Researcher's Companion · Standalone
A Day in the Life of an AGI Researcher: The Complete Operating Manual
The practical follow-up to the Field Guide — an hour-by-hour walkthrough of a real research day, daily reading sources, the tool stack, a sample experiment from hypothesis to write-up, weekly rhythms, interview prep by stage, core concepts every researcher should know, conference prep, and an honest look at burnout and sustainability in this field.
Opening
A Day in the Life
How It Varies by Lane
Daily Reads & Sources
Tools & Software
Sample Experiment
Research Judgment
Writeups
Weekly Schedule
Interview Prep
Success Criteria
Core Concepts
Research Events
Keeping Up
Life Balance
What's Missing
Videos
References
Verdict
From Reading List to Real Life
The Researcher's Field Guide told you what to read and who succeeded before you. This article is about what the actual days look like once you're in the room — the reading habits, the tools, the judgment calls, the writing, the interview questions you'll actually be asked, and the parts of this work that are genuinely hard on a person, not just intellectually demanding.
A note on sourcing: The day-in-life walkthrough and weekly schedule below are a realistic composite built from public reporting on frontier-lab and academic research culture across this series — not a leaked schedule from any specific named individual. The burnout and life-balance section quotes real, sourced researcher statements, clearly attributed.
A Day in the Life: Hour by Hour
7:30-8:30 AM
Morning reading pass
Before anything else: a scan of Hugging Face Daily Papers or an arxiv-sanity-style personalized feed, plus whatever surfaced overnight on X/Twitter from researchers they follow. This is time-boxed deliberately — the goal is triage (what's worth a real read today?), not full comprehension of everything.
8:30-9:00 AM
Check overnight experiment runs
Review loss curves, eval scores, or overnight job logs from experiments left running — the first real decision of the day is often "does this thread continue or die," based on what came in overnight.
9:00-11:30 AM
Deep work block #1 — the actual research
The longest protected block of the day: designing an experiment, writing analysis code, reading a paper closely enough to reproduce a key result, or debugging why yesterday's run didn't behave as hypothesized. Meetings are deliberately avoided in this window at research-heavy organizations.
11:30 AM-12:00 PM
Async written update
A short written note (Slack, Notion, or an internal research doc) on what was found or attempted that morning — consistent with the async-over-meetings coordination style documented across this series' "Inside an AGI Startup Team" articles, especially in Lane 3.
12:00-1:00 PM
Lunch, ideally away from a screen
Often informal — the research-sync conversations that matter happen here as much as in scheduled meetings, per multiple lab-culture accounts referenced throughout this series.
1:00-2:00 PM
Research sync or reading group (2-3x/week)
The primary coordination mechanism at research-heavy labs — sharing findings, getting peer scrutiny on an interpretation, or presenting a paper to the group, as documented for both SSI and Goodfire in this series.
2:00-4:30 PM
Deep work block #2 — iteration
Second protected block, often used for iterating on the morning's findings, running follow-up experiments, or starting the write-up of a result that's stabilizing.
4:30-5:30 PM
Kick off overnight/long-running jobs
Launch the next experiment or training run before leaving, timed so results are ready to review the following morning — a rhythm that recurs constantly in compute-constrained research.
Evening (variable)
Optional: reading, or nothing
Some researchers read papers or write in the evening; many deliberately don't, given the burnout patterns discussed later in this article. This is the single most individually variable part of the day.
How This Shifts by Lane
| Element | Lane 1 (Frontier Scale) | Lane 2 (Applied) | Lane 3 (Narrow Research) |
|---|---|---|---|
| Morning priority | Training-run health check first, reading second | Usage dashboards and support tickets | Reading and overnight experiment review |
| Deep work interruption risk | High — infra incidents page anyone on-call | High — user-facing bugs are same-day priorities | Low — the whole day is protected by design |
| Meeting load | Moderate — cross-team compute allocation syncs | Low — async-first, PM-less structure | Very low — 2-3 syncs/week at most |
| Feedback loop length | Weeks (a training run) | Days (a shipped feature) | Months (a validated research thread) |
Daily Reads & Information Sources
Hugging Face Daily Papers
arxiv-sanity-lite
Papers With Code
X/Twitter research community
Import AI newsletter
Alpha Signal
Lab blogs (Anthropic, DeepMind, Transformer Circuits)
Dwarkesh Podcast & similar long-form interviews
Because arXiv alone publishes hundreds of AI papers daily, most researchers don't read arXiv directly as a firehose — Hugging Face Daily Papers ranks by community upvotes to surface what's actually being discussed, and Andrej Karpathy's arxiv-sanity-lite builds a personalized feed based on papers you've already starred. Newsletters like Import AI specifically summarize papers and explain why they matter, rather than just listing them — a meaningfully different (and faster) way to stay current than reading full papers daily. The practical discipline researchers describe: pick a small set of sources, consolidate them, and read at a fixed time rather than continuously — a roughly 10-minute daily routine is achievable if sources are narrowed deliberately rather than expanded indefinitely.
