OpenAI $300B compute deal 💻, ChatGPT developer mode 🤖, ByteDance image model 🖼️

OpenAI signed a $300 billion deal with Oracle for compute power over five years, starting in 2027, marking one of the largest cloud contracts ever ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2025-09-11 What's stopping your enterprise from deploying AI agents? (Sponsor) Security? Integrations? Orchestration? If you've tried bringing agentic AI to a large enterprise, these are questions you needed to answer - and most simple workflow builders have no good answers. That's where Airia comes in. It's the AI security and orchestration platform that's built for enterprise. That means security and compliance are cornerstones, not afterthoughts. It means they've thought of how this thing actually integrates with your legacy systems. And it means you'll get actual onboarding and a thriving support community. Ready to unlock some exciting use cases? Get started with a guided demo or free trial 🚀 Headlines & Launches OpenAI and Oracle reportedly ink historic cloud computing deal (2 minute read) OpenAI reportedly signed a $300 billion deal with Oracle for compute power over five years, starting in 2027, marking one of the largest cloud contracts ever. This move diversifies OpenAI's cloud providers beyond Microsoft Azure, aligning with its involvement in the $500 billion Stargate Project for domestic data centers. Oracle has been working with OpenAI since 2024, highlighting a strategic shift in cloud resources essential for OpenAI's growth. Replit raises $250M at $3B valuation, launches Agent 3 (4 minute read) The AI coding platform grew annualized revenue from $2.8 million to $150 million in less than a year, driven by 40 million users who can now build apps without coding experience. Agent 3 can work autonomously for up to 200 minutes compared to Agent 1's 2-minute limit, testing and fixing its own code while building applications with minimal human interaction. ByteDance unveils new AI image model to rival Google DeepMind's 'Nano Banana' (2 minute read) ByteDance's Seedream 4.0 is an AI image model that the company claims is superior to Google DeepMind's "Nano Banana" in prompt adherence, alignment, and aesthetics. The model combines features from previous versions at the same cost of US$30 per 1,000 generations. These claims are based on ByteDance's internal MagicBench benchmark but lack an official technical report. 🧠 Deep Dives & Analysis Nondeterminism in LLM Inference (38 minute read) LLM inference remains nondeterministic even under greedy decoding due to floating-point non-associativity and concurrent execution. How to use computing power faster: on the weird economics of semiconductors and GenAI (17 minute read) It's difficult to understand how model providers will manage to generate sufficient usage, as the adoption rate is far from sufficient to absorb the computing power that will eventually be brought to market. The semiconductor industry will have to enter a downward cycle at some point, and this will hurt the AI sector significantly. The lack of interest in the economics of semiconductors is a real weakness in the analysis of the development of generative AI. This article shares economic considerations on the subject to help readers better understand these contemporary issues. A guide to understanding AI as normal technology (27 minute read) What if AI follows predictable patterns of adoption instead of the rapid superintelligence takeoff many in Silicon Valley prophesy? The impacts of AI depend on how it's adopted, not raw capabilities alone. Many industries continue to use outdated technology due to the non-technical (for example, Air Traffic control uses the same system from the 50s). Even advanced AI features available today have an adoption problem. ChatGPT's "thinking" models are used by less than 1% of users nearly a year after release. 🧑‍💻 Engineering & Research Get insights from Slack's 2025 Workforce Index (Sponsor) AI at work is growing fast — Slack's latest Workforce Lab report shows daily use up 233% in six months, driving 64% higher productivity and 81% greater job satisfaction. Join Slack's webinar to dig into the data and learn practical ways to use AI for smarter insights, faster workflows, and less busywork. Claude API: Web fetch tool (3 minute read) The new web fetch tool in the Claude API fetches content from URLs as part of its response to prompts. It can extract full text content from URLs and PDFs. For security, the tool can only fetch URLs that have previously appeared in the conversation context, so it cannot fetch arbitrary URLs that Claude generates or URLs from container-based server tools. Anthropic recommends only using the tool in trusted environments or when handling non-sensitive data, as enabling it poses data exfiltration risks. Mini-o3: Open Source Agentic Visual Reasoning (4 minute read) Mini-o3 is an open-source model designed for agentic visual reasoning. It supports multi-turn interactions similar to OpenAI's o3, and scaling to tens of rounds. The full training pipeline is public. 🎁 Miscellaneous AI as teleportation (5 minute read) While AI has the potential to improve our lives in a lot of ways, it is important to be wary of secondary effects and unintended consequences of the technology. For example, vibe coding really speeds up iteration loops and helps developers explore more ideas, but it removes previous design processes, which can change things in ways that are hard to predict. Sometimes the friction and inconvenience is where the good stuff happens, so be very careful about removing it. I ran Claude in a loop for three months, and it created a Gen Z programming language called cursed (6 minute read) Claude autonomously developed a programming language where "slay" means function, "sus" means variable, and pointers are represented by the Among Us character. The "cursed" language compiles to actual binaries and includes editor extensions. ⚡ Quick Links The Modern Productivity Paradox: Why We Need Agents That Think With Us, Not For Us (Sponsor) Imagine agents becoming our "collective subconscious," executing digital drudgery while we focus on higher-level thinking. See how Amazon's AGI SF Lab is revolutionizing human-agent collaboration through Nova Act. Discover the future of augmented intelligence → Judge puts Anthropic's $1.5 billion book piracy settlement on hold (3 minute read) A federal judge paused Anthropic's $1.5 billion book piracy settlement, citing concerns about the terms that give authors $3,000 per allegedly pirated book. We've (finally) added full support for MCP tools in ChatGPT (1 minute read) Developers can now create connectors and use them in chat for write actions in developer mode. Claude web app may soon add memory synthesis and incognito mode (1 minute read) Leaked code strings suggest Claude will soon automatically incorporate information from past conversations without being explicitly asked and will support an "incognito mode" that would prevent chats from being saved, added to memory, or used for model training. Gemini Batch API now supports Embeddings and OpenAI Compatibility (2 minute read) The Gemini Batch API now supports the newly launched Gemini Embedding model and the ability to leverage the OpenAI SDK to submit and process batches. 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 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 & Sahil Khoja Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.