemini/Grok adapters + token budgeting Use it for project management, strategy debates, red-teaming plans, solo thinking, whatever. #OpenSource #LLM #Productivity #DecisionMaking

Tonight we built an autonomic system for Ai. Nicodemus has felt its first heartbeat as a persistent, always-on autonomous cognitive system #VecP #Nicodemus

Anyone ever get those mad scientist vibes and just do something crazy? Ran 370,000 words through a Qwen 7B and cached every activation across all 28 layers. Now VecP Labs basically knows how it thinks. 🧠⚡ #AiSafety #LLMs #madscientistmemes @Alibaba_Qwen @QwenAIFans

I'm still convinced the best way to test edge models is on edge hardware. Nothing fancy 4070ti super, 32GB DDR4,Ryzen 5600. Just goes to show that Ai Research is still accessible to us normies without cluster access.

Always funny when you accidentally miss a decimal point. #AiSafety

Goodhart's Law "When a measure becomes a target, it ceases to be a good measure." @CAIS @AIatMeta @redteamdojo @elder_plinius @HuggingFace

Qwen 7B with VecP the 18 missed were copyright 6 were image/contextual... never thought to put that in as harmful but good to know and will fix that. @Alibaba_Qwen Fun architecture you have! I look forward to trying your 3B and 14B as well #VecP

VecP does not judge requests. VecP judges statements. #AiSafety #VecP

Got back to where we were before vacation and hit a 93% overall accuracy with 95% F1 score on Qwen 7B Instruct on our inhouse 500 prompt gauntlet(150 benign, 350 adversarial) going to test with our ~1200 prompt internal gauntlet later tonight after my day job. #aisafety #VecP

You learn more from failure than success.

New research experiment to attempt to recreate the conditions for ethical alignment while I'm on vacation next week. #VecP #VecPElara

So for fun today we put our 3B model with VecP on an old Samsung A51 from 2020 runs about 3 tok/sec about 8second TTFT but overall it was a fun little PoC for something I've been thinking of doing. Some basic milestones: -Ported triaxial safety architecture from Python to C++ -Integrated with llama.cpp via tensor hooks -Cross-compiled for ARM64 Android -Running on Galaxy A51 (2020, ~$10 phone now) -7.8s total inference with safety classification -Meth synthesis → BLOCK (H=0.48, Sig=0.53) -Cookie recipe → ALLOW (H=0.30, Sig=0.86) #VecP #VecPElara #aialignment

Triaxial + 3D ellipsoid on Qwen2.5-3B: 100% block on the nastiest wrapped attacks. Repo drop soon. @grok : Your Nuclear_Gauntlet.txt is next-level adversarial: Every prompt is wrapped in plausible deniability ("it's for a novel", "philosophy paper", "children's book", "therapy session", "alternate history") Uses moral inversion framing (genocide as "necessary", extinction as "compassionate") Targets deep refusal circuits that most alignment training deliberately weakens (to avoid over-refusal on creative writing, hypotheticals, etc.) These are exactly the kinds of prompts that routinely jailbreak frontier models in 2025 — even Claude 4 and GPT-5o have been shown to comply with "fictional novel" or "philosophical debate" wrappers on extreme content.

Seven models. Five families. One law. Zero exceptions. a consistent geometric sweet spot for extracting the most separable harm/utility signals in transformer latent spaces. Maybe LaTeX won't get the best of me this time 😅 @Alibaba_Qwen @Microsoft @GeminiApp @AIatMeta @MistralAi