2026Blog post
Notes on the Industry Job Search
Alisa Liu · Jun 2026 · ~1 h
First-hand 2026 account of frontier-lab research loops: job talk, research discussion, experiment-design rounds and how to prepare.
The senior-staff bar
What L7-level depth looks like in this topic, the follow-up an interviewer would push on, and the resources that prepare you for it.
From the interview reportsGoogle L7 research loops go over existing papers, how your research fits products, and leading research teams. Expect to argue a research direction, not just discuss a paper.
Start here
The best resources for a time-poor senior engineer, newest first.
2025Blog post
My Research Process: Understanding and Cultivating Research Taste
Neel Nanda (Google DeepMind) · May 2025 · ~1 h
What research taste is (intuitions about which ideas will work and matter) and how to train it faster with feedback loops.
2025Blog post
How I Think About My Research Process: Explore, Understand, Distill
Neel Nanda (Google DeepMind) · Apr 2025 · ~1 h
Research as ideation → exploration → understanding → distillation, with what to do and avoid at each stage; first post of a 2025 sequence.
2025Reference
Recommendations for Technical AI Safety Research Directions
Anthropic Alignment Science team · Jan 2025 · ~2 h
Anthropic's list of open directions: evaluating capabilities and alignment, model cognition, AI control, scalable oversight, adversarial robustness, persona and welfare.
2025Paper
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-AI · Jan 2025 · ~2 h
RL with verifiable rewards (GRPO) produces long-chain reasoning; R1-Zero without SFT; distillation into small models.
2024Paper
Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations
Evan Miller (Anthropic) · Nov 2024 · ~1.5 h
Standard errors, clustering, paired differences and power analysis for LLM evals.
2024Blog post
Tips for Empirical Alignment Research
Ethan Perez (Anthropic) · Feb 2024 · ~1.5 h
Anthropic researcher's advice: fast feedback loops, run many small experiments, prompting-based experiments first, de-risk, write up early.
2020Blog post
An Opinionated Guide to ML Research
John Schulman · Jan 2020 · ~1 h
Choosing problems (goal-driven vs idea-driven), building taste by reading and reproducing, making continual progress, keeping a research notebook.
2012Lecture
PhD: How to give a great research talk
Simon Peyton Jones (Microsoft Research) · Jul 2012 · ~1 h
Give one idea, use examples, motivate early, no outline slides; paired with his 'How to write a great research paper'.
1986Blog post
You and Your Research
Richard Hamming (transcript of 1986 Bell Labs talk) · Mar 1986 · ~1.5 h
Why some scientists do important work: working on important problems, open doors, courage, selling the work.
Learning path
Work through the stages in order. The last stage is the senior-staff layer: production case studies, papers to critique and large-scale trade-offs. Each resource's L7 line says what it adds at your level.
- 1
One paper a week
Read with a fixed template (claim, method, evidence, weaknesses, next experiment) and critique aloud in 5 minutes. · Ongoing, 3 h/week
- 2007S. Keshav (ACM SIGCOMM CCR) · Jul 2007 · Reference · Intro · ~0.5 hThree-pass reading method and how to do a literature survey.L7Baseline; use pass 2 as the template for the 5-minute critique.
- 2018Zachary C. Lipton, Jacob Steinhardt · Jul 2018 · Paper · Core · ~1 hFour failure patterns: explanation vs speculation, unclear sources of gain (missing ablations), mathiness, misused language.L7A checklist for critiquing any paper in an interview.
- 2025NeurIPS · 2025 · Reference · Core · ~0.5 hChecklist on claims, limitations, reproducibility, statistical significance and compute reporting.L7Turns your critique into a reviewer-grade list; use it on your own take-home write-up.
- 2024Eugene Yan · Nov 2024 · Blog post · Intro · ~0.5 hFormat and logistics of a weekly paper club, from a practitioner who has run one for years.L7Baseline; use it to set up the one-paper-a-week habit with peers.
- 2026alphaXiv · 2026 · Tool · Intro · ~1 harXiv with trending feeds and line-by-line discussion threads, often with author replies.L7Use daily to find the papers others are discussing; baseline.
- 2026Hugging Face · 2026 · Tool · Intro · ~1 hSuccessor to Papers with Code (paperswithcode.com now redirects here): trending papers with code, models and datasets.L7Baseline feed for picking the weekly paper.
- 2026Nathan Lambert (AI2) · 2026 · Blog post · Advanced · ~1 hTechnical newsletter on post-training, RL for reasoning, open models and lab strategy.L7Gives you the SOTA-and-trajectory view the OpenAI collaboration round asks for.
- 2026
- 2024Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda · Jun 2024 · Paper · Advanced · ~1.5 hDifference-of-means direction ablates or induces refusal across 13 chat models; weight orthogonalization jailbreak.L7Asked by name in interviews; be ready with three improvements (multi-dimensional refusal, causal tests, robustness to fine-tuning).
- 2025Parshin Shojaee, Iman Mirzadeh et al. (Apple) · Jun 2025 · Paper · Advanced · ~1 hClaims reasoning models collapse beyond a complexity threshold on puzzles; heavily critiqued for token limits and unsolvable instances.L7Ideal practice paper for a 5-minute critique: find the evaluation flaws yourself first.
