DeepSeek-V3 3️⃣, Microsoft and OpenAI AGI definition 📖, foundation models for music 🎵

Chinese AI startup DeepSeek has released DeepSeek-V3, a 671B parameter model employing a mixture-of-experts architecture, available on Hugging Face ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR TLDR AI 2024-12-31 🚀 Headlines & Launches DeepSeek-V3, ultra-large open-source AI, outperforms Llama and Qwen on launch (5 minute read) Chinese AI startup DeepSeek has released DeepSeek-V3, a 671B parameter model employing a mixture-of-experts architecture, available on Hugging Face. DeepSeek-V3 has outperformed leading models like Meta's Llama 3.1 and rivals closed models such as OpenAI's GPT-4o. The model emphasizes efficient performance with innovations like multi-token prediction, offering substantial training cost savings. Microsoft and OpenAI have a financial definition of AGI (2 minute read) Microsoft and OpenAI define AGI as AI systems generating $100 billion in profits, which OpenAI is far from achieving. OpenAI is reportedly losing billions and doesn't expect profitability until 2029, affecting how long Microsoft will have access to its technology. Speculation on OpenAI declaring AGI sooner is countered by these financial metrics. OpenAI 'considered' building a humanoid robot (1 minute read) OpenAI is considering building its own humanoid robot, leveraging previous investments in robotics firms like Figure and 1X. The company disbanded its robotics division in 2021 and would face challenges re-entering this competitive market. 🧠 Research & Innovation Meta-Learned Optimizer Boosts Continual Learning (18 minute read) Researchers present a meta-learning approach to continual learning where a transformer-based optimizer selectively updates model parameters to prevent forgetting past knowledge. Multi-Scale Super-Resolution (11 minute read) This paper introduces a new approach to Super-Resolution (SR) that challenges the conventional method of training separate models for each scale. Text-to-3D Generation of Complex Indoor Scenes (4 minute read) SceneCraft introduces a method for creating detailed 3D indoor scenes based on user-provided text descriptions and layout preferences. 🧑‍💻 Engineering & Resources Annotation in Medical Imaging (GitHub Repo) Label Critic is an innovative tool designed to streamline medical dataset annotation by using AI-generated labels instead of starting from scratch. A New Metric for Better Cell Tracking (GitHub Repo) CHOTA metric (Cell-specific Higher Order Tracking Accuracy) improves the evaluation of cell tracking methods in biomedical research. Unlike current metrics, which focus on local accuracy, CHOTA unifies the assessment of cell detections, global coherence, and lineage tracking, making it more effective for high-level biological analysis. List of Foundation Models For Music (GitHub Repo) This repository, along with the companion paper, contains a list of services, models, datasets, and systems used to generate music. 🎁 Miscellaneous Chain of Continuous Thoughts (8 minute read) Meta's COCONUT is a novel approach that allows LLMs to reason in continuous latent space rather than discrete language tokens by encoding reasoning steps in continuous vectors. The method improves reasoning abilities but sacrifices interpretability. It could be a valuable addition to future LLMs despite the trade-off. Tenstorrent and the State of AI Hardware Startups (16 minute read) Tenstorrent's open-source AI hardware approach offers a promising alternative to Nvidia's dominance by using a unique CPU and AI core integration strategy. The company is leveraging Samsung Foundry's cost-effective SF4X process and aims to tackle latency issues for scaling out AI workload efficiently. Tenstorrent's recent $2B valuation highlights its potential, especially as an option for high-performance RISC-V IP amidst ARM's pricing pitfalls. Waymo versus Uber (5 minute read) Waymo's self-driving ride-hailing service has expanded to Los Angeles, joining San Francisco and Phoenix. Riders appreciate the smoother and more private experience compared to traditional rideshares. However, while ridership is increasing, Waymo's profitability remains uncertain. ⚡ Quick Links ChatGPT Search can be tricked into misleading users (1 minute read) ChatGPT Search, a new AI-powered search engine, can be tricked into generating misleading summaries by hiding text in websites. Meta is rolling out live AI and Shazam integration to its smart glasses (3 minute read) Meta's Ray-Ban Smart Glasses now feature live AI assistance, language translation, and Shazam music identification. AI helps ID paint chemistry of Berlin Wall murals (5 minute read) SAPNet is a neural network developed by Italian scientists to enhance spectral data analysis from handheld Raman spectroscopy devices. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://refer.tldr.tech/0b6a6dc1/2 Track your referrals here. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us. If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan & Andrew Carr If you don't want to receive future editions of TLDR AI, please unsubscribe from TLDR AI or manage all of your TLDR newsletter subscriptions.