I’ve left the UK civil service, after two years spent creating the
@AISafetyInst
, AI Safety Summit, AI Research Resource, and more. It was an honour and a privilege.
There’s lots left to do, and I’ll continue to work on advancing global progress in AI. Onwards and upwards 🌁
So proud of what the team is putting together here, and particularly excited to have Shahar Avin joining us to lead this alongside Chris Summerfield!
More info here:
We are announcing new grants for research into systemic AI safety.
Initially backed by up to £8.5 million, this program will fund researchers to advance the science underpinning AI safety.
Read more:
Ben has been an exceptional force for advancing the UK’s science and tech leadership.
A huge loss for the country to lose him from the civil service, but a huge gain for whoever gets to work with him next.
After a whirlwind few years, it’s time for me to leave government. Will stay based in London, resuming my Strathclyde professorial role and a few interesting irons in the fire. Onwards! 🚀
From via
@gwern
’s wonderful essay on scaling . Further discussion on the paper, between Moravec and peers (including
@robinhanson
and
@anderssandberg
), at .
Alongside ARIA, our sci/tech team wanted to create a 'twin for ARIA' composed of physical labs organised very differently to conventional research environments to pursue groundbreaking research and invention, putting young scientists first. Read more here
Announcing the work we started on Theano-JAX: Using JAX as a backend for Theano, which is still amazing for its maturity, hackability, graph-simplifications, symbolic diff and much more. Contributions welcome:
@GaryMarcus
@gdb
I asked
#dalle
to try harder and it did not work because the horse was too heavy. I think OpenAI needs to retrain the model with lighter horses?
AI for public benefit
- Few industrial labs pushing frontier AI way beyond capabilities of gov/org
- Limited mechanisms to explore public benefit & 1/risk in these systems.
- UK has talent pool to solve hard (sociotechnical) problems re access, alignment, trust, risk, etc. /5
@boazbaraktcs
@whybansal
These results seem to be more about linear classifiers than SSL - if you used a supervised model as your (frozen) feature extractor, and trained a linear classifier on top, you would likely see low generalization gaps under noise as well right?
2. Public search engines provided by the government.
Equal access to information should be a fundamental right, guaranteed by law. The would continue the role that public libraries had for centuries.
@whybansal
@boazbaraktcs
Ah good point - I was considering a baseline setting in which you’d supervised pretrain with correct labels and then train the linear classifier with increasingly noisy labels.
As we watch Putin destroy Ukraine's democracy, we in the Western left must abandon isolationism once and for all. The thieves, thugs and fascists ruling Russia and China do not respect our ideals and have no interest in letting us improve our own societies in peace.