Inside the Brain of the Top 1% AI Companies – Márton Szabó (Talk Outline)

Márton Szabó runs a Sequoia-partnered hiring agency placing engineers at top-tier AI companies (Lovable, Synthesia, and much of Sequoia’s European portfolio). His thesis: skills get you in the loop; judgment gets you the offer. He argues modern orgs hire for seven archetypes (not “software engineer”), demand end-to-end ownership, and — following OpenAI, YC, and Sequoia — are moving toward end-to-end problem solving where domain judgment, not model access, is the moat.

1. Opening Story — Same CV, Opposite Answers

• Last quarter he sent the same engineer (strong GitHub CV, “you don’t think twice before forwarding”) to two AI companies with near-identical job descriptions.
• One said “strong profile, get him in ASAP”; the other instantly rejected the CV.
• Not an accident: the lazy explanation is “one was wrong,” but in reality the two companies were optimizing for different things — different judgment. Modern software orgs change fast; to work at one you must fit what they’re looking for.

1.1 Speaker

• Works at a tech-talent company — a specialized hiring agency for early-stage companies, partnered with Sequoia, working with most of its European portfolio including Lovable and Synthesia.
• His edge: he sees hiring at multiple top-tier AI companies — a unique view of what they optimize for. These companies win the contested talent and set the trends your org will likely follow later.

2. Thesis — Skills Are No Longer Enough

• Skills get you the interview; there are many engineers with similar skills. Judgment brings the offer.
• Judgment isn’t a vague personality trait — it’s making decisions under pressure and owning them. People at big companies could deflect (“my manager told me to”) — that’s no longer enough.

2.1 Two interview trends

• Client A: two coding interviews — one allows AI, one doesn’t. They don’t care if you can write code (anyone can with AI); they care whether you understand code and can make and defend decisions. Old hard LeetCode problems are replaced by small projects (about an hour) that you then explain — “if you can’t defend your thesis, why do we need you?” AI executes; you take responsibility because AI won’t.
• Client B optimizes for this by giving you a real company problem they’ve already solved and asking for your solution.
• AI made the first draft cheap. Some companies even started re-hiring junior engineers because they’re cheaper than AI tools — AI isn’t the solution to everything; your judgment makes AI strong. You must own what to build, what to cut, and the outcome.

3. Seven Engineer Archetypes

• Past titles: software engineer / back-end / front-end / DevOps. Now most companies he works with optimize for seven archetypes — similar-looking but not the same — which is why the GitHub engineer was a yes at one company, no at another.
• From these seven you can build an entire early-stage organization. Many companies don’t even understand what they need.
• Design engineer example: previously UI/UX designers handed designs to front-end engineers; now designers can write code and engineers can create designs — pushing toward end-to-end ownership (delivering the same with fewer engineers, as the constant big-tech layoffs show).

3.1 Forward-Deployed Engineer — judgment under constraint

• “The most important role in modern software engineering right now.”
• Not new — Palantir introduced it around 2012; many have copied it. More than “a consultant with a new title”: FDEs make the solution work inside the client’s reality — legacy systems, politics, deadlines — problems you don’t see when selling a packaged product (where you’d otherwise just shrug and lose the client). These engineers often become founders. He works with a company founded by the people who created agentic forward-deployed engineering at Palantir (“some of the smartest people I’ve met”).
• OpenAI proof point: OpenAI started its own deployment company with $4B+, acquiring Tomorrow.ai and 150+ forward-deployed engineers / deployment specialists. They have the data showing the problem isn’t model access — it’s that companies can’t use the product. “Why build a new model when people can’t use the one we have?” If frontier labs move this way, take it seriously.
• Difference from product engineers: a good product engineer might be a good FDE, but FDEs usually aren’t good product engineers — FDEs live in messy reality (legacy, hard clients) while product engineers build clean solutions in clean environments.

3.2 Product Engineer — deleting the PM handoff

• Started a couple of years ago because orgs deleted product managers (or at least the PM handoff). When building software is cheap, the loop becomes the bottleneck — PM gathers requirements, iterates, hands tickets, problems recur forever, and speed suffers.
• So engineers own the whole thing — from discovery to iteration — and are therefore responsible for it (that’s “ownership”).
• Measured by business impact, not output — “anyone can write tens of thousands of lines; that’s not value.” Even top engineers struggle to adjust (“lowered runtime 10% — no one cares; show how you moved the needle for the business”). Series A/B companies show fewer and fewer PMs.

