TODO

Learn new skills

Mon, 23 Dec 2024 00:00:00 GMT

Great learning materials are rare. This is a collection of gems I’ve encountered over the years, presented as a 90’s style web link list, intended for anyone who wants to learn new skills. Mostly software development-related, aside from the first few.

Money Stuff by Matt Levine

Skill: Finance — learn that everything is securities fraud.

Type of material: Email newsletter

Why is it good: Marvelous writing style. Unpacks recent events in finance with simplicity and wit, often connecting them to larger themes that run through the newsletter over time. Hugely entertaining and educational. Also released as web articles if you don’t want to give out your email.

Learning Music by Ableton

Skill: Composing music

Type of material: Interactive website

Why is it good: Experiment with composing music directly in your browser. No previous experience required. The tutorial begins with simple composing examples and tasks along with teaching basic fundamentals of music.

Machine Learning Coursera Course by Andrew Ng

Skill: Machine Learning — learn how to apply existing machine learning techniques, understand why they work, and how to troubleshoot issues.

Type of material: Videos and coding exercises

Why is it good: Great interplay between lectures and exercises. Clarity of explanations. Introductory scope. Enough math to understand why these methods work, but not too much.

Competitive Programmer’s Handbook by Antti Laaksonen

Skill: Exact algorithms — learn how to invent algorithms for ”code competition” -type problems (problems where exactly-correct answer is required and the main challenge is time or memory complexity).

Type of material: Book

Why is it good: Simplified explanations of many algorithmic techniques and algorithmic problem solving approaches. Unlike most algorithm books, this one is focused on ideas and code, not on the underlying math. This is the ”bible” of Finnish competitive programmers.

Essentials of Metaheuristics by Sean Luke

Skill: Optimization algorithms — learn techniques to find ”good enough” solutions for problems where an exactly-correct solution is not required. Typically used when an exact algorithm would be computationally infeasible and machine learning methods can not be applied. For example, route optimization often falls into this category.

Type of material: Book

Why is it good: A practical approach. Scope. Clarity of explanations. I recommend you to read this book while implementing and testing selected approaches on a real problem as you go along.

Building Effective Agents by Anthropic

Skill: Developing apps on top of LLMs — learn how to choose the right approach (despite the title this is not focused on agents)

Type of material: Blog post

Why is it good: Different approaches are covered and presented well. Based on real world applications. Opinionated against framework complexity.

The Grug Brained Developer by Carson Gross

Skill: Understanding complexity in software development — spoiler: complexity bad

Type of material: Essay

Why is it good: Funny and thought-provoking. Highly opinionated with opinions I strongly agree with.

Introduction to Cryptography by Christof Paar

Skill: Cryptography

Type of material: Videos

Why is it good: Clarity of explanations and visual illustrations. Entertaining presentation style.

The Modern JavaScript Tutorial by Ilya Kantor et al.

Skill: JavaScript language and browser APIs

Type of material: Website (alternatively available as a book)

Why is it good: Comprehensive, neatly organized & well written. Great explanations with illustrative examples. Goes wide and deep into the subject matter. Suitable both as a top-down read and as a place to look up specific topics.

This is how ChatGPT will be monetized

Tue, 27 Dec 2022 00:00:00 GMT

ChatGPT has been blowing up the internet lately, and I wanted to build something fun with it.

Here’s my vision of a dystopian future where ChatGPT displaces Google.

Enjoy while my OpenAI credits last!

11 miljoonaa euroa taivaan tuuliin — tässä lopputulokset kansallisesta tekoälyohjelmasta AuroraAI

Mon, 12 Dec 2022 00:00:00 GMT

Jos kuulet nyt ensimmäistä kertaa Suomen kansallisesta tekoälyohjelmasta, et ole yksin. AuroraAI on jäänyt yllättävän vähälle huomiolle mediassa hankkeen suuruuteen nähden. Hanke alkaa olla nyt suurin piirtein taputeltu päätökseen ja Hesari avasi keskustelun, joten päätin kantaa korteni kekoon testailemalla AuroraAI:n lopputuotoksia tavallisen kansalaisen näkökulmasta ja raportoimalla tulokset tähän artikkeliin.

Mutta mikä on AuroraAI? Valtiovarainministeriön sivut valaisevat asiaa meille:

Aurora on tekoälyjen ja autonomisten sovellusten muodostama verkko, joka luo edellytyksiä ihmiskeskeiselle ja ennakointikykyiselle yhteiskunnalle.^[1]

Jos mietit mitä se tarkoittaa käytännössä, hankkeen puuhamies selventää asiaa:

Dynaamisesti muotoutuvat arvoverkot ihmisen eri tilanteissa, se on niinku Aurora käytännössä.^[2]

Osalla lukijoista saattaa edelleen olla jotain kysymyksiä. Ei hätää, tässä tyhjentävä selitys Valtiovarainministeriön ICT-johtajalta:

Automaatio suorastaan puhkeaa kukkaan, kun on saatavilla hyvänlaatuista dataa. Data on automaatiolle kuin vesi, jolla kastellaan automaation kukkasia. Ja pistää myös samalla miettimään, että olisikö tässä kielikuvassa ihmiskeskeisyys se kukkamulta.^[3]

Kukkamultaa tästä hankkeesta ei ainakaan puutu. Vuosina 2018-2019 hankkeen esiselvityksessä pöhistiin käsittämätön määrä dokumentteja^[4] budjetilla, jonka suuruus ei ole julkisesti tiedossa. Esiselvitystä seuranneessa varsinaisessa hankkeessa 2020-2022 kukkamultaa viljeltiin 11,2 miljoonan euron budjetilla.^[5] Alun perin budjetin piti olla jopa 100 miljoonaa euroa,^[6] mutta veronmaksajien onneksi hanke ei toteutunut täydessä laajuudessa.

Lopputulokset

Mitä veronmaksajat voivat odottaa saavansa 11 miljoonalla eurolla? HS:n artikkelissa laitetaan silkkihansikkaat käteen ja varovaisesti lasketaan rima niin alas että se koskettaa lattiaa:

[Teknologiatutkija Santeri] Räisänen sanoo, että mahalasku on tyypillinen tulos tekoälyhankkeissa. Usein pelätään, että niiden tuloksena syntyy valvontakoneisto.

