Community
Discuss
Real interview experiences and compensation breakdowns from AI engineers, FDEs, and MLEs.
openai fde loop done, got the offer. writing this up bc theres literally nothing out there about this role
ok so first thing: this is NOT the normal swe loop. theres no leetcode gauntlet, no algorithms screen at all actually. whole thing was like 3.5 weeks, 7 touchpoints. recruiter call first. their opening question was basically "why forward deployed and not just openai" and i could tell from her tone that "bc its the only open role lol" ends your candidacy right there. they want ppl who actually like sitting with customers, not ppl using fde as a backdoor. then the take home. ~5 hrs building something on their apis for a fake customer + you record a video walkthrough. KEY thing nobody tells you: the video is graded as a customer demo, not a code review. i narrated it like i was showing a clients eng lead the product. apparently most ppl just screenshare their ide and talk about their code and get cut here lol. technical deep dive was them picking apart my take home for an hour. every single decision. why this chunking. why this eval. why not fine tune. then rapid fire stuff like "customer says the llm workflow is slow, walk me through debugging it" and "when would u fine tune vs just fix retrieval". solution design: deliberately vague brief, "a bank wants an assistant over internal policies, design it end to end." the trap is architecting immediately. i spent the first 15 min just asking questions (who uses it, latency, data sensitivity, what does success even mean to them) and the interviewer straight up told me after that thats what most candidates skip. hm round: real customer stories. "tell me about a time u told a customer no." "tell me about a deployment that failed." they want actual failure stories not humble brags. values round had "what worries u about deploying powerful ai inside an enterprise" - have a real answer. tbh i think what got me through was mentioning evals, cost, latency and guardrails without being prompted in basically every answer. thats the production judgment theyre scoring for. ama, happy to help. roles genuinely great if u like ambiguity
anthropic applied ai loop - the values round is NOT a formality and i almost prepped wrong
just finished the anthropic loop (applied ai side). ~3.5 weeks. posting bc everyone grinds codesignal for this and then sleepwalks into the round that actually filters ppl. recruiter call is failable btw. they test mission alignment right there and explain the pbc structure. u need a real answer for why anthropic beyond "claude is good." like be able to explain how they differ from openai structurally, not vibes. codesignal take home, 90 min, 4 levels building one system (i got the bank with transaction types one). black box tests, gotta pass ALL of them at one level to unlock the next. ppl run out of time bc they gold plate level 1. just move fast and keep the design extendable. onsite was 5 sessions: - hm round: deep walkthrough of ur past projects. they drill into what YOU did vs the team. know ur stuff cold - coding: build from scratch style. concurrency came up here and AGAIN later, its a theme across their loop. be ready to re-implement stuff without stdlib and defend ur complexity choices. python expected - system design (shared doc): skews to actual anthropic problems. mine was an llm serving api - batching, queuing, gpu utilization tradeoffs. a friend got "design a claude chat service" - second role specific coding round - VALUES ROUND (1 hr). this is where most candidates fail and a recruiter basically admitted that to me. non technical interviewers. reflective stuff - when were ur values tested, when did u change ur mind about something big - plus ethical hypotheticals. and the followups probe ur actual reasoning and FEELINGS not ur conclusion. weirdest advice that turned out true: they reward pushback over alignment signaling. if u just agree with everything and perform "safety person" it reads fake af. reading that actually helped: machines of loving grace, core views on ai safety, darios long interviews. also ai use is strictly banned in their interviews. dont even have it open lol. outcome: passed, in team match rn. consensus hiring means the wait after onsite can stretch. ama
anthropic fde loop (rejected) - still the best interview process ive been through, writing it up
