I'm gonna try to blog under this account or my real name this year.

Pho is spicy and brothy it's a very Christmas coded dish.

grown ass prof emeritus be like "hee hee"

Do I live in a bubble or has anyone seen Grok being used as the large LLM of choice for a paper? It's always ChatGPT, sometimes Claude, or rarely Gemini.

The weirdest thing about the "google AI overview uses billions of water" myth is that the crowd that believes in it should also know well that a corporation would never use such expensive technology for your benefit.

Why does being late on reviews prevent you from being considered as a great reviewer? I understand meeting review deadlines is very important, but given the context that [no other reward exists!!], it creates a weird scenario that incentivizes reviewers to push out garbage reviews to meet deadlines

ICLR 2025 is the conference that keeps on giving

‪How it feels to introduce my own methodology as "a novel methodology" in the introduction.

Hurts to see a dear friend go into the alignment-ist AI safety research direction.

Reinforcement Learning (RL) is a training algorithm that didn't work with LLMs. However, this changed when researchers from Hugging Face developed their own version, called RLHF (Reinforcement Learning with Hugging Face), which has been used to make state of the art models, such as Llama 4.

Having to come out of research retirement at 65 because my chud grandson couldn't publish an ACL before high school smh

very valid counterargument to many woeposts

Is noone mentioning that the authors of the ICLR fabricated best-of-n submission paper has the same last name, and the second author appears to be a "Researcher" at a prep school?

still uncertain whether inviting all of internet to gawk at long-tailed instances of spectacular review outliars is a good productive thing

fellas we need to reject more slop papers

the memification of llm reviews have people posting bad reviews (that are obviously not llm-generated) and everyone assumes its llm-generated

Lots of new insights gained during EMNLP. LLMs will never replace the joys of conversing with fantastic people with their own perspective.

Big hit viral papers are 80% of people are very impressed and the smart 20% that knows the subject are like "this is wrong/misleading"

If OpenAI doesn't care about my privacy and data, why does their chatbot model say they do? checkmate atheists

how to tell if your model is becoming worse: perplexity goes up how to tell if society is becoming worse: Perplexity goes up

I've been noticing an increasing amount of Argument by Karpathy-said-so recently.

It's sentiment classification all the way down.

What does an LLM do when it translates from Italian "amore" to Spanish "amor" or French "amour"? That's easy! (you might think) Because surely it knows: amore, amor, amour are all based on the same Latin word. It can just drop the "e", or add a "u".

There exists a man who has the highest Erdos number in the world.

get out of my head get out of my head get out of my head get out of my head get out of my head get out of my head get out of my head get out of my head get out of my head get out of my head

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

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.

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?

> 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...

METAREVIEWER IGNORED REVIEWER 2

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

Its over

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

WHY DID ARR REVIEW RESPONSES CLOSE???

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.

Anthropic the type of company to stress test their new models by repeatedly asking "would you love me if I was a worm" type questions