Inside the Loft
A Union Square loft ships video-generation models while the rest of the industry debates whether diffusion transformers belong in production. Mirage operates as a high-velocity, founder-led organization where technical depth and rapid execution are the primary currencies of influence, and the employee experience splits sharply between those who find the intensity energizing and those who experience it as unsustainable.
Mirage runs lean (138 people, a team that fits in a large restaurant) and the org chart is flatter than the headcount suggests. Seven open roles cluster in Union Square: Group Product Manager ($200k–$325k), Software Engineers across Agents, Backend, and iOS ($175k–$300k), and Research Engineers in Agentic Systems and Generative Video ($175k–$300k). Median band: $275k. That concentration (product, applied research, and platform engineering all hiring at once) signals a roadmap where model advances and product surfaces move in lockstep, not in sequence.
| Role | Salary Band |
|---|---|
| Group Product Manager | $200k–$325k |
| Software Engineer (Agents, Backend, iOS) | $175k–$300k |
| Research Engineer (Agentic Systems, Generative Video) | $175k–$300k |
| Median | $275k |
Collaboration isn't a wall poster; it's the operating mechanism. The company's life page states that "every team member plays an active part shaping the company" and that "from design and engineering to marketing and beyond, we all help inform new products." In practice, a research engineer validating a new sampling schedule sits in the same review loop as the designer deciding how that latency surfaces in the timeline UI. Dwight Churchill, who runs recruiting, calls the interview process "one of the highest leverage investments a founder makes" and says it's "continuously refined and obsessed over" — hiring as a product-velocity lever, not a people-ops function.
Decisions route the same way. The public org structure shows a single co-founder, Gaurav Misra, on LinkedIn. The Group Product Manager role (the only non-engineering leadership slot open) carries a $125k band, same width as the engineering tracks, signaling product direction as a force multiplier, not a coordination tax.
Pace follows the model release cycle. Generative video research moves quickly; agentic systems run in parallel. Backend and iOS teams must ship consumer-grade reliability on top of research artifacts that may change signature between Monday and Friday. The vision ("redefining what's possible for video, for all kinds of people") only holds if the platform team absorbs that volatility without passing it to users.
Research on culture theatre warns that "when organizations confuse one for the other, the cost is steep" and "People stop asking hard questions, because they've learned that comfort ranks higher than truth." Mirage's test: whether its flat, high-velocity model sustains psychological safety when the model breaks — or whether the speed that attracts builders becomes the thing that drives them out.
The Rules That Actually Run the Place
Mirage's operating principles read less like a culture deck and more like a survival manual for a founder who bootstrapped multiple companies and refuses to confuse activity with progress. Griffin Haddrill founded Mirage Digital LLC after a decade in entertainment and media. His non-negotiables now govern how the firm prioritizes, trades off, and defines success.
The clearest signal sits in the hiring funnel. As of October 2025, roughly one in two thousand engineering candidates who entered the pipeline received an offer — a rate Churchill contrasted with Ramp (one in 435), Y Combinator (one in 100), and Harvard (one in 39). A commenter on Churchill's post wrote: "I have never wished I was less selective and rigid in my interview process. I have only wished the opposite." The result, per Churchill: a team "insanely talented, inventive, and deeply invested in their work."
That selectivity maps to a value the research-engineer guide states explicitly: production-readiness over pure research. Candidates must show "deep expertise in PyTorch, CUDA, and Triton" and "ability to write efficient, hardware-aware code that maximizes GPU performance." The guide adds: "Because Mirage prioritizes production-readiness, you must show you understand the full lifecycle of a model — not just training, but also monitoring, evaluation, and the infrastructure required to keep a system performant in production." The phrase "builder's mindset" appears verbatim — people "as comfortable writing a research paper as they are profiling a CUDA kernel to save 5ms of latency." In-person work in New York City is non-negotiable.
Haddrill's own principles, documented in a 2025 interview, reinforce the same orientation. "Fail fast" — he gives himself three to six months to "bump into walls, talk to potential clients (sometimes without a product), and try my best to think about the issues I might be facing in 2–5 years." "Break down process." "Value learning and create new benchmarks." "Treat everyone like a customer and don't talk about our wins until they are proven." He distinguishes between people who like to win and people who hate to lose: "I firmly belong to the latter category. Failure signifies a temporary setback with a path forward, whereas loss represents an absolute dead end and acceptance of defeat." He insists on documenting principles during prosperous times and adhering to them during challenging ones — a discipline tested in 2022–2023 when advertising restrictions and privacy-policy shifts forced a choice between reinvesting cash reserves or slowly declining. The executive team (Haddrill, John Littell, Robin Bigge) chose to reinvest, recalibrate workflows, and retain the existing team. Two years later, Haddrill says, "the foundational work we established during those challenging times has become the driving force behind our current success."
