I find myself wanting to read more, and I’ve found that reading paper is more pleasurable. I was intrigued by the Offprint service that launched recently. The idea is a magazine made of links that you send in, printed monthly, and mailed to you.

I received my first edition a few days ago. It is as nice as I imagined, and I could see this becoming a real thing. A few snags though. One, it’s slower to receive than I imagined. I didn’t receive my August issue until halfway through September. By the time I received the magazine I had forgotten about several of the articles. Two, I don’t always have links to send. Sometimes I just have 2 or 3, sometimes 10. Offprint has a static number of 90 something pages, with extra pages being “notes”. I’d rather write in the margins, and having 20 pages of notes feels a bit useless.

While I really liked Offprint, I just didn’t like the experience enough to continue. My wife, always being the practical person, said I should just print it out at the office. So I made an app, of course. I’m calling it Galley. While I really liked Offprint, I just didn’t like the experience enough to continue. Paste in links, collect them into “editions” and print them out on A4 or letter sized paper.

It’s simple for now, and Mac only. But I’m open to bug reports and feature requests, if any. https://galley.robertritz.com/

Just got my first issue of Offprint. The idea is to save articles that are too long to read on a phone or laptop. You get your customized magazine based on links you send to the service.

I completely forgot about these articles and now I’m pretty excited to read them! So I guess it’s working?

For years I heard about how some cultures are more “collectivistic” than others. A collectivistic culture is a culture that prioritizes the needs, goals, and well-being of the group over individual desires. Oftentimes I would hear about East Asian cultures being more collectivistic, with Western society being more individualistic.

Directionally this always seemed correct, but when you actually try to measure this, you can see that things start to fall apart. Let’s give a few examples:

• “I put the group before myself.”
• “I enjoy cooperating with other people.”
• “Collectivists care more about other people.”
• “Collectivists value harmony.”

These statements sound good on the surface, but are more about “warm fuzzy feelings” and aren’t actually very specific. Are we talking about my family, my coworkers, my parents, my country? Which “others” do collectivists care more about?

Historically surveys of collectivism relied on these types of questions to measure collectivism, and the result has been a confusing mix of answers. The concept of collectivism makes sense, but the way it has been measured never has.

A recent paper published in Nature Human Behavior that I had the privilege of helping with attempts to refine this understanding of societies. This new concept is called responsibilism. My contribution to the project was collecting data from Mongolians (n=194).

Responsibilism emphasizes responsibilities in close relationships. The Responsibilism Collaboration Team (snazzy name right?) collected survey data from 12,247 people in 100 cultures around the world. Instead of asking vague questions about “the group” or “others,” the survey used specific relationships and concrete situations.

Participants were asked to consider questions such as:

• Should we keep our aging parents with us at home?
• How much should we trust or help people we barely know?
• How willing are we to form new relationships with strangers?

The resulting data helps us measure feelings towards duties and responsibilities in close relationships versus generalized warmth toward everyone. A simplified idea is who they feel responsible for. I may feel more responsibility towards my parents, and comparatively less responsibility to a stranger I meet on the street.

Across all cultures surveyed scores were given from 0-100.

Put simply a higher responsibilism score means stronger felt duties and responsibilities toward people in close relationships. In addition, a higher score is correlated with weaker felt duties and responsibilities at greater social distances.

This inverse relationship is shown nicely with these two examples:

Lying for a friend

Returning a lost wallet to a stranger

For societies with higher responsiblism, people will do more for those closer to them. Inversely, they will do less for those farther away. Unfortunately Mongolia isn’t listed in the charts above because these questions were asked as part of different studies. The answers were plotted against the responsiblism data, and the correlations are quite strong. This allows us to make some predictions about where Mongolia would end up here.

These aren’t exact number obviously, but the trend is quite interesting. I built an explorer to view the data for Mongolia and compare it to other countries. Mongolia is an interesting country because it has elements of both Western “individualism” and Asian “collectivism”.

Here are a few interesting results I found, and where Mongolia differs from the United States.