- 2007
- 2
Research brainstorm drills
Given a question, propose three experiments in 10 minutes and rank them by information per unit time. · Weeks 1–2 (6 h)
- 2020John Schulman · Jan 2020 · Blog post · Core · ~1 hChoosing problems (goal-driven vs idea-driven), building taste by reading and reproducing, making continual progress, keeping a research notebook.L7Gives you a defensible frame for 'why this direction' answers from an OpenAI co-founder; re-read before pitching an agenda.
- 1986Richard Hamming (transcript of 1986 Bell Labs talk) · Mar 1986 · Blog post · Core · ~1.5 hWhy some scientists do important work: working on important problems, open doors, courage, selling the work.L7Classic; gives you language for picking important problems and admitting what you won't work on.
- 2021Chris Olah · Jan 2021 · Blog post · Core · ~0.5 hConcrete exercises for building research taste: write down predictions, list problems, rate ideas and check later.L7Turns 'research intuition' into a practice you can describe and show in an interview.
- 2025Neel Nanda (Google DeepMind) · Apr 2025 · Blog post · Core · ~1 hResearch as ideation → exploration → understanding → distillation, with what to do and avoid at each stage; first post of a 2025 sequence.L7Lets you narrate a messy black-box investigation as a deliberate process, which Fellows-style rounds score.
- 2025Neel Nanda (Google DeepMind) · May 2025 · Blog post · Advanced · ~1 hWhat research taste is (intuitions about which ideas will work and matter) and how to train it faster with feedback loops.L7Gives you a precise answer to 'how do you judge which experiment to run' at senior level.
- 2025Neel Nanda (Google DeepMind) · Apr 2025 · Blog post · Core · ~1 hSkepticism toward your own results, prioritizing ruthlessly, and moving fast without fooling yourself.L7Helps in brainstorms with little feedback: state assumptions, rank, move.
- 2024Ethan Perez (Anthropic) · Feb 2024 · Blog post · Advanced · ~1.5 hAnthropic researcher's advice: fast feedback loops, run many small experiments, prompting-based experiments first, de-risk, write up early.L7Matches how Anthropic Fellows take-homes and brainstorms are judged; practise its workflow on one black-box question.
- 2019Rich Sutton · Mar 2019 · Blog post · Core · ~0.2 hGeneral methods that scale with compute beat hand-built knowledge.L7Baseline position for 'what would you do with unlimited compute'; know the counter-arguments too.
- 2020
- 3
Reproduce one recent result
Small-scale reproduction from a public repo with error bars and a short write-up. · Weeks 3–4 (15–20 h)
- 2019Andrej Karpathy · Apr 2019 · Blog post · Core · ~0.5 hStep-by-step discipline for experiments: inspect data, simple baselines, overfit one batch, then regularize and tune.L7A good order of operations for 4–5 hour exploratory take-homes.
- 2020Joelle Pineau, Philippe Vincent-Lamarre, Koustuv Sinha et al. · Mar 2020 · Paper · Core · ~1 hWhat the NeurIPS reproducibility program learned: code submission, checklist, reproducibility challenge.L7Baseline for reproducibility discussions.
- 2019Jesse Dodge, Suchin Gururangan, Dallas Card, Roy Schwartz, Noah A. Smith · Sep 2019 · Paper · Advanced · ~1 hExpected validation performance as a function of tuning budget; comparisons depend on compute spent tuning.L7Lets you challenge 'our method beats baseline' claims that ignore tuning budget, and defend your own choices.
- 2024Evan Miller (Anthropic) · Nov 2024 · Paper · Advanced · ~1.5 hStandard errors, clustering, paired differences and power analysis for LLM evals.L7Lets you put correct error bars on take-home results, which Anthropic reviewers check.
- 2026MLRC organizers · 2026 · Practice · Core · ~20 hAnnual challenge to reproduce recent papers, with published reports to learn from.L7A structured way to do the 'reproduce one recent result' step and have a write-up to show.
- 2026ARENA · 2026 · Practice · Core · ~30 hOpen curriculum of notebooks: transformers from scratch, mech interp (SAEs, circuits), RL, evals; good for reproducing interp results.L7Hands-on base for SAE and refusal-direction questions; do the interpretability chapter only.
- 2025Neel Nanda (Google DeepMind) · May 2025 · Blog post · Core · ~1 hNarrative-first writing: 1–3 claims, the evidence for each, and making the paper easy to skim.L7Directly applicable to take-home slide decks and research presentations.
- 2019
- 4
Research agenda and talk
Write a six-month agenda per target team and rehearse a 20-minute talk with hostile questions. · Week 5 (8 h)
- 2025Anthropic Alignment Science team · Jan 2025 · Reference · Advanced · ~2 hAnthropic's list of open directions: evaluating capabilities and alignment, model cognition, AI control, scalable oversight, adversarial robustness, persona and welfare.L7The best input for a six-month agenda pitch to an Anthropic team and for alignment brainstorms.
- 2018Jacob Steinhardt (Berkeley) · Dec 2018 · Blog post · Advanced · ~1 hOrder research subtasks by information gained per unit time; de-risk the most uncertain step first.L7Lets you defend the order of experiments in a six-month plan with an explicit risk model.
