Twitxy Saved

Dave Banerjee

Verified account This account is verified. Learn more

@DaveRBanerjee

researcher at @iapsAI, all views my own
davebanerjee.ai Joined July 2020
477 Following 1.2K Followers 1.2K Posts

Latest posts 5

Dave Banerjee Verified account @DaveRBanerjee
As a corollary: if most of your labor force for doing alignment R&D and other def/acc R&D will be from future AIs, then it will be v important to have access to SOTA models as soon as possible

By default, OAI and Ant will prob keep these models internal for a while. So it would be highly desirable to have some kind of project glasswing for alignment and def/acc R&D

And to be clear, I think getting this right is probably more important than most field building at the moment
Dave Banerjee Verified account @DaveRBanerjee 2d
Say you want to increase the labor force working on solving the alignment problem

One thing you could do is set up an Alignment fellowship and build a talent pipeline

But if you think that AI R&D automation is imminent, then your best bet at “building the alignment field” is to train AIs to be more capable in domains useful for alignment (e.g., conceptual research)

However, accelerating certain capabilities can be dual-use and actually be net-negative (e.g., shortening timelines)

So the solution is to avoid advancing capabilities through training data. And instead you focus on building scaffolds/harnesses/data-stores that can only be bought/procured for alignment R&D. This prevents the downsides while harvesting the upside
View quoted post on X
Dave Banerjee Verified account @DaveRBanerjee
Say you want to increase the labor force working on solving the alignment problem

One thing you could do is set up an Alignment fellowship and build a talent pipeline

But if you think that AI R&D automation is imminent, then your best bet at “building the alignment field” is to train AIs to be more capable in domains useful for alignment (e.g., conceptual research)

However, accelerating certain capabilities can be dual-use and actually be net-negative (e.g., shortening timelines)

So the solution is to avoid advancing capabilities through training data. And instead you focus on building scaffolds/harnesses/data-stores that can only be bought/procured for alignment R&D. This prevents the downsides while harvesting the upside
6 replies 0 reposts 31 likes 1.9K views View on X
Dave Banerjee Verified account @DaveRBanerjee
New Epoch post on Huawei's biggest data center yet

To me, the main upshot is that semiconductor export controls work shockingly well

The new Huawei data center is China's biggest data center, yet its compute capacity is only 7% of the biggest US data center!!

Even though the Huawei data center is packed with the best Chinese chips, these chips are far worse than Western chips because Huawei doesn't have access to state-of-the-art semiconductor manufacturing equipment, thanks to the export controls on EUV machines back in 2019

The next step to continue Western dominance in chips is to pass the MATCH Act

Passing the MATCH Act would prevent China from accessing DUV machines, which they currently use to manufacture their best chips
Dave Banerjee Verified account @DaveRBanerjee
One thing about sample-efficient learning algorithms is that, by definition, they require little data, which means that you, as the experimenter, will figure out very quickly whether your learning algorithm is actually working

So maybe you actually get very tight feedback loops in this domain, which would push compute budgets towards this kind of research (as opposed to massive automated RL environment construction, which might not scale as well)