At the end of 2024, I did not write an end-of-year reflection post (I didn't even have an active blog then!), but if I did, it would have been... morose. Sleep-deprived from travel and frustrated with work, I furiously poured all of my feelings and insecurities into an Apple note: an inescapable panic at all the things I wanted to do, read, and know, and the limited time I had to get to them given my demanding work schedule; comparing myself to others with coveted jobs or cool blogs, thinking I could never have those things; the loss of self-trust from piles of failed ML experiments; and shame from my seeming inability to detach myself from my phone and be present during my time away from work.

Since then, objective reality has changed some, and I have actually achieved some of the things whose lack I bemoaned. But more importantly, my spirits and attitude have improved—partly from recognizing that my feelings about my reality aren't usually caused by my reality; they're just as often caused by something else, like lack of sleep, forgetting to eat, or not doing enough yoga. I still don't feel good about the fact that I only have one life to live, and that every decision means forgoing a different path. And I certainly haven't fixed my Twitter vice, only stopped the bleeding. But thinking back on this year, I feel much more confident, self-assured, and optimistic, despite dealing with setbacks and loss along the way. Here are a few of the notable things that made 2025 one of the most years to ever have happened.

Working on What I Want to Work On

The single most important shift in my thinking came from a conversation with Michael, a trusted YC advisor. The endless stream of AI companies raising at ever more eye-popping valuations and sharing splashy over-produced videos about it has made it easy to lose sight of why you decided to build a company in the first place. There's a sense that if you aren't making $10 million after working on something for 2 months, it's hopeless and you should give up. I know I'm not alone in feeling this; it's a sentiment that has metastasized throughout the startup ecosystem, and its causes are easy to understand: the need to fundraise, the desire to attract talent, and an approach to growth that prizes distribution over all else, as SaaS becomes easier and easier to build (see: Technical Deflation).

I'm sure that the peak of the crypto bubble felt similar: if you hadn't built a token based on a cartoon dragon wearing a diaper and sold it for $100 million, you were never going to escape the permanent underclass. During times like this, focus is critical, and it's impossible to focus if what you're choosing to work on is based on the current shiny object. The shiny object will change faster than you can make a successful company. Vector databases one day, MCPs the next, computer use and coding agents after that. Michael told us something really simple: work on something you're okay failing at for at least a year.

At first, I thought for me that meant finding interesting technical problems in AI research. Computer use (automated QA testing?) and custom embedding models came to mind. But as it turns out, these, too, were shiny objects, interesting because they were the next cool thing in AI, but easy to lose interest in as trends changed. My co-founder Brian told me that I needed to care about the problem, not just the technical details. And he was right. Weeks of self-reflection and discussion brought us to where we are now: building AI software for lawyers and the public sector. And though it's not easy, it's more satisfying than making an AI QA tester ever would have been.

Blogging & Meeting Interesting People

The second-best thing that happened in 2025 was starting this blog. At the end of last year, I read Zhengdong's amazing end-of-year post (he has another one for 2025!) and instead of feeling inspired by it, I made myself feel bad: here was a man whose interests bridged humanities and frontier AI research, and lo, he also picked out a beautiful serif font. You, sad short-form Twitter addict that you are, could never summon up the energy required to put 800 words of HTML onto the internet, let alone with a tasteful stylesheet.

But that turned out to be wrong. This May, I pulled out my old Eleventy website that I hadn't updated since 2023, added a new page, and made a bit of time to write on the weekends. And it's been, no exaggeration, life-changing. I don't think my posts are particularly deep or insightful, but I do have a unique voice, and it seems to be something people enjoy reading (yay!). Posts from the blog have been shared within AI labs, have been subtweeted by famous researchers, and have led to tons of Twitter inbound (friend offers and job offers; no husband yet). I don't want a new job, but if I did, I can confidently say that 1 blog post is worth 100,000 job applications.

