jzb

How can we vibe code while still maintaining code quality? Over the past year, I've shifted 95% of my development from manually writing code to using coding agents. I wrote this blog on some tricks I learned to work successfully with agents: all-hands.dev/blog/vibe-codi…

How far are we from having competent AI co-workers that can perform tasks as varied as software development, project management, administration, and data science? In our new paper, we introduce TheAgentCompany, a benchmark for AI agents on consequential real-world tasks.

Today we’re launching our new Copilot experience. I truly believe we can deliver a calmer, more helpful and supportive era of technology, with a Copilot that is now more intuitive, more personalized, and secure. Learn more, download, and enjoy. At Microsoft AI, we are creating an AI companion for everyone.  This is the first step.

🌟Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision 🌟 arxiv.org/abs/2403.09472 How can we keep improving AI systems when their capabilities surpass those of human supervisors? (1/n)

New paper! How to train LLMs to effectively answer questions on new documents? Introducing *pre-instruction-tuning* - instruction-tuning *before* continued pre-training — significantly more effective than traditional instruction-tuning after PT. arxiv.org/abs/2402.12847

@saizhang0 My question is "Is it normal to have clusters with >100 8*A100 nodes in academia"? 😀

Someone is running jobs on 128 GPU nodes (~1000 A100) during holidays. Is this a normal PhD life?

Everyone is using RAG, but most of the retrieved context is noisy! 🚨 Introducing FilCo: “Learning to Filter Context for Retrieval-Augmented Generation” TL;DR: Get rid of the irrelevant content using FilCo, and you'll get better outputs. Preprint: arxiv.org/abs/2311.08377

🚀 Can RLAIF fully replace RLHF to align language models from scratch, enhancing both their alignment and capabilities? SALMON introduces a principle-following reward model in the realm of self-alignment, using just 6 ICL exemplars and 31 principles to outperform LLaMA-2-Chat!

How do you turn a language model into a chatbot without any user interactions? We introduce LIMA: a LLaMa-based model fine-tuned on only 1,000 curated prompts and responses, which produces shockingly good responses. * No user data * No mode distillation * No RLHF

[1/4] Large language models (LLMs) tend to hallucinate, especially when generating long outputs. We present active retrieval augmented generation, in which an LLM actively decides when and what to retrieve throughout the generation process.

Active LLM Retrieval Augmented Generation -Iteratively uses a prediction of upcoming sentence to anticipate future content which is used as query to retrieve relevant docs to regenerate sentence -On 4 long-form generation tasks: superior / competitive arxiv.org/abs/2305.06983

🔥If data is the oil, have we exhausted the mine? Introducing our latest work, 🎉. "From Zero to Hero: Examining the Power of Symbolic Tasks in Instruction Tuning"💪 💡TL;DR: Generate tons of train examples via symbolic tasks to boost data quantity for instruction tuning🚀 1/3

Can the text evaluator be customized for different/new evaluation aspects without training? Our GPTScore achieves customized, multifaceted, and training-free using emergent abilities of PLM, i.g., instruction and in-context learning. Paper: arxiv.org/pdf/2302.04166…

Retrieving information for QA is critical for reliability, but training separate retriever/reader models is cumbersome. At #EMNLP2022 we present Retrieval as Attention (ReAtt), a single retrieve/read model that is effective, generalizable, and adaptable arxiv.org/abs/2212.02027 🧵

📢 New Paper: Program-aided Language models Prompting methods such as chain-of-thought (@_jasonwei) employ LLM for decomposing the problem into steps *and* solving each step. Instead, PaL decomposes the problem into *programmatic* steps and solves using a Python interpreter. 1/4

(1/4) Introducing EditEval, an instruction-based evaluation suite leveraging high-quality existing & new datasets for automatic evaluation of editing capabilities. At present, comprehensive eval of editing capabilities (i.e., fixing wrong info or reorganizing text) is lacking.

I'm excited to share our work on a new sequence modeling architecture called Mega: Moving Average Equipped Gated Attention. Mega achieves SOTA results on multiple benchmarks, including NMT, Long Range Arena, language modeling, ImageNet and raw speech classification.

MEGA is a new method for modeling long sequences based on the surprisingly simple technique of taking the moving average of embeddings. Excellent results, outperforming strong competitors such as S4 on most tasks! Strongly recommend that you check it out: arxiv.org/abs/2209.10655

Internship applications are now open! This year we have opportunities across various teams and offices 🌎🌍 Apply today via dpmd.ai/internships and learn more about the experience below ⬇️ #DeepMindInterns

Happy to announce that I've formed a company, Inspired Cognition (inspiredco.ai) together with @stefan_fee and @odashi_en! Our goal is to make it easier and more efficient to build AI systems (particularly NLP) through our tools and expertise. 1/2