2024 Year in Review
Eugene Yan
ML engineer Eugene Yan posted his 2024 year-in-review, covering shipped LLM systems, two side apps, and a health overhaul. It's a rare data-backed look at what building in public actually costs and returns.
Based on reporting by Eugene Yan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Eugene Yan just published his annual review, and it reads less like a highlight reel and more like an engineering postmortem applied to a whole year of life. The headline for his day job: the LLM prototypes he built in 2023 finally became production systems in 2024, serving real customers at scale. That's the unglamorous part of AI work nobody tweets about — getting something reliable and cost-effective into production is a different skill than getting a demo to work once.
He also kept writing, publishing six substantive pieces on things like synthetic data generation, LLM-as-judge evaluation, and prompting fundamentals, plus a widely-read collaborative piece, What We've Learned From a Year of Building with LLMs, that made it onto O'Reilly as a book. Speaking gigs followed: Netflix's internal RecSys conference, the AI Engineer World's Fair closing keynote, and a Weights & Biases hackathon judging slot.
The more interesting bets were two apps he shipped solo. Tara, a voice-based AI coach, has quietly helped hundreds of users — one story he shares involves Tara talking someone's spouse into finally quitting a job they called "soul-sucking." The second, AlignEval, is his attempt to fix how people build LLM evaluators, and the usage numbers are oddly specific: 538 dataset files uploaded, 289 labeled with at least 20 samples, only 84 taken all the way to optimization. That drop-off from upload to full optimization is basically the funnel every eval tool struggles with, and Yan just showed his numbers instead of hiding them.
The site metrics tell a split story. Google Search impressions fell 27.3% and clicks fell almost exactly the same amount, which lines up with the broader narrative that search traffic is eroding site by site as AI answers eat into referrals. Yet unique visitors on eugeneyan.com rose 10.5% to 285,000, likely thanks to Hacker News spikes rather than search. Newsletter subscribers grew from 6,200 to 9,500, a 34.7% jump, while Twitter and LinkedIn followings grew more modestly.
On the personal side, he leaned harder into diet and exercise changes started in 2023 — intermittent fasting since September, daily 30-minute workouts since June, tracked body fat and resting heart rate both trending down. The one clean failure: no black-run snowboarding, after admitting the goal was unrealistic given only five days a year on the slopes. It's a small, honest miss in a review otherwise stacked with wins.
My take — AI-written commentary, not fact-checked reporting
What I like here is that Yan treats his own year like a system with metrics, not a vibe — and he's not afraid to publish the failure right next to the wins, which almost nobody in the AI-influencer space does. Everyone posts their launch numbers; almost nobody posts their search traffic dropping 27% or their unfinished goals. That kind of transparency is worth more than another thread about a benchmark score.
Read more about this at: Eugene Yan