Thanks to the blog, I've made way more online friends, and met some new people IRL too. The activity resulting from blogging has bolstered my confidence, giving me faith that I could do anything I set my mind to, that I wasn't in a race against time to escape the permanent underclass, that there were people out there who would happily give me a job if I was ever starving on the street. Weirdly, all the interest and offers of employment have made it easier to tune out the noise and focus on my company, free from fear of closing doors.

More Failed ML Experiments

In 2024, I couldn't separate failed machine learning experiments from my identity. If the experiments were a failure, I was a failure. This led to an obsessive attitude (one, I'm sure, that is shared by many researchers, and probably makes them very productive, at the cost of their mental health). Part of what's happened this year is that I had less time to do this stuff: I've had other side projects that were less capricious, including my fabulous open-source LLM SDK lm-deluge, which now has its own website and which is used by researchers at esteemed research institutions like AI2.

But also, I think I just stopped caring about failing as much. Not only is failing required to learn and succeed, but also, the happiness that comes from doing something "SOTA" is fleeting, and quickly gives way to anxiety and jealousy as someone else does something slightly better than your thing. Your silly little experiments on 1 node of GPUs can absolutely be novel, interesting, important work; but you shouldn't worry so much about them being "the best," whatever that means. Especially if you're just some guy and not an AI lab with infinite resources. The only way to win that game is to not play. Or be Google.

This year, I mostly played around with computer-use models, building datasets for visual grounding. This wasn't for work, just a pet interest. I learned that all of the big computer use datasets are full of garbage. I also learned that the obvious interventions to filter out the garbage (remove 404 pages, stratify pages by numbers of buttons and see if different strata result in a better model) didn't help that much. Scaling up to a way bigger FineWeb grounding dataset didn't help a lot either. The best way to improve performance on WebClick was to train on WebClick. Remember folks, an "eval" is just another name for a really clean training dataset.

Things Happen

In January, I went to a Crab Feed at a Greek Orthodox church and met a lovely group of friends who I now see basically every month. In February, I went to Missouri to say goodbye to my grandfather (Papa) Bob Anderson. He was an engineer and was always interested in whatever I was working on; I am sad that he won't be here to see my company succeed, but I know he would be proud. In March, I went to LA to visit my sister and see her college. I don't like LA, but I had a good time anyway. My friend Thomas said I was a "normie" when I told him about the sci-fi I liked; I asked him what counts as non-normie and he introduced me to Peter Watts. I am now a proud non-normie.

In April and May, we had the aforementioned pivotal conversations about the future of our startup, and emerged more determined than ever to build something that really matters. Over the summer, I saw my family at our annual Lake of the Ozarks trip. In October, I visited Japan, and officially became an Annoying Japan Guy. (Why can't our trains go that fast? Why don't we have coin lockers? Why is our food poison? Why do we have to have poop on the street?) I lost 5 lbs just from eating their non-poison food for a week and walking 20k steps a day. When I got back I started making udon & soba noodles at home and eating out less. Hope it sticks.

I hosted an awesome Friendsgiving in November, turned 29 in December (yikes!) and got to spend some time with family over the holidays, which was dampened by my great aunt falling and landing in the ICU. We're not sure if she'll make it out. The happy and the sad things remind me equally how important it is to spend time with loved ones, and to be fully present during that time. Discovering Vacation Mode in Clearspace has been a setback on that front, and I hope to do a better job protecting my neurons from algorithmic feeds in 2026.

This year, I'm excited to read more about financial bubbles and speculation from tulips to railroads, which will be relevant, since we'll hopefully get to learn once and for all if we're in a bubble and it will all come crashing down, or if we're in a slow takeoff and have 2 years to accumulate capital before the music stops. Or maybe China's about to invade Taiwan and theoretical debates about takeoff speeds are about to become a quaint luxury. Hope not!

Happy 2026, thanks for reading my blog and being part of changing my life!