Pinned Tweet
Been seeing a lot of videos and posts of tech influencers talking about how they're tired of the speed of coding, LLMs, new AI models...
It seems to be crashing all at once, but I think that's a good thing not a bad one
1/x
Excellent example of cross dataset application
Here's a new model to predict where to put EV chargers in a city.
You can think of the standard way to site chargers as being basically an optimisation problem. It involves minimising building costs, minimise how far drivers have to go, maybe check the grid can cope. What those models rarely look at is the neighbourhood around the stations already built: busy commercial land, housing, a main road carrying heavy traffic. If existing stations share a recognisable kind of surrounding, that pattern is pretty much a free guide to where the next one should go.
Researchers recently built a model to learn that pattern. They tested it on Bryan–College Station in Texas, a 185 km² college town with 24 public charging stations.
First they turned the city into a grid of overlapping circles. Every 100 metres along the road network they drew a circle with a 500-metre radius, giving 1,971 in total, and filled each one with 22 measurements from public data: registered EVs, how much land is commercial, residential or other, daytime and night-time population, and the length and daily traffic of six classes of road. Only 126 circles, roughly one in sixteen, contain a charger.
The model then works in two steps.
Step 1: It squeezes those 22 numbers into a short summary of what each place is like.
Step2: It compares every circle with its neighbours, so a location is judged partly by what surrounds it.
On held-out circles, it found 97% of the places that really do have a charger, and about four in five of the places it flagged actually had one. That gives an F1 score (a single number balancing those two) of 0.87. The best conventional machine learning baseline, CatBoost, managed 0.74, and the same neighbour-comparison model without the compression step reached 0.71.
Obviously, this is one small US college town. The model only knows where stations are, with no data on how busy they are, so it assumes the current network is well placed and will reproduce its blind spots. Neighbouring circles overlap heavily, which makes the test easier than it sounds. Costs, land availability and grid capacity sit entirely outside the model.
Still, the starting point holds up. Every city that has built chargers already holds a rough template for the next ones, and it's sitting in the streets around the ones it has.
link to full article: arxiv.org/abs/2609.30417
All countries now have patch notes to see new updates, and a history to see what evolves over time daedalmap.com/geometry/canad…
Download packs are now Geolibre and QGIS compatible, get the data and join it easily!
downloads.daedalmap.com/
@giswqs
This is why if you want to trust AI you need to give it access to real data, instead of relying on what it knows, its training data, or anything behind a black box.
Trust but verify, always show your sources or the data is meaningless!
Excellent visualizations
Every river in North America, every day for two years. The combined flow to the sea sinks as low as 53,000 m³/s in winter and swells past 290,000 by late June, more than quadrupling each spring. Data: GloFAS (Copernicus, ECMWF)
#rivers #dataviz #hydrology
Lots of hard to find data here
Been running my blog for 13 years, sharing the hardest to find datasets for free.
Enjoy!
revolutionarygis.wordpress.c…
DaedalMap retweeted
This Demographics & Housing WebGIS, made by Mohammad VahidiBorji, caught my attention because it shows something I believe is becoming increasingly important in GIS:
Turning statistics into spatial intelligence.
DaedalMap retweeted
Here’s what I would do in the next 30 days if I wanted to win a fully funded PhD before the end of 2026.
Days 1–3: Build your research profile.
Write down your 3 strongest research experiences. For each one, record: research question, dataset/material, methodology, your specific contribution, one problem you encountered, how you solved it, and the main result. Then write 3 research areas you could realistically pursue for a PhD. Your output: a one-page research profile.
Days 4–7: Find 30 potential labs.
Search university PhD vacancies, lab pages, funded projects, research group websites and recent papers. Do not shortlist universities first. Shortlist labs and professors. For every target, record: professor, research topic, current project, funding status, required methods, deadline and application route. Your output: a spreadsheet of 30 potential labs.
Days 8–10: Read before you apply.
