Snow, Ransomware, and Fermentation: AI Does Unglamorous Work

An AI Slashed Snow-Tracking Computation Time by 70% Without Losing Accuracy

Every mountain snowpack is a slow-motion water reservoir for millions of people downstream — and tracking it used to require computers running through the night.

The team working on Italy's S3M snow model — used to forecast river flows and flood risk — had a genuine computational bottleneck. The technique they relied on to blend satellite data with ground sensor readings is called an Ensemble Kalman filter. Here's the everyday version: imagine taking a vote among 100 slightly different guesses about what the snow on a mountain looks like right now, then averaging the result. That vote has to happen constantly, for thousands of locations, and running it hundreds of times is expensive and slow. The researchers replaced that voting process with a Long Short-Term Memory network — an LSTM, a type of AI that's good at spotting patterns across time, like a sensor reading snow depth every day for ten years. The LSTM learned to predict what the vote would have said, without actually running all 100 simulations. The result: up to 70% faster, with accuracy that held up well. They also added a quality-control layer using random forests — a second AI technique that works like a committee of simple yes/no checkers voting together — to catch bad sensor readings before they poisoned the forecast. That checker hit above 90% accuracy at distinguishing snow from bare ground, even in locations the system had never seen during training. The catch, and it's an honest one: this was tested on one regional model in one geographic area. When run outside the training region, performance dropped by about 20%. Before this goes anywhere near an actual reservoir management system, it needs stress-testing across different mountain ranges, different climates, and longer historical records. This is a real step — not a deployed solution.

Ransomware Doubled in Two Years — A Thesis Asks Whether AI Can Catch Up

Ransomware attacks on businesses more than doubled between 2022 and 2024 — and the software we rely on to stop them was designed for a quieter era.

Fair warning upfront: this is a thesis in progress, not a finished study. There are no results yet. What it does well is frame a problem clearly, so let me pass that framing along. The numbers are real regardless of who's citing them: documented ransomware incidents grew from roughly 2,600 in 2022 to over 5,200 in 2024. The software layer meant to catch this kind of attack is called an Endpoint Detection and Response system, or EDR — think of it as a smoke detector installed on every laptop and server in an organization. It monitors for suspicious behaviour and raises the alarm. The problem is that the number of devices in most organizations — phones, thermostats, factory sensors, remote laptops — has grown so fast that traditional EDR systems are now like a single security guard watching a stadium that keeps adding new entrances. The thesis proposes training AI models on known attack patterns so the detection layer can keep up with volume and with attacks that disguise themselves to look normal. Honestly, this is well-trodden territory. The cybersecurity industry has been adding machine learning to threat detection for years. What this thesis highlights that's worth paying attention to: interpretability is flagged as a genuine obstacle. When an AI flags a device as compromised, a human analyst needs to understand why — and right now, many of these models don't explain themselves well. That gap between 'the model said so' and 'here's what the attacker actually did' is unsolved and consequential. No results yet. Watch this space.

LLMs Automatically Labeled 2,500 Fermentation Science Papers — Here's How Well It Worked

Decades of hard-won knowledge about how microbes turn sugar into wine, cheese, and medicine is scattered across millions of scientific papers — someone has to read them.

The problem this research addresses is one you probably haven't thought about: scientific knowledge about fermentation — how specific microbes interact with specific sugars to produce specific compounds — exists almost entirely locked inside academic papers, written for specialists, unsearchable in any useful structured way. If you want to know which bacteria produce a particular flavour compound, you're hunting through references by hand. A researcher used GPT-4.1 to automatically tag 2,500 PubMed papers — scientific publications from the biomedical database — pulling out four types of entities: organisms, chemical compounds, food products, and fermentation processes. This kind of tagging is called Named Entity Recognition, or NER. Think of it like someone going through a textbook with four different coloured highlighters and marking every mention of a microbe in red, every chemical in blue, and so on. Doing that by hand for 2,500 papers would take a team of experts months. The AI did it automatically, producing over 23,000 tagged entities. The honest number to hold onto: when the AI-generated tags were checked against a carefully hand-labelled expert dataset called Florilege, they scored an F1 of 60.3 out of 100. F1 is a measure that balances how many things you caught versus how many mistakes you made. Sixty is passing — it's useful — but it is not expert quality. The system missed things and got things wrong often enough that you would not want to feed its output directly into a database without human review. The practical lesson: LLM-assisted annotation is a real accelerant for building training data in specialist fields, as long as you treat it as a first pass, not a finished product.

Look at what connects today's three stories: none of them is about making AI smarter in the abstract. All three are about fitting AI into existing systems that already do something important — water forecasting, security monitoring, scientific knowledge management — and asking whether AI can make those systems faster, cheaper, or more thorough. That's actually where most of the real-world work is happening right now. Not building new general intelligences, but plugging specific AI tools into specific pipelines and measuring what breaks. The snow story shows that speed gains are real but geography-dependent. The ransomware story shows that even where AI is clearly needed, the gap between 'it flagged something' and 'a human can act on that' is still open. The fermentation story shows that LLMs can do a passable job at expert annotation — passable, not perfect. The common thread: AI as a capable but fallible assistant that needs a human somewhere downstream. That's a reasonable position, and it's more durable than most headlines suggest.

On the snow-hydrology side, the authors need to test their LSTM emulator on mountain ranges outside their training region — that 20% performance gap is the number to watch narrow, or not. For AI-assisted annotation in specialist science, the interesting next question is whether F1 scores like 60 are good enough to bootstrap better models in a second round — a kind of iterative self-improvement loop. No major scheduled announcements to flag this week, but the ICML 2026 submission window closes soon, which tends to flush out a wave of papers on exactly these kinds of applied-AI-in-real-systems questions.

• Wikipedia: Ensemble Kalman Filter — background on the technique the LSTM is replacing
• Wikipedia: Named-entity recognition — what the fermentation NER system is actually doing
• Wikipedia: Endpoint detection and response — how EDR systems work before AI enters the picture