Sometimes, we just have to keep in mind that people around us also share the desire to be happy and content.

Lately on my journey of self-discovery, one quote resonates with me most: > “We are here to awaken from our illusion of separateness.” — Thich Nhat Hanh ========== Continuing from my previous post [1], I believe life is a journey of: (1) Looking inward to understand ourselves (2) Looking outward to understand the world and the universe Reflecting on my own life, I spent most of my time and energy on (2), so I ended up learning a great deal about the world, yet knowing very little about myself. It wasn’t until the past few years—when I began devoting more time to exploring my inner self—that I realized how little I actually knew about who I am. There have been many instances when, in moments of stillness, I listened carefully to the voice in my mind, only to discover how I truly think and feel. Looking back on the education I received, almost no one ever taught me to understand myself; instead, it was always about learning new facts about the world. When I consider our recent history, humankind also seems to invest most of its time and energy in (2). This has indeed driven scientific progress and improved our material lives. But has it really made us happier? I can’t claim to have any profound insights into life, but I’ve come to believe that we need to balance our focus on (1) understanding ourselves and (2) understanding the world. Only by knowing ourselves and the world well can we find our proper place in it. [1]

Annaka Harris: Free Will, Consciousness, and the Nature of Reality – Lex Fridman Podcast Ever since I was young, I’ve been especially interested in human consciousness and psychology. I’ve always been fascinated by questions like “Who am I?” and “Why am I, me?” In fact, I believe that life is a journey of (1) looking inward to understand ourselves and (2) looking outward to explore the world. Today, I happened to listen to a profoundly insightful interview. It presented numerous experimental findings and perspectives that are extremely counterintuitive, challenging the way I view my own consciousness and how I perceive the world. Although I really want to write a reflection on it, the content is so deep that I need time to let it all sink in. For now, I’d like to recommend the video to anyone who’s also interested in this topic. * Link to the interview:

RT @BlackHC: NeurIPS 2024 PCs being a bunch of clowns 🤡 the state of ML 🙄 All you get back a month after raising a concern:

I'm also deeply grateful to the companies that share open-weight LLMs despite the enormous cost of pre-training. Now that the era of post-training is coming, I look forward to seeing all the interesting ideas that resource-constrained researchers can bring to the table.

# NeurIPS 2024 Reflections I’m grateful for the opportunity to attend NeurIPS (Neural Information Processing Systems) 2024, where our research team presented our third paper—StreamBench [1]. Here are a few reflections from the experience at the conference: ## Improvements for Poster Presentations During my poster presentation, I noticed several aspects where I could improve: 1. Telling a Good Story In an interactive poster session, the most important thing is to spark the audience’s interest before diving into the research details. To do so, it’s crucial to provide a clear storyline—start with the motivation, context, and significance of your work so that the audience can see the bigger picture. Once they resonate with the context, they’ll be more receptive to your technical explanations and abstractions, which brings us to the next point. 2. Using Examples to Illustrate Abstract Ideas Another lesson is that posters need concrete examples and should not be overly abstract. There are many times when only after I provided real-world examples did the audience’s confusion begin to clear. The reason is simple: a single abstract concept can be interpreted differently by different people, and they might imagine different scenarios in their minds. This means the audience and I are not being “on the same page.” These two points were actually mentioned in one of the tutorials at ACL 2024 called Presentation Matters [2], but I feel I didn’t execute them well this time. I’ll keep working on it! (Highly recommend checking out [2].) ## How to Allocate Time at a Multi-Track Conference Large-scale events such as NeurIPS often have multiple sessions running in parallel, making time allocation crucial: 1. Talks and Presentations Can Be Watched Later Keynotes, invited talks, or oral sessions typically have recordings available. Since these sessions are more one-way communication with limited interaction, one can watch them afterward without missing much. Reserve precious conference time for activities that require in-person interaction, such as poster presentations. 2. Poster Sessions In my view, the poster session is the most in-depth and efficient way to exchange information during a conference. If time permits, arrive a bit earlier than the official start time to visit the posters you’re most interested in. Early on, there’s usually fewer people, and the authors are usually more energetic and ready for more in-depth discussions. 3. Attend Workshops NeurIPS is huge and covers a wide range of topics, which are usually only briefly touched on at the main conference. If one is interested in a specific research area, we'll often gain more by attending high-quality workshops. This applies equally to invited talks, oral presentations, and poster sessions—so don’t skip workshops! ## Reserving Social Energy Beyond contributing research and gaining new knowledge, there’s a third important purpose for attending conferences: networking. However, for someone like me who is introverted, networking can be mentally draining—especially at a conference like NeurIPS, which offers numerous social activities. I often find myself in a state of mental “half-paralysis” if I overextend my social interactions, where my thinking and communication abilities drop significantly. Next time, I’ll make sure to manage my social energy more carefully to stay effective during critical networking opportunities. ## Conclusion Overall, attending NeurIPS was a very fresh experience and led me to reflect on how to present research, allocate time, and manage personal energy more efficiently. I hope that with these reflections, I’ll be able to make even greater strides the next time I attend a conference of this scale! [1] StreamBench: Towards Benchmarking Continuous Improvement of Language Agents [2] ACL 2024 Tutorial: Presentation Matters

