Where to apply + how to be visible
📖 Walk me through it — plain English
This isn't an algorithm lesson — there's no code to trace. It's the closing strategy lesson: now that the rest of the guide has made you ready, this answers two questions. (1) Where should you point all that preparation — which kinds of companies? (2) How do you get those companies to notice you in 2026? The big claim is that the old habit — polish a resume, blast it through job portals, refresh LinkedIn — is just one channel, and for small AI companies it's the weakest one.
Two pieces of jargon up front. A channel just means a route by which a company learns you exist (a cold email, a referral, a blog post, etc.). Cold outreach means contacting someone who has never heard of you — no introduction — usually a direct email. FDE = Forward-Deployed Engineer, a software role that works face-to-face with a customer rather than only writing code internally. YC = Y Combinator, a startup accelerator; its companies come in "batches," and the newest batch is hiring hardest.
Everyday analogy: think of it like fishing, not a lottery. Resume-spam is buying 50 lottery tickets a week — cheap per ticket, but the odds on any one are tiny, and you're indistinguishable from everyone else holding a ticket. The lesson says instead pick 2–3 ponds to fish (the "target tiers" — AI labs, AI-product startups, Forward-Deployed roles, AI features at big SaaS, fresh YC startups, or boring-but-stable enterprise), then use the right bait for each pond. A founder at a 20-person startup reads their own inbox, so a personal, specific email lands. A big enterprise has a hiring pipeline, so a normal application still works there. You wouldn't use deep-sea gear in a small creek — you match the channel to the pond.
How to actually use this lesson, step by step:
- Pick your 2–3 tiers. Don't chase all six. The page lists what each tier hires for so you can match it to your strengths — e.g. startups reward shipping speed and taste; enterprise rewards the classic LeetCode + system-design + behavioral loop and pays a higher base as a safety net.
- Match the channel to the tier. Small AI startup → a personalized cold email to the founder (there's a fill-in template on the page; the rule is to rewrite the middle paragraph every time so it's about their specific work — that earns roughly a 10–20% reply rate, versus near-zero for generic blasts). Bigger company → a normal resume application is fine.
- Build visible signal in the background. "Signal" means public proof you can do the work — a blog post on your project's evaluation method, a small open-source pull request (a "PR," a code change you submit to someone else's project) to a tool your targets actually use, or a substantive comment on AI Twitter/X. One good writeup beats ten LinkedIn status updates.
- Read 30 min/day to build "taste." The reading list (Simon Willison, Latent Space, the Anthropic and OpenAI engineering material, etc.) teaches the vocabulary interviewers quietly test for. Skim daily, deep-read one piece a week, and form opinions — having a view is the whole point.
Why this shape works: it spends your effort where it converts. The closing "cadence" — about 5 personalized cold emails and 2 referral asks per week, one writeup and one open-source PR per month, roughly 15 hours total — is deliberately the opposite of spraying 50+ applications a week. Volume-spam maximizes noise and buries your signal; a smaller number of targeted, evidence-backed touches makes the few people who matter actually notice you. For AI-native roles, fewer-but-sharper beats more-but-generic.
The guide makes you ready. This lesson is about where the ready-version of you should be looking, and the channels that actually work in 2026 for SWE / AI-Engineer / Forward-Deployed / founding-eng roles. The traditional resume-to-LinkedIn loop is one channel. For small AI companies it's not the strongest one.
The tier shapes below are durable. The current company names occupying each tier live in the dated Market snapshot — check there, then re-verify before outreach.
The frontier-model companies. High bar; longer loops; the bar is technical depth + safety judgment. Apply if you want the frontier; the loops are weeks long.
50–500 people building AI-first products. Hire for shipping speed + taste. Founding-engineer or early-eng titles. Cold outreach works.
Customer-facing technical role. Different bar — product judgment + customer empathy + speed. FDE decomposition round matters.
Bigger teams, more process, but real users + real budgets + comp parity with Big Tech.
Look at the YC company directory filtered by recent batch + AI. Many founding-engineer slots filled via cold outreach + demo. Higher risk, higher equity, fastest hiring decisions (often <2 weeks).
Banks, healthcare, big enterprise software. Traditional SWE loop — LeetCode + system design + behavioral. Higher base, slower pace. Safety-net tier.
- Direct cold email to founder / hiring-eng manager — works best at <50-person AI startups. Founders read their inbox. Template below.
- Referrals via warm intros — UMD alumni in target companies; college friends now at AI startups. LinkedIn alumni search is still the fastest way to surface these.
- Public technical writeups — one blog post on your SoloMock evaluation methodology is worth 10 LinkedIn updates. Post on your portfolio site; cross-post to X.
- Open source contribution — small PRs to LangChain, Instructor, Promptfoo, llamaindex, or any tool used by your target companies. Visible technical signal + a name they recognize.
- AI Twitter / X presence — comment substantively on threads from people at target companies. Don't shitpost. Post your own builds when you ship.
- In-person — AI Engineer Summit, local AI meetups — SF if you can travel, otherwise the smaller regional ones. One conversation can shortcut weeks of cold outreach.
- Resume applications — still works for bigger companies (AI labs, enterprise). Lower hit rate at small startups where the eng team reads inbound but rarely opens Greenhouse.
Subject: [Their product] + the thing I built last month Hi [Name], I'm Tony — UMD CS '22, last year I built SoloMock (solomock.com), an AI verbal mock-interview app using the Realtime API + WebRTC. The eval methodology — how I grade verbal precision instead of correctness — is the part I'm proudest of. I'm reaching out because [one specific thing about their product that I actually noticed — a launch, a blog post, a feature]. I think the [retrieval / agent / eval / UX] piece overlaps with what I just built, and I'd love 15 min to ask how you're thinking about [the hard problem in that space]. Happy to send a 2-min demo video. Resume + portfolio: toyinyu.com. — Tony
Personalize the middle paragraph every single time. Generic outreach gets ignored. Specific outreach about their work gets ~10–20% reply rate at small startups.
- Simon Willison — simonwillison.net. Daily LLM news with sharp technical takes. The bar for AI taste.
- Latent Space — latent.space (swyx + Alessio). Long-form interviews with people building AI products. The vocabulary lives here.
- Anthropic engineering blog — anthropic.com/research and anthropic.com/engineering. Frontier model behavior + applied AI patterns.
- OpenAI cookbook — github.com/openai/openai-cookbook. Concrete patterns you can crib.
- Hamel Husain — hamel.dev. Evals + LLM engineering, opinionated and practical.
- Chip Huyen — huyenchip.com. ML systems + LLM productionization.
- Eugene Yan — eugeneyan.com. ML/AI patterns, evals, production lessons.
- AI Engineer YouTube — talks from the AI Engineer Summit. Watch one per week.
Six months of this gives you the vocabulary interviewers indirectly test for. Skim daily; deep-read one piece per week. Form opinions — that's the point.