tal yarkoni

i'm deeply ashamed to report that i really have been posting on linkedin for the past few months. not often, and definitely not unironically, but posting nonetheless

anyway, the problem with posting on linkedin is even the subversive posts are carefully metered. you get one for every ~150 posts in your feed. try to read any more than that and they come to take your soup away

i'm deeply ashamed to report that i really have been posting on linkedin for the past few months. not often, and definitely not unironically, but posting nonetheless

this is interesting, but feels at best incomplete. i think there are at least 3 separate reasons we haven't yet seen consistently good writing from AI, at least in long form (microfiction is arguably mostly solved): 🧵

which is just to say that, even if base model development halted in its current state, we probably *will* still get a good AI novelist in the next couple of years. it's just that it's not trivial to build the right harness, and not a lot of resources are being thrown at that particular problem.

this is interesting, but feels at best incomplete. i think there are at least 3 separate reasons we haven't yet seen consistently good writing from AI, at least in long form (microfiction is arguably mostly solved): 🧵

3. it's arguably more of a harness problem than a base model problem. i think @moultano.bsky.social's point that information needs to come from somewhere is correct, buuut... it doesn't have to be in the prompt! it could be in the weights (hard for reason (2)), or it could be in the harness.

this is interesting, but feels at best incomplete. i think there are at least 3 separate reasons we haven't yet seen consistently good writing from AI, at least in long form (microfiction is arguably mostly solved): 🧵

2. for reasons related to the above, learning to produce coherent and high-quality long form writing via pretraining or RL is probably much harder than learning to generate code or even do math. lack of verifiability, and the inherent subjectivity of writing, really hurts you here.

this is interesting, but feels at best incomplete. i think there are at least 3 separate reasons we haven't yet seen consistently good writing from AI, at least in long form (microfiction is arguably mostly solved): 🧵

1. the problem is just way harder. good writing is mostly an aesthetic judgment, unlike code or science, where correctness is often verifiable. people are much less forgiving! if AI writes 120K words of code that *works*, devs are impressed even if they think it's insecure spaghetti code.

characteristically lovely and reflective piece by @ftrain.bsky.social. if you want to better understand how the ground is shifting under every software engineer's feet right now (and soon, under pretty much *every* knowledge worker's feet), this is a great place to start

excellent piece, and should be required reading for anyone skeptical of the technical capabilities and policy implications of AI. there are certainly exceptions, but by and large, the political left is currently treating what should be a five-alarm fire like a distasteful spectator sport.

i understand being skeptical about AI companies' intentions, but if you think Altman's calls for less regulation are a cynical ploy to limit checks on OpenAI's power, it seems hard to argue in the same breath that Amodei's call for *more* regulation is a cynical ploy to hype the product. pick one!

i mean, sure, one can imagine a world where *all* of these companies are malevolent, and their CEOs just have contradictory opinions about how to be more evil. but the much simpler explanation is that many people who work in AI are completely sincere when they say governments should be very worried

yes. right now a lot of people are high on the idea that all the unpleasant work is being done by the models, leaving humans in charge of all the valuable high-level judgment and decision-making work ("taste"). it won't last. the models will soon have better taste than any human. plan accordingly.

i usually know exactly what my songs are trying to say, and it's almost always nothing

a crucial thing i think is lost in much of the "is AI a net good or bad" narrative is the secondary impact to the vast majority of people who aren't at immediate risk of either losing their job or seeing it dramatically transformed. historically, this is where the bulk of utility comes from!

in isolation, the impact on my life is tiny compared to that of an illustrator who loses their job, or who now can't get hired. but there are maybe 30K professional illustrators in the United States. and tens of millions of people who need or want to create illustrations (or would, if it were easy).

a crucial thing i think is lost in much of the "is AI a net good or bad" narrative is the secondary impact to the vast majority of people who aren't at immediate risk of either losing their job or seeing it dramatically transformed. historically, this is where the bulk of utility comes from!

this is what feels missing to me in the current conversation. it seems clear, for example, that many artists and illustrators will lose their jobs because of nano banana pro, which is a real step change in capability. that's a bad thing! we shouldn't pretend otherwise, and we should be sympathetic.

a crucial thing i think is lost in much of the "is AI a net good or bad" narrative is the secondary impact to the vast majority of people who aren't at immediate risk of either losing their job or seeing it dramatically transformed. historically, this is where the bulk of utility comes from!

if you had only focused on the impact to scribes and writers circa the invention of the Gutenberg press, you might have concluded that the press was either really evil (the scribes all lost their jobs!) or the best thing ever (easier and cheaper for writers to disseminate their writing).