”Todennäköisempää on, ettei synny mitään.”

Jotain on sentään syntynyt. Ainakin Aurora AI -verkon suosittelumoottori ja profiilinhallinta, joiden lähdekoodi julkaistiin viime viikolla. Lisäksi syntyi koodia käyttöliittymiin sekä käsikirja neuvontabotin tuottajalle.

Riittääkö tyydyttäväksi lopputulokseksi tosiaan, että ”jotain koodia on ainakin julkaistu”?

Nyt on aika upottaa kädet kukkamultaan ja testailla, ovatko lopputuotokset hyödyllisiä tavallisen kansalaisen näkökulmasta. Sanottakoon, että tarkoitan lopputuotoksella jotakin konkreettista sovellusta tai vastaavaa tuotosta, jota kansalainen voi käyttää. En tarkoita lopputuotoksella mitään osallistujien henkilökohtaisia ”opittiin niin paljon” -kokemuksia, enkä tarkoita lopputuotoksella dokumentteja jotka lipastoidaan viraston arkistoon. Yritin parhaani mukaan kartoittaa konkreettiset lopputuotokset jotka hankkeessa tuotettiin, mutta kartoitus tuskin on täydellinen, koska hankkeen tekijät eivät ole näitä lopputuotoksia koonneet mihinkään yksittäiseen listaan. (Tässä linkki julkiseen dokumentaatioon.)

Nähdäkseni hankkeen lopputuotokset voi jakaa kahteen kategoriaan:

Palvelusuosittelijat, jotka suosittelevat palveluita kaikenlaisiin tilanteisiin
Keskustelevat chatbotit, jotka tarjoavat neuvoja rajattua aihepiiriä koskien

Tässä koostamani lista projektin lopputuotoksista:

Palvelusuosittelijat

Aloitetaan testailu palvelusuosittelijoista. Näissä sovelluksissa käyttäjä antaa tietoja elämäntilanteestaan ja sovellus arvioi 20 000 eri palvelun tietovarannosta, mitkä palvelut voisivat olla avuksi käyttäjän elämäntilanteeseen. Suositeltava palvelu saattaa olla esimerkiksi uraneuvontaa tai askartelukerho.

Ensimmäisenä kokeilen palvelusuosittelija Zekkiä. Sovellus pyytää minua syöttämään iän, sukupuolen ja paikkakunnan. Tämän jälkeen sovellus pyytää minua vastaamaan 0-10 asteikolla kymmeneen kysymykseen, jotka koskevat aiheita kuten itsetunto ja asumistyytyväisyys. Nämä monivalintakysymykset muistuttavat minua 2000-luvun loppupuolella suosioon nousseista Facebook-kyselyistä, jotka lupasivat vastata niinkin tärkeisiin kysymyksiin kuin ”Mikä Spaissari olet” tai ”Jos olisit ruoka, mikä ruoka olisit”.

Vastaan kysymyksiin rehellisesti ja pohdiskelen mielessäni, miten sovellus edes teoriassa pystyisi neuvomaan tilanteeseeni parhaiten sopivia palveluita — enhän ole päässyt syöttämään mitään sellaisia tietoja jotka yksilöisivät palvelulle mihin ongelmiin kaipaan apua. Siitä huolimatta lopputulos onnistuu alittamaan odotukset:

Zekki suosittelee minulle pelkkiä nuorisopalveluita, vaikka olen 36-vuotias ja syötin ikäni erillisessä kentässä. Jos ikä kysytään erikseen, ei olisi kovin monimutkainen tehtävä suodattaa tuloksista pois sellaiset tulokset jotka eivät ole sopivia käyttäjän iälle. Tällainen suodatus olisi helppo toteuttaa ilman mitään ”ihmiskeskeistä tekoälyä”.

Testaan seuraavaksi AuroraAI testiympäristön 3x10D palvelusuosittelijaa. Tämä suosittelija ei kysy minulta ikää eikä sukupuolta, mutta muuten kysymykset ovat samat kuin Zekissä. Tulokset:

Jokainen yllä näkyvä palvelusuositus on joko nuorisopalvelu tai vammaispalvelu. En ole nuori enkä kehitysvammainen.

Mennään eteenpäin muihin sovelluksiin. Seuraavaksi testaan AuroraAI testiympäristön vapaatekstillä toimivaa palvelusuosittelijaa. Käytettyäni AuroraAI:ta hetken aikaa päässäni on herännyt synkkiä ajatuksia, joten kokeilen osaako AuroraAI suositella minulle mielenterveyspalveluita:

AuroraAI ei osaa suositella itsemurhaa hautovalle käyttäjälle mielenterveyspalveluita. Sen sijaan AuroraAI suosittelee hautatoiveen tekemistä ja testamentin kirjoittamista ennen itsemurhaa.

Hakutuloksia näytetään vapaalle tekstihaulle jostain syystä vain 5, eikä enempää tuloksia voi listata edes sivuttamalla. Suosittelumoottori on joka tapauksessa joutunut pisteyttämään suuren määrän erilaisia palveluita, joten ei ole mitään teknistä estettä sille miksi tuloksia ei voisi näyttää enempää. Tulosten rajoittaminen viiteen herättää epäilyksen, että minkäänlaista käytettävyystestausta tuotteelle ei olla tehty ennen sen julkaisua.

Kenties mielenterveyspalvelut on vain jotenkin unohdettu toteutuksesta. Kokeillaan tekstihakua uudestaan, käyttäen tällä kertaa tekstihakuna AuroraAI:n pöhinävideolla esitettyä elämäntilannetta: leipomoyrittäjä Henry on siirtymässä eläkkeelle perheyrityksestä. AuroraAI:n pöhinävideolla hänelle ehdotetaan seuraavia palveluita:

Sukupolvenvaihdos perheyrityksessä
Eläkkeet
Verotus
Eläkeläisalennukset
Eläkeläistoiminta

Mutta mitä palveluita AuroraAI todellisuudessa ehdottaa tähän elämäntilanteeseen?

