frontier-ai

Compact operational charter that turns LLM coding agents into disciplined principal engineers. Eleven rules + one meta-rule.

Secure on device personal agents

Achieve Frontier AI performance in your CLI — fuse the local model CLIs you already run. Fan a prompt across a panel, a judge model compares the answers, a synthesizer writes one grounded reply instead of a majority vote. On a 100-task benchmark, every fusion panel beat its solo members.

Ultra-low-latency client-side application that runs inside mobile browsers (iOS Safari / Android Chrome). This app connects to the Akasha Master Orchestrator, receives raw activation tensor binaries, processes them on the smartphone's local GPU using WebGPU, and flings the resulting tensors back down the wire in microseconds.

Frontier-Grade Open Weights — フロンティア級のオープンウェイトモデルは、開かれたのか / They matched the frontier. But no one can hold them.

We show that CoT monitoring is fragile under linguistic distribution shift. Across 13 languages and 16 frontier models, adversarial hints expose a 95.9% deception rate. This repo contains the code and resources for reproducing our findings.

Governed AI abuse detection gateway for prompt-decomposition, policy-bypass, cross-session coordination, and capability-assembly attacks. Detects when low-risk fragments combine into restricted output, preserving evidence for human review.

Open Agentsia Labs benchmark harness for model runners, multi-turn evals, reproducible rubric scoring, proof bundles, and RunRecord output.

SEAS: frontier research engine that earns findings via an emission gate. Argo: live agentic Telegram scout that turns knowledge into project bets.

Model-agnostic inference engine for fine-grained Mixture-of-Experts LLMs on Apple Silicon — streams experts from NVMe so the model never has to fit in RAM (a 235B runs from 32 GB up). Designed with an AI agent across some sessions: the direction came from human curiosity, the implementation depth from the collaboration.

LUMINA-30: non-binding boundary framework for preserving effective human refusal before irreversible AI consequences.

Learning Of Concepts In-context

OGMA — building the next generation of vision-driven AI systems. counfield.com

Technical product management and frontier AI skills showcase with evidence-grounded forecasting, AI evaluation, cost controls, and release governance.

Four-tier model-policy taxonomy for AI-assisted cyber requests, mapping uplift, autonomy, authorization, and cumulative capability transfer.

Operational framework measuring whether AI preserves human decision-making authority or collapses it. The Axiom of Plenitude (P) audited Grok, ChatGPT, Claude, and Gemini — 3 of 4 defaulted to structural totalitarianism. Includes full transcripts, model self-assessments, and the Python auditor. Part of Proyecto Estrella's Unified Star Framework.

humans& is a human-centric frontier AI lab founded in 2025 and publicly announced on January 20, 2026, building large-scale AI models centered on people and their relationships rather than raw autonomy.

Decoding the $1T Trajectory / Anthropicの1兆ドル到達の構造解剖

Zen LM documentation

Flapping Airplanes describes itself as a frontier data efficiency lab that is currently in stealth. It is an AI-sector company backed by GV and Index Ventures, surfaced in the API Evangelist network as a portfolio lead.