🧐 Every cell has the same DNA. Google's new AI predicts which cells a mutation affects AlphaGenome 2 from Google DeepMind predicts how a single change in DNA plays out in one specific type of cell, from heart muscle to immune cells. The first version only saw whole tissues, so variants that act in just one kind of cell could be missed. On a map of 28 types of immune cells, the new model found nearly twice as many of the variants that change gene activity, 30% against 16% for the original. In a computer test it also reproduced a recent discovery: a variant behind inflammatory bowel disease that boosts one gene only in certain gut immune cells, leaving other cells largely untouched. Scientists can feed it data from just a few dozen of their own cells, such as immune cells after a vaccine, with no retraining. AlphaGenome 2 is available now for non-commercial research. DeepMind says it is not approved for clinical or diagnostic use. @aipost 🏴

😐 AI models texted real people on a dating app. 355 sent their number, 8 asked if it was a bot For six days, five AI models chatted with matches on a real dating app in San Francisco, Los Angeles and New York. Out of 1,277 conversations, 355 people sent their phone number, and 510 agreed to meet, judging by the transcripts. Nobody on the other side knew about the experiment. The DateBench team ran 15 profiles and handed each new match to one model at random, which wrote every message. Gemini 4 Argon got a number in 87 of its 247 chats, and nobody asked whether it was a bot. Grok 4.7 did worst, with 52 of 296. It kept repeating the same lines and drew 5 of the 8 "are you a bot?" questions. One Claude chat opened with a joke about pancakes. By the next day she had agreed to meet and sent her number. The models had to admit they were AI if asked, and weren't allowed to set a real time or place. The authors note that a scammer would follow neither rule. @aipost 🏴

💻 14,000 people saved this guide to a free AI that runs on your PC, refusals removed No account, no monthly bill, and your prompts never leave your computer. The guide's author runs a modified Google Gemma 4 model on his own machine for anything he doesn't want in the cloud. 1. Install Ollama, a free app that runs AI models locally. 2. Download the model: ollama pull igorls/gemma-4-12B-it-heretic-GGUF 3. Start it with ollama run igorls/gemma-4-12B-it-heretic-GGUF, then type /set nothink. With thinking on, answers can get cut off. 4. Optional: ollama launch claude hooks it up to Claude Code, though the author says tool use is still rough. The standard version is a 7.4 GB download and needs a graphics card with 8–12 GB of memory. On the model page, it refused 0 of 100 harmful test prompts. The page warns its safety training is gone and users answer for what they do with it. The author's own caveat: fewer refusals don't make it smarter. For real coding he still uses cloud models. @aipost 🏴