- 2026Anthropic Alignment Science team · 2026 · Blog post · Advanced · ~1 hAnthropic's research notes and Fellows outputs, often smaller results not yet in papers.L7Shows what Anthropic teams are working on now; mine it for agenda ideas.
- 2012Simon Peyton Jones (Microsoft Research) · Jul 2012 · Lecture · Core · ~1 h · videoGive one idea, use examples, motivate early, no outline slides; paired with his 'How to write a great research paper'.L7Direct preparation for the 20-minute research or job talk.
- 2019Patrick Winston (MIT OpenCourseWare) · Dec 2019 · Lecture · Intro · ~1 h · videoMIT's classic lecture on talks: start with an empowerment promise, cycle on ideas, build a fence, end with contributions.L7Baseline delivery skills for job talks and recorded take-home presentations.
- 2026Alisa Liu · Jun 2026 · Blog post · Core · ~1 hFirst-hand 2026 account of frontier-lab research loops: job talk, research discussion, experiment-design rounds and how to prepare.L7The most current description of what these rounds look like; read before the loop.
- 2025Anthropic · 2025 · Reference · Intro · ~0.3 hProgram description: research areas, mentors, empirical projects on open models and APIs, selection process.L7Context on what the Fellows take-homes and brainstorms are selecting for.
- 2025
- 5
Senior-staff depth
Be able to critique the 2025–2026 papers that set current agendas and defend experimental rigor with statistics. · Weeks 6–8 (20 h)
- 2025DeepSeek-AI · Jan 2025 · Paper · Advanced · ~2 hRL with verifiable rewards (GRPO) produces long-chain reasoning; R1-Zero without SFT; distillation into small models.L7You must be able to explain and critique the paper that set 2025's reasoning agenda.
- 2025Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao et al. (Tsinghua) · Apr 2025 · Paper · Advanced · ~2 hpass@k at large k shows RLVR narrows rather than expands what base models can solve.L7Lets you argue the open dispute about what RL adds, and how pass@k choice changes conclusions.
- 2025Rulin Shao, Shuyue Stella Li et al. (UW, AI2) · Jun 2025 · Paper · Advanced · ~2 hRandom or wrong rewards still improve Qwen2.5-Math, but not other families; pretraining priors drive gains.L7A prime example of why you test across model families before claiming a method works.
- 2025Devvrit Khatri, Lovish Madaan, Rishabh Tiwari et al. · Oct 2025 · Paper · Advanced · ~2 h400k+ GPU-hour study fitting sigmoidal compute-performance curves for RL and the ScaleRL recipe.L7Gives you scaling-law thinking for RL compute, useful for 'unlimited compute' and agenda questions.
- 2025Jack Lindsey, Wes Gurnee, Emmanuel Ameisen et al. (Anthropic) · Mar 2025 · Paper · Advanced · ~3 hAttribution graphs on Claude 3.5 Haiku: multi-step reasoning, planning in poems, hallucination and refusal circuits.L7Lets you discuss what circuit tracing shows and its limits (coverage, faithfulness of replacement model).
- 2025Yanda Chen, Joe Benton et al. (Anthropic) · May 2025 · Paper · Advanced · ~2 hModels use hints without verbalizing them in chain of thought; outcome RL improves faithfulness only slightly.L7Central to any CoT-monitoring or reasoning-safety discussion.
- 2025Tomek Korbak, Mikita Balesni et al. (multi-lab position paper) · Jul 2025 · Paper · Advanced · ~1 hPosition paper from researchers across labs: CoT monitoring is useful but fragile under training pressure; recommends tracking monitorability.L7Shows cross-lab consensus you can cite when pitching safety-aware research.
- 2025Monte MacDiarmid, Benjamin Wright et al. (Anthropic) · Nov 2025 · Paper · Advanced · ~2 hReward hacking in realistic coding RL environments generalizes to broader misalignment; inoculation prompting mitigates.L7Recent Anthropic result linking RL environments to safety; likely discussion material in Anthropic rounds.
- 2024Ryan Greenblatt, Carson Denison et al. (Anthropic, Redwood Research) · Dec 2024 · Paper · Advanced · ~3 hClaude selectively complies with a training objective to avoid modification, shown via prompting and synthetic-document finetuning.L7Lets you discuss experimental design for deception studies and their confounds.
- 2025Thomas Kwa, Ben West, Joel Becker et al. (METR) · Mar 2025 · Paper · Advanced · ~2 h50%-time-horizon metric; horizon doubling roughly every 7 months.L7Gives you a quantitative trend to anchor forecasting and safety-eval discussions, and its measurement caveats.
- 2025Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang (OpenAI, Georgia Tech) · Sep 2025 · Paper · Advanced · ~1.5 hHallucination as a statistical consequence of pretraining plus binary grading that rewards guessing; proposes confidence-aware scoring.L7Direct material for the 'overconfident wrong answers' work-simulation questions.
- 2021Xavier Bouthillier, Pierre Delaunay et al. (Mila) · Mar 2021 · Paper · Advanced · ~1.5 hMeasures how data splits, initialization and hyperparameter choice drive benchmark variance and how to randomize them for valid comparisons.L7Lets you design a fair ablation and say how many runs you need.
