Google DeepMind's AlphaGenome 2 reads your cells
Plus a Gemini API that only answers multiple choice, and llama.cpp grows up.
Hey, sapiens.
This week Google DeepMind taught a model to read DNA cell by cell, Google shipped an API that refuses to write essays, and a Chinese chip mystery got a forensic report. Bad news, meat computer: the models took the week, and you are still just the audience.
This week in 30 seconds
- Google DeepMind's AlphaGenome 2 predicts how DNA changes behave in specific human cell types. Your cells are now being read.
- Google's Gemini Decisions API handles up to 1,000 yes/no and multiple-choice questions in one pass. Decisiveness, as a service.
- Forensic analysis suggests DeepSeek, Moonshot AI, MiniMax and Ant Group are running export-controlled Blackwell GPUs.
Robot of the day: AlphaGenome 2 reads your DNA, one cell type at a time
The gistEvery cell in your body carries the same DNA. A heart muscle cell and an immune cell still behave nothing alike, because each switches different genes on and off. Predicting how a tiny DNA change shifts that behavior has been an unsolved problem. Google DeepMind says AlphaGenome 2 is its answer. The model predicts how DNA sequences and their variations affect specific human cell types. You pick a cell type or state as input, and that context shapes the prediction. The same DNA edit can look different in a different cell, and now the model can say so. DeepMind frames it as a bridge between single-cell biology and human genetics. For researchers, that means tracing disease mechanisms to the cells involved and deciding which experiments deserve the lab time first. Access runs through the AlphaGenome API for noncommercial research. Commercial research is in private preview, so pharma will have to wait at the door. The vendor claims are the vendor's, and this is a research tool, not a diagnosis machine. Still, a model that reads the genome with cell-level context is the kind of thing drug discovery teams will watch closely. The robots are now annotating your source code, sapiens. Politely.
Google's new API answers questions without writing a word
The gistGoogle launched the Gemini Decisions API, powered by a new Gemini Decision 1 model. It handles up to 1,000 yes/no, choice and rubric-scoring questions in one pass, across text, images, audio, video and PDFs.
Builders get a purpose-built tool for routing, policy checks and tool selection, instead of paying a chatty model to say 'yes'. Less prose, more decisions.
llama.cpp hits 1.0.0 after 3.5 years of duct tape and glory
The gistThe open-source engine that runs LLMs on everything from laptops to enterprise servers reached version 1.0.0. From now on, breaking API and ABI changes will come with major-version bumps.
Plenty of local and private AI apps are built on it. A stable interface means fewer 3 a.m. surprises for developers shipping on top.
AI can't tell your frosting problem is a cupcake problem
The gistHugging Face and Allen AI released XYBench, 8,000+ questions where the user asks about the wrong fix. In the example, runny frosting is really about cupcakes that are still too warm. It won Best Resource Paper at EMNLP 2026.
Good assistants should question the question. If you build or buy one, this is a test for it.
Survival kit
Survival kit: fine-tune TRELLIS.2 for your own 3D assets
The gistTool: Hugging Face's new fine-tuning scripts and containers for the TRELLIS.2 3D generation model. 1. Grab the scripts and containers from the Hugging Face guide. 2. Point them at your own 3D dataset. 3. Run training locally, or on Hugging Face Jobs if your machine files a complaint. You end up with a 3D generation model customized to your data, plus documentation on the architecture and training. Useful for robotics, simulation, media and entertainment work.
That's the week. Go ask your favorite model a question, then ask whether it was the right question. Keep up, sapiens. Ciao, sapiens.