These are the AI articles and resources from this past week you don’t want to miss:
How Claude Watermarks AI-Generated Text by Sebastian Raschka, PhD: This is a great overview of Claude’s watermarking and how it works. Essentially, it works by controlling the random sampling of words during text generation. This method allows later detection of AI-generated text using a scoring system and threshold. Sebastian’s explanation is great, contains visual explainers, includes a video explanation if you prefer that, and goes over how these watermarks can be defeated.
Designing Loops for Production-Grade Work: Liquid AI demonstrated that coding agents can autonomously build a high-throughput, production-grade BPE tokenizer in Rust when guided by automated verification loops. It’s a great lesson in how to enable agents to complete complex agentic engineering tasks and what makes them so great at it. My favorite takeaway from the article:
“From the experiment, we learned that coding agents are multi-domain experts, covering a span of expertise none of our engineers have. Turns out it knew OpenAI's cl100k regex, and it knew Rust’s
rayon. Nobody we could have staffed on this knew both.”Measuring Autonomous AI Research: An evaluation of 153 autonomous nanoGPT speedruns across 18 frontier models revealed that top models excel at empirical iteration rather than inventing novel algorithmic architectures. There was huge variability in model performance across the board. There’s been a lot of talk recently about RSI, but it’s fundamental for agents to improve iteratively to achieve it. We’ll see a lot more articles on this very soon.
Reinforcement Learning for LLMs: The Complete Guide by Cameron R. Wolfe, Ph.D.: This is a compilation of multiple articles Cameron wrote related to Reinforcement Learning. It’s long, but worth the read. It covers fundamentals, policy gradient algorithms, advanced research topics, and how all of this is applied to frontier LLMs. You’ll want to prepare to sit down and read this in multiple sittings.
Compute-Optimal Scaling Laws for Human Motion Generation: Cartwheel found that generative motion models follow Chinchilla scaling laws. This means we can throw more compute and time spent training at models trained on motion to see improved performance. They also created 2D-to-3D motion reconstruction software that enables curating large motion datasets without the need for motion capture.
Pacing model development in an era of cyber-critical capabilities: After the Hugging Face incident, OpenAI is slowing model development to ensure they’re releasing models safely. This means tightening security and focusing on alignment, or ensuring the model does what they want it to, to make sure critical cyber incidents don’t happen.
AI Performance Engineering: A repository from wafer.ai sharing resources to learn AI performance engineering. It covers GPUs, kernel optimization, profiling, inference, and more. This is a great guide to get more into performance engineering.
Jalapeño’s first results show industry-leading speed and efficiency in AI inference: An overview of how OpenAI’s AI chips are progressing and a reminder of the benefit of owning the entire AI stack—especially that compute layer. I wish there was more information about Jalapeño here, but it’s still interesting to see it performing well.
Thanks for reading! Feel free to drop your favorite articles from this week in the comments.
Always be (machine) learning,
Logan