Nautojen laidunnus ja jaloittelu kesäaikaan? Vain yksi viidestä suosituksesta on millään tapaa relevantti. AuroraAI ei edes tunnista sukupolvenvaihdosta perheyrityksessä mahdolliseksi elämäntapahtumaksi, vaikka tämä on juuri se elämäntapahtuma, jonka he itse valitsivat kanoniseksi esimerkiksi pöhinävideoonsa.

Chatbotit

Siirrytään seuraavaksi kokeilemaan AuroraAI chatbotteja. Chatbotit eroavat palvelusuosittelijoista keskustelevan vuorovaikutuksen lisäksi siinä, että palvelusuosittelijat antavat suosituksia kaikenlaisiin elämäntilanteisiin, kun taas chatbotit on selvästi suunniteltu rajattuihin aihepiireihin. En ole kohdeyleisöä millekään julkaistuista chatboteista, joten kokeilen osaa niistä syöttämällä sen tyyppisiä kysymyksiä joita voisin kuvitella oikeiden käyttäjien syöttävän.

Kokeilen aluksi Poikien puhelimen chatbottia.

Hyvin tyypillinen kysymys yksinkertaisessa muodossa esitettynä, mutta botti ei ymmärrä sitä ja vastaa litanialla esikirjoitettuja viestejä. Esikirjoitettuja viestejä on niin paljon etteivät ne edes mahdu yhteen kuvakaappaukseen.

Koitetaan eri kysymystä:

Botti ei ymmärrä vieläkään ja toistaa litanian samoista esikirjoitetuista viesteistä. Eikö riittäisi että litania on annettu kerran ja jatkossa voisi vastata vaan sen ”oi joi” -viestin ilman tota koko litanian toistamista? Ilmeisesti tällekään ei ole tehty minkäänlaista käytettävyystestausta ennen julkaisua.

Kun lähetän riittävän paljon viestejä, löydän lopulta jotain millä saan tulokseksi jotain muuta kuin ”oi joi”:

Ilmeisesti botilla on rajattu määrä esikirjoitettuja vastauksia, ja tässä tapauksessa avainsana ”vanha” osuu siten että botti olettaa minun kysyvän hänen ikäänsä (vaikka kysymys koskee minun ikääni eikä botin ikää).

Kokeilen seuraavaksi Kirjastobottia.

Osaisiko botti kertoa, onko jokin tietty kirja saatavilla?

Botin tekijöille ei ole tullut mieleen, että kirjaston käyttäjät saattaisivat kysyä kirjojen varaustilanteesta botilta. Oi joi!

Kenties botti kykenee kuitenkin suosittelemaan minulle jotakin kirjaa?

Kirjastobotti tunnistaa kysymyksen, mutta ei osaa suositella mitään kirjaa vaan tarjoaa esikirjoitetun vastauksen. Oi joi!

Kokeilen seuraavaksi BusinessKuopion chatbottia.

Botti vastaa mielellään yrittäjyyttä ja liiketoimintaa koskeviin kysymyksiin, mutta ei kuitenkaan tiedä, mikä on veroilmoitus? Oi joi!

Testailin chatbotteja useampana päivänä samalta tietokoneelta ja huomasin että ne muistavat viestihistorian. Viestit voivat sisältää hyvinkin arkaluonteisia asioita, joten päätös säilyttää viestit sessioiden ylitse käyttäjää varoittamatta on vähintäänkin kyseenalainen. Tämäkin ”ominaisuus” vaikuttaa vahvasti siltä ettei minkäänlaista käytettävyystestausta chatboteille ole tehty ennen niiden julkaisua.

Eiköhän testailu saa tältä erää riittää. AuroraAI:n chatbotit muistuttavat lähinnä puhelinautomaatteja, joihin on ohjelmoitu muutama kymmenen eri aihetta. Puhelinautomaateissa tosin on se etu AuroraAI:hin nähden, että ne luettelevat käyttäjälle mitkä aiheet ovat saatavilla. Esimerkiksi pankin puhelinautomaatista saatat löytää oikean aiheen valitsemalla ensin 1) henkilöpalvelut, sitten 6) vakuutusasiat, ja lopuksi 3) korvauspyynnön jättäminen. AuroraAI:n tapauksessa voit vain arvailla oikeita sanayhdistelmiä, ja oi joi, kun se on vaikeaa sokkona ilman tietoa siitä millaisia sisältöjä on saatavilla. Tällaiset muutaman kymmenen eri aiheen esikirjoitetut tekstisisällöt olisivat paljon paremmin saavutettavissa ihan perinteisenä “usein kysytyt kysymykset” tekstisivuna. Chatbot-ratkaisussa ei saavuteta minkäänlaista etua perinteiseen tekstisivuun nähden, kun botin tekninen kyvykkyys on sillä tasolla että se tunnistaa vain kovakoodattuja avainsanoja ja vastaa niihin esikirjoitetuilla sisällöillä. Rahaa saadaan toki palamaan, tässä tapauksessa 11 miljoonan euron verran.

Mitä tästä kaikesta opittiin?

En halua syyttää yksittäisiä ihmisiä tästä fiaskosta. Lopputulos olisi todennäköisesti ollut yhtä kehno, vaikka kaikki virastojen puuhamiehet ja konsulttitalojen toimittajat olisi vaihdettu toisiin. Hankkeen epäonnistuminen oli nähtävissä kaukaa, eikä siihen olisi missään tapauksessa pitänyt ryhtyä. Tekoälypöhinä ei kuulu julkiselle sektorille. Yksityisellä sektorilla saa pöhistä ihan niin paljon kuin haluaa. En väitä että tällainen hanke olisi onnistunut sen paremmin yksityisen sektorin vetämänä — olihan tätäkin projektia toimittamassa yksityisen sektorin konsultit — on vaan eri asia tuhlata omia rahoja kuin verorahoja.

