On the Optimal Trajectory from Tehran to Toronto: A Multi-Objective Study in Citations, Diffusion Models, and Converting a Google Brain Salary into Startup Equity

Background & Motivation

The literature on optimal career trajectories is vast and largely useless. Most studies focus on incremental improvements — a promotion here, a lateral move there. This paper documents a far rarer phenomenon: the compound interest career, in which each position creates exponentially more optionality than the last.

The subject begins in Tehran, Iran, where he completes his undergraduate studies at Sharif University of Technology — an institution whose CS department has a disturbing habit of producing people who go on to reshape entire fields. He then relocates to the University of Toronto for a PhD under David Fleet, where he is awarded a Google PhD Fellowship in Machine Learning — a signal that Google was already investing in this particular human before the graduation ceremony.

His PhD thesis on scalable similarity search quietly laid the groundwork for everything that followed: if you can find similar things quickly, you can build systems that generate entirely new things. We suspect he knew this at the time. The thesis committee did not.

Experimental Setup: The Google Brain Era (2016–2023)

Upon completing his PhD in December 2015, the subject was absorbed into Google Brain, Mountain View in January 2016. He transferred to the Toronto office in 2018, presumably because he missed winter and/or poutine. He rose to the rank of Senior Staff Research Scientist — a title that, within Google's leveling system, roughly translates to "we would prefer you never leave."

The experimental procedure during this period can be summarized as:

while True:
  idea = sample(brilliant_ideas)
  paper = write(idea, coauthors=[Hinton, et al.])
  citations += thousands
  if bored: break

The key projects — our "methods" — were as follows:

Quantitative Analysis

Table 1 presents a comparative analysis of the subject's output against common benchmarks for "a productive career."

Departure from Google Brain

This retraction follows a well-documented pattern in the literature, commonly referred to as the "Google Brain to Startup Pipeline." Notable prior work includes the departures of the entire team that went on to found [redacted for competitive reasons]. The common causal factor appears to be: showing researchers what's possible, then not letting them ship it.

Ideogram AI — The Startup Paper

In 2023, the subject co-founded Ideogram AI in Toronto with a team of ex-Google Brain researchers. The thesis was deceptively simple: what if people could generate images where the text actually says what they typed?

This may sound trivial. It was not. At the time, every major image generation model — Midjourney, DALL-E, Stable Diffusion — produced beautiful images with text that looked like it had been written by a drunk doctor. "Happy Birthday" would emerge as "Hpqy Birhty." "Open" would render as "Oepn." The entire field had collectively given up on text rendering.

Ideogram launched and immediately nailed it.

The product launched as a free tool in August 2023. Within the first month, it had 500,000+ users. By 2025, Ideogram had shipped versions 1.0, 2.0, 2a, and 3.0 — each setting new benchmarks for text rendering, photorealism, and style coherence. Bloomberg, Globe and Mail, VentureBeat, and every AI newsletter you subscribe to covered it.

The company is based in Toronto, which the subject has now established as a legitimate AI hub rather than just "the city where Hinton lives."

The Hinton Coefficient & Other Observations

A common critique in career studies is the "Hinton Coefficient" — the suspicion that anyone who co-authors with Geoffrey Hinton automatically receives a citation boost via academic gravity. We address this directly.

Yes, the subject co-authored multiple papers with Hinton. But he also co-authored papers without Hinton that received thousands of citations (Imagen, Palette, WaveGrad, Neural Combinatorial Optimization). More importantly, Hinton himself chose to collaborate with Norouzi repeatedly — and Hinton does not co-author out of charity.

We further note that the subject has published with Samy Bengio, Quoc Le, Jimmy Ba, David Fleet, Honglak Lee, Dale Schuurmans, Nando de Freitas, and approximately half of everyone who has ever won a NeurIPS Best Paper Award. This is either evidence of exceptional collaborative ability or proof that he sits in a very specific chair at a very specific cafeteria.

The subject's research also spans an unusually wide range: computer vision, NLP, reinforcement learning, audio generation, 3D reconstruction, medical imaging, hashing, and program synthesis. Most researchers pick a lane. Norouzi picked all of them.

Reviewer Comments

Final Verdict

Acknowledgments

The authors thank David Fleet for supervising the PhD that started it all, Geoffrey Hinton for co-authoring the papers that made the citation counter break, and Google Brain Toronto for providing the research environment and free snacks that fueled seven years of groundbreaking work.

We thank a16z and Index Ventures for writing checks large enough to make leaving Google feel like a reasonable life choice.

We thank Sharif University of Technology for continuing to be an unreasonable source of talent in the global AI ecosystem, and Iran for exporting one of its finest minds to Canada, which Canada will definitely not give back.

We thank the Ideogram team for joining a startup founded on the radical premise that AI-generated text should be legible.

Finally, we thank Reviewer 2, whose tireless negativity keeps the entire peer review system honest, and whose own citation count we decline to look up.

^† This paper was not generated by Ideogram, though we suspect the figures could have been.   ^‡ No Google Brain researchers were harmed in the making of this startup. Several were recruited.

Selected Bibliography