did the anthropic fde loop this cycle. made it to the final stages, no offer, but honestly the best designed process ive seen so heres the map. ~5 weeks: recruiter → technical screen → applied ai build round → customer discovery case → enterprise sys design → values → team match. recruiter screen: same as their swe loop, failable, mission calibration. technical screen: practical python/ts, NOT dsa. building an api client with pagination + retries then extending it. very "day in the life" flavored. applied ai build round is THE round. u build with claude - tool use, an mcp style tool server, subagents - then u write evals for what u built and debug a failing trace step by step. if uve never actually built an agent with tool calling and stared at a broken trace, go do that this weekend. reading about it is not the same thing. customer discovery case: ambiguous enterprise scenario. workflow discovery, scoping a pilot, security review, rollout, adoption metrics. theyre testing if u can run a real discovery call. enterprise design: mine was a fraud review agent w/ auditability requirements. the bar is what id call safety operationalized: action boundaries (read only vs draft vs write), human approval for high risk stuff, acl filtering, audit trails. "id just prompt claude better" is a failing answer lmao. they want engineering controls. values round: same reflective style as swe loop. where i think i lost it: the customer case. i jumped to solution architecture too early instead of staying in discovery mode. classic fde mistake and literally what the round is designed to catch. recruiters feedback was "wanted more discovery depth" which tracks. also the postings say 25-50% travel and they mean it. ask early if thats a problem. retaking in a year. gl everyone
gdm research engineer loop done - fair warning the math is real
wrapped the deepmind re loop (mle-ish track). first thing: this process is separate from regular google, its longer, and the ml fundamentals bar is way higher than normal mle loops. total: ~7 weeks including hiring committee. resume screen is brutal. tiny acceptance rate, referrals genuinely matter. if u know anyone inside, use it. recruiter intro also routes u research vs applied. be clear which u want, it changes the loop. then 2 coding rounds (gated). lc medium-hard BUT the code has to actually run in coderpad. no pseudocode handwaving. ~20 min per problem so pattern recognition speed matters. i got a graph one and a dp variant. pass both to unlock the ml rounds. ml fundamentals "quiz" is where gdm diverges from everyone. its math intuition not definitions. i got asked to walk through WHY l1 induces sparsity - not "bc it does" but the actual gradient dynamics argument. plus gradients vs weights probing, an elbo adjacent derivation sketch, bayesian reasoning. my interviewer had co authored papers in the area. u cannot bluff these ppl lol. ml design: mine was short form video recs - full pipeline: data, features, arch, training, serving, metrics. they push on every tradeoff. widely considered the hardest round and i agree. structure everything: clarify → frame → high level → drill. final: team lead, senior team lead, behavioral w/ people & culture partner. then hiring committee. strong no hire → strong hire scale means one weak round can sink u even if the rest were great. and level often isnt discussed until after committee so dont expect a number early. notes: jax/pytorch fluency expected. ai assistants banned. the wait after final round was 2.5 weeks of silence and genuinely horrible for my mental health lol. outcome: offer. ama
meta e5 mle loop (genai org) - the new ai assisted coding round is a thing now
just did the meta e5 mle loop for a genai team. process has changed a lot, most prep material is outdated so heres the current state. loop: recruiter → 2 screens → onsite (2 coding, 1 ml sys design, 1 behavioral) → team match. the big change: the ai assisted coding round. one of my screens was a toy codebase w/ unit tests and an integrated ai assistant in coderpad. u debug and extend it WITH the assistant. what theyre actually scoring: do u plan out loud first, are ur prompts tight, do u explain WHY u accept or reject the ai output, do u benchmark before optimizing. full screenshare required, no background blur. ppl who just vibe code and paste fail. ppl who ignore the assistant entirely also do badly - collaborating with the tool is literally the skill. funny moment: the assistant wrote me a subtly broken retry loop mid round. i narrated exactly why i was rejecting it and what invariant it violated. interviewer told me later that moment scored me more than any correct code i wrote lol. classic coding round: pace is everything. 45 min = 1 easy + 1 medium or 2 mediums. ideal rhythm is instant pattern recognition, 2-3 min explain, 5-8 min code, dry run out loud. ml sys design: ranking funnels are table stakes now - candidate gen → l1/l2 ranking → re-rank w/ value model. u differentiate on domain depth. mine drifted into generative retrieval for recs and the interviewer lit up. know where the frontier is not just the 2019 playbook. ml fundamentals rapid fire in the screens: 3-5 quick questions, 2 wrong can genuinely sink u. mine covered regularization, calibration, bias-variance. the frustrating part: team match. cleared the loop then... waited. only a couple openings per org per cycle and ur re-evaluated against other cleared candidates at match. took 3.5 weeks. total process 9 weeks. outcome: matched, accepted. comp in comments if ppl want it