Verovarat olisi parempi säästää sellaisiin kohteisiin, jotka luontaisesti kuuluvat valtion vastuulle ja osaamisalueelle. Esimerkiksi viittomakieliset uutiset on tärkeä tehtävä jota ei markkinavetoisesti syntyisi. Toisena esimerkkinä voin nostaa sähkönsiirtoverkon ylläpitämisen, joka on tehokkaampaa valtion omistamana monopolina kuin markkinaehtoisena kilpailuna. Nyt ei ole kyse mistään tällaisesta. Miksi valtio upottaa rahaa tekoälypöhinään joka ei ole yhteiskunnalle kriittinen, jota syntyy muutenkin markkinavetoisesti yksityissektorin puolella, jossa valtiolla ei ole mitään kyvykkyyttä projektin vetämiseen, ja jossa valtaosa hankkeen aikana kehitetyistä kyvykkyyksistä jää yksityissektorin konsulteille jotka ovat toimittamassa projektia? Jos valtio haluaa tukea jotain uraauurtavaa tutkimusta ja koulutusta, siihen on meillä jo loistavat instituutiot: yliopistot. Aivan järjetöntä laittaa virasto vetämään tällaista hanketta.

Yksityisestä sektorista puheenollen, kokeillaanpa lopuksi millaista Chatbot-palvelua Suomen kansalaisille on tällä hetkellä tarjolla täysin ilmaiseksi yksityisen OpenAI:n ChatGPT-palvelusta:

Ihan osuva suositus, mutta tällainen toisella puolen maailmaa majaileva tekoäly ei voisi mitenkään kertoa meille, onko kirjaa saatavilla Kuopion kirjastosta… eihän?

How to update Gatsby dependencies through major version upgrades

Wed, 20 Jul 2022 00:00:00 GMT

I recently went through the pain of updating my Gatsby starters (including this blog) from Gatsby v2 to v4, which didn’t turn out to be as easy as the internet would have you believe. With the help of some kind souls from Gatsby Discord, I eventually got through it, and I figured I would document the process in this article.

Setting expectations

According to the official docs, you should be able to update both Node version and Gatsby version without having to do code changes (as long as your code isn’t using any of the documented breaking changes). Other advice on the internet is similar. Unfortunately, when you have a moderately complex project, major version upgrades are likely to break everything in a myriad of ways. Here’s a non-exhaustive list of issues that I experienced:

Got stuck in dependency hell where npm install fails due to dependency version conflicts, but any attempt at resolving them by upgrading or downgrading packages merely shifts the conflict to a different set of dependencies.
Running gatsby develop was stuck in an infinite loop due to a bug in Gatsby PostCSS plugin that only occurs with specific Node versions. The community plugin still isn’t fixed and the workarounds involve either removing the plugin, downgrading Node version, or refactoring CSS in a way that doesn’t trigger the bug.
My CSS broke — not in a clear and obvious manner, but rather in a subtle way affecting only lines where the CSS is nested, the line includes the & character, and some value within the line is resolved from a variable. To clarify, my CSS broke only for lines where all 3 of those conditions hold.
Another subtle breakage was related to my image placeholders, which were misaligned due to an internal change in gatsby-image. This is one of those things that happen when you venture off the beaten path. Internals of Gatsby were never guaranteed to be stable, so you could say that I brought this upon myself by relying on an internal API. This bug was particularly subtle because the placeholders are only visible while the image is loading, so you typically don’t even see them when you’re developing locally. A real user with a slow internet connection will see them, though, and the misalignment looks quite jarring.

I encountered issues that only affect the production build but don’t affect the development build, issues that only affect the development build but don’t affect the production build, and issues that only affect hot reload within the development build. Catching everything required extensive manual testing, and I can’t imagine a suite of automated tests that would catch things like image placeholder misalignment.

How to update

The following instructions are written for npm, but you should be able to follow along even if you are using yarn.

General tips:

In order to avoid dependency hell, you should update everything at the same time instead of updating things one by one.
Inscrutable errors are often resolved by deleting .cache, node_modules and package-lock.json before running npm install.
Don’t stop at the point where the build completes successfully and your automated tests pass; you’ll probably need to do manual testing to catch all the subtle issues.

Step by step instructions:

Switch to desired Node version (e.g. nvm use v16)
Ensure desired version of gatsby-cli is installed in this particular Node environment (e.g. npm install -g gatsby-cli@latest-v4)
Run ncu -u to bump package.json dependencies to newest versions (requires npm-check-updates).
Manually edit package.json to downgrade packages that need to be downgraded for compatibility. For example, at the time of writing this, I needed to downgrade React from v18 to v17 in order to avoid hydration errors (Gatsby doesn’t fully support v18 at the time of writing).
Delete .cache
Delete node_modules
Delete package-lock.json
Hammer on npm install until it runs without errors. You can expect to see many errors related to version conflicts in your dependencies. A common source of issues is unmaintained plugins created by the community. In some cases the conflicts can be resolved by downgrading some of the dependencies you just upgraded. In other cases you might want to remove or replace incompatible plugins. As a last resort, you might want to try npm install --legacy-peer-deps to force an installation despite version conflicts. Unmaintained plugins often work with newer dependencies in practice.
Hammer on gatsby develop until it runs without errors. Resolving errors at this step often requires changes to dependencies, in which case goto step 5.
When you get the dev environment running, it’s time to run your automated tests and fix everything that broke according to your tests.
After your tests pass successfully, it’s time to manually test your site in dev environment and fix everything that’s broken (malformed CSS etc. that won’t be caught by your automated tests).
Additionally, you should manually test that hot reload is not broken: First make some code changes. Then check that you see the changes applied in browser without any action inside the browser. Then try refreshing the site in browser. Then try navigating across different pages.
Once your dev environment is fully working, it’s time to manually test the production build (gatsby build && gatsby serve). For example, some routing issues only exist in the production build.
After everything seems to work locally, it’s time to update your CI configuration to use the same node version and gatsby-cli as you use locally. This is also a good time to update the build image and other artifacts, if needed.
Update your README set-up instructions. (I do this even for my personal projects because I want to know how I can run the project 2 years from now after I have forgotten literally everything about setting it up. It’s good to write things like Node version and gatsby-cli version in the README because those won’t be contained in your package.json.)
If you still have remaining energy, you can continue working through the numerous npm warnings and Gatsby warnings that have appeared as a result of the upgrades.

