Atlas Computing updates: Q3
Evan Miyazono
Hello Atlas community,
I realize we're about halfway through Q3, but here’s our a quick update (since our late Q2 update posted in April):
Project updates
Most exciting: Atlas is back, and better than ever! We have a new take on our strategy:
We’re focusing on making it easy for specialists to tackle the most important, neglected problems related to powerful AI. That means we’re going to be making a list of problems, possible solution directions, skills needed, etc.
We’re taking an approach that’s agnostic to organization form, so expect to find things that could be solved by start-ups, frontier AI labs, nonprofits, think tanks, policymakers, and/or others.
We’ll work to curate this list and try to make it easier for a small group of specialists with the right skills to start making progress on a specific solution.
If you want to contribute ideas, please send your idea to probl...@atlascomputing.org. These should ideally be an idea where one of the people working on AI whom you respect most might agree with you that this is a very important problem that someone should be working on.
We also officially got our IRS 501(c)(3) determination letter, so we’re our own nonprofit now!
Other updates:
I wrote a proposal on red- and blue-teaming AI systems to prevent AI Sleeper Agents that was published today! I think that China hawks and AI Security advocates might agree more on this proposal more than anything else I’ve come across.
Ask: if you have a moment, please like+repost IFP’s twitter announcement or my quote.
I'm now a research fellow at the nonprofit Convergent Research! This is particularly exciting because the formal verification* and hardware governance mechanism efforts I’ve been trying to drive both make more sense at the scale of a focussed research organization (FRO). These efforts will continue, but many of them will take place with my Convergent Research hat. Since that's much of what I was doing at Atlas, the other exciting update is…
I had a bit more writing, including these two top blog posts on Substack since Q2 update
This post about what it would be like to have a specification-based workflow for generating software (where engineers may never actually see code)
I’ve also written a piece with the Institute for Progress submitted to the Senate’s ASAP call, which will be published as part of IFP’s launch sequence. I think my proposal might be something that AI safety people and accelerationists mostly agree on…
I was on Humans on the Loop, an interview series by Michael Garfield (previously host of the Santa Fe Institute's amazing podcast Complexity)
Events:
I was recently at AI for Good, speaking about hardware governance, the need for formal specifications, and other risks. I participated in two sessions, and presented these slides at one of them.
I also joined Convergent Research’s quarterly offsite at Edge Esmeralda in Healdsburg.
Growth:
We’ve brought on Tzu Kit Chan as Chief of Staff at Atlas. He’s already helped clarify our competitive advantage and helped form the new strategy, and has also increased my output by owning urgent, lower priority items and helping me maintain focus on priorities.
Travel updates
Upcoming events: catch us in person or let us know if we should meet anyone
Anyway, thanks for reading!
- Evan
p.s. What Evan’s reading…
You can tell I had my first vacation in a while, plus some travel time
Meditation for Mortals definitely has some nice perspectives. I hope and plan to revisit this one, despite not agreeing with all of it.
A Brief History of Intelligence is a great description of five breakthroughs that caught life from no neurons to human brains. I also had a delightful conversation with a researcher last month used these breakthroughs to track AI progress, emphasizing both (1) that current architectures don’t really seem to accommodate model-based reinforcement learning, so AI systems may need explicit models or training mentalizing before they we can unlock aspects of alignment, and that (2) AI systems don’t need these to be catastrophically dangerous. Skim some of these articles for a taste of what’s in the book.
Emperor of all Maladies – fascinating history of cancer and cures. Nice reminder that the way we view problems is very dependent on the historical context at the time.
Moral Ambition was recommended to me a few times and I totally see why. I’m hopeful that the book can start a movement around working hard to do good. Obviously Effective Altruism was discussed, and while I haven’t attended an EA event or consider myself to be a part of that community, I think it seemed like a fair coverage.
How We Got to Now was a fun book full of anecdotes about technological development (each chapter is humanity's progress with a different category of technology, like glass or the ability to engineer “cold”). I think my favorite was probably the fact that fiberglass was first created by attaching a piece of molten glass to a crossbow bolt.
I hope that Abundance leads to some interesting political changes. I do think that degrowth is not a viable political solution to things like climate change, and it’s interesting to understand better how it came about that US politics became focused on procedure over outcome, and legitimacy by process.