What design decisions in LLM training affect the final performance of LLMs? Scaling model size and training data is important, but it's not the only thing. We performed an analysis of 90+ open-weights models to answer this question. 🧵 arxiv.org/abs/2503.03862 (1/12)

"So, Bradley and I started Pangram, the best, most accurate, most advanced AI alert system possible. Today, we’re helping companies spot spammy and fraudulent AI-created content." Read Max Spero of Pangram Labs interview with #AuthorityMagazine here: medium.com/authority-maga…

Join our #UIST2025 workshop in Busan! 🇰🇷 📅 longitudinal-workshop.github.io we, HCI people, build systems to address user challenges but beyond that “1-hr single-session” testing, how can we truly show tools’ novelty, UX, impact? how users learn, adapt? 🧩 long-term eval helps 🌱 1/

NEW: A major AI copyright legal showdown just took a huge twist today. Facing a class action on behalf of book authors that could've seen it pay over a TRILLION in damages for alleged piracy, Anthropic has agreed to settle instead: wired.com/story/anthropi…

We have more evidence that some AI detectors ( in particular Pangram Labs ) work more than others and we should use them

Always feel this work was ahead of its time :) And people who are realizing GPT5s writing quality as garbage, its always been that way

Fun weak supervision approach to improve model response :)

Good news, Ethan Mollick! We finally got our independent study of FPR. Alex Imas and Brian Jabarian studied Pangram alongside other AI detectors and found that Pangram had zero false positives (at a threshold of 0.5) among their dataset of 7,968 human writing samples.

Hieu Pham You don't fix problems by declaring existing institutions completely unfixable and in need of destruction and then have zero plan for how anything's supposed to be better after that. You're literally just doing the meme

Pretty bad take from an @openai employee. Who do you think is responsible for AI slop ?

LLM-as-a-judge should be approached from prediction-powered inference. otherwise, we will suffer the wrath of bias in the future.

Still valid

NotJustKama✈️COLM&EMNLP2025 Conference on Language Modeling NeurIPS Conference Sorry, but this is not how conferences work. The PCs of a conference are responsible for their program, and for the decisions. As much as the Conference on Language Modeling PCs appreciate and look up to NeurIPS Conference PCs, they can't (and must not) outsource program decisions to them

Seen this (v good) article making the rounds. *AI detectors have to be part of conversation* Almost 100% accuracy and ~0 false positives. But policy must be compressive and come from top down, so educators use them effectively. We propose policy framework for this.

Thanks for the “recommended” tag, Larry! Excited to share my job-market paper, which uses the recent religious charter school case to show how faith-based entities secure public funds while sidestepping constitutional limits—a phenomenon I term “private disestablishment.”

The Indians are in many ways the new Jews, hated for their success rather than their perceived failings. Indians take away all the nativist excuses: legal, skilled, productive, law-abiding, family-oriented, English speakers.

Many people are confused in this way about education. "If an LLM can write an essay, why are you still assigning essays, the whole ed system is broken lol" No— you don't write an essay to produce an essay. You write an essay to level up the brain that produced the essay. /end

The bar is low but every time a sincere student sends a long email thats not #ChatGPT written and is thoughtful and detailed, faith in humanity is restored.

There were approximately 2.9 million non-fiction records, and 2.3 million in the fiction dataset.This goes on to prove that there is some economic value in fiction books for training LLMs As always lots of academic non fiction books which help directly on benchmarks such as MMLU

Am I the only person who thinks this $1.5bn Anthropic books settlement counts as a win for Anthropic?