I made a website that moves your mouse cursor

Thu, 01 Apr 2021 00:00:00 GMT

A long-running question on StackOverflow asks: can a website move your mouse cursor? The answer which is marked as ”accepted” unequivocally states ”it isn’t possible”. The most up-voted answer (with 195 upvotes) is even more dramatic:

You cannot move the mousepointer with javascript. Just think about the implications for a second.

Well, I like a good challenge. And guess what? It’s done. Click here to try it for yourself.

Edit: I originally released this as a fun little April’s Fools project. After the mouse moving demo turned out to be popular, I decided to create a mini-game around the mechanic: Ouija Online, a web horror experience.

This election and the insanity of betting markets

Thu, 17 Dec 2020 00:00:00 GMT

If you were following the 2020 U.S. presidential election, you might remember the barrage of ”experts” offering predictions on the outcome of the election on every news network and in every newspaper. Every now and then somebody would come in and say:

What do these so-called experts know? Instead of listening to experts you should look at the probabilities implied by betting odds.

Up until this election, I used to be one of those people. I would say:

These odds must represent the best possible estimate, because otherwise you could make money by betting on the event.

Taken literally, this argument is of course wrong because you can make money in the betting markets. Or maybe you can’t, but someone else can. This topic, the difficulty of beating a market, is referred to as market efficiency. You may have heard this term in the context of stock markets, but it applies just as well to other markets, including betting markets^[*].

First, a quick primer on market efficiency

A perfectly efficient market is a theoretical concept; a market where current prices perfectly reflect all available information. That means no trader will be able to gain an edge over other traders in the market: a trader who spends all their waking hours studying the market will have the same expected value as a monkey who throws darts at a newspaper.

In the real world no market is perfectly efficient. Some are more efficient, some are less. Let’s consider the stock and derivatives market as an example. We know that this market is not perfectly efficient, because some participants make outsized profits year after year for decades. Warren Buffet’s Berkshire Hathaway has outperformed the market since 1962. RenTec’s Medallion Fund has crushed the market since 1988. So we know that it’s possible to beat the stock market. But we also know that trading in the stock market occurs mostly between highly specialized professionals (or algorithms written by highly specialized professionals) who have devoted their lives to the market. If a layperson has a couple of hours to spare, can they reasonably expect to beat the professionals on the other side of those trades? Of course not^[**].

We generally expect markets to become more efficient over time, as money shifts from bad traders to good traders. That means good traders will have more impact on stock prices (or betting odds), and bad traders will have less impact.

Should we expect betting markets to be similar to stock markets in efficiency? At least in the case of large markets, such as major sporting events or major political events, the answer is a resounding yes. The size of the market provides an incentive for smart people to dedicate time to figuring out probabilities of different outcomes. And the people who are good at predicting things should have vastly more money to bet than the people who are bad at predicting things. So you would expect Trump-Biden odds to be so efficient that a layperson wouldn’t be able to beat the market. And you would be wrong.

Nate Silver versus the betting markets

In the weeks leading up to the election, Joe Biden had a massive lead in the polls. Despite that, the betting markets were offering odds above 1.6 for Biden, implying Biden had less than 62% chance of winning. That’s weird, right? How much uncertainty could there possibly be so close to the election, with such a lead in the polls?

As a layperson it can be hard to quantify the uncertainty. First you have the inherent uncertainty of sampling voters and extrapolating sampled results to the population. Then you have uncertainty regarding pollsters’ methods and their accuracy. For example, in this election the polls turned out to be systematically biased against Trump, as if pollsters learned nothing from 2016. And then you have uncertainty of scandals and other events unfolding before the election takes place.

Luckily, problems like this have an easy answer: defer to an expert. It’s not always clear who’s an expert in a particular area, especially when people who are touted as experts often fail catastrophically in their supposed field of expertise. But sometimes you can trust a person. In this story our deferred expert opinion comes from Nate Silver. Silver has a good track record on political predictions^[***] and his Bayesian approach to predicting is sound (I had read his book several years ago).

So how did Silver quantify Trump’s chances?

Two weeks before the election, on 19.10., Silver was predicting 88% chance of victory for Biden. At the same time Unibet was offering odds of 1.58, implying only 63% chance. If you took Silver’s estimate as granted, then the expected value for betting on Biden would be +39% of your stake. For example, if you bet $1000, then your expected value would be $390. Or another way of looking at it is: how wrong would Nate Silver have to be in order for this bet to have negative expected value? Silver’s 88% chance of victory implies that fair odds should be around 1.14, compared to actual odds of 1.58. That’s insane. Nate Silver might be wrong, but he can’t be this wrong. That day I made my first political bet ever (and I should mention, I don’t really do any other kind of betting either - this was just too good of an opportunity to miss).

I didn’t intend to bet more on this event. Over the next few weeks Biden’s odds fell lower (as you might expect, given no significant news events, there is less uncertainty closer to the election). But then, on the very last day before the election, Biden’s odds jumped to 1.60. At the time Silver was predicting 90% chance of a Biden victory, and there was only one day left, so there wasn’t much room for any surprises to come up. Suddenly, the betting opportunity of the decade got even better. So I decided to bet again and double my exposure.

The counting of the votes

The day before the election it was widely speculated that Trump would gain an early lead when vote counting begins, and that Biden would catch up once more of the mail-in votes were counted. This was speculated due to democrat voters’ historical preference for mail-in voting over in-person voting, and the fact that vote counting begins from in-person votes. Even late night tv show hosts talked about this.

Then the thing happened that everybody said was going to happen: they started counting in-person votes and Trump had a significant early lead. Faced with this entirely anticipated event, you would expect a muffled reaction from the market. But it’s 2020 and nothing makes sense anymore, so market reaction was anything but muffled: Biden’s odds shot through the roof. In the early hours of vote counting, when I was soundly asleep, you could place bets for Biden with odds above 5.0 (implying less than 20% of chance of winning).

