How Yair Alon went from building Portrait Mode at Google to founding researcher at Poetiq

Yair Alon spent a decade at Google Research and DeepMind. He built the deep learning models behind Portrait Mode on Pixel phones. He published papers on efficient deep learning. He designed models for Search, Ads, and YouTube. His last project was VideoPoet, a video generation system. Everything was great, by his own account. But he felt that small startups had become where the most exciting innovations were happening. So when former DeepMind collaborators told him they were starting a company, he knew immediately he wanted in.

That company is Poetiq (YC S26, San Francisco). In one sentence: Poetiq is building an AI system that can optimize any other AI system, including itself.

Leaving Google after a decade

Yair had been at Google since his PhD internship at CMU, where he studied machine learning and computer vision. He knew how to be productive and successful inside that ecosystem. Leaving felt scary. "I was there for so many years," he says. "I knew how to be productive and successful in that environment so leaving felt like a scary new thing." He already knew the founders and trusted the technical direction, so the real hurdle was personal. He had to find the courage to walk away from something comfortable toward something unproven.

What he actually does all day

Yair is one of a handful of people who started on the first day of the company. His official title is Founding Research Scientist. Unofficially, he calls himself Coffee Evangelist. In practice, being early means working on everything, sometimes in a single day: "You could be working on a new ML technique in the morning, fixing issues with the DB later in the day and then in the afternoon meeting a realtor to look at a potential new office."

One project he is particularly proud of is the experimentation infrastructure. At Google, entire teams handled experiment launchers, dashboards, and visualization. At Poetiq, none of that existed, and the work is novel enough that off-the-shelf tools were not relevant. He and the team built a first-tier experimentation system from scratch, with detailed analysis of how their optimizer is performing.

Building an AI that improves itself

Poetiq's core bet is on recursive self-improvement: an optimizer that can improve any AI system — harnesses, agents, skills — and that can also improve itself, so the rate of improvement accelerates. Yair puts it simply: "The main thing we are building at Poetiq is an AI that can improve other A.I.s and itself. I think nothing can top that."

Early on, the work was split between two tracks: figuring out the algorithmic foundations of the meta-optimizer and building the experiment-running infrastructure to validate it quickly. More recently, his focus has shifted to scaling — identifying what is missing so the system can run longer without human intervention, finding what would cause friction if they wanted to ten-times their experiments, and spending significant time on hiring and prototyping Poetiq's interview process.

People, challenge, autonomy

Yair describes three things that matter most to him in a company: people, challenge, and autonomy. On people, he says the team is genuinely enjoyable to spend time with. On challenge, he frames it as the rare combination of creating a new domain of AI research while simultaneously inventing a new product category. On autonomy, the founding team — all with deep experience in industry and academic research — avoids micromanagement. Instead, they run frequent deep-dive syncs to stay aligned on priorities and converge quickly when needed.

He is candid about tools, too. He has used VS Code, Anti Gravity, and now mostly Claude Code, but he thinks people put too much emphasis on the setup. "The key thing is knowing what you want to build, having a clear design for it in your mind, and knowing what success looks like. Then whether you use Claude Code, Codex or whatever doesn't matter that much."

Who would thrive here

Yair's advice for anyone considering an early-stage role is grounded: "Realistically most startups do not become Anthropic. So you need to go into it with the correct mind set. The goal is to learn a lot, be part of a small team executing together, and be able to contribute across the board. Anything more is just icing on the cake." If you are an engineer or researcher who wants ownership over the full lifecycle of what you build — and you are drawn to a problem no one has solved before — Poetiq is hiring.