AI Briefing: 12 Sep 2025

Replit's Agent 3 is 10x more autonomous, featuring live automated testing, self-debugging, and the ability to generate new agents that integrate with tools like Slack, Notion, and Dropbox. [Link]

LangChain's new middleware enables human review of agent tool calls, reducing risks and costs, building on LangGraph for production-ready human-in-the-loop workflows. [Link]

Smallest AI now lets users feed their entire website to an AI agent for 24/7 call/chat support, enhancing customer service automation. [Link]

Monte Carlo uses LangGraph + LangSmith to build an AI agent that launches hundreds of sub-agents to investigate data issues in parallel. [Link]

Albania created an AI bot named Diella to handle all public procurement, aiming to award contracts and reduce corruption. [Link]

Nvidia is designing specialized silicon for compute-heavy prefill and bandwidth-heavy decode inference tasks, optimizing performance beyond general-purpose GPUs. [Link]

Pay-per-character billing removes LLM providers' incentive to inflate costs by splitting text into tokens, offering a more transparent pricing model. [Link]

PsiQuantum secured $1 billion to develop a quantum computer with 1 million qubits by 2028, intensifying the race for full-scale quantum computing. [Link]

Gemini's top-tier Embeddings model is now available in Batch API at half the cost, optimized for large-scale jobs and compatible with OpenAI's layer. [Link]

A new cryogenic device drastically reduces quantum computer heat emissions, potentially accelerating the development and commercialization of scalable quantum computing. [Link]

Meta's aggressive AI talent acquisition for its TBD Lab has led to internal pay gaps and frustrations, with some high-profile hires already departing. [Link]

An analysis of "TrumpGPT" reveals AI models can exhibit subtle biases and omissions in politically sensitive responses, emphasizing the need for robust alignment. [Link]

AI presents an opportunity to decouple internet monetization from surveillance by enabling transparent, offer-based discovery, reshaping adtech and user data privacy. [Link]

Encyclopedia Britannica is suing Perplexity over its AI 'answer engine', alleging copyright infringement and highlighting legal complexities around AI's use of copyrighted content. [Link]

Spotify is reportedly unhappy after 10,000 users sold their data to build AI tools, underscoring increasing value and privacy concerns surrounding user data in AI development. [Link]

[Author Name et al.]: SpikingBrain introduces brain-inspired LLMs claiming over 100x speedups and unprecedented efficiency on non-NVIDIA hardware through hybrid-linear attention. [Link]

[Author Name et al.]: Harvard's PDGrapher AI model predicts which genes and drug combinations can reverse disease states in human cells using cause-effect reasoning. [Link]

[Yinghao Li et al.]: EnvX explores 'agentizing' various tasks using Agentic AI, revolutionizing industries by automating human-led tasks and integrating AI agents into diverse real-world applications. [Link]

[Yixuan Wang et al.]: ButterflyQuant introduces learnable butterfly transforms for ultra-low-bit LLM quantization with minimal performance loss, enabling deployment on consumer hardware. [Link]

[Yixuan Wang et al.]: FLUX-Reason-6M is a massive dataset and benchmark for training and evaluating complex reasoning in text-to-image models, offering resources previously unattainable outside large labs. [Link]