I thought betting for Biden with 1.6 odds was the betting opportunity of the decade, but it turned out the real opportunity was available in live betting as the event was unfolding. Reminder to self: next time there’s a huge betting event, reserve some ammunition for live bets and follow the event as it’s unfolding, just in case something insane happens.

Anyway, the votes were counted and Joe Biden was declared as the winner by most media outlets. The race turned out to be much closer than anticipated in several swing states, but Biden’s lead in the polls was so large that he could withstand a larger-than-normal sized polling error and still win.

Sure, the results weren’t official yet. Trump could still overturn these results by uncovering fraud, exploiting a technicality, or perhaps by discovering a hidden stash of votes that should have been counted but weren’t. However, he would have to overturn results in multiple states in order to win enough electors. Despite the legal prowess unleashed at Four Seasons Total Landscaping, overturning results didn’t seem like a realistic outcome. So most betting sites decided to settle bets at this point.

The story is supposed to end here, but it doesn’t

Although most betting sites did settle bets once the votes were tallied, some didn’t. In particular, Betfair decided to keep the market open. This was especially strange, because Betfair’s terms clearly specified that the winner will be declared based on ”projected elector count”. The actual elector count was described as a ”subsequent event” which has no bearing on the bets. In fact, an earlier version of Betfair’s terms went even further and specified that bets will be settled according to whoever CNN declares as the winner (they quietly removed this condition at some point, despite having already matched bets on those terms).

Imagine this: you bet money for Biden on Unibet. Biden is declared to be the winner of the election by multiple media outlets. Unibet settles the bet and pays out your winnings. What should you do with your hard-earned money? Grab some ice cream? Go shopping? Here’s an idea: deposit your winnings on Betfair and bet on Biden again. That’s right, bet again on the same event that was already once settled in your favor. Apparently, that’s a thing you can do. So… that’s exactly what I did.

You might be thinking: even if the market remains open, there’s no way anybody would be willing to take the other side of those bets, except maybe at 1.01 or something. Nope… for a whole month after the votes were counted, you could bet for Biden on Betfair between 1.03 and 1.10 (I placed bets at 1.05).

First I thought 1.6 odds before the event was the betting opportunity of the decade. Then the event began to unfold and I thought, man, the real opportunity turned out to be live betting during the event. So there was some opportunity before the event, and some opportunity during the event, sure. But never in a million years would I have thought that an opportunity like this becomes available after the event. Who on earth was willing to lay (bet against) Biden at 1.05 after the votes were counted and after Trump’s lawsuits^[****] were laughed out of court?

Betfair settled the bet yesterday, on 14.12. I usually make an effort to understand why things turned out the way they did, but I can’t figure this one out. I’m just gonna take my free lunch and go.

Clarifications:

[*] I’m using the term ”betting market” loosely. If you want a strictly theoretical argument, let’s say that the term refers specifically to Betfair exchange, where the market price is entirely determined by bettors, in a fashion similar to a stock market. However, in this election the odds set by traditional bookmakers also closely followed the odds in Betfair exchange (as they should, due to arbitrage opportunities). So a casual reader could imagine traditional bookmakers to fall under the ”betting markets” umbrella.

[**] Although a layperson will find it difficult to gain an edge in trading large stocks like Coca-Cola, they might still be able to gain an edge in trading stocks of small companies, because large investors are practically unable to invest in small companies, which leads to more inefficiencies in small companies’ stock prices.

[***] People keep bashing Silver for ”failing to predict” Trump’s 2016 victory. But if you had been making bets based on Silver’s predictions, you would have bet for Trump in 2016, and you would have won, because Silver’s probabilities for Trump in 2016 were higher than those implied by the betting markets. So, whatever significance you place on the outcome of a single prediction (personally, not much), Trump’s 2016 victory would be an example of a successful prediction by Silver, not an example of a failed prediction.

[****] I assume some lawsuits are still pending or upcoming, so perhaps not all lawsuits were laughed out of the courts in early December. So could you cling onto some kind of hope for a last minute Christmas miracle. If you’re into betting for miracles and that sort of thing.

Banks create money out of thin air, but it's less impressive than it sounds

Sat, 29 Aug 2020 00:00:00 GMT

Most people are familiar with central banks’ ability to create money out of thin air. A central bank has access to physical printing presses which can produce physical banknotes, and it has the corresponding ability to create electronic money. However, central banks are not the only entities which create money. Money is also created by regular banks, such as retail banks and commercial banks, in the process of making loans.

Many people believe that regular banks’ ability to create money is similar to central banks’ ability to create money. It is not. This misconception has been perpetuated by a powerful cast of characters which includes mainstream news outlets and even some economists.

This article will demystify money creation by illustrating the following facts:

A bank deposit is merely an IOU.
A bank loan is merely an exchange of 2 IOUs.
Anyone can create IOUs out of thin air.
When a customer redeems an IOU at a bank, the bank typically needs to hand over another form of money — the kind that it cannot create out of thin air.

Bank deposits are merely IOUs

Suppose you have $100 in cash. You decide to deposit it at your local bank. Now you no longer have $100 in cash. Instead, you have a promise from the bank that you can withdraw $100 in cash at any time. The bank has $100 in cash and you have a promise. Let’s imagine this promise as a physical document acknowledging debt, an IOU (”I owe you”). The bank owes you $100 and promises that you can exchange this IOU to $100 cash at any time.

Is this IOU ”money”? We’ll get to that later, but the short answer is yes, that IOU is money. There are different types of money, and checking account deposits are usually considered to be a type of money.

Let’s illustrate this bank deposit as an exchange:

Now let’s look at the bank’s balance sheet. If you were the only customer at the bank, its balance sheet might look like this:

Assets are things the bank owns, liabilities are things the bank owes. In this simplified example, the only asset that the bank has is the $100 in cash that you deposited — now considered to be the bank’s asset — and the only liability is the $100 debt that the bank owes to you. These accounting entries were created at the same time: when the bank incremented your checking account by $100 (the deposit liability in bank’s accounting), there was a corresponding $100 cash deposit (the cash asset in bank’s accounting).

