Last Friday was my last day as CTO at Wave, capping an incredible ~8 years filled with more professional and personal growth, joy, and meaning than I could have hoped for.
This was the hardest decision I’ve made. I’m still just as excited about Wave’s mission and trajectory. And most of the important work is still ahead of us: as of yet there’s only one of our potential markets where we move over 70% of GDP.✻ Note that the ceiling for this metric is much more than 100%, because some types of money transfer (e.g. gifts) don’t count towards GDP. Most wrenchingly, I’ll be saying goodbye to a huge number of dear friends, inspiring colleagues, and incredible mentors.
It’s been the opportunity of a lifetime—helping grow Wave’s engineering team has taught me a tremendous amount, and by helping build something that’s important to millions of underserved people, I’ve been lucky to have more positive impact than I ever expected to be able to. It’s hard to describe how much Wave has meant to me, though I hope to capture more in a future post.
But over the last few months, I’ve had two key shifts that made me think about working on something else.
First, it finally feels like Wave is at a point where I’m not critical for engineering work to go smoothly. We’ve hired and grown lots of other amazing people who can take over every part of my job—building product, making it scale, hiring, and building teams and organizations—without needing much input.
Second, recent progress in AI has made me pretty worried about whether that transition is on track to go well for society. For example, large language models are becoming human-level at new tasks faster than anyone expected, but we don’t know how to provide almost any guarantees about their behavior; our best known safety measures, like those on ChatGPT, are easy to bypass. For a model with a ChatGPT-ish level of capability, this is concerning but not catastrophic. But if you extrapolate the scaling laws out a surprisingly short time, things start to look scary.† For a more rigorous, in-depth overview of the reasoning behind this, I recommend Holden Karnofsky’s series of articles on The Most Important Century and its implications. Holden is one of the main people who convinced me of the importance of mitigating risks from advanced AI: he’s one of the most careful and rigorous thinkers I know, so when he updated from “this seems wild and unlikely” to “I’m spending most of my time worrying about this,” it really got my attention.
Because of this, while there’s never a good time to change direction, now felt like the least-worst time to take a leap.
When I looked at what organizations were working on AI alignment, and filtered down to the ones where it looked like the engineering leadership skills I’ve built at Wave could be most useful, the clear standout ended up being Anthropic. Anthropic builds large language models, similar to the one behind ChatGPT,‡ Anthropic doesn’t have a comparable public product, but they do have one (“Claude”) in private beta, and you can find people posting examples on twitter. and does safety research on them—for example, training them to be more helpful, less harmful, and more honest.
Sidebar to avoid contributing to any potential information cascade: I want to emphasize that this doesn’t mean I personally strongly endorse everything Anthropic does, and that much of my excitement about Anthropic at this point comes from the vetting of other people I trust, not from my own first-principles reasoning.
In particular, for example, Anthropic probably has at least some effect of accelerating the “race through the minefield” to develop transformative AI. They try to be thoughtful about how to balance this trade-off (e.g. holding back the release of their chat model), and most of the people I got advice from thought their existence was worth the trade, but some were quite uncertain.
I intend to get to the point where I can have my own first-principles takes on what I do/don’t endorse, but there are a lot of moving parts here and I’m not very familiar with the space, so I expect it to take a while for me to flesh out my understanding.
Aside from their mission, the thing that makes me most excited about Anthropic is the quality of the team: I originally heard of them because they employed more people I thought were super cool than any other company.§ Besides Wave, which has an unfair advantage since I knew everyone on the product/engineering team. That high bar has let them do an impressive amount for their size in a short time (~2 years)!
While Anthropic’s team is relatively small right now, it’s growing quickly, and I’m excited to try to help navigate that rapid growth using what I’ve learned at Wave. I don’t know exactly what I’ll end up doing, but my plan is to show up, start working on whatever seems most useful, and iterate from there, the same way I did last time.
I’ll also be moving to the Bay Area (likely Berkeley) later this year to work onsite, so will be in the market for new friends—say hi if you’re in the area and interested in meeting up!
All the best, Ben!
I’ve also been considering applying to Anthropic for a while now, since I’ve been intensely obsessed with AI research for the last 8 months (with a more casual interest since BERT came out ~2018), but I feel like I’m underqualified on paper since I’ve been optimizing for knowledge/understanding/investigation above personal branding/concrete outputs/results so far.
I want to put some legible publicly visible contributions/publications/work up first so that I can point to those accomplishments for signaling competence/talent/expertise.
It’s also possible that I’m in a position where I may be able to contribute more to AI safety by staying where I am instead of jumping ship, so the impact calculus is slightly different, although that may simply be rationalization.
Thank you Ben.
Through your posts about your work at Wave, I’ve learned a lot about Fintech, but more importantly about how to think and work (two areas where I still have a lot of progress to make).
I look forward to reading your next posts (I hope) about AI and the work you’ll be doing at Anthropic.
Bon vent! In this new direction.