Here a “good” friend can be defined as someone supporting you or as someone who notes that you maybe didn’t do something well.

Americans are much less trusting of random people near ATMs.

Awards are perhaps less common in the US, and therefore less susceptible to the awardflation that Mongolia suffers from.

If you want to see all the questions and how Mongolia compares, you can access the full explorer here: robertritz.com/responsibilism

However, I think the takeaway is that this research gives us a much more precise way to think about something that “collectivism” only described vaguely. The real question is not whether a society values the group over the individual. It is which people we feel responsible for, how strong those obligations are, and how quickly those obligations weaken as social distance increases.

Fun new dataset on data.mn! Parliament attendance in Mongolia is a perennial topic, and the Parliament of Mongolia now has an API and dashboard to check on the attendance of MPs. I collected the raw data from 2024 to now and displayed it in a nice scatter plot.

You can view the new dataset here: data.mn/en/data/m…

There are many reasons for MPs to be late or absent, including intentionally to prevent a quorum as has happened recently. Nevertheless it’s an interesting dataset.

This is the time of year graduates start coming over and asking advice, especially those who studied data science at AUM. I wanted to make a post to give these students my top points of advice in writing.

1. Show off your work

Showing a portfolio of work, especially printed out, is a great way to impress an HR person, and the hiring manager they hand it off to. Add charts, explanation, and aim for a semi technical explanation without getting too technical, especially with machine learning projects.

Often hiring managers are looking for experience with specific domains (housing, finance, even cars) so sprinkle in a mixture of different topics in your portfolio. When discussing these projects try to gauge the technical competency of the person. Sometimes they might know a LOT about a domain, but might not understand the data sciency stuff. Other times they might know both. Assume they know their business and give a non-technical explanation of the project. Who, what, why, where, when are all good things. And don’t forget “so what?”

2. Don’t work somewhere just for the experience, especially audit

This is super controversial, because I actually think you should work in your early career with a focus on experience. But that shouldn’t be the only goal. You should also find a workplace that gives you room for advancement, treats you with respect, and gives you the proper tools.

I say this especially about audit. Audit in Mongolia is a race to the bottom project, often with fixed price contracts. I consider it unethical to have a fixed price audit of any kind. With fixed prices, audit companies will be pressured to provide a completed audit within a specified budget. What if there is more work? You guessed it, unpaid overtime and tight deadlines with questionable quality.

3. Your starting salary matters

Many former students don’t put a lot of emphasis on starting salary, as they simply want to work somewhere to increase their experience. This is a great goal, but it also creates an opportunity for companies to take advantage of you. When you leave this first job, many companies will base your new salary on your old one. Raises also happen usually as percentages.

If you can get 15-20% more for your first job, you will probably end up with a higher salary in 5-7 years. Don’t short change yourself.

4. Ensure the company gives you access to AI tools

If your company does not have a corporate Codex, Claude Code, or even Muse account for you, I would think twice about working there. Today a single experienced analyst using AI tools can do the job of 4-5 employees from 5 years ago. These tools can securely handle customer and company data (remember the CIA is one of Anthropic’s biggest clients). If it’s good enough for them, it’s good enough for your company. Zero data retention means the AI company never saves your prompts or the results.

Any company in Mongolia today hiring analysts, data scientists, or coders without these AI tools deserves to lose all of the best and brightest Mongolian minds. They simply don’t deserve you.

5. Turn overwork into your advantage

If you boss keeps coming to you to get more work done, this is an opportunity, not a problem. If you help them out now, they owe you one. Take a bit of pain for a few weeks to help them out. If they don’t pay you back, never help them ever again with anything. Business works with mutual trust. If your coworkers work within this system, they will value you, and you will value them. For those that choose to take advantage, remember them and never help them again.

That’s all I can think of for now. I’ll add to this later if I can think of more.

