TLDR AI 2025-11-27

Anthropic CEO Dario Amondei has been asked to testify at a hearing on December 17 on how Chinese state actors used Claude Code in a wide-reaching cyber-espionage campaign. Google Cloud CEO Thomas Kurian and Quantum Xchange CEO Eddy Zervigon were also asked to testify at a hearing on the future of AI and cybersecurity next month. Lawmakers are increasingly interested in hearing more about how nation-state hackers could use AI tools to conduct attacks and how AI can be used to defend organizations.

Harmonic, co-founded by Robinhood CEO Vlad Tenev, has raised $120 million, valuing the AI startup at $1.45 billion. The company focuses on AI's math and reasoning, aiming to eliminate "hallucinations" and is exploring commercial applications in safety-critical industries. Investors include Ribbit Capital, Sequoia, Kleiner Perkins, and Emerson Collective.

Perplexity introduced personalization features that allow its AI assistant to retain preferences, interests, and context across conversations. This memory system improves answer quality and reduces context engineering by automatically loading relevant information for continuity.

This post explains how continuous batching works. Continuous batching maximizes throughput by processing multiple conversations in parallel. Optimizing attention and KV caching enables faster and more scalable LLM serving under heavy user load.

This post discusses a shift from models that learn by predicting the next work to models that gain experience via interaction with environments. There will soon be a shift towards AI models being increasingly trained via interaction with environments via reinforcement learning. The post details the technical process underpinning that shift. There will be significant opportunities for new businesses to be built around recent developments in reinforcement learning in the next year.

EverMind is the long-term memory OS for next-generation AI applications. With EverMind, agents can continuously improve with every interaction: remembering, learning, and adapting over time. Memory is persistent, consistent, and retrievable - unlocking AI that's more coherent, personalized, and capable. Check out the open source repo on GitHub or join the waitlist.

Developers are increasingly requiring agents to take on complex tasks that require work that spans hours or even days. Getting agents to make consistent progress across multiple context windows remains an open problem. Agents need a way to bridge the gap between coding sessions as context windows are limited, and most complex projects can't be completed within a single window. This article proposes a two-fold solution, with an initializer agent setting up the environment on the first run, and a coding agent tasked with making incremental progress in every session while leaving artifacts for the next session. Code examples are provided.

workmux is an opinionated workflow tool for managing git worktrees and tmux windows as isolated development environments. It is perfect for running multiple AI agents in parallel without conflict. workmux makes managing parallel development tasks or multiple AI agents as simple as managing tmux windows.

Gemini CLI is an AI assistant that brings Google Gemini directly into the terminal. It functions as a conversational agentic command-line tool. Gemini CLI can reason about requests, choose tools, and execute multi-step plans. It acts like a supercharged pair programmer and command-line assistant. The assistant excels at coding tasks, debugging, content generation, and system automation.

The US holds a 10x compute advantage over China. Proposed safety regulations might add 1-2% to training costs through disclosure requirements and capability evaluations, which won't meaningfully erode the lead. What could catastrophically undermine it: chip export lobbying that would collapse the advantage to 2x or application-layer regulations that strangle deployment with compliance burdens.

If xAI's model can successfully compete in League of Legends, that means that it would be able to read and understand any computer interface without needing a specialized API. The model will also need to be able to reason and produce a robust and complex plan. Such a model would be a game-changer for AI capabilities and the global economy.

Faster models might not be as 'smart', but they're capable of many tasks and can iterate quickly if they make mistakes.

Chinese companies are training their AI models through lease agreements with foreign data centers owned and operated by non-Chinese entities.

Andrew Ng introduced an agentic paper-review system trained on ICLR 2025 reviews, showing AI-human review correlation slightly higher than human-human.