TLDR AI 2025-12-18

Warp just launched its biggest Agents update yet, propelling it to #1 on Terminal-Bench and #5 on SWE-bench Verified, ahead of Claude Code and Codex CLI.

Warp's #1 Agent now includes:

✅ Full terminal use: agents can interact with long-running commands like servers, debuggers, and more.

✅ Steerable planning: review and edit plans as agents work to stop them going off the rails.

✅ Full lifecycle: planning → coding → deployment

With over 700K users and 56% of the Fortune 500 on board, Warp is emerging as the platform of choice for developers.

Try the world's best coding agent, free →

xAI has opened up the voice stack powering Grok in Tesla vehicles to all developers. The API ranks #1 on Big Bench Audio with sub-1-second time-to-first-audio at $0.05 per minute (half the cost of OpenAI's Realtime API).

Google has introduced Gemini 3 Flash, a fast, lightweight model designed to deliver frontier-level intelligence with lower latency. It's built to help users quickly learn, plan, and build across a range of everyday tasks and applications.

Google and Meta are working together on a new initiative to make Google's Tensor Processing Units better at running PyTorch. The move is aimed at weakening Nvidia's longstanding dominance in the AI computing market. TorchTPU will remove a key barrier that has slowed the adoption of TPU chips. It will make existing infrastructure built using PyTorch software fully compatible with TPUs. Google is considering open-sourcing parts of the software to speed uptake.

AI agents, like those launched by OpenAI, threaten marketplace moats by taking over discovery, transactions, and supply workflows, thus collapsing traditional marketing funnels. Marketplaces must adapt by optimizing brand experiences, personalizing interactions, and integrating supply to maintain value. Founders should focus on high-frequency markets, fortify customer connections, and explore new monetization strategies like subscriptions or data licensing to stay competitive.

Oracle is down 43% from its September high despite reporting $523 billion in "remaining performance obligations". About $300 billion of that comes from a five-year OpenAI contract, but analysts are skeptical that OpenAI, which only has ~$20B in annualized revenue, will be able to pay. Nvidia's stalled $100 billion OpenAI investment (announced in September but still unsigned) compounds the anxiety over the circular deals plaguing AI infrastructure.

Current models vastly underutilize hardware: DeepSeek-V3 and Llama-4 achieve only ~20% FLOP utilization during training, and inference runs at single-digit utilization because autoregressive models are bottlenecked on loading weights from memory, not compute. Models we see today are also lagging indicators, trained on last-gen hardware that's unoptimized for the massive size of modern clusters and the latest training methods.

AI is all the rage, but are you really using it to your advantage? This practical guide from You.com guides you through the process of identifying and prioritizing opportunities where AI can best serve your org. Explore where AI can drive the most impact in your organization, both internally and externally.

Learn more with this AI Use Case Discovery Guide.

Opal is a tool for building AI-powered mini apps. It is now directly available in the Gemini web app. Users can use it to create experimental Gems to unlock even more customized Gemini experiences. The tool can be found in the Gems manager.

Plan mode is a feature in Claude Code that pauses execution until users approve of a plan. A plan in Claude Code is a markdown file written into Claude's plans folder by Claude in plan mode. Users can review and edit the plan before Claude executes whatever task it needs to. The tool brings in more complexity into the user interface, taking away some of the magic. It replicates something that natural language can already do, making the user experience a little less natural.

Chatterbox is an open-source state-of-the-art text-to-speech (TTS) model with multilingual support, emotion control, and zero-shot voice cloning.

Peter DeSantis, SVP of Utility Computing at AWS, is taking two of his teams from AWS and putting them into a unit called 'AGI'. The unit will report directly to Andy Jassy, Amazon's Lead Product Market for GenAI and CEO. AI and AI-related compute is now an all-Amazon thing, not just an AWS thing. The unit is building things that are so important that Amazon is taking one of its most senior AWS leaders and putting him in charge of all of it.

Chinese scientists have built a working prototype of an EUV lithography machine, the $250 million tool that only ASML knows how to make and which the US has blocked from being sold to China since 2018. A team of former ASML engineers reverse-engineered the machines in a high-security Shenzhen lab, with recruits given fake IDs and signing bonuses up to $700,000. The prototype hasn't yet produced working chips, but sources expect it will by 2028-2030.

Cached prompts are 10x cheaper and 85% faster. What actually gets cached isn't the response, but the K and V matrices from the attention mechanism. See the exact math from ngrok

ChatGPT's daily active users in India surged 600% year-on-year to 73 million, more than double its US base, after OpenAI made its Go plan free for a year.

In another circular deal, OpenAI is looking for a $10 billion investment from Amazon alongside an agreement to use AWS' Trainium chips.

Opera has globally launched its Neon browser with paid early access at $19.90/month, targeting AI power users.