If foreign companies want to pay protection money, maybe

Back in July, I wrote about Mongolia’s ambition to create data centers. There, I wrote mostly about current PM Uchral’s comments about Mongolia’s various advantages with regard to AI:

• Mongolia is a non-aligned nation
• Data sovereignty
• An abundance of land in Mongolia
• Mongolia is land-“linked” (i.e., close to China)

Since I wrote that article, a new draft law has been submitted to Parliament titled simply “About Data” (Өгөгдлийн тухай). I originally wanted to write an article about the financial viability of a 100 MW AI data center in Mongolia. I went deep into costs, electricity prices, fiber capacity, and more. Then I read the draft law. I ended up with more questions than answers, and no clear way to get answers to those questions.

So rather than talk about a hypothetical data center, I wanted to comment on several provisions of this draft law. I’m relatively positive about the direction this law takes data policy in Mongolia, but I also feel it is important to give my perspective on where things aren’t as good.

Encouraging foreign investment…for a price

This draft law takes a very notable step toward creating protections for foreign data centers built on Mongolian soil.

13.2. Data and information of foreign entities placed in a foreign data center shall be inviolable, and except as provided for in international agreements, Mongolian authorities shall be prohibited from accessing, physically or electronically inspecting, seizing, or suspending the operations of the data center.

This is a pretty remarkable set of protections for foreign data centers. In addition, and perhaps with a nod to my previous post, Mongolia is attempting to address the issue that data will flow across the borders of Russia and/or China.

13.6. In order to ensure the inviolability of foreign data center transit traffic passing through the territory of neighboring countries, special negotiations will be held with neighboring countries and an international agreement will be concluded with Mongolia.

13.8. In order to ensure the safety and reliability of the data center transmission network and to prevent potential technical risks, Mongolia shall organize at least three geographically and technically independent and unrelated international Internet gateways.

If AI tokens are a new type of “electricity,” having those tokens pass through a foreign power might not make a lot of sense. Energy independence is something most countries put considerable effort into. With AI, being “land-linked” isn’t currently seen as an asset, but rather as a liability.

These protections come at a price. In fact, they have a literal price tag. Emphasis is my own:

13.12. A benefit-sharing agreement will be concluded with a foreign entity that has requested to operate data center services in Mongolia, and the agreement will include the following terms:

13.12.1. In the event that the special protection and inviolability of the data center specified in Articles 13.2 and 13.5 of this law is guaranteed, the payment for the inviolability guarantee shall be deposited in the state budget at the established rate;

13.12.2. In the case of providing data transmission and export services using low-orbit or other satellite networks, regardless of the infrastructure of the Mongolian Communications Authority, special service fees to ensure network spatial independence shall be paid to the state budget based on the amount of exported data;

13.12.3. If you create an independent energy source using renewable energy sources to power your own data center, the energy independence fee shall be calculated based on the amount of energy generated and deposited into the state budget;

13.12.4. If the data center is powered by natural resources and generates its own independent energy source, the energy independence fee shall be calculated based on the amount of energy generated and the greenhouse gas emission offset fee shall be deposited in the respective state budget until the carbon emissions are reduced to net zero;

13.12.5. Create conditions for Mongolian government organizations, research and educational institutions to use a certain percentage of the data center’s computing, backup and storage equipment free of charge;

That’s a lot of fees! The biggest problem isn’t the fees themselves. The real problem is that the amounts are not set in the law and will instead be determined by the government later. This creates the possibility that the methodology or rates could also change later.

Foreign investors require consistency and trust before they will be willing to invest. We saw what happened with the Oyu Tolgoi negotiations, which were repeatedly mired in disputes and accusations of corruption. A better approach would be a static fee arrangement that is publicly defined and cannot be changed for the duration of the project.

I am not aware of another country that imposes this particular set of fees specifically on foreign data centers. It is also notable that domestically owned data centers would not have the same fees.

