#HackathonTakeaways Part 5: Multi-Token Prediction's 3x speedup with no quality loss maps perfectly to SolarHive: less power per query on solar-run edge hardware, and reclaimed time for deeper grounding. Here's why it fits: github.com/youshen-lim/...

#HackathonTakeaways Part 4: One rule shaped my SolarHive build: no number in an answer comes from the model. Each figure is grounded in a live API call, because a language model gives plausible numbers, not correct ones. github.com/youshen-lim/...

#HackathonTakeaways: I built SolarHive for the Gemma 4 Good Hackathon (Google DeepMind x Kaggle) during a family visit to Singapore. An open layer helping a community plan and use solar energy. Learnings in repo: github.com/youshen-lim/...

#HackathonTakeaways Part 2: What I was most careful to ground in SolarHive: when my fine-tuned Gemma 4 models call a tool and when they don't.
I built training data around the When2Call paper. My 26B model scored 3/3 on held-out probes, the smaller 2/3.
github.com/youshen-lim/...

#HackathonTakeaways: I built SolarHive for the Gemma 4 Good Hackathon (@googledeepmind.zpravobot.news.ap.brid.gy x @kaggle.com), even during a family visit to Singapore. An open intelligence layer that helps a community use solar production. Sharing what I built (Dataset and models).