🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

"Human Feedback is not Gold Standard" was accepted at ICLR 2024 🥳 I'd love to chat about the limits of human feedback wrt LLM alignment (and about cohere) if you're going to be at the conference! 🇦🇹 Thanks again to Max Bartolo for making it an awesome internship experience ❤️

Alignment is now defined so broadly that all of AI, all of ML, and the entire history of technology is—and always has been—"alignment research".

CS Theory prof during an on-site interview: so Upol, why do you think HCI is part of CS? Me: I'd argue you cannot have CS without HCI Him: How? Me: Name one computing system in this planet or even in space that works in a vacuum without human interactions Him: (ponders for a few

*Trying to type set my posters*

Call for Student Volunteers for ASE 2024 is now open. See you in Sac! conf.researchr.org/track/ase-2024…

Let's spread positivity. Don't want to see more Reflection posts in my timeline. Such high numbers on gsm8k (beyond annotation noise) and other benchmarks will always demand more scrutiny. Hopefully people learn to share social media model releases more carefully now.

every time you: • say a model is "highly accurate" but have no evals • finetune a decoder LLM for classification without trying BERT-style classifiers • use only embeddings without trying text for retrieval, matching, dedup • optimize for sota/complexity instead of solving

one key to combat this problem is to look beyond the hype and not work on transient stuff as an academic. this is much easier said than done, and im guilty of chasing the hype train a bunch of times myself (its so fun!) but to truly do research and not feel anxious about it, I

Many people believe that AI advances will dramatically increase inequality. In a paper with two Nobel laureates, Daron Acemoglu and Simon Johnson, plus 30 multidisciplinary experts, we argue that it’s more complex than a simple “rich-get-richer” story. For example, we coined

Google Translate now as Inuktut (Latin and Syllabics), Crimean Tatar (Latin script), Santali (Ol Chiki), Tshiluba, and French (Canada)!

hot take to debate🧐: "human-AI collaboration" studies are starting to seem more like "human supervision of AI" studies and less like "humans augmented by AI"

ASE 2024 has officially begun! 🎉 Dive into five days packed with inspiring sessions, hands-on workshops, and opportunities to connect with brilliant minds from around the world. Don’t miss out—join us for an unforgettable experience!

the transformer is good for big tech because it means the only way to win in AI is to have the biggest data center the bitter lesson matters, but yelling about it with suboptimal asymptotics is bad… the incumbents aren’t incentivized to change this, which is where we come in

Open Source Initiative (OSI) says AI models aren’t “open source” unless data, weights, hyperparameters, and executable code to build and run the model are released

It feels like US academia, esp. in the social sciences and humanities, but really in all fields of higher ed, seem to have no sense of what a catastrophe could be starting in about a week

Study of 554 resumes & 571 job descriptions evaluated by AI (3 LLM's) finds they favored white names 85% of the time; female names only 11%. The AI preferred white men even for female-dominated roles, like HR workers. Lisa Stiffler h/t Jess Calarco geekwire.com/2024/ai-overwh…

rbinom(1, size=80000, prob=0.50)

There are numerous leaderboards for AI capabilities and risks, for example fairness. In new work, we argue that leaderboards are misleading when the determination of concepts like "fairness" is always contextual. Instead, we should use benchmark suites.

Feeling the same tbh