I've done a bunch of AI Safety programs, some as staff, some as participant. So I wanted to share a take on a very special one that just finished its 1st edition: CAMBRIA.
Participant selection: spot on. The attendance was diverse and also everyone was on the same page, which is hard to achieve. There were generous travel allowances for international participants, one guy flew from Australia, one girl from Vietnam and myself from Argentina. Most of the cohort was from different top schools in the US (MIT, Brown, Stanford, Harvard) and some San Francisco folks.
Knowledge level of participants before the program: there was a lot of variance. And that was part of the beauty of it. The program is centered around pair programming and everyone was pretty cracked but with different backgrounds so it was great to teach each other our little tricks and to also ask the stupid questions that we all have on the back of our heads.
Use of LLMs: almost 0 during the programming exercises, as much as you want for the final capstone project. Old school coding was very refreshing to me, it felt like a detox period or something like that. Gives you a sense of much more ownership over what you are building and learning.
Office: it's right in front of Harvard Computer Labs and it has everything you might need to have a great time.
Intensity: it's a lot. If you are juggling with many other things apart from this, I wouldn't recommend it. It's great if you are in a transition period or something like that. This will require your full attention if you expect to make the most out of it.
Curriculum: it's a compressed version of ARENA. Very cool overall, but it's very packed and notebook-based learning has its limitations. Prioritization is hard and it's easy to get lost in implementation details and lose sight of the big picture. The content is definitely high quality, though I'd recommend tweaking the notebooks a bit for this experience in order to highlight / prune some of the parts.
TAs: they actually cared a lot about your progress and actively engaged. I often prefer talking to an LLM or figuring things out myself because TA interaction is usually disappointing. This was not the case at all. They were much better than other TAs I've seen in top schools. Thanks @Dmitrii Troitskii, @Claudio Mayrink Verdun and @Jiaming Yu!
Capstone project: Compressed as it can be. A bit too compressed in my opinion.
Potential Improvements: It would be good to have recap sessions to digest all the content. Also nice to have: ~1 social activity organized by the program each week.
Summary: I wholeheartedly recommend it.
Keep an eye on CBAI for future opportunities like these.
I've done a bunch of AI Safety programs, some as staff, some as participant. So I wanted to share a take on a very special one that just finished its 1st edition: CAMBRIA.
Cambridge, Massachusetts. 3 weeks. 20 nerds. Hands-on technical upskilling.
Participant selection: spot on. The attendance was diverse and also everyone was on the same page, which is hard to achieve. There were generous travel allowances for international participants, one guy flew from Australia, one girl from Vietnam and myself from Argentina. Most of the cohort was from different top schools in the US (MIT, Brown, Stanford, Harvard) and some San Francisco folks.
Knowledge level of participants before the program: there was a lot of variance. And that was part of the beauty of it. The program is centered around pair programming and everyone was pretty cracked but with different backgrounds so it was great to teach each other our little tricks and to also ask the stupid questions that we all have on the back of our heads.
Use of LLMs: almost 0 during the programming exercises, as much as you want for the final capstone project. Old school coding was very refreshing to me, it felt like a detox period or something like that. Gives you a sense of much more ownership over what you are building and learning.
Communication and Logistics: Impeccable (thanks @Alex Semendinger).
Office: it's right in front of Harvard Computer Labs and it has everything you might need to have a great time.
Intensity: it's a lot. If you are juggling with many other things apart from this, I wouldn't recommend it. It's great if you are in a transition period or something like that. This will require your full attention if you expect to make the most out of it.
Curriculum: it's a compressed version of ARENA. Very cool overall, but it's very packed and notebook-based learning has its limitations. Prioritization is hard and it's easy to get lost in implementation details and lose sight of the big picture. The content is definitely high quality, though I'd recommend tweaking the notebooks a bit for this experience in order to highlight / prune some of the parts.
TAs: they actually cared a lot about your progress and actively engaged. I often prefer talking to an LLM or figuring things out myself because TA interaction is usually disappointing. This was not the case at all. They were much better than other TAs I've seen in top schools. Thanks @Dmitrii Troitskii, @Claudio Mayrink Verdun and @Jiaming Yu!
Capstone project: Compressed as it can be. A bit too compressed in my opinion.
Potential Improvements: It would be good to have recap sessions to digest all the content. Also nice to have: ~1 social activity organized by the program each week.
Summary: I wholeheartedly recommend it.
Keep an eye on CBAI for future opportunities like these.