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

Yair Alon spent a decade inside the belly of the beast. At Google Research and DeepMind, he was instrumental in the tech we now take for granted, like the Portrait Mode algorithms on your Pixel phone. But after ten years of seeing how the sausage is made at a trillion-dollar giant, Alon did what a lot of high-level researchers are starting to do. He left the comfort of Big Tech to join a startup called Poetiq.

The Golden Cage of Big Tech Research

Working at DeepMind is often seen as the pinnacle for an AI researcher. You have unlimited compute, the world's best datasets, and a paycheck that makes most founders jealous. But there is a hidden cost: velocity. In a massive organization, your work is often buried under layers of product management and risk aversion. You might spend two years optimizing a single feature that may or may not see the light of day.

For Alon, the transition to Poetiq represents a shift from incrementalism to foundational problem-solving. He isn't just trying to make a camera lens blur look slightly more natural anymore. He is tackling the recursive loop of AI development. The goal is simple to state but incredibly difficult to execute: building a system that can actually learn from its own mistakes without a human holding its hand every step of the way.

The Problem with Static Models

Right now, we are living in the era of frozen intelligence. We train a massive model, we freeze its weights, and we ship it. If the model gets something wrong, it stays wrong until the next training run. This is inefficient and, frankly, a bit primitive. If you're a builder in this space, you know the frustration of seeing a model hallucinate the same error over and over again despite your best prompt engineering efforts.

The vision Alon is chasing at Poetiq is the self-improving loop. In the startup world, we call this the 'flywheel.' In AI research, it's the holy grail. If a model can identify its own failures, generate synthetic data to correct those failures, and then integrate that knowledge back into its core, the speed of progress moves from linear to exponential.

Why Builders Should Care

As an editor and a founder, I talk to a lot of people who think the AI game is already over. They see OpenAI, Google, and Anthropic and assume the 'foundational layer' is a closed shop. Alon’s move proves the opposite. The big players are excellent at scale, but they are often clumsy at the architectural level of self-improvement. They are essentially building bigger and bigger engines, while startups like Poetiq are trying to invent a fuel that never runs out.

For those of us building on top of these APIs, this shift matters. If the next generation of models can self-correct, the overhead of 'vibe-checking' and manual QA goes down significantly. We are moving away from 'prompting' and toward 'teaching.' That is a massive paradigm shift for how we design software.

The Talent Migration

We are seeing a massive drain of talent from the 'MAGMA' companies into the seed-stage ecosystem. When people like Alon leave, they aren't just taking their resumes; they are taking a decade of intuition about what doesn't work. In the early days of any tech cycle, knowing what to ignore is more valuable than knowing what to do. Alon likely saw the limitations of the current LLM trajectory while at DeepMind and realized the solution wouldn't come from a corporate committee.

• Speed over Scale: Big companies have the GPUs, but startups have the agility to test wild architectural theories.
• Focused Mission: At a startup, you aren't building a chat interface; you are solving a specific mechanical failure in AI logic.
• Ownership: For researchers, the ability to see their math translate directly into a live product in weeks rather than years is the ultimate drug.

Building at a startup isn't about having the most resources; it's about having the fewest distractions.

The Skeptic's Corner

I’ve seen plenty of 'research-heavy' startups burn through millions of dollars without ever shipping a product people actually want to use. There is a risk here that Poetiq becomes a laboratory rather than a business. The 'self-improving AI' promise has been around for a while, and we haven't seen a truly autonomous loop that doesn't eventually degrade into gibberish or 'model collapse.'

However, Alon’s background gives this specific attempt more weight than your average YC pitch. If you can figure out the edge detection and depth mapping for Portrait Mode at a global scale, you understand the nuance of training data. You know that the real world is messy, and a model that can't handle messiness is useless.

The Takeaway for the Rest of Us

If you're a founder or a developer, don't just wait for the next GPT-5 drop. Watch the researchers. The fact that top-tier talent is fleeing the giants to work on 'self-improvement' tells you exactly where the bottleneck is. The future of AI isn't just bigger parameters; it's better feedback loops.

Alon is betting his career that the next breakthrough won't come from a bigger server farm, but from a smarter way to learn. Whether Poetiq succeeds or fails, the direction is clear: the era of static, hand-fed AI is coming to an end. We're about to see what happens when the machines start helping us build them.

Read the original at Y Combinator Blog →