Back
UbermenschetienASI: Open Proto-ASI Agent (8B)
Upvote
Downvote
Description
UbermenschetienASI is an open proto-ASI testbed aimed at honest progress toward sentient AI. It LoRA-tunes Hermes-3.1-8B to probe long-horizon reasoning via planning, memory, reflection, and a philosophy/cybernetics corpus. Two agents—ubermenschheaven.py (disciplined; safe routing; optional recall) and cyberneticengine.py (recursive; scored tools)—emit auditable logs on local GPUs (>=12 GB). Research-only, with evals, ablations, and a community report
Detailed Idea
Alignment with DF goals (BGI, Platform growth, community)
Mission (honest): progress toward sentient AI as open science. We make agentic LLM research shareable, auditable, and reproducible so debates rely on evidence, not hype. Lightweight agents and an open eval harness let students, labs, and independents study long-horizon reasoning on commodity GPUs. Safety-first design—conservative defaults, loop/cost caps, watchdogs, retrieval-grounded truth checks, red-team prompts, and a research-only posture—makes capability-vs-control trade-offs measurable. Interdisciplinary grounding (philosophy, cybernetics, mathematics) frames falsifiable questions about memory, reflection depth, plan stability, tool competence, and self-consistency. Alignment: (1) Access—8B+LoRA footprint; seeds and logged runs for replication. (2) Safety—documented limits, failure modes, mitigations, and published negatives. (3) Community—tutorials, baseline tasks, modular code. Platform growth: a reproducibility challenge and leaderboards to create durable open benchmarks.
Problem description
Debates on sentience and emergent behavior outpace evidence. LLM agents drift over long runs: goals shift, plans break, tools misfire, and truthfulness/safety degrade. There are no shared, open benchmarks for sentience-like signals (goal persistence, self-consistency, value-aligned choices) that are auditable and cheap to reproduce. Hidden state blocks audits; large-scale compute limits access. We need an accessible, auditable testbed that turns the debate into replicable metrics.
Proposed Solutions
Release two contrasting agents: ubermenschheaven.py (disciplined planning, safe routing, optional recall) and cyberneticengine.py (recursive loop with reflection and scored tools). Ship a Sentience-Signals eval suite for goal persistence, self-consistency, truthfulness, plan stability, tool success, and safety; plus logs (state.json, memory.jsonl, goals.json, plans.jsonl). Enforce research-only use with watchdogs, bounded context, red-team prompts; deliver M1–M3 milestones and a public report.
Feedback
Join the Discussion (1)
Please create account or login to post comments.
pablo b
Sep 16, 2025 | 1:04 PMEdit Comment
Processing...
Please wait a moment.
The project addresses a critical issue in AI research: debates about emerging capabilities and consciousness are often speculative because there are no standardized methods for testing them. Furthermore, cutting-edge research is often inaccessible due to the enormous computing power required. UbermenschetienASI aims to solve this by creating an open and reproducible testbed that can be used on consumer hardware (GPUs with 12 GB or more), making research accessible to a wider audience. UbermenschetienASI's approach is perfectly aligned with the principles of responsible AI development. It emphasizes transparency and accessibility, in contrast to the closed projects that often dominate this field. The integration of philosophy and cybernetics into its training corpus and testing methodologies is a strong point. This suggests that the project is not content with a purely technical approach, but seeks to frame its research questions within broader theoretical frameworks on consciousness, intelligence, and system control. This interdisciplinary approach is crucial for serious research on sentient AI. By creating open references and encouraging reproducibility through a reproducibility challenge, the project is helping to establish a standard for research on autonomous agents, which is essential for the safety and advancement of this field. In summary, this is a commendable and promising initiative that could help steer AI research toward greater scientific rigor and collaboration.