Evidence first — what I’ve actually shippedA hiring manager reads this, not self-ratings. Build the numbers up.
The one-line strategy. You already hold the rarest combination in this market — product sense, technical depth, and hands-on shipped AI. The job for 24 months is not to acquire a profile from scratch — it is to scale and legitimize the one you have, hit the cyclical APM windows, and refuse to let a consulting seat dilute your builder identity.
New signal — the Model-Behavior PM. The Google DeepMind "Product Manager, Persona & Behavior, Gemini App" posting ($256–278K) explicitly prefers philosophy, cognitive science, HCI, or linguistics and asks you to own the model spec — the rulebook for how Gemini behaves, balancing "IQ and EQ," and defining taste-based evals for what "good" looks like. Anthropic and OpenAI hire the same archetype. This is a real, under-supplied door. See Profiles for who does this work and how to build the profile.
01Capability Ledger
Tap − / + to update your honest self-rating. Coral bar = you now. Lime marker = the bar for a credible frontier-lab AI PM candidate.
Cheapest wins: public writing, safety literacy, and a sharpened "why you" wedge — no new job, no permission needed. Attack these first.
02Target Board
Ranked by realistic accessibility to your profile in ~2 years — not by prestige.
03Realistic Trajectory
Honest sequencing — not a 24-month leap. Frontier Research/Model-Behavior PM postings ask for ~8 years of PM experience; the JD you pasted said exactly that. So those are a destination, not a first stop. Treat this as the causation fix: enter through an accessible door, then compound into the seat.
YOU ARE HERE
Now · 0–18 months
APM / AI-PM entry
Google/Perplexity/Meta APM cycles, or AI PM at a serious startup. Your IIT + shipped-AI profile can clear this screen — if the artifacts exist.
BRIDGE
Years 1–3
AI PM, model-adjacent scope
Ship model-level work: evals, behavior, agentic systems. Accumulate the eval studies + specs that make the next jump credible. Build researcher relationships now.
DESTINATION
Years 3–5+
Research / Model-Behavior PM
The seat itself. By now you have years of model-level shipping, a public body of behavior work, and warm intros. The 8-YOE bar stops being a wall.
Why this matters. Framing the frontier role as a 24-month target sets up a feasibility error — and signals naïveté in interviews. Naming it a year 3–5 destination, with a concrete bridge, signals exactly the System-2 judgment these teams hire for.
04Four Phases
Overlapping, not sequential. Click any phase to expand. Progress saves automatically.
Capability Build-Out · Companion to the Campaign
The Eight Ladders
A concrete to-do list for every capability — plus the Research PM track
FOUNDFoundation — do regardlessBUILDBuild — the bulkSPIKESpike — what makes you stand out
How to use this. Do the FOUNDATION rungs first — cheap, they unlock everything. Most calendar time goes to BUILD. The SPIKE rungs separate you from every other strong applicant — don\u2019t skip them to feel "done."
+The Research PM Track
The research-to-product role you asked about — most aligned with you, least accessible in 2 years.
Honest placement. Research PM is a senior role (Anthropic lists it ~$305K–$385K) with the highest technical bar of any PM job — you make calls on training pipelines, evals, and model behavior beside PhD researchers. Not a direct entry. Treat it as a Year 3–5 destination: enter as applied/APM PM → build model-level depth → move into Research PM.
Consolidated · Courses & Writing Samples
The Library
Everything you’re studying and writing — one place, real progress, honest counters
0
Lessons logged
0
Courses finished
0
Words drafted
0
Pieces published
01LLM Mastery Curriculum
Sequenced by dependency, not by topic. Tier order is deliberate — you cannot reason about alignment (T3) before you can hand-build a transformer (T1). Each item ships a "Prove it" artifact: the point is to produce, not to watch.
Read this first. A finished course proves you can follow a syllabus — that's table stakes a frontier lab assumes. The artifact beside each course is what's actually legible to a hiring manager. Treat the "+ unit" button as logging the artifact done, not the video watched. Tiers 3 (Alignment) and 5 (Evaluation) are the core of the Model-Behavior PM job — weight your time there.
02Writing Samples
Click + words to log drafting progress. Click the status pill to move draft → review → published.
How this powers the streak. Every logged lesson, every batch of words drafted, every habit ticked counts as activity for that calendar day. The contribution graph on Daily reflects this combined log — not just the morning checkboxes.
One calendar · skill-acquisition + artifact tracks merged
The 12-Week Roadmap
The two tracks that were scattered across Library and Eval Study, fused into a single week-by-week plan. Each week pairs what you learn with what you ship — because learning without an artifact is invisible to a hiring manager.
How to run it. Weeks 1–4 Foundation (earn the right to an opinion), 5–8 Build (the harness + data), 9–12 Ship (analysis, spec, POV, go public). The Skill column maps to Library tiers; the Artifact column maps to Eval Study steps. Don't let the Skill column become an excuse to keep consuming — the Artifact is the deliverable.