Tools & Software Stack
PyTorch / JAX
Weights & Biases / MLflow (experiment tracking)
Jupyter (reproducible analysis)
Notion (research notes, decisions, half-formed hypotheses)
Claude Code / GitHub Copilot (infra & tooling code)
Litmaps / Paperguide (literature review)
Slack (async coordination)
Anthropic Inspect AI / HELM (eval)
Notion's specific value for research, as documented in current 2026 tooling coverage, is that research generates far more context than clean findings — interview notes, decisions made and later reversed, half-formed hypotheses that might matter again in six months — and a flexible notes tool is what keeps that context retrievable rather than lost in someone's memory. Literature-review tooling has shifted from a single search-and-summarize step to a structured multi-stage workflow (plan → search → screen → extract → generate), cutting a task that used to take 80+ hours manually down to roughly 12-15 hours for a large search.
A Sample Experiment: Hypothesis to Write-Up
1
Day 1 — Hypothesis: Notice a pattern in eval failures (e.g., a model handles a class of reasoning tasks inconsistently) and form a specific, falsifiable hypothesis about why — not "the model is bad at X" but a concrete mechanistic guess.
2
Day 1-2 — Cheap test design: Design the cheapest possible experiment that could falsify the hypothesis quickly — a small ablation, not a full-scale run — following the "falsify fast" discipline documented in this series' Lane 3 build pipeline.
3
Day 2-4 — Run and monitor: Launch the experiment, monitor via W&B/MLflow dashboards, and resist the urge to over-interpret noisy early results before the run stabilizes.
4
Day 4-5 — Peer review the interpretation: Bring the result to a research sync before drawing conclusions publicly — per this series' documented pattern at SSI and Goodfire, peer scrutiny (not managerial sign-off) is the actual quality gate.
5
Day 5-7 — Decide: scale up, pivot, or kill: Based on the cheap test's result, either commit more compute to a larger-scale version, revise the hypothesis and retest, or write down why the thread is dead and move on — explicitly documenting negative results, not just positive ones.
6
Ongoing — Write-up: Whether the result is positive or negative, a written record goes into the team's internal research doc — the actual "product" of a research week is very often this document, not a shipped artifact.
Research Judgment: How Researchers Decide What's Worth Pursuing
As covered in the Frontier Map's "dead ends" analysis, the single hardest skill in this work isn't running experiments — it's recognizing early when a promising-looking direction is actually a dead end, and doing so before sunk-cost thinking sets in. A few heuristics that recur across researcher accounts referenced throughout this series:
Falsify Fast, Not Confirm Slowly
Design experiments to try to break your own hypothesis quickly, rather than running larger and larger confirmatory tests — the fastest route to a true negative is often more valuable than a slow route to a false positive.
Trust Peer Scrutiny Over Solo Conviction
A result that survives a skeptical peer's questions is worth far more than one that only survives your own review — the research-sync habit exists specifically to catch this before a result gets over-invested in.
Set a Kill Date Before You Start
Decide in advance how long an experimental thread gets before it must show a specific signal — without a pre-committed kill date, almost every thread starts to feel worth "just a bit more time."
Writeups & Documentation Practices
Research writing happens at three distinct levels of formality, and knowing which one a given result deserves is itself a judgment call:
1
Internal research docs (daily/weekly): Short, informal, written primarily for your future self and immediate teammates — the goal is retrievability, not polish. This is where most research writing time actually goes.
2
Blog posts (occasional): Neel Nanda's career-defining month of daily public blog posts (covered in the Field Guide) shows the format at its most impactful — public, informal, but rigorous enough to build a field's foundational vocabulary.
3
Formal papers (rare, high-stakes): Reserved for results significant and validated enough to withstand peer review — the least frequent form of writing by volume, but the one most visible externally and most tied to career advancement in academic-adjacent tracks.
A Weekly Preparation Schedule
| Day | Primary Focus |
|---|---|
| Monday | Review weekend/overnight results; set the week's experimental priorities; async update to the team on what's planned |
| Tuesday-Wednesday | Core experimentation — the week's longest uninterrupted deep-work stretch, before Thursday/Friday's coordination load rises |
| Thursday | Research sync/reading group; peer review of the week's emerging results; course-correct before the week ends |
| Friday | Write-up day — turning the week's findings (positive or negative) into a documented record; light experimentation only |
| Weekend | Optional reading; some researchers keep experiments running but avoid active monitoring, consistent with the sustainability practices discussed later in this article |
Interview Preparation: Questions to Prepare at Each Stage
Building directly on the documented process from the Field Guide, here's what to actually prepare for at each stage:
1. Recruiter Screen (~30 min)
Prepare: A clear, specific answer to "why this lab, why this research direction" — generic AGI-safety enthusiasm reads as weaker than a specific opinion about the lab's actual research bet.