A concrete trade-off illustrates the quality-over-volume principle. Early on, Mirage leaned heavily on contract labor. "As we scaled, we saw mixed results — more volume, but the quality didn't always match up," Haddrill said. "I made the call to shift focus entirely to full-time employees. We slightly dialed back on output and doubled down on quality control. That's been the guiding principle at Mirage for the past three years, and the results have been nothing but positive — both revenue, client satisfaction, and company culture."
Trust and cognitive diversity round out the founder's list: "Surround yourself with people who challenge and think differently than you. Trust your team." These aren't aspirational posters. They show up in the interview rubric, the compensation structure (base salary plus RSUs, no cash bonus), and the decision to keep the team lean and in-person while competitors chase headcount. The principles are specific enough to be falsifiable, and the company's trajectory — over 350 creators and major media organizations, more than a million daily monetized views across 30 platforms as of mid-2025 — suggests they're being stress-tested in real time.
Who Clears the Bar
Mirage's interview guides for Research Engineer and Research Scientist roles, published as of August and September 2026, lay out a bar that filters heavily for applied depth over theoretical breadth. The baseline for Research Engineers is explicit: two-plus years of professional industry experience, expert-level command of PyTorch, CUDA, and Triton, and a proven track record with distributed training frameworks such as FSDP. Candidates must demonstrate they can profile and optimize models for low-latency inference. The interview process tests this — deep-dive technical discussions paired with hands-on engineering challenges that mirror the R&D-heavy, fast-paced environment. The company values people who move quickly from prototype to production and feel ownership over the systems they build.
Research Scientists face a different but equally concrete threshold. An advanced degree (MS or PhD) in ML, CS, or a related quantitative field is required, along with demonstrated experience in modern transformer-based architectures and proficiency in Python and PyTorch. But the credential is the floor, not the ceiling. The guide emphasizes that "a great candidate doesn't just know the theory — they have an opinion on what works and what doesn't based on their own experimental history." Being able to defend technical choices with data is described as "the ultimate differentiator." The role is "highly applied": scientists are expected to move their models into the product, meaning they must write clean, production-grade code alongside research experiments.
Both tracks converge on the same signal: experimental judgment backed by shipping evidence. Mirage is an AI-native video platform built on natural-language-driven creative automation, per its own description, and the roles on Zero G Talent's board reflect that specialization. The compensation spread signals seniority — these are not entry-level positions, and the interview bar reflects that.
The broader hiring literature offers a cautionary frame. MIT Sloan Management Review notes that what laypeople call "merit" is referred to as validity by industrial-organizational psychologists — meaning a selection practice must be demonstrated as job-related, not assumed. Assessments with only "face validity" (appearing job-related to the layperson) often lack predictive power, and busy managers frequently adopt them anyway. Meanwhile, University of Georgia research covered by Mirage News found that candidates facing AI evaluators increase deceptive embellishment, and the AI systems tested failed to penalize it — human evaluators, by contrast, detected and downgraded inauthentic behavior. Mirage's process, built around live technical discussion and hands-on challenges, appears designed to sidestep that trap by keeping evaluation human and evidence-based.
What the bar selects for, in practice, is a narrow profile: engineers and scientists who have already operated at the intersection of research and production, who can articulate why they made specific architectural choices, and who treat latency, throughput, and deployability as first-class concerns. The filter is high. The signal is specific. The compensation reflects the scarcity of people who clear it.
The Split in the Reviews
The public record on Mirage's employee sentiment is fragmented across platforms that don't always distinguish between the Las Vegas casino and the New York–based frontier tech firm. What can be separated points to a split experience mirroring the company's stated intensity.
On Indeed, the "Mirage" employer page shows a 3.7 out of 5 overall rating from 100 reviews as of March 2026. SimplyHired reports a 3.8 average across 43 reviews, with 64 percent of 11 respondents saying they're satisfied with pay. Glassdoor lists 19 reviews for "Mirage" — a count more consistent with a 138-person tech company than the 115 reviews for "The Mirage," which almost certainly belong to the casino. The board's first-party salary bands align with the pay-satisfaction figure, though the sample is small.
