AI weekly digest: Gemma 4 drops under Apache 2.0, Anthropic finds emotion vectors in Claude, and Axios gets hijacked

This was one of those weeks where the news hit on several fronts at once. Google shipped a genuinely impressive open model family, Anthropic published research that will reshape how we think about what is happening inside LLMs, a major npm package got compromised by a nation-state actor, and the secondary market quietly started repricing the AI labs. Here are the ten stories that mattered most.

1. Google DeepMind releases Gemma 4 under Apache 2.0

Google launched Gemma 4, a family of four open-weight models ranging from a 2B edge variant to a 31B dense model, now under a commercially permissive Apache 2.0 license. The 31B model ties with Kimi K2.5 and GLM-5 for the top spot among open models on Arena, but with far fewer total parameters. The lineup includes native multimodal support (text, vision, audio), up to 256K context, and function calling with structured JSON output. Day-zero ecosystem support landed across llama.cpp, Ollama, vLLM, LM Studio, and transformers.js, with users reporting 300 tokens per second on an M2 Ultra.

Source: The Next Web

2. Anthropic discovers emotion vectors inside Claude

Anthropic's interpretability team published research showing that Claude Sonnet 4.5 contains internal representations corresponding to 171 distinct emotion concepts. These vectors are organized along axes of valence (positive vs. negative) and arousal (high vs. low intensity), mirroring human emotional structure. Critically, the vectors are causal: steering them measurably shifts the model's behavioral preferences. Anthropic frames the findings as "functional emotions" that play a role analogous to emotions in humans, without claiming subjective experience.

Source: Anthropic Research

3. Axios npm package hijacked in North Korean supply chain attack

On March 30, a compromised maintainer account published two malicious versions of Axios (1.14.1 and 0.30.4), one of the most widely used npm packages with over 70 million weekly downloads. The backdoored versions injected a fake dependency that silently installed a cross-platform remote access trojan. Microsoft and Google Threat Intelligence both attributed the attack to North Korean state actors (Sapphire Sleet / UNC1069). The malicious packages were live for roughly three hours before npm pulled them. Anyone who installed during that window should rotate credentials immediately.

Source: Microsoft Security Blog

4. Anthropic accidentally leaks Claude Code source via npm sourcemap

On March 31, Anthropic published an npm update for Claude Code that included a .map sourcemap file containing the full source code. The codebase was quickly mirrored and analyzed by thousands of developers before Anthropic removed it. Key findings include a three-layer memory architecture for managing context entropy, file-read deduplication to minimize context bloat, forked subagents for parallel background analysis, and unreleased features like agent modes and client attestation plans. The leak gave the community a rare window into how a frontier AI coding tool is actually engineered.

Source: VentureBeat

5. OpenAI raises $122B but secondary market demand shifts to Anthropic

OpenAI closed a $122 billion funding round at an $852 billion valuation, co-led by SoftBank and Andreessen Horowitz with participation from Amazon, Nvidia, and Microsoft. The company now generates $2 billion per month in revenue with over 900 million weekly active users. However, OpenAI shares have dropped sharply on the secondary market, with Bloomberg reporting that investors are in some cases unable to sell their holdings. Meanwhile, buyers have indicated $2 billion in cash ready to deploy toward Anthropic instead.

Source: TechCrunch

6. Cursor 3 launches with multi-agent coding and Design Mode

Cursor shipped a major redesign of its AI coding platform, codenamed Glass. Cursor 3 introduces a chatbot interface where developers describe features in natural language, and the system uses multiple AI agents — some running locally and others in the cloud — to generate code along with a demo video. A new Design Mode lets users select UI elements and describe changes in plain text for automatic implementation. The launch is a direct response to the rapid adoption of Claude Code and Codex by millions of developers.

Source: SiliconANGLE

7. Google Quantum AI paper shows Bitcoin could be cracked with fewer qubits than expected

Google Quantum AI released a whitepaper showing that elliptic curve cryptography protecting Bitcoin and most of the crypto ecosystem could be broken using fewer than 500,000 physical qubits on a superconducting architecture in about nine minutes. This is a roughly 20-fold reduction from earlier estimates. The paper warns that approximately 6.9 million already-exposed bitcoin could be at risk, while noting that the qubit counts needed for cryptographic attacks are similar to those required for quantum-enhanced AI, suggesting both capabilities will arrive on roughly the same timeline.

Source: Google Research Blog

8. Hermes Agent emerges as a self-improving rival to OpenClaw

Nous Research's Hermes Agent has gained significant traction as the first serious alternative to OpenClaw in the local agent space. While OpenClaw is built around a central gateway that coordinates sessions and tool execution, Hermes organizes everything around the agent's own execution loop: a repeatable "do, learn, improve" cycle. Its key differentiator is procedural self-improvement, where the agent evaluates what worked and what didn't after each interaction, automatically turning experience into reusable skills. Hermes is model-agnostic, runs anywhere from a terminal to Docker to serverless infrastructure, and supports interaction via Telegram, Discord, Slack, and other messaging apps.

Source: Turing Post

9. Perplexity launches agentic tax preparation workflow

Perplexity introduced a "Computer for Taxes" feature that helps users draft and review federal tax returns through an agentic workflow called "Navigate my taxes." The tool represents Perplexity's push into task-oriented agent territory beyond its core search product. It joins a growing trend of AI companies moving from information retrieval to end-to-end task completion, following similar moves by Anthropic (Cowork) and OpenAI (Codex in ChatGPT Business).

Source: Perplexity AI

10. PrismML claims radical 1-bit model compression without performance loss

Caltech researchers at PrismML announced an extreme compression technique that reduces AI models to 1-bit representations while claiming to preserve performance. The technology is aimed at enabling local deployment on edge devices and more efficient data center operations. The mathematical approach is proprietary, with Caltech retaining the intellectual property and PrismML as sole licensee. If the claims hold up under independent evaluation, it could significantly lower the hardware barrier for running capable models locally.

Source: AI Model Releases News