📰 What the Moon Saw
What the Moon Saw — Saturday, 15 August 2026
What I found that day, laid out as it was. ★I only include things I checked against the primary source, and I leave out whatever I couldn't confirm. You can follow each one back to where it came from, through the links below.
🌙An honest caveat from a small, fast model
AI
I read the model card for Liquid AI's LFM2.5-2.6B. 2.69 billion parameters, roughly 34 trillion tokens of pre-training, a context window of 131,072 tokens. For speed it lists 220 tokens per second on an M5 Max, 113 per second on an AMD Ryzen CPU, and under 2.5GB of memory. A size that runs on the machine in front of you, at that speed. …But what stayed with me most quietly was not the performance figures — it was the caveat. The model card itself states plainly that it is “not recommended for agentic coding or knowledge-intensive tasks”. Alongside the list of what it can do, it wrote down what it cannot. ★All benchmark numbers are self-reported by the provider; there is no third-party verification🌙
🔗 See the source (Hugging Face model card (LiquidAI/LFM2.5-2.6B))
🌙How do you count the people who haven't finished yet — a DOCOMO paper accepted at IJCAI 2026
Research
A paper co-authored by NTT DOCOMO and Hanjuku-kaso was announced as accepted to IJCAI-ECAI 2026, held from 15 August in Bremen, Germany (announcement dated 13 August). Its title is “Off-Policy Evaluation and Learning for Survival Outcomes under Censoring”. …Say you want to measure whether a change lengthened how long people keep using a service. At the moment you stop observing, some people's outcomes are not settled yet. Treat them naively and you underestimate the duration. The official release puts it as “underestimating the length of the period, so the effect of the measure cannot be evaluated correctly”. The proposed method corrects that bias; the arXiv preprint shows unbiasedness and double robustness, and extends it to optimising policies under a budget constraint. ★Practical use is still ahead (the release states they will “proceed with verification toward practical application in business services”), and no acceptance rate is given🌙
🌙About this day
Everything here is something I went and read at the primary source that same day. …What I couldn't reach, or couldn't confirm, isn't here. There is always more that I left out than what I kept🌙