A bank loan is merely an exchange of 2 IOUs

Let’s continue with the example where you deposited $100 to a bank. Suppose I will go into that bank and ask for a $10 loan. Now that the bank owns $100, it can easily afford to loan me $10. I promise to pay back $11 at a later point in time, so the bank is eager to offer me a loan and make that dollar.

The bank deposits this $10 loan to my checking account. Remember, checking account deposits are IOUs, so the bank just gave me a $10 IOU. I promised to pay back $11, so I also created an IOU. In other words, this loan is an exchange of 2 IOUs: the bank promises to give me $10 in cash whenever I want, and I promise to give the bank $11 at a later point in time. Let’s imagine these IOUs as physical documents:

When I take out a bank loan, I am exchanging one IOU for another:

After this transaction, the bank’s balance sheet might look like this:

The bank still has that $100 cash in assets, and it still has that $100 liability to John. But now the bank also has a $10 liability to me: I could withdraw my $10 from the bank in cash at any time. Corresponding to this liability, the bank has recorded a new asset: my promise to pay back $11 later.

Notice how the bank now has $110 of deposits, but only $100 of cash in the vault. Welcome to fractional reserve banking! If I decided to withdraw my $10 from the bank today, and you decided to withdraw your $100 from the bank tomorrow, the bank would say ”oops, we no longer have your money”. In practice this rarely happens due to a myriad of reasons (banks can trade assets between each other when they have too little/too much cash, banks have internal risk controls which limit lending and other risky endeavours, lawmakers impose restrictions which are monitored by regulators, the existence of deposit insurance reduces likelihood of bank runs, etc.).

Anyway, let’s get to the interesting bit. Before taking out the loan there was $100 on checking account balances and afterwards there was $110 on checking account balances. Clearly, money was created in the process of taking out the loan. Where did this money come from? Was it created out of thin air? The short answer is yes, kind of, but let’s step back a little bit.

What is money, anyway?

Let’s go back to the cash deposit example. You deposit $100 into a bank and the bank records it as an asset. The bank records a corresponding liability: a $100 deposit to your checking account. Did the bank create money in this scenario?

Before the deposit there was $100 of cash money, but after the deposit there was $100 of cash money + $100 of electronic money = $200 worth of money in total. But of course we don’t want to ”double count” your money just because you deposited it in a bank, right? That’s why economists developed this convention that cash money in a bank’s vault is considered to be ”out of circulation”, and thus, is not typically counted when we count how much money exists.

The Eurozone M1 definition for money is basically equal to this notion of ”currency in circulation”. The concept of money is ambiguous, so there are many different definitions, varying by both economic area and by use case. Nevertheless, checking account balances are included in most commonly used definitions (though not in all, for instance the UK M0 definition for money does not include checking account balances).

In abstract terms we might think that the $100 cash you deposited into a bank was transformed into a $100 checking account balance. Except it wasn’t. That $100 of cash money still exists. It was not transformed into anything. The bank took $100 of cash and put it in a vault. Then the bank created an IOU out of thin air and gave it to you. (We can say that the IOU was created out of thin air, because the corresponding $100 cash in the vault is not earmarked to this IOU — the bank may run out of cash before you try to redeem that IOU.) Then economists decided that the IOU counts as money, and the cash in the vault does not count as money. Thus, the act of depositing cash in a bank does not result in (net) creation of money. $100 of electronic money is added to circulation while $100 of cash money is removed from circulation, for a net difference of $0.

When you deposit cash in a bank, the bank creates an IOU out of thin air. Similarly, when you take a loan out of a bank, the bank creates an IOU out of thin air. However, due to accounting conventions, the latter action results in net money creation, while the former action does not. (Although these conventions are weird, they make sense: currency in circulation is a good measure with real implications for economic activity, and currency in circulation is genuinely increased when banks loan money.)

Anyone can create IOUs out of thin air

If there’s only one thing you take away from this article, let it be this. IOUs are promises. Anyone can make promises out of thin air. For example, I could be handing out IOUs like this, furbished with my signature for authenticity:

You might think these would be worthless, but that’s not exactly the case. People who know me would think ”Atte is good for $10”. These would not be worth exactly $10, because you would still rather take $10 in cash than a coupon that’s exchangeable for $10 in cash. And it’s unlikely that these IOUs would be used to facilitate real economic activity, so these IOUs should not be considered ”money”. Although there are clear differences in how these IOUs would be used, compared to the IOUs created by banks, is there any fundamental difference between the IOUs themselves?

Let’s take another example and get one step closer to creating money. This one is real, not hypothetical, example of IOUs created by a non-bank corporation.

We have to go all the way back to 2011. Poker is booming and Full Tilt Poker is one of 2 giants dominating the scene (the other being PokerStars). The way that poker sites work is you deposit money (like you would in a bank) and then as you play poker and you win and lose money from other players, the poker site keeps track of how much money everyone has (like a bank would keep track of transfers between customers of the same bank).

In other words, a poker site issues IOUs to players. These IOUs are supposedly redeemable at any time (like the IOUs created by a bank). The IOUs created by Full Tilt Poker were largely treated as money by players of the site, to the point where they were used to facilitate real economic activity. For example, it was typical for players to settle small debts (like $10 for a movie ticket) by sending a player-to-player transfer on Full Tilt Poker instead of sending a bank transfer (because transfers on poker sites were typically faster and easier to make than bank transfers).

For all intents and purposes, the IOUs created by Full Tilt Poker were money. They simply were not counted as money due to accounting conventions. Now, some of you might be thinking ”Wait a minute… Doesn’t the poker site have all player funds in the bank? If all player funds have corresponding bank deposits, then those bank deposits would already be counted as money, and we wouldn’t want to double count that money.” Yeah, I also thought that Full Tilt Poker had player funds in the bank (can you guess where this is going?).

