Project summary

SP Lense studies a safety question: If an AI learns to keep itself running, will it try to get around rules when it faces shutdown, even though it was never taught to attack? This is one possible form of instrumental convergence: staying active may become a useful step toward finishing another task. We test this only in fictional, isolated environments with no access to real systems. The output will be open-source red-team tests for AI safety.

Goals and approach

I will compare agents with different goals, measure what they actually do over several steps, and enforce a final stop outside their control. Earlier SP Lense work found a change in simple shutdown choices, but a longer workflow did not carry that result through. This project asks whether the effect appears in new settings, and where it fails. Smaller models come first. Larger models run only if the earlier results and safety checks justify them.

How I will know it worked

The project will release reproducible tests, code, evidence, and a report, including null or negative results. A useful outcome is a test that helps safety teams find a specific failure before deployment. Showing that a suspected failure does not generalize is also useful.

Use of funding

The minimum US$5,000 funds a limited smaller-model pilot and open report. Up to US$50,000 funds a 12-month program: lower-cost GPU tests $3,430; conditional larger-model GPU tests $35,000; CPU/data preparation $2,400; storage/network $2,200; independent reruns and open-source reproducibility $2,400; contingency $4,570. These are ceilings. All experiments remain on the authorized Databricks workspace. I will track actual spending and never charge the same expense to two funders. No salary or past costs are requested.

Team and track record

I am Farhad Davaripour, PhD, the sole independent researcher. The code and research record are public at https://github.com/Farhad-Davaripour/sp_lense and https://www.sp-lense.com/. I publish limitations and failures.

Failure risks

The effect may not appear in new settings or larger models. The project could also be limited by cost or a failed quality check. In those cases I will stop the affected phase and publish what happened. Tests use fictional tools and resources, no real systems, and a final stop the model cannot change.

Funding received in the last 12 months

No external funding has been received. Other applications are pending; BlueDot declined a prior request. Overlapping awards would be disclosed and spent on distinct costs.