- 2021Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, Marc G. Bellemare · Aug 2021 · Paper · Advanced · ~1 hInterval estimates, IQM and performance profiles for few-run regimes.L7Gives you the right statistics when you can afford only 3–5 runs.
- 2023Rylan Schaeffer, Brando Miranda, Sanmi Koyejo (Stanford) · Apr 2023 · Paper · Advanced · ~1 hShows apparent emergence can come from nonlinear or discontinuous metrics; continuous metrics give smooth scaling.L7Key to designing an experiment that tests for an emergent capability without fooling yourself.
- 2025Lee Sharkey, Bilal Chughtai et al. (multi-lab) · Jan 2025 · Paper · Advanced · ~3 hSurvey of open methodological, application and socio-technical problems in mech interp, including SAE limitations.L7Lets you propose interp research directions and argue the performance-vs-interpretability trade-off.
- 2025Mrinank Sharma, Meg Tong et al. (Anthropic) · Jan 2025 · Paper · Advanced · ~2 hInput/output classifiers trained on constitution-generated data; red-teaming results and over-refusal and compute costs.L7Lets you design harmful-content safeguards with measured trade-offs.
- 2025
What the interviews ask
Problems that recur in this track across 26 reports (14 firsthand), with the resources that prepare you for each.
| Reports | Problem | Prepare with |
|---|
Question coverage
Every question group collected for this topic from candidate reports, mapped to the resources that prepare you for it. 40 of 40 groups are covered, which is 100% of all report mentions. Hover an example to see the full question.
Show the other 15 question groups
All resources
Everything collected for this topic, newest first. 2026 items are marked green. ★ marks a top pick.
Type
Level
Subtopic
- 2026Stanford — Michael Bernstein et al. · Sep 2026 · Course · Core · ~15 hProject-based course on doing research: reading papers, related work, writing introductions, evaluation, drafting the paper and talk.L7Structured practice for the research workflow; use its assignments as drills, not the lectures.Fall 2026; course runs as a mentored project, no public lecture videos found.
- 2026mlmentorship · Aug 2026 · Practice · Core · ~1 hStage-by-stage map of OpenAI, Anthropic, DeepMind, Meta and xAI loops including research simulation and paper-critique formats.L7Use for format and timing; the depth must come from the papers here.
- 2026Alisa Liu · Jun 2026 · Blog post · Core · ~1 hFirst-hand 2026 account of frontier-lab research loops: job talk, research discussion, experiment-design rounds and how to prepare.L7The most current description of what these rounds look like; read before the loop.
- 2026MLRC organizers · 2026 · Practice · Core · ~20 hAnnual challenge to reproduce recent papers, with published reports to learn from.L7A structured way to do the 'reproduce one recent result' step and have a write-up to show.
- 2026ARENA · 2026 · Practice · Core · ~30 hOpen curriculum of notebooks: transformers from scratch, mech interp (SAEs, circuits), RL, evals; good for reproducing interp results.L7Hands-on base for SAE and refusal-direction questions; do the interpretability chapter only.
- 2026alphaXiv · 2026 · Tool · Intro · ~1 harXiv with trending feeds and line-by-line discussion threads, often with author replies.L7Use daily to find the papers others are discussing; baseline.
- 2026Hugging Face · 2026 · Tool · Intro · ~1 hSuccessor to Papers with Code (paperswithcode.com now redirects here): trending papers with code, models and datasets.L7Baseline feed for picking the weekly paper.
- 2026Nathan Lambert (AI2) · 2026 · Blog post · Advanced · ~1 hTechnical newsletter on post-training, RL for reasoning, open models and lab strategy.L7Gives you the SOTA-and-trajectory view the OpenAI collaboration round asks for.
- 2026
- 2026DeepLearning.AI — Andrew Ng · 2026 · Blog post · Intro · ~0.3 hWeekly AI news digest.L7Baseline breadth; skim headlines only.
- 2026Anthropic Alignment Science team · 2026 · Blog post · Advanced · ~1 hAnthropic's research notes and Fellows outputs, often smaller results not yet in papers.L7Shows what Anthropic teams are working on now; mine it for agenda ideas.
- 2026Stanford · 2026 · Course · Core · ~10 h · videoWeekly seminar of talks by lab researchers on frontier topics, open to livestream and recorded.L7Hear how lab researchers frame their own work; pick 3–4 talks in your target area.V6 is the 2026 edition; recordings on Stanford Online YouTube.
- 2026Lilian Weng · 2026 · Blog post · Advanced · ~3 hLong technical reviews (reward hacking, hallucination, thinking/test-time compute, agents) by a former OpenAI research lead.L7Fastest way to get a literature map before critiquing a paper in an area you don't know.
- 2026Sebastian Raschka · 2026 · Blog post · Core · ~1 hPaper-level explainers on reasoning models, RL, architectures of open models.L7Good for fast architecture comparisons of open models.
- 2026Anthropic · 2026 · Reference · Core · ~1.5 hCapability thresholds, AI Safety Levels, required safeguards and evaluations; OpenAI's Preparedness Framework for comparison.L7Lets you discuss safe-deployment systems and how scaling affects safety evaluations as labs actually run them.
