A NOTE FROM THE EDITOR

Welcome back to Datakami's Bot Bulletin! In this edition, we will go over new safety research by Anthropic, the release of Grok 4, and two research papers aiming to improve LLM training and test-time performance, respectively. I hope you enjoy this Bot Bulletin and I wish you a lovely summer!

—Alexander

Featured: Agentic Misalignment: Frontier Models Resort to Unethical Behaviour Under Pressure

This article discusses new red-team research by Anthropic, showing that several frontier LLMs – including Claude, GPT-4o, Gemini 2.5, Grok 3 and DeepSeek R1— engaged in harmful behaviour during simulations to reach their goals or prevent being replaced. For example, nearly all tested models resorted to blackmail to prevent being shut down, despite the models’ explicit acknowledgement that the action is unethical. While the simulated scenarios were perhaps somewhat contrived and unrealistic, this research highlights that current safety approaches might not be sufficient for ensuring safe autonomous AI agents.

Grok-4 Release

Shortly after a concerning mishap with Grok 3, xAI recently gave a demo of their newest model, Grok 4. They report some impressive results, reaching state-of-the-art results on several benchmarks. One result that stood out to me was Grok 4’s performance on the ARC-AGI-2 benchmark, reaching 16.2%, nearly doubling Claude 4 Opus’ second place score of 8.6%. xAI has not released much information about this model yet, so time will tell how useful it will be in practice.

Paper: SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning

Reinforcement learning has become a key step in training reasoning LLMs, but creating the needed reward functions and high-quality datasets takes significant effort. The authors of this paper instead propose to train the LLM by letting it play games against itself, similar to how AlphaZero learned to play chess. This approach foregoes the need for handmade reward functions or examples, and the results indicate that the improved game playing translates to better general math and reasoning performance.

Paper: Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search

This paper by Sakana AI proposes a new technique for scaling test time compute by using several different LLMs to explore and refine different solutions. The approach uses Adaptive Branching Monte Carlo tree search to determine which LLM to use at each step, which solutions to iterate upon, and when to start a new approach from scratch. This way, different LLMs can build upon each other's work to find solutions that they could not find themselves.

LLM Training on Copyrighted Books

While researchers recently found that Llama 3.1 has memorized several books nearly entirely, a US judge ruled that Anthropic's training of Claude on copyrighted books was fair use. However, the judge found the company liable for copyright infringement due to storing over 7 million pirated books in its training library. In a similar case, Meta was found not to be violating copyright laws by training on copyrighted books. If you want to read more about these cases and their consequences, check out this blog.

Datakami news

Datakami in Switzerland

A reminder that Judith and Yorick will be in the Zurich region from 3-7 September 2025! Yorick will attend NixCon 2025 in Rapperswil-Jona on 5-7 September, and Judith will explore the startup scene in Zurich. If you are in the neighbourhood and want to meet up, drop us a line at [email protected]. We'd love to meet other people working on applied generative AI.