Lightnews — Scholar-powered news

Ari @ari-holtzman.bsky.social · 1d

FYI that UChicago CS & Stats is hiring at all levels via the Data Science Institue:

Postdoc: uchicago.infoready4.com#freeformComp...
Assistant Professor: apply.interfolio.com/174766
Associate Professor: apply.interfolio.com/174768

3 8

Ari @ari-holtzman.bsky.social · 2d

LLM Literalism is the most annoying parts of interacting with an LLM. If I give an example of Xand then ask it to invent its own version, it will use that example almost exactly is painful. It's like attempting to work with a lazy middle schooler that doesn't want to be there.

2

Ari @ari-holtzman.bsky.social · 2d

hrsimulator.communicationgames.ai

HR Simulator™: Be the Person You Hate

A game that will change how you write emails.

hrsimulator.communicationgames.ai

Ari @ari-holtzman.bsky.social · 2d

Just wrote a blog post about what my interest was in putting together HR Simulator™️

I can't speak for my colleagues...but hopefully they agree a bit :)

cichicago.substack.com/p/in-search-...

In Search of Tacit Knowledge

What do we mean when we say things?

cichicago.substack.com

1 2

Ari @ari-holtzman.bsky.social · 4d

hrsimulator.communicationgames.ai

HR Simulator™: Be the Person You Hate

A game that will change how you write emails.

hrsimulator.communicationgames.ai

Ari @ari-holtzman.bsky.social · 4d

Sora is leading a whole new kind of AI slop and, it is with great regret, that I have to admit I'm kind of into it

1 1

Ari @ari-holtzman.bsky.social · 6d

dreaming that these talks and discussions will be thought of as the birth place of Communication and Intelligence as a field

chenhaotan.bsky.social @chenhaotan.bsky.social · 6d

@ari-holtzman.bsky.social kicking off @uchicagoci.bsky.social seminar this year! Full house!

4

Ari @ari-holtzman.bsky.social · 6d

So...are LLMs just better than the median translator for high-resource languages? Is this just a huge shift that didn't get that much attention while everyone was debating whether AGI was a philosophical zombie that hungers for brains or something??

1

Ari @ari-holtzman.bsky.social · 12d

Marco,

Thank you for your candid feedback. While “thanks, I guess” may not align with our gratitude policy, it’s been logged under Provisional Appreciation – Pending Clarification™. Please circle back in 2–3 business days with a confirmed stance.

Warm regards,
Ari

1 1

Ari @ari-holtzman.bsky.social · 12d

We made a game out of corporate email hell 😈

Dang Nguyen @divingwithorcas.bsky.social · 12d

HR Simulator™: a game where you gaslight, deflect, and “let’s circle back” your way to victory.
Every email a boss fight, every “per my last message” a critical hit… or maybe you just overplayed your hand 🫠
Can you earn Enlightened Bureaucrat status?

(link below!)

1 2

Reposted by Ari

Dang Nguyen @divingwithorcas.bsky.social · 12d

HR Simulator™: a game where you gaslight, deflect, and “let’s circle back” your way to victory.
Every email a boss fight, every “per my last message” a critical hit… or maybe you just overplayed your hand 🫠
Can you earn Enlightened Bureaucrat status?

(link below!)

2 5 4

Reposted by Ari

chenhaotan.bsky.social @chenhaotan.bsky.social · 22d

As AI becomes increasingly capable of conducting analyses and following instructions, my prediction is that the role of scientists will increasingly focus on identifying and selecting important problems to work on ("selector"), and effectively evaluating analyses performed by AI ("evaluator").

2 8 10

Ari @ari-holtzman.bsky.social · 25d

Slack & co. need temporary subset chats. Pick people from a channel, hash things out privately, auto-delete when done. This already happens (#channel-minus-boss anyone?) but it's clunky. Sometimes groups coordinate better when subsets align first. Make the backchannels official!

Ari @ari-holtzman.bsky.social · 26d

I doubt we will find a better representation than natural language for explaining AI. Instead of looking for a formal representation we can prove things about with our beautiful mathematical tools I think we should figure out what stories help people generalize approprirately.

2

Ari @ari-holtzman.bsky.social · 27d

try your best work on the thing you actually want to create and think about instead of the thing that has the right sounding name, which everyone mentions when you try to explain what you really care about even though you can tell they don't really understand what you're on about

4

Ari @ari-holtzman.bsky.social · 29d

most researchers probably don't research to be automateable (for reasonable reasons!) but personally I'd love it if most research today turned out to be easily automateable within the next 5 years—I don't think we'd have any trouble finding deeper, funkier, delightful questions

2

Ari @ari-holtzman.bsky.social · Sep 7

hypothesis: LLMs have a good predictive model of 'good writing' (based on features strongly associated with good communicators wo think deeply), but when you ask them to write something good they just produce those features decoupled from the actually interesting thinking process

Ari @ari-holtzman.bsky.social · Sep 5

testing a game we're building who's main mechanic is writing tricky HR emails, and noticing that LLMs have a built-in secret handshake with users to bypass safety guardrails. This seems both necessary to make LLMs actually useful and like they make guardrails essentially useless.

1

Ari @ari-holtzman.bsky.social · Sep 5

in retrospect, I think a reasonable rule would have been to disallow companies from making a public API available for an LLM without publishing a cheap, accessible, 3rd-party verified CAPTCHA for it...

I think that cat's already out of the bag, sadly

Ari @ari-holtzman.bsky.social · Sep 5

it will take a few years, but eventually I will stop finding this kind of thing funny

2

Ari @ari-holtzman.bsky.social · Sep 4

thinking today about how LLMs aren't yet bicycles for the mind, but I think they're starting to look like trains for the mind: they get us all to the same stations very, very fast—but we're on foot once we get there

1

Ari @ari-holtzman.bsky.social · Sep 4

how do I convey to my students that they have permission to try something daring and fail?

2

Ari @ari-holtzman.bsky.social · Sep 1

today I'm thinking about an LLM asked to follow a fictional premise to its natural conclusion it tends to regress to the distribution of possible outcomes of fictional stories rather than following the premise in spirit

Ari @ari-holtzman.bsky.social · Aug 30

room temp take: I haven't seen an LLM application that goes 'deep'. It's impressive 'how little LLMs can do with so little guidance' but any extra guidance one does provide appears to make things *worse*

Claude Code is the only counterexample I can think of, what am I missing?

1

Ari @ari-holtzman.bsky.social · Aug 26

today I'm thinking about if there's a branch of knowledge that I didn't even think about learning about because I assumed no one was around to explain it to me and just passive reading wasn't enough that I could now learn by reading and discussing with an LLM. any suggestions?