Hi I'm Anton!
I work on inference and pretraining stuff at anthropic.
Story
I am pretty monomaniacal about AI.
In undergrad (Boston University 2014-2018) I got the idea that I was going to go into hardware design. I focused on digital logic classes and eventually implemented my own CNN accelerator on FPGA. But it turns out that the market for software engineers is way better than for hardware engineers.
Out of college, I joined
Omniscience
as a general software engineer, which eventually became general ML infra. This was the era of scikit-learn, CNNs, and tensorflow. When someone went and implemented an attention-based neural net, it was very much a new and fancy thing.
In September 2019 I left to go backpack Asia for 6 months. I stayed in hostels for ~a week or two at a time, bouncing around:
- Thailand
- Hong Kong (I got spicy-water-cannon'd by the Chinese police!)
- Singapore
- Taiwan
- The Philippines
- Japan
Around this time I wrote my
citrine
inference server, based on what I think was a correct intuition that distributing models to non-specialist end users would be a popular project. Sadly, I got scared off by the many others doing similar things.
Eventually I joined
CrowdAI
as a backend engineer. The thesis there was that the best way to "democratize" AI would be to give everyone the tools to train their own (small) models.
In late 2020 I moved to Milwaukee as a convenient covid bunker.
...but then in February 2021 I learned that my father had developed
Glioblastoma
-- stage 4 brain cancer. I moved back to the Peninsula to spend time with him, and eventually left CrowdAI for the same reason. He passed away in January 2023, after watching all of his kids graduate college. I wish he could have seen more.
This was not a good time.
In March I started to try to move on, and moved from the Peninsula to San Francisco. At this point it was plain to see that small models were not the future, and any impact I might have would be at a highly-resourced scaling lab. I made a
Manifold Market
to forecast my job search, and the marketeers correctly predicted that I would end up at Anthropic.
Since July 2023 I've been on the Anthropic inference team, all over the stack but generally trending towards infrastructure.
I've mostly been neglecting life besides this, with the intent that I'll have plenty of time to make it up post-singularity.
Vibes
I value when systems are effective, independent of the actual goals of the system.
I like being able to do things myself. In software / at work, this manifests as a horrible itch whenever there's a part of the stack I can't contribute to, and a kind of "bulldozer" energy about making things work and discovering information despite a lack of primary sources. I've done an unusual amount of yak shaving, which isn't efficient in the moment, but it does mean I've accumulated a wide repertoire. I'm used to being the backstop for technical issues.
Outside of software, the DIY energy manifests as a stream of hobby projects. These include:
- Soap making
- Japanese kotatsu tables (woodworking!)
- Leatherwork
- Electronics
I'm a decent baker and cook. I've done a few batches of various alcohol, but in general I'm terribly unreliable at fermentation. I'd like to pick up chemistry and metalworking at some point. Post AGI things.
I'm somewhere on the rat spectrum -- I've read lots by
Gwern
,
Yudkowsky
,
Scott Alexander
, et al. I was very active on manifold for a while, but have tapered off since joining Anthropic (I no longer make public predictions about AI). I'm not part of the core Berkeley rat community, but I do show up at lighthaven from time to time.
I'm a bit of a weeb -- I read a lot of manga, and can speak barely-passable Japanese (picked up from anki kanji flashcards).
Mostly I read nonfiction, but I also enjoy a lot of ratfic and hard sci-fi.