NVIDIA & AMD chips ⚡, profit per token 💰, GRPO tricks 💡

Nvidia has revealed its latest AI server systems. The Vera Rubin chips are set to go on sale in the second half of this year ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2026-01-06 Stop Winging AI Rollouts. Start Proving Value by the End of Q1 (Sponsor) Without a clear roadmap, AI adoption often fizzles, leaving ROI unproven and teams frustrated. The 90-Day AI Adoption Playbook from You.com gives you an effective, phased roadmap to move from pilot to real, measurable ROI—in only 90 days. What you'll get: A week-by-week rollout plan for secure, scalable AI adoption Key actions to build user skills, certify competency, and drive consistent usage Certification steps to ensure every user delivers impact Make your first 90 days count. Download the playbook. 🚀 Headlines & Launches Nvidia Unveils Faster AI Chips Sooner Than Expected (5 minute read) Nvidia revealed its latest AI server systems at the Consumer Electronics Show in Las Vegas. The Vera Rubin chips are set to go on sale in the second half of this year. They are designed to handle the enormous computing loads needed to create simulations of reality for use in model training. Rubin delivers a 10-fold reduction in cost compared with Blackwell chips. AMD Unveils New Chip For Corporate Data Centers, Talks Up Demand (3 minute read) AMD's new MI440X chip is designed for use in smaller corporate data centers. The company's new Helios system, based on the new chip, will go on sale later this year. AMD's MI500 series of processors will debut in 2027. Those chips will deliver up to 1,000 times the performance of the MI300 series, which was first rolled out in 2023. NVIDIA introduces Alpamayo (4 minute read) NVIDIA launched the Alpamayo family of open AI models, datasets, and simulators to address long-tail challenges in autonomous driving. The models use reasoning-based vision-language-action architectures to support safer, end-to-end AV systems that generalize to rare scenarios. 🧠 Deep Dives & Analysis Gross Profit per Token (2 minute read) Meta's $2 billion acquisition of Manus highlights the importance of gross profit per token, with Manus showing 40x GP multiple. DeepSeek and Together AI have the lowest multiples because they resell inference, unlike Perplexity, which achieves the highest at 222x as an application. Investors prioritize token monetization over raw volume, as indicated by a 0.71 correlation between gross profit per token and valuation. AI Capex: Built on Options, Priced as Certainty (6 minute read) Letters of intents are real options that move markets, guide counterparties, and shape capex decisions before obligations harden. Options can be enough to pull forward tens of billions in irreversible investment. If the option doesn't convert, the stranded capital becomes a restructuring problem. 🧑‍💻 Engineering & Research GRPO++: Tricks for Making RL Actually Work (72 minute read) Group Relative Policy Optimization (GRPO) is the RL optimizer used to train most open-source reasoning models. It is popular due to its conceptual simplicity and practical efficiency. The vanilla GRPO algorithm has subtle issues that can hinder the RL training process, especially at scale. This post provides an overview of the work done to solve the shortcomings of GRPO. M2.1: Multilingual and Multi-Task Coding with Strong Generalization (17 minute read) MiniMax-M2.1 matches or surpasses the level of global top-tier models on multiple internal and external benchmarks. The open source model has exceptional performance in code generation, tool usage, instruction following, and long-range planning. This post discusses how the model was trained and shares insights gained during the process. Deep Delta Learning (2 minute read) A new approach to neural architecture, Deep Delta Learning generalizes residual networks using a single scalar gate that interpolates between identity, projection, and reflection. Tencent's Youtu-Agent Framework for Autonomous Agents (2 minute read) Tencent's Youtu-Agent offers a modular framework for building and evaluating AI agents with open-source models. It emphasizes clean abstractions for tools, environments, and agent logic to support extensibility and robust deployment. KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators at Meta (1 minute read) Making deep learning recommendation model (DLRM) training and inference fast and efficient presents three key system challenges: model architecture diversity, kernel primitive diversity, and hardware generation and architecture heterogeneity. KernelEvolve is an agentic kernel coding framework that tackles heterogeneity at scale for DLRM. It is designed to take kernel specifications as input and automate the process of kernel generation and optimization for a recommendation model across heterogeneous hardware architectures. KernelEvolve is designed to optimize a wide variety of production recommendation models across generations of Nvidia and AMD GPUs, as well as Meta's AI accelerators. It significantly mitigates the programmability barrier for new AI hardware. 🎁 Miscellaneous OpenAI to Buy Pinterest? Strategic Analysis (8 minute read) OpenAI may acquire Pinterest, a visual search app disguised as a social platform. ChatGPT is missing Pinterest's level of visual efficiency because that is not a core capability of large language models. Pinterest's ability to go from intent to product, or from inspiration to intent to product, will be very important for OpenAI's commerce aspirations. ChatGPT must go visual to create a successful UX and user experience for agentic commerce. Gemini-Powered Features Coming to Google TV (3 minute read) At CES 2026, Google previewed Gemini integrations for Google TV, including visual, interactive topic exploration, natural language settings control, and creative tools like Photos Remix and Veo. ⚡ Quick Links NVIDIA Unveils Personal Agent Robotics at CES (6 minute read) At CES 2026, NVIDIA introduced Reachy Mini, a personal robotic AI agent powered by DGX Spark, capable of real-time multimodal interaction. OpenAI loses top AI researcher Jerry Tworek after seven years (1 minute read) Jerry Tworek was a key player in building GPT-4, ChatGPT, and OpenAI's first AI coding models. Thoughts on Claude Code (17 minute read) Claude Code is a phenomenal coding companion that dramatically amplifies the fun side of coding by eliminating the annoying parts. Is hallucination-free AI code possible? (11 minute read) While verification methods will soon be able to pick up almost all logical flaws in some fields, and foundational models will get better, they may still struggle to produce 'correct' code. 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. Want to work at TLDR? 💼 Apply here, create your own role or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! TLDR is one of Inc.'s Best Bootstrapped businesses of 2025. If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian, & Jacob Turner Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.