palantir fdse loop - the decomp round is genuinely unique (offer, declined)
did the palantir fdse loop. got the offer, ended up declining, but the decomp round is unlike anything else in the industry and worth prepping for specifically. loop was fast: ~3 weeks. recruiter → hackerrank oa → technical screen → onsite (coding + decomp + learning round + hm). decomp is the signature round. vague ambiguous real world problem, u decompose it out loud while they watch ur structure. mine was a variant of the classic: "ur cousin drives for a food delivery app and has historical + live order data - design a system to help him pick which orders to accept." then once u solve the cousin version: "ok now scale it to the whole platform." what theyre grading is NOT the right answer bc there isnt one. its whether u clarify constraints first, whether u structure into clean sub problems, whether ur data model makes sense, whether the scaled version follows logically from the small one. i literally narrated a framework: users → goals → data → core entities → simplest thing that works → what breaks at scale. the interviewer engages like a collaborator not an examiner which threw me off at first. warning: dont relax and stop narrating when it gets conversational, thats a trap. debugging round: broken api service, real ish logs, trace it, find the bottleneck, fix a memory issue. very "u will actually do this job." coding: deliberately not about optimality. mine was a multiplayer card game scoring simulation. they just want working readable code that runs. do NOT waste time deriving the optimal solution. learning round: they hand u unfamiliar material, u digest it, then apply it. tests learning velocity which is honestly the core fde skill. why i declined: comp. offer was solid for palantir (~210k tc, liquid pltr rsus which is genuinely nice - no liquidity discount) but i had a frontier lab fde offer at nearly double. calibrate expectations before investing in the loop if comp matters. process itself: respectful, fast, well run. would recommend the experience even as practice.
databricks senior mle loop - 8 rounds and the double coding is where ppl die
finished the databricks mle loop (senior). 8 rounds, ~6 weeks, mostly virtual. heres the map. 1. recruiter screen. standard. 2. coding screen (60 min). lc medium-hard, graphs/dp flavor. 3. hm screen. project deep dive on ur resume. they go SPECIFIC - arch choices, what ud redo, ur personal contribution vs the teams. 4-8. onsite, 5 rounds: - coding #2: THE TRAP. the second coding round pivots to concurrency/multithreading at hard level. everyone preps more lc mediums and gets blindsided. prep concurrent data structures and rate limiter style stuff instead - system design: distributed systems, shared doc, lakehouse flavored. mine had point in time feature joins w/ delayed labels, watermarking, offline/online skew - ml & modeling: heavy genai weighting now. i got a rag debugging scenario (retrieval metrics improving but answer quality regressing - diagnose) and a prompt vs lora decision question where they wanted actual metrics reasoning (groundedness, citation support, deflection rate) not vibes - behavioral: six core values probed directly. truth seeking and first principles come up a lot - bar raiser: interviewer from a totally different team. inherited from the spark open source culture - ownership, ship it mentality, "would this person raise the bar" judgment what filters ppl: the double coding. recruiter basically confirmed most candidates bleed out there bc they prep one style. happened to someone i know lol. other notes: references weighted HEAVILY in the final decision, have 2-3 recent ones ready. equity is private rsus so if ur comparing against public offers the illiquidity delta is ur negotiation lever. framing it explicitly with a dollar figure moved my grant ~15%. vague "but its illiquid" gets u nothing. outcome: offer, accepted. team match happened BEFORE the onsite for me which was nice, no post loop limbo. ama
scale ai fde loop - the customer simulation round needs its own warning label
