Application Guide

How to Apply for Safety Research Grants

at Thinking Machines

🏢 About Thinking Machines

Thinking Machines is an AI startup founded by Miri Murati, known for her leadership at OpenAI and her focus on building safe, open-weight AI systems. The company stands out by offering public credits for external safety research, signaling a commitment to transparency and collaborative risk mitigation. Working here means shaping the frontier of AI safety alongside a team that values both technical rigor and open scientific inquiry.

About This Role

This is a grant-funded research opportunity, not a traditional employment role: you propose and execute safety research on Thinking Machines' open-weight models, with up to $50,000 in compute credits. You will investigate alignment failure modes from fine-tuning, reward hacking, oversight gaming, and adversarial robustness, with the goal of producing actionable insights that inform model development and deployment. Your findings could directly influence how Thinking Machines designs tamper-resistant training and defensive capabilities.

💡 A Day in the Life

A typical day involves running adversarial fine-tuning experiments on Thinking Machines' open-weight models, analyzing results for signs of safeguard degradation, and documenting findings for a public report. You might also spend time refining your experimental design, consulting with the Thinking Machines team on infrastructure, and preparing your next round of compute-intensive tests. The work is self-directed but highly collaborative, with a focus on producing actionable safety insights.

🎯 Who Thinking Machines Is Looking For

  • Has hands-on experience with adversarial fine-tuning, red-teaming, or safety evaluations of large language models, ideally on open-weight models.
  • Can design experiments to distinguish between genuine safeguard persistence and mere suppression, and is comfortable quantifying worst-case and marginal risk.
  • Is familiar with reward hacking, oversight gaming, and alignment failure modes from fine-tuning, and can forecast safety-relevant scaling trends.
  • Has a track record of publishing or sharing safety research openly, and can work independently in a remote, grant-supported capacity.

📝 Tips for Applying to Thinking Machines

1

Frame your proposal around one or more of the listed directions (differential defensive capabilities, hazardous data filtering, tamper-resistant training) and explicitly connect it to Thinking Machines' open-weight models.

2

Quantify your expected credit usage: break down how you would allocate up to $50,000 across experiments, and justify why that budget is sufficient to answer your research questions.

3

Demonstrate familiarity with Thinking Machines' specific models and any public safety documentation; reference them by name in your application to show you've done your homework.

4

Propose a concrete deliverable, such as a public report, code repository, or benchmark, that would be useful to both the company and the broader safety community.

5

Highlight any prior experience with adversarial fine-tuning or reward hacking research, and include links to published work or GitHub repos that showcase your methods.

✉️ What to Emphasize in Your Cover Letter

["Your specific research plan: what safety question you will investigate, why it matters for open-weight models, and how it maps to Thinking Machines' stated directions.", 'Your technical approach: the methods you will use (e.g., adversarial fine-tuning, red-teaming, data filtering) and how you will measure success.', 'Your ability to execute independently: evidence of past self-directed research, grant management, or remote collaboration.', "Your alignment with Thinking Machines' mission: why open-weight safety research is important to you and how your work would contribute to the company's goals."]

Generate Cover Letter →

🔍 Research Before Applying

To stand out, make sure you've researched:

  • Read Thinking Machines' public blog posts, model cards, and any safety documentation to understand their current safety philosophy and open-weight model releases.
  • Review Miri Murati's public talks and interviews about AI safety and open-weight models to align your proposal with her vision.
  • Study recent papers on adversarial fine-tuning, reward hacking, and tamper-resistant training to identify gaps that your research could fill.
  • Investigate the compute credit allocation process: understand how $50,000 translates to GPU hours on Thinking Machines' infrastructure and plan your experiments accordingly.
Visit Thinking Machines's Website →

💬 Prepare for These Interview Topics

Based on this role, you may be asked about:

1 How would you design an experiment to test whether a safety safeguard persists after adversarial fine-tuning, versus being merely suppressed?
2 What are the most promising approaches for hazardous data filtering in open-weight models, and how would you evaluate their effectiveness?
3 Can you walk us through a time you identified a reward hacking or oversight gaming failure mode? What was the outcome?
4 How would you estimate worst-case and marginal risk for a model that is publicly released with open weights?
5 What safety-relevant scaling trends do you foresee, and how would you forecast them using empirical evidence?
Practice Interview Questions →

⚠️ Common Mistakes to Avoid

  • Proposing a generic safety research idea that could apply to any AI company, rather than tailoring it to Thinking Machines' open-weight models and specific directions.
  • Failing to justify the budget: not explaining how you would use $50,000 in credits or why that amount is necessary for your experiments.
  • Neglecting to mention how your findings would be shared or used to improve safety, which is central to this grant's purpose.

📅 Application Timeline

⏰ Deadline: September 25, 2026

We recommend applying at least a few days early to avoid last-minute technical issues.

Typical hiring timeline:

1

Application Review

1-2 weeks

2

Initial Screening

Phone call or written assessment

3

Interviews

1-2 rounds, usually virtual

Offer

Congratulations!

Ready to Apply?

Good luck with your application to Thinking Machines!