Prepare: A concise walkthrough of your most relevant project or paper, focused on your specific contribution, not the whole team's.
2. Technical Screen(s)
Prepare: Core ML/RL fundamentals cold — backprop, attention mechanisms, optimization basics — even for senior roles, since these are baseline filters, not the differentiator.
Prepare: Direction-specific depth from the Field Guide's core papers — be ready to discuss a recent paper in your target direction in enough technical detail to defend or critique its methodology.
3. Take-Home (Anthropic-style, ~48hr window)
Prepare: Practice working through a genuinely ambiguous, underspecified problem rather than only clean textbook exercises — the take-home is explicitly designed to test rigor under ambiguity, not correctness on a well-posed question.
Prepare: A habit of documenting your reasoning as you go, not just your final answer — the write-up quality is part of what's being evaluated.
4. Research Presentation / Deep-Dive
Prepare: A presentation of your own past research that anticipates hard questions — practice defending your methodology choices, not just presenting results.
Prepare (Anthropic-specific): A genuine, considered position on AI safety questions relevant to your research area — this round explicitly tests safety-research fit, not just technical competence.
5. Onsite: Coding, Research Design, Behavioral
Prepare: Examples of times you changed your mind on a research direction based on evidence — this directly probes the "reasoning under ambiguity" trait labs explicitly screen for.
6. Final Round with Research Leadership
Prepare: A specific, honest answer to "what would you actually work on here in your first six months" — vague enthusiasm is a common failure mode at this final gate.
Success Criteria by Career Stage
| Stage | What "Doing Well" Looks Like |
|---|---|
| First 6 months | Land one clearly documented result (even a negative one) and demonstrate you can work independently on an ambiguous problem without excessive hand-holding |
| Year 1-2 | Own an experimental thread end-to-end; contribute a finding that changes the team's direction, however small |
| Year 2-4 | Develop a recognizable research "taste" — the specific judgment about what's worth pursuing that senior researchers are valued for; potentially publish or blog independently, per the Nanda pattern |
| Senior/Lead | Arbitrate compute allocation and thesis-continuation decisions for others, not just your own work; recruit and mentor; carry external credibility that helps the lab fundraise or hire |
Core Concepts Every AGI Researcher Should Know
1
Transformer architecture & attention mechanisms — the baseline architecture underlying nearly all frontier work today.
2
Scaling laws — how loss, data, parameters, and compute relate, and why Sutskever's own critique of "pre-training as scaling" specifically targets this relationship.
3
RLHF and Constitutional AI — the two dominant post-training alignment techniques covered in the Field Guide.
4
Mechanistic interpretability basics — circuits, sparse autoencoders, and the general idea of reverse-engineering a model's internal computation.
5
World models & self-supervised representation learning — JEPA-style architectures and why they're positioned as an alternative to pure next-token prediction.
6
Generalization & distribution shift — why a model can excel on benchmarks yet fail on subtly different real tasks (the "jagged intelligence" concept covered in the SSI investigation).
7
Meta-learning & continual learning — "learning to learn" and the challenge of updating a model without catastrophic forgetting.
8
Evaluation methodology — the difference between a benchmark score and a genuinely informative eval, and why eval design is treated as its own research discipline.
Research Events & Conference Preparation
Conferences serve three distinct purposes for a researcher: publishing/validating work, staying current on the field's frontier, and building the professional network that — per the Field Guide's recruiting analysis — matters more at frontier labs than cold applications. Preparation differs by purpose:
1
If submitting a paper: Work backward from the submission deadline (typically 4-6 months before the conference itself) — ICML 2026's historical deadline pattern lands around late January for a July conference, and NeurIPS's around May for a December conference.
2
If attending to learn: Pre-select a shortlist of papers and workshops from the program in advance — the volume of concurrent sessions at NeurIPS-scale conferences makes unplanned attendance highly inefficient.
3
If networking: Workshops and poster sessions are where direct conversations with paper authors actually happen — far more valuable for relationship-building than main-track talks, which are one-directional.
April
ICLR
Main conference Apr 23-25, workshops Apr 26-27, 2026 — a leading venue for representation learning and generative model research.
June
CVPR
The flagship computer vision conference — highly relevant for world-models and multimodal research directions.
July
ICML
Jul 6-12, 2026, Seoul — one of the two largest general ML venues, broad coverage across every direction in this guide.
August
ACL
The leading NLP/language-model venue — central for generalization and alignment research specifically tied to language models.
December
NeurIPS
Dec 6-12, 2026, Sydney, with satellite events in Atlanta and Paris — the largest and most influential general AI/ML conference by volume and prestige.