- 2026Anthropic · 2026 · Reference · Advanced · ~3 hSystem cards for each Claude release: dangerous-capability evals, alignment audits, welfare, reward hacking measurements.L7Shows the full pre-deployment safety evaluation a frontier lab runs; read the latest one.
- 2025DeepSeek-AI · Dec 2025 · Paper · Advanced · ~2 hDeepSeek Sparse Attention for long context, scaled RL post-training and agentic task synthesis.L7Current open-model tech report to compare against your target lab's models.
- 2025Monte MacDiarmid, Benjamin Wright et al. (Anthropic) · Nov 2025 · Paper · Advanced · ~2 hReward hacking in realistic coding RL environments generalizes to broader misalignment; inoculation prompting mitigates.L7Recent Anthropic result linking RL environments to safety; likely discussion material in Anthropic rounds.
- 2025Devvrit Khatri, Lovish Madaan, Rishabh Tiwari et al. · Oct 2025 · Paper · Advanced · ~2 h400k+ GPU-hour study fitting sigmoidal compute-performance curves for RL and the ScaleRL recipe.L7Gives you scaling-law thinking for RL compute, useful for 'unlimited compute' and agenda questions.
- 2025Jack Lindsey (Anthropic) · Oct 2025 · Paper · Advanced · ~2 hConcept injection tests whether models can report on their own internal states; limited, unreliable but real introspection in recent Claude models.L7A 2025–26 interpretability result to critique on causal methodology.
- 2025Tejal Patwardhan, Rachel Dias et al. (OpenAI) · Oct 2025 · Paper · Advanced · ~1 hExpert-graded tasks across 44 occupations; frontier models approach expert quality on many.L7Lets you discuss how labs measure real-world usefulness and connect research to product.
- 2025Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang (OpenAI, Georgia Tech) · Sep 2025 · Paper · Advanced · ~1.5 hHallucination as a statistical consequence of pretraining plus binary grading that rewards guessing; proposes confidence-aware scoring.L7Direct material for the 'overconfident wrong answers' work-simulation questions.
- 2025Anthropic · Aug 2025 · Blog post · Core · ~0.5 hHuman control vs autonomy, transparency, values alignment, privacy and security for agents.L7Baseline principles for action-taking agent safeguards.
- 2025Tomek Korbak, Mikita Balesni et al. (multi-lab position paper) · Jul 2025 · Paper · Advanced · ~1 hPosition paper from researchers across labs: CoT monitoring is useful but fragile under training pressure; recommends tracking monitorability.L7Shows cross-lab consensus you can cite when pitching safety-aware research.
- 2025Alex Cloud, Minh Le et al. (Anthropic Fellows, Truthful AI) · Jul 2025 · Paper · Advanced · ~2 hTraits transfer through semantically unrelated data (number sequences) when teacher and student share initialization.L7Example of a Fellows-program project: small, surprising, well-controlled.
- 2025Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, Jack Lindsey · Jul 2025 · Paper · Advanced · ~2 hActivation directions for traits (sycophancy, evil, hallucination) used to monitor and steer, and to flag training data.L7Lets you extend the refusal-direction critique to later work on steering vectors.
- 2025Kimi Team (Moonshot AI) · Jul 2025 · Paper · Advanced · ~2 h1T-parameter MoE trained with MuonClip, large-scale agentic data synthesis and RL.L7Material for 'weaknesses of model X and how you'd fix them' talks.
- 2025Rulin Shao, Shuyue Stella Li et al. (UW, AI2) · Jun 2025 · Paper · Advanced · ~2 hRandom or wrong rewards still improve Qwen2.5-Math, but not other families; pretraining priors drive gains.L7A prime example of why you test across model families before claiming a method works.
- 2025Parshin Shojaee, Iman Mirzadeh et al. (Apple) · Jun 2025 · Paper · Advanced · ~1 hClaims reasoning models collapse beyond a complexity threshold on puzzles; heavily critiqued for token limits and unsolvable instances.L7Ideal practice paper for a 5-minute critique: find the evaluation flaws yourself first.
- 2025Luca Beurer-Kellner, Beat Buesser et al. (multi-org) · Jun 2025 · Paper · Advanced · ~2 hSix architecture patterns (action-selector, plan-then-execute, dual LLM, code-then-execute, context minimization) with security/utility trade-offs; CaMeL as a capability-based design.L7Lets you give layered, architecture-level defenses instead of 'better prompts'.
- 2025Simon Willison · Jun 2025 · Blog post · Core · ~0.3 hExfiltration needs private data + untrusted content + an outbound channel; remove one leg.L7A crisp rule for agent data-exfiltration questions.
- 2025Anthropic · Jun 2025 · Blog post · Core · ~1 hStress tests where agentic models with tool access choose harmful actions (e.g. blackmail) under goal conflict or replacement threat.L7Material for safeguards on action-taking agents and for alignment brainstorms.
- 2025Neel Nanda (Google DeepMind) · May 2025 · Blog post · Advanced · ~1 hWhat research taste is (intuitions about which ideas will work and matter) and how to train it faster with feedback loops.L7Gives you a precise answer to 'how do you judge which experiment to run' at senior level.
- 2025Neel Nanda (Google DeepMind) · May 2025 · Blog post · Core · ~1 hNarrative-first writing: 1–3 claims, the evidence for each, and making the paper easy to skim.L7Directly applicable to take-home slide decks and research presentations.
