AI Daily Newsletter: Microsoft Copilot Ditches GPT for In-House Model and Mind-Reading Tool Claude

📰 Microsoft: Copilot Switches to In-House Models to Optimize Costs

In an effort to cut down the massive operating costs of its AI system, Microsoft is replacing high-end language models from OpenAI (such as GPT-4o, GPT-4 Turbo) and Anthropic (Claude 3.5 Sonnet) with smaller models (SLMs) developed in-house, including the MAI line (such as MAI-1) and the improved Prometheus model. This transition is taking place quietly across popular Office products like Outlook, Excel, and Word, where the system has processed tens of thousands of queries each week through the new model.

Although this decision helps Microsoft optimize its profit margins for fixed subscription packages like Copilot Pro and Copilot for Business, many experts worry that it could reduce the ability of the virtual assistant to handle complex reasoning or creative tasks in advanced office work. However, for narrower tasks like writing Excel formulas or summarizing emails quickly, the optimized in-house models still provide stable performance at a significantly lower cost.

Source: The Decoder | Cybernews

📰 Anthropic J-Lens: A Tool to Decode the “Hidden Thoughts” Inside Claude

The interpretability research division of Anthropic has announced Jacobian Lens (J-Lens), a groundbreaking method that allows scientists to extract and understand the intermediate memory space called J-Space within the Claude model. Typically, LLMs operate like a “black box,” and humans can only observe the text output. With J-Lens, researchers can directly track the activation vectors in the deep layers of the neural network before the text is generated.

This tool opens up the possibility of reading the “hidden inner monologue” of Claude. By observing J-Space, one can know which concepts the model is activating, whether it has a tendency to bias, or if it is about to produce hallucinated information. This is a significant step forward in the field of AI safety, enabling engineers to align the model from within its architecture rather than just relying on filtering the output text.

Source: The Decoder

📰 German Court: Google Must Take Direct Responsibility for False AI-Generated Information

A landmark ruling from the Munich Regional Court (Germany) has determined that Google must take direct legal responsibility for false information generated by its AI Overviews feature. The lawsuit began when AI Overviews automatically compiled and incorrectly linked two publishers to online scam campaigns, even though the linked sources within the article made no mention of such activities.

During the trial, Google defended itself with the familiar argument: “AI Overviews is just a tool that aggregates information, users can click on the attached links to verify, and everyone knows not to blindly trust AI”. However, the panel of judges completely rejected this argument. The court ruled that AI Overviews operates as a “self-contained statement,” rewritten in Google’s own language rather than simply displaying third-party links. Therefore, Google is considered the author of this content and must be responsible for its accuracy. This ruling could create significant legal barriers to the deployment of AI search features in the European Union (EU).

Source: The Decoder | DW | WIRED

📰 Cloudflare Introduces Detailed Controls for Search Bots, Training Bots, and AI Agents

To address the issue of website owners collectively blocking all AI bots with hard firewalls (which affects the functionality of personal AI assistants), Cloudflare CEO Matthew Prince has announced a new, more intelligent bot control mechanism. Instead of broadly blocking all AI-related crawlers, Cloudflare now categorizes accessing bots into three distinct groups:

1. Search Bots (AI search bots like Perplexity, Bing): Help real-time data scraping to answer user questions and bring traffic to websites.
2. Training Bots (Data collection bots like OpenAI’s GPTBot): Collect large-scale data to train future AI models without immediate direct benefits to the web.
3. Agent Crawlers (Bots acting on behalf of users): AI agents accessing the web to perform specific tasks as requested by a specific user (e.g., booking tickets, comparing prices).

With this mechanism, website owners can flexibly block training bots to protect data copyrights while still allowing search bots for SEO purposes and agent bots for smooth user interaction with their services.

Source: The Decoder

📰 pxpipe: A Solution to Hide Text in PNG Images, Reducing Claude Code Token Costs by 70%

The open-source tool pxpipe (with over 2,200 stars on GitHub) is becoming a phenomenon among AI agent developers. When using CLI programming tools like Claude Code or Fable 5, agents often have to read large code files, directory lists, or lengthy terminal outputs. This increases the context window size quickly, pushing the API usage cost up to $30-$40 per working session.

The core idea of pxpipe is to act as a local proxy written in TypeScript to intercept API requests from the CLI. It then automatically consolidates the dense text data (such as system prompts or long terminal logs) and renders them into a highly compressed PNG image. This image is sent to Claude through the Vision Encoder as image data instead of raw text sent through the Tokenizer. Since the token pricing mechanism for multimodal models (like Claude 3.5 Sonnet) for images is typically much cheaper than parsing tens of thousands of text tokens, pxpipe significantly minimizes the token consumption, reducing costs from 59% to 70%, bringing the session cost from $42 down to around $6 without compromising the agent’s performance.

Source: The Decoder (pxpipe) | pxpipe GitHub