Negative reviews on Indeed cite concrete grievances. A March 2026 reviewer wrote: "They dont pay employees very well for time and efforts and was an issue among many employees that they constantly go through. They lack proper training for new employees." A June 2022 review in French alleged psychological harassment and zero respect for employees. An August 2023 review complained about unfair tip distribution — a detail suggesting at least some Indeed reviews bleed in from hospitality roles. Without employer responses or verified role tags, it's impossible to assign each complaint to the tech side with certainty.
The clearest signal comes from a February 2026 TechCrunch investigation into a 200-person tech company that conducted more than 40 in-depth interviews with employees embracing AI coding agents. The piece describes a dynamic that reads like Mirage's operating mode: expectations tripled, stress tripled, and actual productivity rose only about 10 percent. Employees reported fatigue, burnout, and a growing sense that work is harder to step away from as organizational demands for speed and responsiveness climb. A separate account by Benj Edwards at Ars Technica captured the same tension: "I have not had this much fun with a computer since I learned BASIC on my Apple II Plus when I was 9 years old." He later burned out. The central finding — that the people who adopt AI tools most aggressively are the first to hit the wall — matches the profile of a high-velocity, founder-led shop operating on those same terms.
A parallel study published by Mirage News in August 2026, covering 283 workers across 23 industries, found that intensive behavioral measurement correlates with lower engagement and higher quit intentions, especially where staff don't perceive processes as fair. Non-measured work (mentoring, flagging risks, helping colleagues) gets deprioritized rationally, not lazily. The researchers' prescription: "Organisations should be asking which parts of the job their dashboard can't see." That applies directly to a culture that rewards visible output velocity.
Professor Martin Edwards of the University of Queensland, co-author of that study, said: "The more intensively an organisation measured employee behaviour, the more likely the work not being tracked was de-prioritised. This isn't laziness, it's a rational response to the signals the system is sending about what actually counts." Professor Tyler Okimoto added that perceived fairness (confidence that valuable work gets recognized even when invisible on dashboards) mitigates the effect.
The research is thin on named, current Mirage tech employees speaking on the record. No Glassdoor verbatims from the 19 "Mirage" reviews are available in the source data, and the casino's "Day One" team members (137 people who opened the property on November 22, 1989) dominate the Review-Journal's July 2024 coverage with remarks about familial bonds and an inability to articulate their grief. Those voices belong to a different era and a different business. What remains for the tech firm is a pattern: strong compensation, a measurable cohort that finds the work energizing, and a visible minority describing unsustainable pace, insufficient onboarding, and a measurement culture that may crowd out the unwritten work that holds teams together.
Surviving the Pace
In an AI-native shop, the measurement pressure compounds. Gary Marcus reported that "LLMs write code wicked fast, but some coders are starting to report burnout, and only relatively modest gains relative to that burnout," citing Connie Loizos at TechCrunch. The security and reliability of autogenerated code remain "very much in question," and no system "can reliably do every five-hour long task humans can do without error, or even close." Engineers who treat AI output as draft rather than deliverable, who maintain rigorous review habits, and who can articulate where the model fails will outlast those who chase speed alone.
The casino Mirage's "Day One" cohort (137 employees who stayed from the 1989 opening through the 2024 closure) described a culture of intense proximity: "These people are like our family. We spend more time with them here than we do at home... We argue, we fight, and then five minutes later, we're family again." One employee said, "I cannot find the word to express how I feel today because I am so sad." That depth of bond forms under sustained shared pressure. The tech Mirage's board footprint (12 salaried roles posted, all in one NYC location) suggests a similarly dense, in-person, high-cadence team.
Candidates who thrive: senior engineers who have shipped complex systems end-to-end, who default to writing over meeting, who treat metrics as feedback not identity, and who have a track record of calling out risk early. Candidates who burn out: those who need explicit process to feel safe, who optimize for visibility over utility, who equate hours with output, or who cannot operate when the dashboard misses the work that matters most.
The next hire walks into the Union Square loft. They'll clear the bar or they won't. The ones who stay will be the ones who already know that 5ms of latency isn't a metric — it's the difference between shipping and stalling.
Working in AI? Zero G Talent tracks the openings: see every open Mirage role, browse AI jobs, the companies hiring, and the people building the field.