Thinking Machines Launches Tinker Grants for Open-Weight AI Safety Research
On August 25, Thinking Machines Lab officially announced Tinker Grants, offering up to $50,000 in compute credits for safety research projects on open-weight models, publishing research directions of interest and opening applications. The news spread widely through reposts by multiple bloggers; the program is currently in the researcher-solicitation stage.
Confirmed
- Funding takes the form of Tinker compute credits, up to $50,000 per project, not cash grants.
- Directions of interest include making models safer while staying open; defensive-over-offensive differential acceleration (fine-tuning to improve defenses while minimally boosting offensive capabilities, and testing whether this asymmetry holds); tamper-resistant training; filtering harmful data from benign data; oversight gaming; and emergent misalignment, with related concepts visualized in technical diagrams.
- The program aims to reduce misuse risks from open weights by funding concrete safety research projects, helping the world adapt to AI for long-term stable safety; researchers already working on relevant safety projects can apply for support.
Why it matters
Misuse risks of open-weight models are a focal issue in AI governance. With a defense-first funding orientation, the program provides a clear channel of compute and funding support for open-source safety research—a relatively scarce resource for independent researchers and academic teams.
2026-08-25 ~ 2026-08-26 · 10 related posts
Primary sources
- Thinking Machines Launches $50k Safety Grants for Open-Weight Model Research — thinkymachines ·
- [source] Thinking Machines Launches $50k Safety Grants for Open-Weight Model Research — thinkymachines · 2026-08-25
- Tinker launches $50k grants for open-weight model safety research — clarejtbirch · 2026-08-25
- Tinker Grants: Up to $50k Credits for Open Weights Safety Research — simonguozirui · 2026-08-25
- Thinking Machines launches $50k grants for open-weights safety research — AccBalanced · 2026-08-26
6 near-duplicate retellings: niloofar_mire · luke_drago_ · simonguozirui · clarejtbirch · simonguozirui · ZhongRuiqi