MATS 9 Retrospective: 14-hour days, $40k compute budgets, and job hunt advice
A MATS alum reveals the brutal work ethic and massive compute costs behind AI safety research.
A LessWrong post by MATS 9 alum ‘beyarkay’ (who joined Team Shard with Alex Turner and Alex Cloud) offers an unvarnished look at the Machine Learning Alignment & Theory Scholars program. The author worked 10–14 hours a day, often running experiments overnight using up to 32 GPUs in 30-minute bursts — a tactic enabled by MATS covering the compute costs. The program budgeted $1k per week per fellow (doubled to $2k in MATS 10), but some RL-focused fellows burned through $40k+ over three months. The takeaway: serious AI alignment research demands far more compute than individuals can fund out-of-pocket.
Beyond compute, the retrospective warns against dividing attention by interviewing for jobs during the 3-month main program. For newcomers without PhDs or first-author papers, earning a strong recommendation from a MATS mentor is the highest-leverage move. The author suggests reserving job applications for the optional MATS extension, which is explicitly designed for that purpose. Overall, the post paints MATS as an intense, compute-rich environment where focused effort can rapidly transform a novice into a publishable researcher — but only if you’re willing to burn the midnight oil and max out your GPU allocation.
- Typical workdays were 10–14 hours, with experiments scheduled overnight to maximize GPU usage (e.g., 32 GPUs for 30 min vs 2 GPUs for 8 hours)
- MATS budgeted $1k/week per fellow (now $2k in MATS 10), but some RL experiments exceeded $40k over 3 months
- Job hunting during the main program is discouraged; the extension period is the right time to apply, as a mentor reference is often more valuable than a CV without prior research experience
Why It Matters
Insights from MATS reveal the true cost and workload of AI alignment research, guiding aspiring scholars on resource allocation and career strategy.