You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$20,000 left to reach the funding goal.
I propose a six-month, openly documented study of feedback loops between visual understanding and image/video generation on personally owned local NVIDIA hardware. I want to determine whether retrieved corrections improve held-out task performance beyond simply giving a generator more attempts, and identify cases where memory or an unreliable visual evaluator makes outputs worse. This proposal was prepared with AI assistance. It describes proposed work, not completed Nemotron results.
Use a local Nemotron vision-language evaluator to inspect frames from locally generated images and short videos. Compare a single-pass generator, iterative feedback without memory, and feedback with reviewed memory, using the same generation settings and maximum compute budget. Keep development tasks and held-out evaluation separate. Add tests with stale or irrelevant memory and record all attempts. The primary study concerns inference-time feedback and memory, not frontier-scale training. Nemotron provides interpretation; a separate model generates media.
Many demonstrations change models, prompts, resolution and memory together. A fixed setup can help separate those effects. The public benefit would be reusable evaluation code and evidence about reliability, failure and resource cost. No claim is made that these tests directly reduce catastrophic AI risk.
Freeze a scoring rubric before final experiments. Report task success, unsupported critique, retries, latency, peak GPU memory and energy where available. Seek blinded human assessment for a feasible subset with appropriate consent and any required review. If it is unavailable, say so and limit the conclusions. The same model's self-score will not count as independent proof. Report uncertainty and negative findings.
During months one and two after funding and procurement, publish setup instructions, build the experiment runner, pilot tasks and freeze the core evaluation. Use months three and four for matched experiments, month five for memory failure tests, and month six for a reproducible release and final report. Release original code, protocols and permitted outputs. Third-party models and data remain subject to their licenses. Load models sequentially when needed; document offloading, quantization, frame sampling and resolution.
US$18,000 planning allowance for one complete NVIDIA GPU workstation, including GPU, host, RAM, storage, cooling and power supply. US$2,000 planning allowance for tax, delivery, power protection and procurement contingency. Total request US$20,000. No salary or cloud rental is requested. Research time and continuing operating costs require confirmation before accepting funding.
Prefer a single 96 GB NVIDIA RTX PRO 6000 Blackwell GPU if a suitable all-in system fits. This is not a vendor quote. Obtain a written configuration and total price before spending. If necessary, agree a revised budget or reduced-memory scope with the funder.
I am asking to purchase equipment in my name and retain it after the study for continued independent research. I have not established that the award agreement permits this. Before accepting funding or signing, I need written confirmation of title, allowable continuing use, transfer/resale restrictions and any return or repayment obligations. A research grant is not unrestricted spending money.
I am Roan Weigert, an AI researcher and developer relations engineer based in San Francisco. Relevant prior work includes an IEEE Access study of video-model cost-efficiency and a public benchmarking repository. My supplied video discusses practical agent memory. These artifacts support my interest and experience; they do not establish that the proposed local system already works. No NVIDIA sponsorship or endorsement is claimed. My professional role at GMI Cloud is separate from this proposed individual project; employer or university IP restrictions would be resolved before accepting funds.
IEEE paper: https://ieeexplore.ieee.org/abstract/document/11525424
Research code: https://github.com/roangws/-Evaluating-Computational-Efficiency-in-Generative-AI-Video-Models
Agent-memory video: https://www.youtube.com/watch?v=W9Es-16bXs8
Public profile: https://roanweigert.com/
The evaluator may reward its own errors; human review is useful and disagreement is a result. Short tasks may not generalize to longer video. Memory may overfit or store false corrections. The target GPU may consume most of the budget and force a revised configuration. A useful negative outcome is an openly documented failure or resource limit, rather than an exaggerated success claim.
In the last 12 months, I received US$15,000 in research funding from Sofia University. The allocation of that funding and any overlap with this equipment request still need to be confirmed before an award is accepted. The same underlying equipment budget is being presented to other funders. I will disclose awards and reduce, decline or agree separate scopes for overlapping support. No double payment for the same workstation is requested.
There are no bids on this project.