Sean Brynjólfsson

@A10Networks／@Cornell '25／views randomly sampled from the posterior／opponent color theory supporter／SweeTango apple enjoyer

[Sean Brynjólfsson @jolfss]

@adamimos @kylejray @RiechersPaul Nonergodicity is a word I've needed for a very long time! Gorgeous blog, I'll be thinking about it

[Sean Brynjólfsson @jolfss]

@RespanAI Take a look at our paper investigating redaction https://t.co/a1tijtOHmi; there's a lot of work to do before we can hill climb on privacy.
@madhavaggar Another one

[Sean Brynjólfsson @jolfss]

@LiuZiyin10 Thanks for working on this, excited to read

[Sean Brynjólfsson @jolfss]

@thsottiaux Swap between personal and work accounts. Massive pain having to pair remote control/mobile every time.

This is the only reason I have an anthropic subscription.

[Sean Brynjólfsson @jolfss]

@JCovetti @max_spero_ @PradyuPrasad @snoopy_dot_jpg hey guys, maybe the point of legislation is to encourage a society we want to live in and less about some emanant moral truth of dynamic pricing

I hate nuance too, but you can actually want discounts for students and oppose dynamic medical pricing at the same time. thanks

[Sean Brynjólfsson @jolfss]

Let's say we both roll dice to determine which token to generate from the distribution at each step. We give our roll to the model and it gives us a token based on our number.

Do you agree we are both drawing from the same distribution? If yes, continue.

Now, I've been writing down the numbers I've been rolling. If I give the model my numbers and my sequence, the model can say "oh yes, your numbers would generate the rest of the sequence". If you agree, continue.

Now I give the model your sequence and my numbers. The model will say, "hmm, no, if I used your random sequence of numbers I would not generate this sequence--you would need to roll a 3 here to get that token," etc.. If this makes sense, continue.

We have therefore demonstrated that the model can recognize its own outputs if it knows the dice that were used at the time of generating. Let's call the set of dice rolls the 'watermark'.

But, you object, you prefer your sequence anyway.

That's fine. We'll simply use your dice, and now our watermark gives the "preferred" response!

This is the gist of the "distortion free" watermark. The model's behavior is not changed, Anthropic simply knows the dice now.

[Sean Brynjólfsson @jolfss]

@packsmarginalia @cain151714 @bayeslord I think “human suffering” is the key to his point. There are people who can’t feel pain so we understand what it takes for “human suffering”.

There could be a sub- or other-conscious system that’s constantly suffering in your brain right now, but I wouldn’t lose sleep over it.

[Sean Brynjólfsson @jolfss]

@ramin_m_h @liquidai This is great, I think throughput is a huge consideration so I’m glad you shared. We crudely approximate with params to avoid presenting hardware/configuration decisions

[Sean Brynjólfsson @jolfss]

@skipper4848 @liquidai I'd be careful regardless, as you can see in the table the scores are all <= 70%

[Sean Brynjólfsson @jolfss]

@liquidai *Our benchmark only evaluates English documents

[Sean Brynjólfsson @jolfss]

@liquidai It's a reasonable model, but considering that openai/privacy-filter is 50M (active) params it's still probably the public choice.

Evaluated on our RedactionBench, happy to follow up

[Sean Brynjólfsson @jolfss]

@burkov https://t.co/dD9QAfxZaq

[Sean Brynjólfsson @jolfss]

@psiloycne @max_spero_ Can't comment on the burnt out portion, but I'll push the other way in saying that this behavior is precisely why I trust Pangram a priori, despite having not used them yet.

[Sean Brynjólfsson @jolfss]

@VictorTaelin GPT has "grass is always greener" syndrome and assumes the current implementation is bad

this is generally a good approximation, but fails to converge ~~ sucks at gathering evidence for taste

keeps me in the loop, but can't rely on it to be my PI

[Sean Brynjólfsson @jolfss]

@Sauers_ Because the authors are phenomenal

[Will Huey @willhuey9]

It’s becoming increasingly common to evaluate policies in sim. Policy improvement is harder. We show why vanilla RL learns to exploit the sim2real gap, and provide a simple and principled solution.

[Sean Brynjólfsson @jolfss]

@iltenahmet @Kodurubhargav1 @chamath Even your second analogy works. Pharmaceutical companies, if not given access to computers, would eventually lose to tech companies that hire pharmacologists or can specify the same objectives

[Sean Brynjólfsson @jolfss]

@iltenahmet @Kodurubhargav1 @chamath This take will age about as badly as saying a computer will never be better than humans at chess

[Sean Brynjólfsson @jolfss]

In honor of @arxiv becoming an independent non-profit:
I applied to @Cornell thinking that every paper on arXiv was from Cornell