Take your best 10 labs. Read 2 recent papers from each professor. Do not read them passively. For every paper, write: research question, methodology, key finding, limitation, and what question remains unanswered. Then compare those questions with your own research experience.
Your output: 10 research-match notes.
Days 11–13: Build your match matrix.
For each of those 10 labs, create five columns:
Lab needs → Required skill → My evidence → Research connection → Possible contribution
For example:
Lab needs: computational modelling
Your evidence: developed and evaluated predictive models during MSc research
Research connection: same modelling framework applied to a different dataset
Possible contribution: investigate whether the approach transfers under different conditions
If you cannot fill these columns with real evidence, remove the lab from your priority list.
Days 14–17: Rebuild your CV.
Stop describing activities. Rewrite your strongest research experiences around problem → method → decision → result.
Instead of:
“Used Python for data analysis.”
Write:
“Developed a Python-based pipeline to analyse X dataset, compared two modelling approaches, identified performance limitations and modified the preprocessing procedure to improve prediction accuracy.”
Your output: one research-focused master CV.
Days 18–20: Write targeted cover letters.
Choose your 5 strongest opportunities. For each one, answer four questions:
Why this research problem?
Why this lab?
What evidence shows I can contribute?
What could I investigate next?
Days 21–23: Prepare professor outreach.
Choose the 10 professors with the strongest research match. For each, identify one specific paper, problem, limitation or methodological connection worth mentioning.
Then write a short email built around:
Their research → your relevant evidence → the research connection → your question about an opportunity.
Your goal is not to tell them your life story. Give them a reason to continue the conversation.
Days 24–26: Submit the strongest applications.
Apply to the opportunities where you have the strongest evidence of research fit. Before submitting, check every document against the advertisement. Every major required skill should either have evidence in your CV or a clear explanation in your cover letter.
Days 27–30: Build the interview pipeline.
For every application and professor contact, prepare answers to:
What exactly did you research?
Why did you choose that method?
What went wrong?
What did you change?
What did you learn?
What would you investigate next?
Why this lab?
If you want to work with me on writing strong scholarship applications using my 30-day strategy; DM me “30 DAYS”
Cleaning up the MCP tools for entering the DaedalMap loc_id universe, and also for getting data within.
Things have gotten a bit messy, but now the boundaries are clear and easier for others to use!
Start here: daedalmap.com/agents
Lots of other good tools here!
I’ve been building Awesome Autonomous GeoAI, a gift to my PhD journey and the GIScience community. 🌍🤖 Today: 16⭐
A curated guide to papers, tools, courses, conferences, and opportunities—built with care.
github.com/AutoGeoAI4Sci/awe…
#GeoAI #GIScience #AI4Science
International teamwork
The Fourteenth International Conference on Geographic Information Science #GIScience2027 will be held in Shanghai, China, on 18-22 October 2027. The conference theme is "GeoAI and Multi-Sensing for a Sustainable Future". Please help spread the words! giscience2027.org
A billion more buildings than the next largest dataset, with higher resolution (3m vs 90m grid)
𝐎𝐩𝐞𝐧𝐒𝐭𝐫𝐞𝐞𝐭𝐌𝐚𝐩 𝐯𝐬 𝐆𝐥𝐨𝐛𝐚𝐥 𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐀𝐭𝐥𝐚𝐬 𝐄𝐱𝐩𝐥𝐚𝐢𝐧𝐞𝐝
A regular reminder to keep your building data in check - from OSM maps to the GBA
Tutorial: lnkd.in/ge3tFSqD
Paper: lnkd.in/dYwagBc4
𝐌𝐨𝐫𝐞: thenewscienceofmaps.com/
Im happy coding away, building out DaedalMap, organizing data, preparing for the big moves later, and to me the moving industry doesn't matter.
It removed the moat of technical cofounders, which just lowers barrier to entry, which I think is a win for creativity
9/x

















