Want to learn how to USE AI technology to make money and/or your life easier? Join our FREE AI community here: https://www.skool.com/ai-with-apex/about
Today in AI: Google’s Phone-Powered Cloud, Smarter Skin Search, and the New Benchmarking Stack
Today’s AI story is less about bigger models and more about the systems around them. The interesting shift is happening in infrastructure, user experience, and the tooling used to judge model quality.
TL;DR
- Google is backing a UC San Diego project that repurposes retired Pixel phone motherboards into a planned 2,000-device compute cluster.
- The project is aimed at low-cost, lower-carbon computing for lighter workloads rather than replacing conventional AI datacenters.
- Google also published research showing an AI-assisted skin-condition tool helped users identify possible conditions better than standard search alone.
- In that study, condition-name accuracy rose from 8% with standard search to 23% with AI assistance, while a dermatologist-guided positive-control setup reached 36%.
- Ai2’s OLMo-Eval repository has been archived and replaced by OLMES, highlighting how open-model evaluation is becoming core infrastructure.
Google backs a low-carbon compute platform built from retired phones
What happened
Google Research highlighted a UC San Diego effort to build a low-carbon computing platform from retired Pixel phone motherboards. The team plans to deploy a 2,000-phone cluster for student and research workloads, turning old consumer hardware into a second-life compute resource.
Why it matters
This is a practical sustainability story inside AI infrastructure. Instead of focusing only on cleaner electricity, the project targets embodied carbon by reusing existing hardware, especially the motherboard, which Google says accounts for roughly 50% of a phone’s embodied carbon in its internal assessments.
Key details
- The project reuses smartphone motherboards rather than full phones, removing parts like batteries, displays, cameras, and chassis that are not useful in a datacenter setting.
- UC San Diego plans a 2,000 Pixel smartphone deployment for coursework and research computing.
- Google says benchmark results suggest 25–50 phones can equal a modern server on some workloads, which it presents as workload-dependent rather than universal.
- The system replaces Android with a general-purpose Linux distribution and uses containers and Kubernetes for orchestration.
- Target workloads include small cloud instances, grading backends, Jupyter notebooks, and other jobs that fit within a phone’s compute and memory limits.
Source links
https://research.google/blog/a-low-carbon-computing-platform-from-your-retired-phones/
Google publishes research on AI tools for understanding skin conditions
What happened
Google Research published a summary of new JAMA Dermatology research on a prototype AI tool designed to help consumers better understand possible skin conditions. The study compared standard search, an AI-assisted interface, and a positive-control version populated with dermatologist-provided differentials.
Why it matters
The strongest takeaway is not that AI can replace clinical judgment. It is that structured, guided presentation may help users interpret health information better than open-ended search results alone, which makes this a user-experience story as much as a model story.
Key details
- The AI interface returned a carousel of 3 to 7 matching conditions with images and explanatory information.
- More than 62% of participants using the AI tool were willing to try naming the condition, compared with 41% in the standard-search control group.
- Condition-name accuracy was 23% in the AI arm versus 8% in the unassisted control arm.
- The study’s “Wizard of Oz” positive-control arm reached 36% accuracy using dermatologist-provided ground-truth differentials.
- Google also reports higher user confidence and satisfaction with the AI-assisted format.
Source links
https://research.google/blog/research-into-how-ai-can-help-users-understand-skin-conditions/
Ai2’s OLMo-Eval gives way to OLMES as open-model evaluation matures
What happened
Ai2’s OLMo-Eval repository, which supported evaluation workflows for open language models, has now been archived. The repository states that it has been superseded by OLMES, the Open Language Model Evaluation System.
Why it matters
Evaluation is becoming a major layer of AI infrastructure in its own right. As open models multiply and benchmark claims become more central to model launches, reproducible scoring pipelines and shared evaluation tooling matter more than ever.
Key details
- The OLMo-Eval GitHub repository is marked as archived.
- The repository explicitly says it has been superseded by OLMES.
- Ai2’s broader OLMo ecosystem documentation continues to reference evaluation tooling as part of the stack around its open models.
- Recent Hugging Face model documentation for OLMo models still points back to OLMo-Eval-related ecosystem materials, underlining how central evaluation remains to open-model releases.
Source links
https://github.com/allenai/OLMo-Eval
https://github.com/allenai/olmo-core
https://huggingface.co/allenai/Olmo-3-7B-Instruct
The common thread across all three stories is that AI is becoming more operational. Hardware reuse, better information design, and stronger evaluation pipelines are less flashy than raw model demos, but they are increasingly where real progress shows up.
---
Want to learn how to USE AI technology to make money and/or your life easier? Join our FREE AI community here: https://www.skool.com/ai-with-apex/about