@jamellebouie.net talks about a revival of civic responsibility, which goes hand in hand. your occupation is your biggest civic contribution, and you should think of it that way. it won't work if everyone spends 40 hrs/wk screwing each other over and 2 hrs/wk volunteering.

anti-social industries pay you a premium to ignore ghastly shit, meaning there's effectively a tax on pro-social labor. while that genuinely sucks, it's still your choice what to do.

here i was naively thinking that LLMs would help bureaucracy work well, making it easier to maneuver complex regulations. not so fast, i hadn't considered the "be a bigger barrier to progress" side of the equation

I think the comparison is useful and I’m worried about how much is being bet on future capabilities of a single industry.

Moissanite! Carborundum! the hard stuff

dissecting a cow eye and getting a load of the lens was a formative experience for me

enjoy the eyeballs

.. confidently provides wrong answers, and generally lacks deep understanding

that's the opposite of what we expect from people, so we assume corresponding intelligence commensurate with capabilities. but that's just not the case.

this is hard to reconcile with results from zero- and few-shot learning, ARC-AGI, etc

and ya, LLMs are uncanny. they are very powerful and it is very strange to interact with a system so seemingly intelligent that also uses terms improperly, gets tripped up by basic logic ..

i keep saying it:

1. hallucinations and 2. lack of a world model

are serious problems fundamental to LLM architecture

and related symptoms of an underlying syndrome, that LLMs only learn and think at a shallow level, or only imitate learning and thinking.

i'm told that their borrowing rates discourage spending, but i suspect structural and political challenges are just as limiting

this YOY increase in postdoc apps tells many stories. one is that Europe trains a lot of scientists! more than it can accommodate. they don't really need the US (especially our labor) to increase EU scientific output

the window will remain open for years! more than an American exodus i think we will see a lack of interest from international talent. we're gonna find out how much American science has been coasting on solid institutions and immigration

marie-sklodowska-curie-actions.ec.europa.eu/news/msca-po...

that's a traditional American strength, but i don't see people proactively choosing to stay in the US atm. lack of opportunities abroad and high barrier to an intercontinental move explain why few have left.

My favorite enzymes! will read with interest

Im thinking dataset procurement and curation, but I don’t know any inside details. I don’t think these models are particularly expensive to train? Maybe hyperparameter optimization (and budget) making a difference?

the biggest problem i see is that 'intelligence' is proving very hard to evoke. LLMs sit in a squishy middle between procedural and intuitive action. yuck. maybe it will take a huge effort to get a usable system. meanwhile, that effort dilutes ROI and leaves opportunity for a temporary crash