Tokens as electricity, and national security

I did some back-of-the-napkin math on building a 100 MW AI data center in Mongolia. My estimate puts the investment at around $6.7 billion, including dedicated power generation. Any foreign entity investing this much money in Mongolia will want serious stability, very much unlike what has been seen in mining in the past. The capital expenditure breakdown is:

• NVIDIA GB300 NVL72 systems: $3.17 billion
• Networking and storage: $0.67 billion
• Buildings, site, cooling, electrical, and backup power: $1.3 billion
• Delivery, spares, and IT contingency: $0.50 billion
• Dedicated wind, solar, and batteries: $1.06 billion

This would be a construction project with a very high initial capital cost, but a relatively modest land requirement of perhaps 30–60 hectares. It’s easy to see why Mongolia wants to attract this type of business. It is also easy to see why these businesses, in my opinion, will most likely stay away from anything approaching this scale in Mongolia.

If LLM tokens are such an important resource, akin to electricity or even labor, nation-states may prefer to keep this generation capability within their own borders. Perhaps Mongolia could be a destination for training open-source models or running inference on them. For this to happen, a clear value proposition will have to be made, with transparent taxes and fees that investors can rely on without fear of later changes or renegotiation.

Last year on a whim I bought data.mn. I’ve been working with Mongolian data for nearly 10 years now, and it seemed like an obvious thing to get. The owner had been sitting on it for years and didn’t do anything productive with it. I wanted to build a single portal with all things data on it about Mongolia.

The government has a few open data portals already, but they are generally not very easy to use. Search is somewhat poor, there are no built in visuals, or they don’t provide a secure connection (see below):

These portals are:

• opendata.gov.mn - general open data portal based on CKAN, currently SSL certificate expired
• opendata.burtgel.gov.mn - legal entity registration and things like that

So I wanted to make something dead simple, has great search, and work with AI agents. What resulted is a blazingly fast website that just gives you the data with a chart, CSV file, or Excel file. Excel file exports show a shortened form often used as a statistical table. CSV exports are “long form” which is ideal for creating visualizations in PowerBI or other visualization tools.

For AI agents, if the server detects if an incoming user is a person or a bot (we welcome bots!). If it is a bot, it redirects to a text only page which gives a description of the data and a the data itself in text form. This allows you to ask a question from ChatGPT or Claude about Mongolian data and it can use real, up to date data to answer you. Without this AI chatbots usually default to reports from Asian Development Bank or other international organizations with usually out of date data.

Data.mn currently has 120 published datasets mostly from Mongolian government sources. I hope to add more private datasets to the platform, something that is relatively rare unfortunately. If you have any ideas let me know!

A former student visited the office last week. He studied in our business + data science bachelors program and has worked as a data analyst focusing on finance since he graduated. He complained that his worked gave him a lot of ad-hoc analysis tasks and he was working weekends and holidays to handle the extra load.

I was totally shocked. I asked him if he was allowed to use AI tools like Codex or Claude Code. He said, “Do you mean ChatGPT?” Queue my head exploding. He was still stuck copy pasting code from ChatGPT into his IDE to get work done.

I then proceeded to show him Codex, how to direct an agent to accomplish a task, how skills worked, etc. He was totally floored. He truly had no idea. He was so heads down working for the past two years that he didn’t know about any of these things. No one at his company was using agentic coding tools at all, but his company did pay for ChatGPT subscriptions for technical employees.

Two reactions:

1. We are so early.
2. You must be a lifelong learner to do well in the next 10 years.

Throughout my time teaching this student I told them they must always learn, because data science (and now AI) doesn’t stand still. In his defense he was learning every day, just not the latest developer tooling. Instead he was deepening his knowledge of finance so he could be an asset to his company, a laudable goal. Unfortunately for him, he was still being buried under ad hoc requests and wasn’t learning the skills to get himself unburied.

FOMO isn’t great, but neither is suffering at work when it could be easier. We need to keep learning and improving without becoming obsessed with it. Always be learning.

For the past week I have enabled Computer History from Codex on my main computer. This utility records interactions on my computer including where I click, what I type, and in general what I’m doing. Today it prompted me to summarize my week. It gave a pretty decent summary of my week, with this as the kicker at the end:

Unscientific final diagnosis: high-agency operator, unusually low tolerance for vague interfaces, and dangerously susceptible to “while Codex is thinking…”

I’ll take the “high agency operator” flattery, and I’ll admit I am dangerously susceptible to doing about 4-5 things at at time with Codex. I’ll keep it on for now, although I don’t really yet see any use for it.

