@TomasMann1878 - Meta Research Engineer Year:
@TomasMann1878 on X A Meta research engineer describes a year working on recursive self-improvement, where AI agents help train the next generation of models. The post doubles as a practical guide to breaking into AI research, covering the skills, projects, and judgment the author thinks matter.
- Recursive self-improvement can work two ways: updating the foundation model itself, or improving the prompts, tools, memory, and code around it.
- Research engineering is roughly 90% coding and 10% idea generation, testing, and repeating until something works.
- Benchmarks can mislead, so read the methodology and inspect agent traces alongside leaderboard scores before trusting a result.
- A good starting project is a small agent harness on one task: run a baseline, change one thing, inspect traces, and write up what failed.
Tomas Mann, who started out fixing printers in technical support and has no academic background, describes a year as a Research Engineer at Meta working on recursive self-improvement (RSI), where AI agents train the next generation of models and improve themselves. His team mixes Research Scientists, Software Engineers, Research Engineers, and Data Scientists who largely do the same work, and he says the title matters less than whether you can help answer the research question. The day-to-day split is roughly 90% writing code and 10% generating ideas, testing them, and repeating. The self-improvement he works on runs along two routes: updating the foundation model itself, or improving the scaffolding around it, such as prompts, tools, memory, and code.
For getting in, his advice is concrete. Public papers, a benchmark, a Kaggle win, or a small agent harness are what recruiters and hiring managers can actually assess. His suggested starter project is to pick one task, build a small agent harness around it, run a baseline, change one thing, inspect the traces, and write up what failed. He lists the skills to build: understanding LLM training stages (pre-, mid-, and post-training) well enough to explain why a benchmark moved or a run broke; orchestrating agents through tool use, context management, and recovery from failure; using coding agents like Codex or Claude while still defining tasks, reviewing output, and verifying results; reading evaluations critically, using examples like PostTrainBench (small base models, one H100, ten hours, scored on GPQA, GSM8K, and HealthBench) and checking what the agent was allowed to do; and inspecting data for quality, coverage, and train/eval overlap, since coding agents don't remove that need.
He also makes a point about "taste," which he learned the hard way. He once ran an impressive-looking experiment, couldn't explain why the result had shifted when questioned, and found a bug that had inflated the numbers. Afterward he started preparing answers to likely questions: what he changed, why he expected it to help, and what evidence supported his explanation. His first paper, on Research Preference Models, addresses choosing which experiments to run when each is expensive. Its key figure compares inference-only and agentic selection against a no-guidance baseline with 95% confidence intervals. He recommends the work to people who like testing approaches and learning quickly from failure, and his closing advice is simply to run a small experiment, check the baseline, look at what happened, and say what you'd try next.