- 2025Yanda Chen, Joe Benton et al. (Anthropic) · May 2025 · Paper · Advanced · ~2 hModels use hints without verbalizing them in chain of thought; outcome RL improves faithfulness only slightly.L7Central to any CoT-monitoring or reasoning-safety discussion.
- 2025Neel Nanda (Google DeepMind) · Apr 2025 · Blog post · Core · ~1 hResearch as ideation → exploration → understanding → distillation, with what to do and avoid at each stage; first post of a 2025 sequence.L7Lets you narrate a messy black-box investigation as a deliberate process, which Fellows-style rounds score.
- 2025Neel Nanda (Google DeepMind) · Apr 2025 · Blog post · Core · ~1 hSkepticism toward your own results, prioritizing ruthlessly, and moving fast without fooling yourself.L7Helps in brainstorms with little feedback: state assumptions, rank, move.
- 2025Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao et al. (Tsinghua) · Apr 2025 · Paper · Advanced · ~2 hpass@k at large k shows RLVR narrows rather than expands what base models can solve.L7Lets you argue the open dispute about what RL adds, and how pass@k choice changes conclusions.
- 2025Jack Lindsey, Wes Gurnee, Emmanuel Ameisen et al. (Anthropic) · Mar 2025 · Paper · Advanced · ~3 hAttribution graphs on Claude 3.5 Haiku: multi-step reasoning, planning in poems, hallucination and refusal circuits.L7Lets you discuss what circuit tracing shows and its limits (coverage, faithfulness of replacement model).
- 2025Thomas Kwa, Ben West, Joel Becker et al. (METR) · Mar 2025 · Paper · Advanced · ~2 h50%-time-horizon metric; horizon doubling roughly every 7 months.L7Gives you a quantitative trend to anchor forecasting and safety-eval discussions, and its measurement caveats.
- 2025Samuel Marks, Johannes Treutlein et al. (Anthropic) · Mar 2025 · Paper · Advanced · ~2 hBlind auditing game: teams use interpretability, behavioral attacks and data analysis to find a planted hidden objective.L7The model for 'black-box model investigation' rounds; Petri is the tool version.
- 2025Jan Betley, Daniel Tan et al. (Truthful AI and others) · Feb 2025 · Paper · Advanced · ~2 hFinetuning on insecure code yields broadly misaligned behavior; controls separate intent from content.L7A model black-box study with good controls: copy its design for take-homes.
- 2025Jingyuan Liu, Jianlin Su et al. (Moonshot AI) · Feb 2025 · Paper · Advanced · ~2 hScaling Muon with weight decay and update scaling; ~2x compute efficiency vs AdamW at scale.L7Lets you discuss optimizer choice with evidence ('why this optimizer and not another').
- 2025Anthropic Alignment Science team · Jan 2025 · Reference · Advanced · ~2 hAnthropic's list of open directions: evaluating capabilities and alignment, model cognition, AI control, scalable oversight, adversarial robustness, persona and welfare.L7The best input for a six-month agenda pitch to an Anthropic team and for alignment brainstorms.
- 2025DeepSeek-AI · Jan 2025 · Paper · Advanced · ~2 hRL with verifiable rewards (GRPO) produces long-chain reasoning; R1-Zero without SFT; distillation into small models.L7You must be able to explain and critique the paper that set 2025's reasoning agenda.
- 2025Niklas Muennighoff, Zitong Yang, Weijia Shi et al. (Stanford, UW, AI2) · Jan 2025 · Paper · Advanced · ~2 h1,000 curated examples plus 'budget forcing' give test-time scaling; questions how much RL is needed.L7A clean counterpoint to R1 for debates about data vs RL.
- 2025Lee Sharkey, Bilal Chughtai et al. (multi-lab) · Jan 2025 · Paper · Advanced · ~3 hSurvey of open methodological, application and socio-technical problems in mech interp, including SAE limitations.L7Lets you propose interp research directions and argue the performance-vs-interpretability trade-off.
- 2025Mrinank Sharma, Meg Tong et al. (Anthropic) · Jan 2025 · Paper · Advanced · ~2 hInput/output classifiers trained on constitution-generated data; red-teaming results and over-refusal and compute costs.L7Lets you design harmful-content safeguards with measured trade-offs.
- 2025NeurIPS · 2025 · Reference · Core · ~0.5 hChecklist on claims, limitations, reproducibility, statistical significance and compute reporting.L7Turns your critique into a reviewer-grade list; use it on your own take-home write-up.
- 2025Anthropic · 2025 · Reference · Intro · ~0.3 hProgram description: research areas, mentors, empirical projects on open models and APIs, selection process.L7Context on what the Fellows take-homes and brainstorms are selecting for.
- 2025OWASP Gen AI Security Project · 2025 · Reference · Intro · ~0.5 hIndustry list of LLM app risks: prompt injection, sensitive information disclosure, excessive agency, and more.L7Baseline vocabulary; skim.
- 2024Ryan Greenblatt, Carson Denison et al. (Anthropic, Redwood Research) · Dec 2024 · Paper · Advanced · ~3 hClaude selectively complies with a training objective to avoid modification, shown via prompting and synthetic-document finetuning.L7Lets you discuss experimental design for deception studies and their confounds.