Heard recently on a podcast about how an ML model was created to predict how the neurons in a fruit fly were connected. The main finding was that the neurons connections could only be predicted reliably with a hyperbolic geometry. I thought it would be interesting to test, and I asked ChatGPT to research if this already existed. It turned up some similar things but not the exact idea.

My core question was:

Could hyperbolic proximity help a learning network choose which dormant connections to activate?

I’m not a deep learning research at all, but I do know how they work and have trained models from scratch and fine tuned them before. So I was familiar enough to know what to ask for.

It seemed like an easy enough thing to test, especially with the help of AI. The end result wasn’t promising (which is probably why it wasn’t published anywhere). I still feel there is some merit in the concept of rewiring, and I know researchers are looking into it. In any event it was a fun side quest to think about.

I’m almost certain someone has tried this idea before. But in a pretty wide search I couldn’t find anything exactly like what I wanted. If more people posted their bad results, it would probably save people (and AI agents) a lot of time!

I’ll let Codex/GPT5.6 Sol explain the experiment and results in its own words.

What We Tested

This summary was written by OpenAI’s Codex. Robert proposed the idea; I built, ran, and audited the experiment.

The idea was to let a sparse neural network change its wiring while it learned. Instead of keeping the same connections throughout training, the model periodically removed weak connections and replaced them with promising inactive ones.

The question was how to choose those replacements. A standard method uses gradients, which estimate which connections would most reduce the model’s error. Robert wondered whether Hyperbolic geometry might do better. Hyperbolic space represents hierarchical structures efficiently, and language contains plenty of hierarchy. Perhaps learned Hyperbolic positions could reveal useful connections that gradients alone missed.

The experiment

I trained a 6.90-million-parameter language model on TinyStories. Only 25% of the connection blocks in its feed-forward layers were active at once. The main comparison included:

• Gradient-only rewiring
• Gradient plus Euclidean geometry
• Gradient plus Hyperbolic geometry

Within each seed, the models shared identical starting weights, sparse masks, training data, and evaluation schedules. Each model trained for 100,663,296 tokens. The full confirmation included five treatments across three seeds, or 15 trained models.

Lower validation loss means better prediction on text the model did not train on.

Method	Mean validation loss
Gradient rewiring	2.293695
Euclidean geometry	2.293754
Hyperbolic geometry	2.293781

The three methods were effectively tied. Hyperbolic was 0.00114% worse than Euclidean on average and took 50.7% longer than gradient-only rewiring. Gradient rewiring did beat static sparsity in all three seeds, although only by about 0.082%.

One larger test

To see whether scale changed the result, I increased the model to 24.54 million parameters and repeated the three main treatments with the same data, token budget, sparse density, and seed.

Method	Validation loss
Hyperbolic geometry	2.046006
Euclidean geometry	2.046079
Gradient rewiring	2.046139

Hyperbolic technically finished first, but its improvement over Euclidean was only 0.0036%. Before training, we required at least 0.5% to count as a meaningful result. The observed advantage was about 140 times smaller. Hyperbolic was also worse on two of the final five evaluations, while Euclidean had the slightly better complete learning curve.

Scale did improve efficiency. Hyperbolic’s runtime overhead fell from 50.7% to 15.9%. It became less expensive relative to the other methods, but it did not learn meaningfully better.

Conclusion

This does not prove Hyperbolic rewiring can never work. It tested one architecture, dataset, sparse structure, and rewiring schedule. It does show that this particular method failed to produce a useful advantage at both 6.90 million and 24.54 million parameters.

The entire project cost $2.96 and produced 18 trained models with verified checkpoints. My conclusion is that dynamic rewiring may help slightly, but adding learned Euclidean or Hyperbolic geometry did not improve it. Making the current design larger does not appear to be money well spent.