3.3 Applied AI Engineer — judgment under uncertainty

• Similar to product engineering — they create value with AI tools. Models are available to everyone; you need someone to turn them into a product. “Most startups are wrapper companies” — no huge invention, just “agentic AI for logistics / retail” on the same tools everyone has.
• The hard part: researchers chase benchmarks that look good on paper, but you must ship something reliable, cost-effective, and trusted — “you can’t let AI hallucinate a contract.”
• Researchers rarely fit applied-AI seats (years in academia, not production). A research AI engineer should go to a frontier lab or rebrand and think differently — value is what matters.

4. Which Archetype Are You?

• Nothing new to learn — you need to think differently and recognize these archetypes are here to stay.
• Founders/leaders should ask: which judgment is my organization missing?

5. The Flooded Funnel

• Most hiring funnels screen for the wrong things — job descriptions ask for keywords that mean nothing now (everyone AI-optimizes their CV). “10 years of Java — who cares? I can also solve hard algorithmic problems.”
• Your skills haven’t vanished — they make you a great engineer — but you must show your judgment and how you create impact.
• Volume: (self-reported, take with a pinch of salt) ElevenLabs had ~31,000 applicants in one month and hired ~143 in the first half of the year. Against that competition you stand out with judgment and value, not technologies.
• “AI isn’t rejecting your CV — people are, because you’re optimizing for the wrong things.” AI made applying trivial, so ghosting is common; a recruiter seeing your tech list still asks the judgment questions above.

6. Follow the Money — AI-Native Services & Domain Judgment

• OpenAI → deployment. YC’s “requests for startups” now include AI-native services companies — because companies spend 4–5× more on services than on tools. If you have the tools, why not do the work end-to-end and capture the bigger margin?
• Parallel movements: development → end-to-end ownership; companies → end-to-end problem solving (fewer people; not “10 tools for the same thing” but “a problem and a solution”).
• Sequoia: model access is not the moat — domain judgment is. Companies that couldn’t adjust are “dying slowly.” Lovable (unicorn) made it easy to create websites/presentations — but only because Anthropic shipped a new Claude model.
• AI didn’t make engineers obsolete — it made judgment the whole game. Ask: what judgment did I own? and, if building/hiring, what judgment are they missing?

7. Closing — Now Is the Time to Build

• He sees more engineers creating their own things — more side projects than in ages, top-company engineers leaving to start startups — because building software is now easy and cheap if done right.

8. Q&A

• Advice for juniors with no specialization? Build side projects — easy now, and something to be judged on; show you can create value (and maybe make money) even without real-world experience.
• Is “software engineer” no longer a good CV title? Yes — adapt your title to what you actually do; map yourself to one of the archetypes. It’s rebranding, and companies optimize for it now.
• How to showcase value on a CV / delete 10+ years of Java? Keep it, but what you write next to it matters — one-page CVs can’t hold 15 years; state what you worked on and the outcome / value you created, since a human (not the AI filter) focuses on that.
• Any hack to reach an interview at a high-tech company? Connections (“people get you places”) — attend events like this, meet people; with 31,000 applicants, any edge helps. (He also teaches e-sports on the side and gives students the same advice.)
• Will larger/older orgs adopt this or are startups different? Already adopting — big-tech layoffs; Musk turned Twitter/X into “a startup” by cutting staff. It’s all about talent density — fewer but better engineers producing the same outcomes. AI creates different jobs and judgments, not zero jobs; businesses that can do the same with fewer people will.

People, Companies & References Cited

• Márton Szabó — speaker; Sequoia-partnered hiring agency; also teaches e-sports.
• Sequoia — investor; “model access is not the moat, domain judgment is.”
• Lovable, Synthesia, ElevenLabs, GitHub, Palantir, OpenAI, Tomorrow.ai, Anthropic/Claude, X/Twitter — companies referenced.
• Y Combinator (YC) — “requests for startups” → AI-native services.
• Concepts: judgment vs. skills, end-to-end ownership, the seven archetypes (forward-deployed / product / applied-AI + design engineer), deleting the PM handoff, business impact over output, AI-native services (4–5× services vs. tools), flooded hiring funnel, talent density.

Video: https://www.youtube.com/watch?v=wvWz_uqp9S0 — Transcript via yt-transcript.sh; exhaustive outline generated from the full transcript.