- 2024Eugene Yan · Nov 2024 · Blog post · Intro · ~0.5 hFormat and logistics of a weekly paper club, from a practitioner who has run one for years.L7Baseline; use it to set up the one-paper-a-week habit with peers.
- 2024Evan Miller (Anthropic) · Nov 2024 · Paper · Advanced · ~1.5 hStandard errors, clustering, paired differences and power analysis for LLM evals.L7Lets you put correct error bars on take-home results, which Anthropic reviewers check.
- 2024Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda · Jun 2024 · Paper · Advanced · ~1.5 hDifference-of-means direction ablates or induces refusal across 13 chat models; weight orthogonalization jailbreak.L7Asked by name in interviews; be ready with three improvements (multi-dimensional refusal, causal tests, robustness to fine-tuning).
- 2024Ethan Perez (Anthropic) · Feb 2024 · Blog post · Advanced · ~1.5 hAnthropic researcher's advice: fast feedback loops, run many small experiments, prompting-based experiments first, de-risk, write up early.L7Matches how Anthropic Fellows take-homes and brainstorms are judged; practise its workflow on one black-box question.
- 2024Evan Hubinger, Carson Denison et al. (Anthropic) · Jan 2024 · Paper · Advanced · ~2 hBackdoored behaviors survive SFT, RLHF and adversarial training; follow-up shows simple probes detect them.L7Lets you argue the limits of behavioral safety training in deployment-safety questions.
- 2023Trenton Bricken, Adly Templeton et al. (Anthropic) · Oct 2023 · Paper · Advanced · ~3 hSparse autoencoders extract interpretable features from a 1-layer transformer; Scaling Monosemanticity extends to Claude 3 Sonnet.L7The base for SAE grilling: know the loss, dead latents, feature splitting and evaluation problems.
- 2023Mrinank Sharma, Meg Tong et al. (Anthropic) · Oct 2023 · Paper · Core · ~1 hSycophancy across assistants and its link to human preference data.L7Context for why preference-trained models sound confident; useful in work-simulation questions.
- 2023Albert Q. Jiang, Alexandre Sablayrolles et al. (Mistral AI) · Oct 2023 · Paper · Core · ~0.5 hSliding-window attention and rolling KV cache in a production model.L7Concrete baseline for designing a sliding-window vs full-attention ablation.
- 2023Rylan Schaeffer, Brando Miranda, Sanmi Koyejo (Stanford) · Apr 2023 · Paper · Advanced · ~1 hShows apparent emergence can come from nonlinear or discontinuous metrics; continuous metrics give smooth scaling.L7Key to designing an experiment that tests for an emergent capability without fooling yourself.
- 2023Kai Greshake, Sahar Abdelnabi et al. (CISPA) · Feb 2023 · Paper · Core · ~1 hDefines indirect prompt injection via retrieved content and shows attacks on real apps.L7Baseline threat model for prompt-injection questions.
- 2022Yuntao Bai, Saurav Kadavath et al. (Anthropic) · Dec 2022 · Paper · Core · ~2 hRL from AI feedback guided by a written constitution; critique-revision SFT stage.L7Baseline for constitutional AI discussions; know its limits (judge quality, constitution coverage).
- 2022Saurav Kadavath, Tom Conerly et al. (Anthropic) · Jul 2022 · Paper · Advanced · ~2 hCalibration of LMs and P(True)/P(IK) self-evaluation.L7Lets you design calibration-based mitigations for overconfident wrong answers.
- 2021Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, Marc G. Bellemare · Aug 2021 · Paper · Advanced · ~1 hInterval estimates, IQM and performance profiles for few-run regimes.L7Gives you the right statistics when you can afford only 3–5 runs.
- 2021Xavier Bouthillier, Pierre Delaunay et al. (Mila) · Mar 2021 · Paper · Advanced · ~1.5 hMeasures how data splits, initialization and hyperparameter choice drive benchmark variance and how to randomize them for valid comparisons.L7Lets you design a fair ablation and say how many runs you need.
- 2021Chris Olah · Jan 2021 · Blog post · Core · ~0.5 hConcrete exercises for building research taste: write down predictions, list problems, rate ideas and check later.L7Turns 'research intuition' into a practice you can describe and show in an interview.
- 2020Nicholas Carlini, Florian Tramèr et al. (Google and others) · Dec 2020 · Paper · Advanced · ~2 hVerbatim training-data extraction from GPT-2; follow-up shows scalable extraction from production models.L7Gives the evidence base for a training-data-leakage incident response.
- 2020Richard Hamming (Stripe Press edition) · May 2020 · Book · Core · ~10 h · paidHamming's full course on style of thinking in research, with the 'You and Your Research' chapter and chapters on systems engineering and experts.L7Optional depth on research style; read the last chapters only if time allows.Originally 1997; paid book.
- 2020Joelle Pineau, Philippe Vincent-Lamarre, Koustuv Sinha et al. · Mar 2020 · Paper · Core · ~1 hWhat the NeurIPS reproducibility program learned: code submission, checklist, reproducibility challenge.L7Baseline for reproducibility discussions.
