Since I’ve been quite inconsistent, I will be switching to monthly summaries covering anything interesting in that month in no particular order.

IT

• Google introduced DiffusionGemma as name suggest uses diffusion architecture instead of standard transformer approach. The advantage is faster token generation for the sake of model hallucinations¹.
• I finally watched The Thinking Game, a short movie about how Google DeepMind tackled the protein folding problem. It’s definitely worth a watch.
  • Watching it reminded me how my colleagues from the lab got excited and wanted to run it out on our GPU server. It took me few days to set it up correctly².
• Bloomberg discusses AI and safety with Dario and Daniela Amodei in Inside Anthropic, the $965 Billion AI Juggernaut.
• How to Setup a Local Coding Agent on macOS by Kyle Howells compares token generation using gemma-4-26B-A4B (58.2 tokens/sec) vs adding MTP Draft mode (72.2 tokens/second). Interestingly, llama.cpp was clearly faster in all the tests compared to mlx which was a surprise.
• Papers With Code is back and offers curated papers from arXiv and Hugging Face.
• WWDC26 presented apple/container used to run linux containers. I am curious to see benchmarks comparing docker and podman runtime³.
• Scripting good practices in Python by Bite code! provides some nice examples. Here are some of my favorites
  • Storing secrets using keyring library instead of relying on environment variables
  • argparse.RawDescriptionHelpFormatter for cli with no arguments
  • Custom exit function applied to sys.excepthook
• Yolov5 is out and it’s impressive as usual, I suggest you try the iOS app and walk around your area.
  • I am also including a nice example by Marco Franzon where YOLO was finetuned to track immune cell chasing bacteria.
• Fernando Irarrázaval was curious if people could break his openclaw instance (assistant Fiu) using prompt injection to leak the secrets thought an email summarized in What happened after 2,000 people tried to hack my AI assistant.
  • He chose Claude Opus 4.6 which already comes with strong guard rails. Interestingly, Fiu recognized over time that this was not an organic traffic and stored this info in its memory.
  • Although Fiu didn’t leak the secrets, was it worth the 500$ cost?

Science

• A gene-therapy trial focusing on partial rejuvenation of older cells in eyes has been administered to first patient executed by Life Bioscience⁴ (co-founded by Prof. David A. Sinclair). Using inducible adenovirus the goal is to activate 3 reprogramming genes (OCT4, SOX2, KLF4 aka OSK genes from Yamanaka factors). The system only activates if antibiotics called doxycycline is administered⁵⁶.
• China deployed artificial human embryos (on batch cultured on uterine cells and second placed inside a microfluid chip) to Tiangong space station, while keeping control cells on Earth. The objective is to study the impact of space environment on the embryo development. The objective is to see if in future life can be started on other planets⁷.
• Daraxonrasib⁸ a newly proposed drug prolongs the lifespan of patients suffering from advanced pancreatic cancer which suffer from RAS-mutations.
  • The study shows that the treatment was able to double the life expectation (however no patient survived to my understanding). It is still a promising new treatment with positive feedback from the community⁹.
• A nice write-up by Eaton about Exploiting vulnerabilities in Johnson & Johnson web apps.
• Cultrera di Montesano et al., 2026 focuses on improving cell type prediction by leveraging hierarchical ontologies. As cell type annotation is primarily manual, sometimes different granularity level are assigned to the cells (lv 0: lymphocytes, lv 1: B cells). This could be tackled by ontologies, where cell state would be assigned based on cumulative score of hierarchical representation of ontology rather than just a single-predicted probability score. In nutshell you use parent/children relationship when determine cell type, giving you higher confidence score #paper.
  • It is intuitive and elegant implementation termed hierarchical cross-entropy loss by the authors. The approach showcases moderate improvements in cell type prediction on quite a large datasets.
• DenAdel et al., 2026 showcased that large number of cells ≠ better performance. The authors conclude that ~20,000 cells might be the sweet spot for training dataset. What does this mean for the future? Maybe we don’t need Billion cell project, but rather better understanding on how to sample enough cell diversity and choosing the right model architecture with optimized parameters #paper.
• A short interview Why AI Hasn’t Cured Anything…Yet, According to Jennifer Doudna.
• Merchant, Guo and Viggiano et al., 2026 introduce a generative programming language Proto for synthetic biology. The main objective is to provide infrastructure for designing DNA, RNA and protein sequences with defined generative and optimizer modules. Basically a constrained generative system for quick assembly. As the authors point out in the discussion, the main challenge was to standardize multiple tools required for this project. Overall, I can imagine Proto being utilized by both wet and dry-labs combined with AI to facilitate faster discovery, removing the installation/infrastructure barrier shifting the focus and energy to research instead of figuring out why tools are not working on HPC. Lastly, I appreciate the disclosure transparency of AI usage for developing and writing the manuscript.
• QED Science, which estimates quality score for scientific papers took X/Twitter by storm with interesting debates and commentaries¹⁰.