did the scale fde loop. 2.5 weeks, fast and intense. loop: recruiter → hm call → codesignal oa → take home → onsite. hm call is business heavy for an eng role. "what does the marginal return curve for high quality llm training data look like?" "customer wants a feature u think is misguided, how do u handle it?" theyre checking if u think like a business partner. take home: build a data quality evaluator for rlhf style prompt-response pairs. jsonl in, multi axis scoring via an llm api, error handling, concurrency, readme. graded roughly on: does it run, are the eval dimensions sensible, eng quality, docs. IMPORTANT: overspending time hurts u. they grade the 8 hr version against the 4 hr version on a curve. ship a clean 5 hr solution not a gold plated 12 hr one. onsite: - coding: lc medium + applied variants - sys design: annotation pipeline - task splitter → worker pool → quality gate → consensus engine. know inter annotator agreement stuff - CUSTOMER SIMULATION: the signature round. the interviewer role plays as a pm from a frontier lab with a vague ask. mine was "we need more reasoning data." if u start proposing solutions in the first 5 min, big deduction. ur graded on clarifying questions, then 3 plans w/ cost/time tradeoffs, recommend one, surface risks unprompted. they stay in character for 40 min INCLUDING being impatient and giving vague answers when ur questions are bad lol. then 5 min out of character for meta discussion at the end - cross functional + leadership round decision in about a week. comp: 210k base + sign on, series f equity on top. the equity is paper - private, low liquidity - so weight the base accordingly. heavy sf onsite preference too. outcome: offer, still deciding. the role is basically "applied ai engineer embedded with the biggest ai customers in the world" which is either ur dream or ur nightmare depending on how much chaos u enjoy
openai fde offer numbers - base, ppus, and how the equity actually works (its confusing)
posting my offer details bc fde comp at frontier labs is a total black box and ppus confused the hell out of me too. level: mid-senior fde, sf. - base: 245k - equity: ppu grant valued ~500k at current valuation, vests 25%/yr over 4 yrs - sign on: 30k (partially framed as covering forfeited equity from my last role) - paper tc: ~370-400k/yr at grant valuation w/ real upside if ppu valuations keep climbing how ppus actually work (made the recruiter explain twice): they are NOT rsus and NOT options. profit participation units = contractual claim on future profits. no ownership, no strike price, value tied to periodic valuations. key mechanics: 25%/yr vesting and reportedly no cliff for new employees now, ~2 yr lockup after vest, and liquidity happens thru periodic tender offers (roughly semi annual) where investors buy ur vested ppus directly. so it IS liquid-ish in practice but on openais schedule not urs, and the capped profit structure means theoretical upside is bounded in a way public rsus arent. negotiation notes: - base moved almost nothing. ppus moved ~15% when i put a competing anthropic fde number on the table. equity is the negotiable part, base isnt - the competing offer mattered more than any argument i made. get a second offer if u can - timeline compression is real: mentioning the other deadline got my loop done in under 3 weeks for calibration my competing anthropic fde offer was ~250k base + equity, and anthropic famously doesnt negotiate within a level so the number was the number lol. risk framing that helped me decide: if u believe tender valuations keep rising, ppus are great. if u want boring certainty, public rsus win. i took the upside but eyes open about the lockup. hope this helps someone negotiating. numbers are my actual offer, anonymized only slightly
chose between anthropic and deepmind mle offers this cycle - numbers + how i decided
had the good problem of picking between anthropic and gdm (l5). posting both packages and my reasoning bc the structures are wildly different and headline tc hides it. anthropic: - base: 320k (they publish bands in listings, mine was inside the posted range, zero surprises) - equity: private rsu style grant, 4 yr vest w/ 1 yr cliff, tender/secondary liquidity - sign on: modest, framed around forfeited equity - tc: ~600k at current valuation - THE critical detail: anthropic does NOT negotiate within a level. everyone at a level gets paid the same, its confirmed policy not a tactic. ur only lever is level placement so fight for the level not the number gdm (l5): - base: 290k - gsus: ~230k/yr at grant, front loaded vest (33/33/22/12) - bonus: 15% target - tc: ~560k, ALL liquid from day one the actual comparison: - headline tc favored anthropic slightly. risk adjusted it was closer than it looks: anthropic equity is private w/ tender liquidity, gdm equity is goog stock u sell on vest - anthropic no negotiation thing means the offer is genuinely the offer. honestly refreshing after the usual dance. gdm moved on equity w/ the competing letter (~12%) - gdms front loaded vest means years 1-2 are way richer than the 4 yr average suggests. nobody models this and its a six figure difference how i chose: anthropic, and not for the money. work scope difference (small team, direct line to deployed model behavior) outweighed the liquidity discount for me. if id optimized for risk adjusted comp, gdm was the correct answer and i wanna be honest about that lol. practical advice: get both loops onto overlapping timelines. everything in this market runs on competing offers. ama within reason