Robert Beard

Recovering tax lawyer, fledgling tax SaaS solo founder

[Robert Beard @rbeard0330]

@patio11 If you read on Kindle, you can get even better recs by exporting your reading history and giving it to the LLM to digest

[Robert Beard @rbeard0330]

@jbarro Fair, but you have to remember that Newsom doesn’t accord as much respect to the institution of marriage as most people do.

[Robert Beard @rbeard0330]

@MattGlassman312 I have been a little taken aback at how vociferously people seem to believe whatever it is they believe about the capabilities/moral status of AI, at least on twitter. It is definitely going to be ugly as this becomes more pressing

[Robert Beard @rbeard0330]

@zeynep I don't see the conflict. We just need a regulation that says that before you can train or deploy a frontier model, you need to convince the regulator that you have adequate safeguards to prevent damaging things the model might do.

[Robert Beard @rbeard0330]

@NoemiTitarenco They are inarguably next-token *generators*, but prediction implies that they're generating tokens to match a target distribution, but that's less true the further you are from pretraining.

[Robert Beard @rbeard0330]

@NateMeyvis A lot of the time you want both--a scheduler that pulls items from a database to put into the queue, and then the queue that conveys items to the destination. IMO, the worker having a dependency on the DB schema is a lot worse than both sides having a dependency on the queue

[Zvi Mowshowitz @TheZvi]

https://t.co/WEzD9MCL63

[Robert Beard @rbeard0330]

@TheZvi This was worth reading: https://t.co/znpCUM1M06. Really interesting suggestion that eval awareness prevents models from generalizing RL cheating to deployment. Maybe somehow related to innoculation prompting?

[Robert Beard @rbeard0330]

@TheZvi Yeah, the fact that these models were being trained to cheat for months is much more alarming than the hack itself. Also CoT monitoring seems totally MIA?

[Robert Beard @rbeard0330]

@pickover @bryceguy72 This is underspecified. There's a noninertial reference frame where B is stationary and A is also stationary. There's an inertial reference frame where A is stationary and B has a velocity of 2000.

[Robert Beard @rbeard0330]

@Gamingtronium Missing the forgot password link

[Robert Beard @rbeard0330]

@jbarro “minor quality of life infractions like…running red lights”

[Robert Beard @rbeard0330]

@joseph_h_garvin @OutSqureCapital @foo @bar Right, post-bar buzz is the function it applies to, so it’s at the same level as bar

[Robert Beard @rbeard0330]

@joseph_h_garvin @OutSqureCapital Decorators apply to the function that starts the line below, not to a block of code. They’re at the same level of indentation as the function.

[Robert Beard @rbeard0330]

@NateMeyvis @mattg312cards I don't think 7NT is ever makeable after a double, since the heart lead would be so obvious. But a flight to 7S doesn't seem totally impossible, and potentially makeable if opener is void in H. The upside to doubling is small too. 7H down 1 won't be common in the field

[Robert Beard @rbeard0330]

@CartoonsHateHer Some animals take a Guess Who loss personally

[Robert Beard @rbeard0330]

@TheZvi Missing an option for obviously no, but nonobviously really really no

[Robert Beard @rbeard0330]

@exQUIZitely The 2019 sequel is obviously quite a bit different, but very much worth playing

[Robert Beard @rbeard0330]

@TheZvi But I also don’t think *practically* making a decision in NP is challenging. It’s theoretically hard though because Omega is punishing standard optimal decisionmaking

[Robert Beard @rbeard0330]

@TheZvi Can you give an example of the kind of thing you’re thinking about? The problematic parts of the problem are pretty weird