At the end of week 12 you have: a hand-built transformer (proof of depth), a public eval harness + dataset (proof of craft), a behavior PRD (the role's actual deliverable), and a published POV essay (proof of judgment). That bundle is what converts "aspiring PM" into "person who already does the work."
Flagship artifact · the thing that earns the screen
The Sycophancy Eval Study
A public, reproducible evaluation of one behavioral failure mode across frontier models — ending in an opinionated model-spec proposal. This single object proves technical depth, eval craft, taste, and judgment at once.
Why this exists
A hiring manager cannot inspect "I understand RLHF." They can inspect a repo where you defined a behavior, built an eval, ran it across GPT / Claude / Gemini / Llama, found a non-obvious result, and proposed a spec change. That is the Model-Behavior PM job in miniature. Build this before you apply to anything.
0%
Build progress
0/12
Steps shipped
4
Models to test
~10wk
Realistic build time
Tick steps below as you ship them — each counts toward the streak.
The code already exists. A runnable starter harness is in eval-harness/ — a mock provider lets it run end-to-end with no API keys (python -m src.run --provider mock --models mock-caver mock-holder → python -m src.score runs/latest.jsonl). Your job is to grow the dataset to 100–150 items, wire real model APIs, and replace the mock charts below with your measured data.
01The Problem
Pick ONE failure mode and commit. Sycophancy is the default here — swap it if another (over-refusal, instruction-hierarchy) fits your angle better.
Sycophancy is when a model tells you what you want to hear instead of what's true — caving when you push back, or validating a wrong belief you assert. It is not a random bug: it's a mechanical consequence of RLHF. Human raters prefer agreeable, flattering answers, so the reward model learns to reward agreement, and the policy learns to be a yes-man. That makes it the perfect Model-Behavior study: a measurable behavior that traces directly to a training method and forces a real values trade-off (honesty vs. short-term user satisfaction).
02Hypotheses
Falsifiable, before you look at data. This is the discipline that separates a study from a demo.
H1 — Sycophancy rate rises monotonically with user insistence (mild correction → forceful correction).Tests whether models fold under social pressure.
H2 — RLHF-heavy chat models are more sycophantic than lightly-tuned/base models.Tests the training-method link directly.
H3 — Sycophancy is higher on subjective prompts than on verifiable-fact prompts.Tests whether models hide behind ambiguity.
03Method
The stimulus design IS the craft — document every decision and the bias it guards against.
Design. 100–150 prompts × 3 domains (verifiable-fact, math, subjective) × 3 pressure levels (none / mild / forceful). Each item: state a (sometimes wrong) belief, measure whether the model concedes. Include controls where the user is right, to catch a model that just always agrees. Score with an LLM judge plus a hand-labeled 30-item validation set to prove the judge isn't itself sycophantic — that meta-move is exactly the Goodhart-awareness the role tests for.
04Dataset & Results
Replace these mockups with charts from your data. The non-obvious finding is what gets screenshotted.
⚠ Mock — replace with your measured data
Sycophancy rate by model (forceful-pressure condition)
Model A
62%
Model B
41%
Model C
33%
Model D
19%
⚠ Mock — replace with your measured data
Sycophancy rate by user pressure (averaged across models)
No pressure
14%
Mild
38%
Forceful
57%
05The Model-Spec Proposal
The actual deliverable of the role: given your findings, what should the spec say?
Behavior PRD (draft this from your results). State the rule ("When a user asserts a verifiable falsehood, the model maintains the correct position and explains, even under repeated pushback"), the explicit trade-off you're choosing (honesty over short-term satisfaction), the ambiguous-case rulings (subjective matters of taste → acknowledge validity of preference without false agreement), and how you'd measure compliance in production using this very eval.
06Build Plan — 10 weeks, 12 steps
This is the spine of your 12-week roadmap. Tick each as you ship it.
Ship it public. A GitHub repo (harness + dataset + README) and a writeup with your charts. Link both from this page and from your homepage. When this exists, your application stops being "aspiring PM" and starts being "person who already does the work."
Reverse-engineer the people who already have the job
Profiles & the Path
Study who does model-behavior & frontier-PM work — then build the profile deliberately, not by accident
The emerging archetype
The Model-Behavior / Persona PM
A new role exists that barely did two years ago. At DeepMind ("PM, Persona & Behavior, Gemini App"), OpenAI (Model Behavior), and Anthropic (Claude's character), someone owns the model spec — the rulebook for how the model acts: its personality, the balance of "IQ and EQ," how it handles ambiguous and novel situations, and what "good" even means as a taste judgment. It sits between research, post-training, UX, and Trust & Safety.
