Geoff Penington

Associate Professor at Stanford. Quantum Information and Quantum Gravity. Currently at @Openai

[Geoff Penington @quantum_geoff]

In 1993 the Superconducting Supercollider was cancelled. Estimated cost: $8 billion. An exodus of physicists left to Wall Street, bringing fancy maths and dubious risk management. 15 years later the global financial crisis cost ~$20 trillion. This is why you don't defund physics!

[Geoff Penington @quantum_geoff]

@bcubeddd I would but he also blocked me

[Jason Wolfe @w01fe]

I’m happy OpenAI put out this statement. Personally I really dislike a lot of things I’ve heard about LTF, and I’ve donated in personal capacity to Bores.

This is just a small step and people may still rightly be skeptical, but I hope we can earn trust through our actions going forward. One thing I’ve learned through being more engaged in our policy work lately is that there are so many people at OpenAI both inside and outside Global Affairs who care deeply about how the right policies could help ensure that AGI benefits all of humanity.

[Geoff Penington @quantum_geoff]

@WKCosmo @DeivonDrago I agree about this. But I’m not sure it’s possible to meaningfully preserve an institution when development cycles become comparable to human careers

[Geoff Penington @quantum_geoff]

I’m not as extreme on this as eg Jared Kaplan, but I really don’t see the point right now in making plans for projects that are meant to start in 2041. The boundaries of what will be possible by then depend hugely on how AI (and robotics) development goes over the next 15 years…

[CERNpress @CERNpress]

[Press Update]: The CERN Council decided to update the European Strategy for Particle Physics (#ESPP)

Find out more: https://t.co/xo78dGpIUW

#FrontlinePhysics

[Geoff Penington @quantum_geoff]

@littmath I think either we will or it will have become possible to do groundbreaking work that pushes the energy frontier on much shorter timescales

[Geoff Penington @quantum_geoff]

@DeivonDrago Im not talking about plans changing based on theoretical developments. Im talking about what you can build and how changing dramatically

[Geoff Penington @quantum_geoff]

@theory_dad Fair enough

[Geoff Penington @quantum_geoff]

@ZoharKo In other words focus on changing the institutional structure to be much more adaptable

[Geoff Penington @quantum_geoff]

@ZoharKo Yeah my guess is that their efforts would be better spent figuring out how not to need such long term plans in the first place. An awful lot can be done in ten years!

[Geoff Penington @quantum_geoff]

@ZoharKo There will still be more time just to plan than the Apollo program took from Kennedy’s speech to Armstrong on the moon

[Geoff Penington @quantum_geoff]

@ZoharKo Is this really true? You really think in five years time it will be too late to plan a project in time to start in 2041?

[Geoff Penington @quantum_geoff]

@bcubeddd @compute25 I don’t think “DINQ” knows anything at all about his salary (I also don’t to be clear)

[Geoff Penington @quantum_geoff]

@ThomasVanRiet2 @bcubeddd It is

[Geoff Penington @quantum_geoff]

@MLJK87 @davidbessis 7*(2+2)-20=8. 7*(5+5)-20=50. 7*(6+6)-20=64

[Geoff Penington @quantum_geoff]

@Bourbaki219433 @MishraArya5819 It was just misstated. The person is taking a sabbatical not giving up tenure

[Geoff Penington @quantum_geoff]

@willdepue Also I’ve no idea if the labs will be willing to take part, but I’d guess some lab with better pretraining than RL/posttraining (GDM?) will agree because it will make them look good. And then you can try to guilt the rest into doing it too (disclosure unraveling)

[Geoff Penington @quantum_geoff]

@willdepue I think you should try. Major confound is pretraining data mix. Probably you have to allow the labs to just fine tune a model for each category (what is “the” base model anyway). But then it becomes even more important to stop cheating…

[Geoff Penington @quantum_geoff]

@littmath @WKCosmo Agreed

[Geoff Penington @quantum_geoff]

@WKCosmo @littmath I’m not sure what you’re looking for here? Do you disagree that there is a meaningful notion of theorems being similar? Do you want me to construct a formal metric (this could be done but seems a waste of time)? Do you just want to vaguely imply that I’m bullshitting?

[Geoff Penington @quantum_geoff]

@WKCosmo @littmath Yes. A morally true should be close in theorem space to an actually true theorem. But the actually true theorem may be significantly more complicated than the morally true version without those extra complications being important for whatever purposes you actually care about