How to Become an Outstanding AI Researcher: Lessons from an Anthropic Scientist

Choosing Your Own Problems

Richard Hamming famously asked colleagues at Bell Labs what the important problems in their fields were and why they weren’t being pursued, prompting many to change tables. The author argues that most researchers simply inherit problems from mentors, labs, or trending papers without understanding the underlying reasoning, leaving them ill‑equipped to assess why a direction might be abandoned.

Two Modes of Machine‑Learning Research (John Schulman)

Schulman distinguishes between (1) improving existing work after reading papers and (2) starting from a personally desired outcome and reverse‑engineering the experiments needed to achieve it. He advocates the second mode because it forces originality and pushes researchers into unexplored territory.

Training the “Research Muscle”

Research ability is built from many small, deliberately trainable skills: selecting questions, reading papers, writing, and accelerating experiment loops. Each skill is illustrated with concrete pitfalls the author has encountered.

Upgrading Your Input

Relying on shared reading lists or trending arXiv pages leads to convergent, low‑value conclusions. The author stresses the undervalued worth of older literature—e.g., expert models from 1991, LSTM from 1997, back‑propagation mainstreamed in 1986—and warns that “the bitter lesson” (Sutton, 2019) predicts field trajectories better than lengthy surveys.

Writing Everything Down

Paul Graham notes that ideas feel fully formed before they are articulated; writing exposes hidden gaps. The Feynman technique—being the first person you could fool—serves as a cheap defensive tool. Recording hypotheses, setups, expectations, results, and updated insights creates a personal audit trail that outperforms reviewer memory.

Closing the Feedback Loop

Speed of research depends on how quickly one discovers mistakes. Efficient tooling—single‑command runs, automated plotting, reproducible configurations—turns development into a core research activity. Andrej Karpathy’s tip: overfit a single batch before large‑scale training to eliminate half the bugs within 30 seconds.

Analyzing Results Properly

A descending loss curve is merely comforting; true analysis extracts information from logs, failure cases, and tail‑distribution anomalies that are often ignored.

Systematic Failure Mining (Andrew Ng)

Collecting and categorizing a hundred failure cases, then tackling the largest class, yields more insight than marginal accuracy gains. Understanding benchmark records at a textual level uncovers hidden lessons.

Purposeful Exploration

Early sub‑field choices are often accidental; researchers should deliberately explore explainability, evaluation, reinforcement learning, and systems to discover niche advantages. Running throw‑away experiments quickly helps prune ideas, and rigorous ablations reveal the single component driving results.

Broad Knowledge as Insurance

All sub‑fields eventually saturate; those who have breadth can continue producing work during transition periods, especially after a topic peaks on Twitter.

Finding Like‑Minded Peers

Hamming observed that closed‑door colleagues tend to accomplish more important work, as interruptions provide indirect information about what the world needs. Keeping communication channels open (e.g., an inbox) can surface valuable signals.

Generosity and Compound Returns

Reproducing results, publishing tools, and explaining complex ideas in plain language generate long‑term benefits—collaborations, citations, or unexpected job offers. Sharing half‑baked ideas early reduces the cost of mistakes compared to formal publications.

Long‑Term Game

Louis Pasteur’s maxim that “chance favors the prepared mind” underpins Hamming’s philosophy: knowledge and productivity compound like interest. Consistent daily habits—reading, recording, iterating, debating—accumulate advantages that appear modest in isolation but become decisive over years.