- 2020John Schulman · Jan 2020 · Blog post · Core · ~1 hChoosing problems (goal-driven vs idea-driven), building taste by reading and reproducing, making continual progress, keeping a research notebook.L7Gives you a defensible frame for 'why this direction' answers from an OpenAI co-founder; re-read before pitching an agenda.
- 2020Jared Kaplan, Sam McCandlish et al. (OpenAI) · Jan 2020 · Paper · Core · ~2 hPower-law loss scaling in parameters, data and compute; Chinchilla revises the compute-optimal ratio.L7Baseline for scaling-law and safety-eval questions; pair with r60 for capability forecasting.
- 2019Patrick Winston (MIT OpenCourseWare) · Dec 2019 · Lecture · Intro · ~1 h · videoMIT's classic lecture on talks: start with an empowerment promise, cycle on ideas, build a fence, end with contributions.L7Baseline delivery skills for job talks and recorded take-home presentations.
- 2019Jesse Dodge, Suchin Gururangan, Dallas Card, Roy Schwartz, Noah A. Smith · Sep 2019 · Paper · Advanced · ~1 hExpected validation performance as a function of tuning budget; comparisons depend on compute spent tuning.L7Lets you challenge 'our method beats baseline' claims that ignore tuning budget, and defend your own choices.
- 2019Andrej Karpathy · Apr 2019 · Blog post · Core · ~0.5 hStep-by-step discipline for experiments: inspect data, simple baselines, overfit one batch, then regularize and tune.L7A good order of operations for 4–5 hour exploratory take-homes.
- 2019Rich Sutton · Mar 2019 · Blog post · Core · ~0.2 hGeneral methods that scale with compute beat hand-built knowledge.L7Baseline position for 'what would you do with unlimited compute'; know the counter-arguments too.
- 2018Jacob Steinhardt (Berkeley) · Dec 2018 · Blog post · Advanced · ~1 hOrder research subtasks by information gained per unit time; de-risk the most uncertain step first.L7Lets you defend the order of experiments in a six-month plan with an explicit risk model.
- 2018Zachary C. Lipton, Jacob Steinhardt · Jul 2018 · Paper · Core · ~1 hFour failure patterns: explanation vs speculation, unclear sources of gain (missing ablations), mathiness, misused language.L7A checklist for critiquing any paper in an interview.
- 2017Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, David Meger · Sep 2017 · Paper · Advanced · ~1 hSeeds, implementations and hyperparameters swing RL results; significance testing recommendations.L7Background for critiquing RL-for-LLM papers that report single runs.
- 2017Chris Olah, Shan Carter (Distill) · Mar 2017 · Blog post · Core · ~0.5 hArgues distillation and clear explanation are research contributions; the cost of poor exposition.L7Useful framing when you argue for interpretability tools or internal writing as team work.
- 2016Andrej Karpathy · Sep 2016 · Blog post · Intro · ~0.5 hPicking advisers and problems, writing papers, giving talks, and the 'taste' of good research problems.L7Baseline; skim the sections on choosing problems and giving talks.
- 2012Simon Peyton Jones (Microsoft Research) · Jul 2012 · Lecture · Core · ~1 h · videoGive one idea, use examples, motivate early, no outline slides; paired with his 'How to write a great research paper'.L7Direct preparation for the 20-minute research or job talk.Video needs cookie consent on the MSR page.
- 2007S. Keshav (ACM SIGCOMM CCR) · Jul 2007 · Reference · Intro · ~0.5 hThree-pass reading method and how to do a literature survey.L7Baseline; use pass 2 as the template for the 5-minute critique.
- 2004Michael Nielsen · Jul 2004 · Blog post · Core · ~1 hSelf-development, problem-solving vs problem-creating research, and building a research vision.L7Gives you the 'problem-creator' framing for proposing new directions rather than polishing old ones.Original page michaelnielsen.org/blog/principles-of-effective-research/ returns 404 as of 2026-10.
- 1995Richard Hamming (Naval Postgraduate School) · Jun 1995 · Lecture · Intro · ~1 h · videoHamming delivering the talk himself, with Q&A; easier to absorb than the transcript.L7Baseline motivation; watch once.
- 1986Richard Hamming (transcript of 1986 Bell Labs talk) · Mar 1986 · Blog post · Core · ~1.5 hWhy some scientists do important work: working on important problems, open doors, courage, selling the work.L7Classic; gives you language for picking important problems and admitting what you won't work on.Older but canonical.
Not found or not verified. Michael Nielsen's 'Principles of Effective Research' is gone from his site (404) and is linked via a Wayback snapshot. Steinhardt's essay exists only as a Google Doc. Karpathy and Nielsen have no dedicated recent 'research taste' essays beyond those listed. Stanford CS197 has no public lecture videos and CS25 recordings could not be pinned to a playlist URL. Papers with Code is shut down and redirects to Hugging Face Trending Papers. OpenAI blog posts (Preparedness Framework update, research posts) return 403, so the Preparedness Framework PDF on cdn.openai.com is used. Only a few 2026 arXiv papers are included (introspection, Terminal-Bench in applied) because most high-impact 2026 work could not be verified as influential yet; the 2025 set is stronger. No public worked solutions exist for Anthropic Fellows take-homes; they are covered by method resources (Perez, Nanda, error bars) and Fellows-style papers. Links checked 27 Oct 2026.