Rolling
Lab-hosted events & workshops
Anthropic, DeepMind, and OpenAI increasingly run their own smaller workshops and fellows programs (e.g., the Anthropic Fellows Program, covered in the Field Guide) — a lower-competition venue for building visibility than the largest conferences.
Exact deadlines shift year to year — always confirm against the official conference site (iclr.cc, icml.cc, neurips.cc) rather than relying on historical patterns alone, since dates above reflect 2026-specific scheduling that will not repeat identically in future years.
Keeping Up With a Field That Moves Weekly
1
Time-box triage, don't chase completeness. The researchers who stay current sustainably narrow their sources deliberately and read at a fixed time, rather than trying to track everything as it happens.
2
Depth over breadth within your direction. You cannot track all seven directions from the Field Guide at research-paper depth simultaneously — pick one as your primary and follow the rest only at a survey/newsletter level.
3
Let community curation do the filtering. Community-upvoted feeds (Hugging Face Daily Papers) and expert-curated newsletters (Import AI, Alpha Signal) exist specifically to save you from reading arXiv's full daily volume yourself.
Personal Life & Sustainability — An Honest Look
This is the part of the job least discussed in recruiting materials and most discussed once researchers leave. Public reporting from early 2026 documents a real pattern of senior departures across OpenAI and Anthropic, often accompanied by public statements about pace and personal sustainability.
"I cannot believe I would say this one day, but I am burnt out. All the mental health deteriorating that I used to scoff at is real, miserable, scary, and dangerous."
— A former OpenAI and xAI staffer, in a 2026 interview
The structural pattern reported across frontier labs: roughly two-to-three years of high intensity, followed by a sabbatical, a relocation, or a pivot to less demanding work — with burnout, not compensation, increasingly described as the dominant retention conversation inside these organizations. One documented contributing factor is technical rather than cultural: as the systems researchers build have grown more refined and numerically sensitive, the work itself has become harder to execute without pushing closer to personal limits, compounding the intensity independent of any lab's stated culture. Some researchers, per this same reporting, describe working long hours by genuine choice rather than external pressure — feeling that falling behind the field's pace is itself the risk they're managing against. Labs have begun building more formal sabbatical programs and on-call rotations in response, though multiple accounts describe the underlying work as "structurally hard to make sustainable" regardless of these policy additions.
The honest version of this section is not "here's how to balance it perfectly" — it's that the field's own senior practitioners are still actively working out whether sustainable long-term careers in frontier research are compatible with the field's current pace, and several credible, senior people have concluded, publicly, that for them the answer was no.
What's Missing From This Guide
1
No lab publishes an official day-in-life account — everything in the "day in the life" section is a reasonable composite built from adjacent public reporting, not a confirmed schedule from SSI, Goodfire, or any specific named lab.
2
Compensation-to-performance mapping remains undisclosed — as flagged in the Field Guide, no lab publishes how a specific interview or early-career performance level translates into offer or promotion decisions.
3
Sustainable long-term career patterns are still an open, unresolved question even among the researchers living it — this guide reports the honest current state of that conversation rather than a settled answer, because one doesn't yet exist publicly.
🔗 References
- arXiv Daily / arxiv-sanity
- Readless — How to Keep Up With AI News in 2026
- Paperguide — Best Literature Review AI Tools 2026
- ICLR 2026 Dates
- ICML 2026 Dates
- NeurIPS 2026 Dates
- IT Pro — AI Isn't Making Work Easier, It's Intensifying It
- Yahoo Tech — Former OpenAI/xAI Staffer on Burnout
- Interconnects — Nathan Lambert, "Burning Out"
- This site — The Researcher's Field Guide
- This site — The Research Frontier Map
🎥 Recommended Videos
Ilya Sutskever — We're Moving From the Age of Scaling to the Age of Research
We're All Addicted to Claude Code
Inside Claude Code With Its Creator
Agentic AI Projects 2026: Guardrails, Governance & Evals
Misha Laskin — From Physics to Superintelligence
$75M Founder Reveals His Agentic Engineering Setup
🧭 Closing — The Job Behind the Job Title
🎯 The Bottom Line
Strip away the funding numbers and lab prestige covered elsewhere in this series, and the daily reality of AGI research is a fairly ordinary-sounding set of habits done with unusual discipline: a time-boxed reading routine, protected deep-work blocks, fast experiments designed to fail cheaply, honest peer scrutiny, and consistent documentation of both successes and dead ends. What this article adds that a reading list alone can't: the same field that rewards this discipline has also, by its own senior practitioners' public accounts, made sustaining it over a full career genuinely difficult — the tools, the schedule, and the interview prep in this guide will get you into the room, but staying in it on your own terms is a separate, ongoing decision every researcher in this field is still learning to make well.