Processor of Natural Languages

Keynote at #COLM2025: Nicholas Carlini from Anthropic

"Are language models worth it?"

Explains that the prior decade of his work on adversarial images, while it taught us a lot, isn't very applied; it's unlikely anyone is actually altering images of cats in scary ways.

everyone focused on superintelligence

they wont be ready for duperintelligence

the same people that are convincing you that AI is frighteningly smart and will replace your job in 2 weeks are now saying endless slop feeds for monetization is a good thing actually

few years ago I would have thought this was an overreaction to dull/unimaginative usage of tools that are actually legitimate/cool

but it is now quite clear that the providers of the technology are settling for a culture of sloposphere as a goal rather than a side effect.

very sad.

Of course the city with Irish heritage does McIntyrep

Previous work on sandwiches has investigated using meats (Yang et al., 2014; Fisher et al. 2019), cheeses (Stevens et al., 1998; Meng et al., 2023), and sauces (Hellmann et al., 1989; Geiger et al., 2008). However, little is known about using liquefied petroleum gas. Our work closes this gap by ...

wait everyone hates utf-8 now??

Can you really call your model Omni if it can't smell?

I feel we perceive the world in similar ways.

> Superintelligence Labs
> 91084 bajillion dollars spent on GPUs
> 178904723 bajillion dollars spent on recruitment
> First paper is out!!!
> Looks inside

It's fucking RAG

I'm a PhD student now.

Better support networks now so I will hopefully complain significantly less on here.

The "Clanker" as a slur for AI joke is really catching on.

Some people seem to enjoy it a bit too much though...

a man is hugging another man on a soccer field with a france lanyard around his neck . ALT: a man is hugging another man on a soccer field with a france lanyard around his neck .

METAREVIEWER IGNORED REVIEWER 2

"Em dash" sounds like something you'd do in Super Smash Bros to dodge a nair

But if they arent how will you quantify them and put into a nice table where number go up

Its over

This happened twice now and I'm wondering if I'm too easily impressed by counterintuitive results.

But then again, can you blame me, when most other assignments are about getting 1 percent increase on sentiment classification datasets.

Paper: We find this really interesting counterintuitive behavior. We characterized it, looked at the dynamics and possible causes, and offered solutions and discussion on its consequential impacts.

Me: wow based this is 4.5/5

That one other reviewer: lol whom cares 2.5/5

Bro beefing with openreview

It got fixed hooray

I spent 10 hours stressing out over a response and I CANT SUBMIT IT??

WHY DID ARR REVIEW RESPONSES CLOSE???

OpenReviw has gone woke against free speech?? OpenReview more like ClosedReview.

I thought OpenReview was down but the site works on incognito lol

ARR reviews are out

Strengths:
Rigorous experiments thoroughly demonstrate performance improvements.

Weaknesses:
The first paragraph failed to mention a sentence on how the advent of LLMs have revolutionized the field of NLP.

Soundness: 1.5/5

Strengths:
Proposed framwork is novel, highly effective, and supported by extensive experiments.

Weakness:
Framework name is a forced acronym that doesnt work (what does the second E stand for??)

Soundness: 2/5

My social media feeds: AI has now become PhD-level experts and will replace all human workers as early as this year.

My ARR review papers: We create an innovative approach, harnessing the power of LLMs, to perform sentiment classification on movie reviews.

Apple's reasoning paper is attracting lots of bad takes from people who didn't and won't read it.

I mean I didn't and won't read it either but still.

Can you tell that I really wanted to visit Vienna