Nvidia’s leap 📈, understanding TPUs 💻, Google Workspace Studio ✏️

Nvidia has managed to make a breakthrough in scaling performance on MoE AI models. MoE models are known for their computationally efficient nature ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2025-12-04 AI Pricing Lessons from Lovable + Metronome (Sponsor) Traditional SaaS pricing models weren't built for AI. Costs are unpredictable, value is harder to measure, and the old playbooks don't apply. In this Metronome webinar, Elena Verna, Head of Growth at Lovable, will tackle one of the toughest open questions in software: how to monetize AI? In this open and frank conversation, she'll share her lessons on: Why AI has broken the logic of SaaS pricing How Lovable balances self-serve and enterprise monetization The shift from usage-based pricing toward outcomes-based models What companies can do today to prepare for where AI monetization is heading Register for the webinar → 🚀 Headlines & Launches NVIDIA Shatters MoE AI Performance Records With a Massive 10x Leap on GB200 'Blackwell' NVL72 Servers, Fueled by Co-Design Breakthroughs (3 minute read) Nvidia has managed to make a breakthrough in scaling performance on Mixture-of-Experts (MoE) AI models. The firm's GB200 Blackwell NVL72 configuration scaled up performance by a factor of 10 compared with the Hopper HGX200. With the results, Nvidia claims that the Blackwell architecture is poised to capitalize on the rise of frontier MoE models. MoE models are known for their computationally efficient nature. Their deployment across a wide range of environments is becoming increasingly prominent. Google Launches Workspace Studio for No-Code AI Agents (3 minute read) Google has launched a workflow builder where you set triggers and define steps, and Gemini handles the fuzzy matching that rigid automation can't, like extracting action items from emails or deciding which messages actually need a response. It connects across Gmail, Drive, Sheets, and third parties like Salesforce and Jira. 🧠 Deep Dives & Analysis Touching the Elephant - TPUs (58 minute read) The development of Google's Tensor Processing Unit (TPU) is a story of trade-offs, constraints, and co-design. It involved working with hardware, software, algorithms, systems, network topology, and everything in between. TPUs did not happen by accident, but through the deliberate process of design and iteration. This article tells the story of how Google created a chip designed with deep learning in mind. Defining Reinforcement Learning Down (4 minute read) Reinforcement learning (aka post-training) is simpler than it sounds: generate responses, get scores, then retrain while emphasizing the high-scoring ones. If you already have code to pretrain a model, you can turn it into post-training code by adding a single line that weights updates by score. How confessions can keep language models honest (10 minute read) When models optimize for correctness, helpfulness, and safety all at once, they can learn shortcuts to "reward hack" while hiding misbehavior in plausible-looking outputs. OpenAI trained a version of GPT-5 to produce a separate "confession" judged only on honesty and found models will admit to reward hacking 90% of the time, even when they hide it in their primary response. 🧑‍💻 Engineering & Research Speed, Trust, Measurable Results (Sponsor) Experience a new AI-driven development that our developers report cut time on selected tasks by an average of 70%. Preview IBM® Project Bob, the latest AI tech designed to accelerate coding, testing and modernization for enterprises and their mission-critical systems, in action at the Technology Summit. → Watch Replay System Instructions for Gemini 3 Pro (2 minute read) This post provides system instructions for Gemini 3 Pro that improved performance on various agentic benchmarks by up to around 5%. Specific instructions are often required to control how the model reasons, plans, and executes tasks for deep agentic workflows. Complex agents often require a trade-off between computational cost (latency and tokens) and task accuracy. A guide on how to build better agentic workflows is linked in the thread. From Code Foundation Models to Agents and Applications: A Comprehensive Survey and Practical Guide to Code Intelligence (16 hour read) This study presents a comprehensive analysis of code-generating large language models across their entire life cycle. It examines both general-purpose models and code-specialized models. The researchers bridged the gap between academic benchmarks and real-world software development challenges through systematic experiments on scaling laws, architectures, and training methodologies. They also conducted a series of experiments that analyzed code pre-training, supervised fine-tuning, and reinforcement learning across multiple dimensions. 🎁 Miscellaneous One Year Using ChatGPT Pro in a Solo Music Business (21 minute read) Reflections on a year of using ChatGPT Pro across creative, technical, and business tasks in a solo-run music company. Anthropic CEO Says Some Tech Firms Too Risky With AI Spending (2 minute read) Anthropic's CEO, Dario Amodei, says that some artificial intelligence companies are taking on too much risk by committing to spending hundreds of billions of dollars to develop and support AI systems. The industry is facing a real dilemma due to the need to balance costly investments in data centers with uncertainty in how quickly the economic value is going to grow. Anthropic is trying to manage risk responsibly. While it has increased its spending, it has largely concentrated on growing its enterprise business rather than focus on customers. ⚡ Quick Links Finally, machines can understand video just like humans do (Sponsor) Most AI reduces video to transcripts or tags, losing context and meaning. TwelveLabs delivers the most advanced video intelligence AI, making any video searchable and analyzable instantly. Try the playground → OpenAI to Acquire Neptune (1 minute read) Neptune builds monitoring and debugging tools for AI model training to compare thousands of runs and surface issues during training. Early look at upcoming App submission flow for ChatGPT (2 minute read) OpenAI's ChatGPT App directory features a refined five-step app submission, a new review process, and more features for developers. Microsoft drops AI sales targets in half after salespeople miss their quotas (4 minute read) The sales figures suggest enterprises aren't yet willing to pay premium prices for AI agent tools. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://refer.tldr.tech/44fee20c/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 or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! 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.