Ilya’s words sounded straightforward and obvious when I first heard his talk. However, when I replayed the talk and pondered over what he was trying to convey, it is indeed thoughtful and not as trivial as it seemed. Ilya’s test-of-time award talk:

LMAO

So glad to see @ilyasut himself here at @NeurIPSConf to receive the test-of-time award 🏆 #NeurIPS2024

RT @lexfridman: I hope to interview both Zelenskyy and Putin, likely in January. For this, I will travel to both Kyiv and Moscow. My hope…

LMAO Talk about how to present results

A very interesting work with comprehensive experiments. I always love to see research projects with connection to human cognition or psychology.

I think the question is a bit wrong here. New methods (e.g., scaling) make old benchmarks obsolete, which leads to requirements of better benchmarks. New benchmarks then motivate better methods. Both are extremely important.

RT @ScholarshipfPhd:

# Shaping the Landscape of AI Today: NeurIPS 2024 Test of Time Paper Award Next Sunday, I’ll be heading to #Vancouver with my colleague @zraytam from the Appier AI Research Team, to participate in one of the most prestigious annual events in machine learning—#NeurIPS2024. We will also be showcasing our team’s research, StreamBench: Towards Benchmarking Continuous Improvement of Language Agents [1]. Earlier this week, NeurIPS announced the recipients of this year’s Test of Time Paper Award [2], which honors two groundbreaking papers published a decade ago that have had a profound impact on the field [3][4]. One of the awarded papers, Sequence to Sequence Learning with Neural Networks, was authored by Ilya Sutskever @ilyasut. His remarks during the oral presentation at NeurIPS 2014* [5] left a lasting impression on me: > (09:17) There you go! This is the main technical contribution of this work. We use minimum innovation for maximum results. > (11:47) The main emphasis of our work is simply to show that this simple, uniform, and uncreative model can achieve good results if the model is LARGE. > (19:47) The real conclusion now is that if you have a large dataset, and you train a very big neural network, then success is guaranteed! Thank you very much! Ten years ago, long before scaling law is actually a thing, Ilya already expressed strong confidence in the potential of big models combined with large datasets. Fast forward to today, the success of large language models across downstream tasks has validated his intuition. Problems once deemed unsolvable by neural networks are now being tackled one by one. Looking back, his foresight is truly remarkable. However, at the time, there were skeptics in the audience who doubted his “grand claim,” viewing it as overly simplistic. This skepticism likely stemmed from a natural human tendency—even among scientists and researchers—to resist ideas that challenge established beliefs. • How could simply scaling up models and datasets be effective? • How could the Earth possibly be round? • How could time and space be relative? This reflection serves as a reminder to myself to maintain an open mind in research. It sounds simple but is often the hardest part of the process. I also hope that @ilyasut will attend NeurIPS this year to accept this well-deserved award in person—it would be an honor to meet him. * Note: At the time, NeurIPS was known as NIPS 2014. [1] [2] [3] Sutskever, I. (2014). Sequence to Sequence Learning with Neural Networks. arXiv preprint arXiv:1409.3215. [4] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27. [5]

RT @svpino: The first place I always look at whenever a Machine Learning model is not performing as well as I'd like: Labels. Most people…

RT @AdamMGrant: Insecure people strive to prove their intelligence. They're quick to assure us that they have the answer—they're determined…

@zraytam What do you identify as? Me: I identify as a writer 🤣