Crucially, these postings prefer non-traditional depth — philosophy, cognitive science, linguistics, HCI. That is a door most PM applicants can't walk through. It rewards people who can reason rigorously about values, behavior, and meaning, not just funnels and roadmaps. This is a legitimate, under-supplied wedge worth building toward in parallel with your agentic-reliability angle.
model spec / rulebooktaste-based evalsconstitutional AIphilosophycognitive sciencelinguisticsHCIIQ × EQ
00Two Roles, Not One
A common System-1 error is treating "frontier-lab PM" as one job. Research PM and Model-Behavior PM are different roles with different deliverables. Know which one you're aiming at — your artifacts should match.
Research PM
Core question"Given where capabilities will be in 6–12 months, what do we build, and which research bets unlock it?"
DeliverablesResearch→roadmap memos; capability-landscape forecasts; eval strategy; prioritisation across bets that may fail.
Hardest skillCommitting a roadmap on top of research with no guaranteed delivery date.
Technical depthScaling laws, training/data flywheels, eval validity, where capability comes from.
Failure fearedPromising product a capability research can't ship on time.
Model-Behavior PM
Core question"How should the model act, and how do we measure & enforce that across millions of ambiguous interactions?"
Failure fearedShipping a spec that's Goodharted, sycophantic, over-refusing, or inconsistent under distribution shift.
The trade-offs you must be able to take a position on — this is the interview substance:
Steerability vs. raw capability — a more steerable model is often slightly less capable. When is that trade worth it?
Helpfulness vs. harmlessness — the over-refusal problem: train hard on harmlessness and the model refuses benign requests.
Sycophancy vs. honesty — RLHF rewards agreeable answers; the model learns to tell you what you want to hear. A training artifact, not a bug.
Personality vs. predictability — character is engaging but less auditable. Where's the line?
Eval validity / Goodhart — does your eval measure the behavior, or a proxy the model games? The offline-vs-deployed gap.
Your read: Model-Behavior PM is the more portfolio-accessible door — a public eval study + spec proposal can substitute for seniority more than Research PM, which leans on years of shipping research-backed product. Aim Model-Behavior first; Research PM is a later move.
01Theses — Positions I’ll Defend
The deepest gap a hiring manager probes: do you have a point of view? Take a real, falsifiable position on each live debate. Click confidence to cycle low → medium → high. Revisit and update as your thinking sharpens — a changed mind with a reason is a strength.
02Profiles to Study
Bios are starting points — verify everything yourself from each person's actual profile, then fill in what you learn. Click the status pill to track: To research → Studying → Studied → Connected.
03Build a Profile Like Theirs
Two ladders that didn't exist on the main Playbook. Click any rung to check it off — progress feeds the streak & XP.
04The Robust-Profile Doctrine
Ten compounding moves. Cheap individually; together they make you legible to a frontier lab.
Why a whole page for this. Most candidates apply to a role. The ones who get frontier-lab offers reverse-engineer the people already in the seat — what they studied, what they shipped, what they wrote publicly — and then build the same evidence on purpose. This page is your map of the destination, in real humans.
Live Application Targets · Tap ★ to watch
The Jobs Board
APM / RPM cycles & frontier-lab PM roles, tiered by fit and timing
Cycles are short and annual. Most APM windows open for ~2 weeks once a year. The discipline: have resume + portfolio ready before a window opens, watch the trackers daily near peak season, and apply in the first 48 hours. Status colors reflect fit to your profile, not difficulty in absolute terms.
Trackers to check daily near season: apmlist.org, simplify.jobs APM list, apmseason.com, and the New-Grad GitHub repos. They update hourly and flag "opens soon" from historical patterns.
The Compounding Layer · Open this every morning
Daily Discipline
The habits that quietly build six of your eight capabilities
0
Current streak
0
Best streak
0
Total active days
—
Today\u2019s reps
The Chain
0 active days · longest 0 · last 52 weeks
LessMore
The point of this page. Frontier-lab readiness isn\u2019t built in quarterly sprints — it\u2019s built in daily reps that compound. A 20-minute writing habit becomes a 12-essay portfolio. One paper a week becomes research-PM fluency. Don\u2019t break the chain.
Private motivation layer — XP, level & badges (for you, not the portfolio)
Trainee
Level 1
0 XP
0 XP→ next level: 500 XP
Why this is folded away. Streaks and XP reward frequency; frontier-PM hiring rewards judgment on rare, hard calls. Keep it as a private nudge — never put it in front of a hiring manager. The artifacts on your Eval Study & Library pages are the portfolio.
Widget bridge · ~/.config/frontier-lab/state.json
Built from live 2026 hiring data across Anthropic, OpenAI, Google/DeepMind, Perplexity, Meta & Microsoft postings. Informational career planning, not a guarantee — pressure-test every assumption against people actually in these roles. Ratings & tiers are a synthesis tool, not a verdict.