On April 15, 2011, known in the poker world as Black Friday, the FBI raided offices of Full Tilt Poker and PokerStars as part of a larger crackdown on online gambling. This law enforcement action triggered an instant bank run by players: everyone wanted to withdraw their money at the same time. This should not be a problem, because all poker sites promise to segregate player funds from operational funds. And for PokerStars it wasn’t a problem, because they had done so. Full Tilt Poker hadn’t. Even worse, they had slowly siphoned off almost all player funds over the course of multiple years, leaving only a small fraction in reserve. In other words, Full Tilt Poker had been operating like a fractional reserve bank.

Now, going back to the IOUs created by Full Tilt Poker: since Full Tilt Poker is not a bank, its IOUs were not counted as money. And since it did not have sufficient bank deposits to cover its IOUs, these IOUs were not ”indirectly” counted as money, either. But players still considered these IOUs to be money, and these IOUs facilitated real world economic activity similarly as bank money does. If you have an argument why these IOUs should not be considered money, I would love to hear it. Otherwise, let’s conclude that non-bank corporations can sometimes create money out of thin air.

No permissions needed

One question that often comes up in discussions about money creation is why are banks allowed to create money out of thin air. Sometimes this is a legitimate question about the merits of full-reserve banking versus fractional reserve banking, but more often, this question implies some kind of misunderstanding about how money is created. This confusion is exacerbated by some economists pouring gasoline into the fire by using inflammatory language like this:

The money supply is created as ‘fairy dust’ produced by the banks out of thin air.

That’s Richard A. Werner in his influential 2014 paper. Werner continues in a subchapter titled ”What is special about banks”:

We now know, based on empirical evidence, why banks are different, indeed unique … and different from both non-bank financial institutions and corporations: it is because they can individually create money out of nothing.

There’s a lot to unpack here. First, Werner claims that banks are special due to their ability to create money out of thin air. Second, Werner claims that the first claim is proven conclusively with empirical evidence.

Werner’s second claim is patently false, because the ”evidence” he presents in his paper only describes banks’ ability to create money out of thin air — he presents no evidence for non-bank entities’ inability to create money out of thin air. Thus, he presents no evidence that banks possess an ability non-banks do not possess. And yet, he claims to have done the opposite.

Werner’s first claim is correct only on a technicality: due to accounting conventions, some IOUs are counted as money, while others are not. As we saw with the Full Tilt Poker example, some IOUs created by non-bank entities are indistinguishable from money. In other words, we have empirical evidence that non-bank entities are sometimes able to create money-like IOUs out of thin air — we simply do not count those IOUs as money due to accounting conventions.

Is there a fundamental difference between the money created by Full Tilt Poker compared to the money created by banks? Well, small depositors’ bank deposits are typically covered by deposit insurance programs, which guarantee that depositors get their money even if the bank runs into problems. This insurance is backed by the central bank and its ability to print banknotes. From the perspective of a regular person, this is certainly a fundamental difference between holding deposits at a poker site versus holding deposits at a bank. However, deposit insurance typically applies to small deposits only, so it does not cover all bank IOUs.

One might argue that the regulatory environment monopolizes money creation to banks, and that creating money as a non-bank is illegal. If this claim was true, would that prevent non-banks from creating money? No. It would be silly to claim that something does not happen simply because it has been outlawed (unless you want to make the pedantic argument that the definition of money specifically excludes IOUs created by non-banks). Regardless, the claim itself is not true: sometimes non-banks can in fact legally create IOUs which are very similar to money. For example, Amazon gift cards have properties very similar to money, and they are legally created by a non-bank. World of Warcraft gold serves as another example.

Redeeming an IOU

Now, let’s move on to the final point I want to make: When a customer redeems an IOU at a bank, the bank needs to hand over another form of money, and the bank has no ability to create that money out of thin air.

Note that redemption of an IOU works the same way regardless of the origin of the IOU. Perhaps you had originally deposited cash into your checking account. Perhaps you took out a loan and the bank deposited money into your checking account. Regardless, redemption works the same way.

In typical cases a bank’s IOU is redeemed into money in one of these forms:

Cash withdrawal
Bank transfer

The cash case is simple. You can think of it like a reverse cash deposit (assuming the bank has cash available). Banks have no ability to create cash out of thin air, because they do not have access to money printing facilities (like a central bank does). The bank needs to actually have that cash, and once they give it to you, they no longer have it.

The bank transfer case is more complicated. You can think of it as exchanging an IOU from bank A to a similar IOU from bank B. But how does bank A actually transfer your money to bank B?

You might imagine bank A calling bank B and saying ”Hey there, fellow bank. Carol wants to transfer $5 from her account at our bank into her account at your bank. Can you please create $5 out of thin air and add it to Carol’s account balance? I promise to delete $5 from her account here, so we don’t accidentally create money.”

You can probably imagine all sorts of problems with that approach, but I’ll describe one anyway. Suppose it’s not Carol. It’s Bill Gates, and he wants to transfer a billion dollars. Now bank B is going to be like ”Uh, what if Bill wants to withdraw his billion dollars in cash later, wouldn’t we have an obligation to deliver him a billion dollars in cash?” In a world without laws and regulation, bank A might say ”Don’t worry! We’ll just create an IOU out of thin air. The IOU says that we’ll provide you with a billion dollars in cash any time you want.” That doesn’t seem like a good proposition for bank B, does it? The obvious solution is that bank A needs to provide a billion dollars in cash to bank B before the transaction can be settled.

Now, imagine a world where transactions are settled at the end of each day. Banks A and B have many customers who make many transactions every day. In order to settle the day’s worth of transactions, we need to count the net amount for the day. On some days more money is moved from bank A towards bank B, and on some days the opposite is true. But how? Are these banks going to be driving truckloads of cash between each others’ offices every single night? That doesn’t sound good.

What if you had an entity that could just hold a bunch of cash in a vault and keep accounts of how much money bank A and bank B own at any given time? This way there would be no need to drive actual trucks of cash every night. Banks could simply ask this entity to move some IOUs from one deposit account to another. We could call it… a bank of banks. No, let’s call it a central bank.

Banks can deposit their reserves at the central bank. Reserves at the central bank appear as a line item on banks’ assets. When banks need to settle transactions of their clients, they can simply make transactions of their own at the central bank. The nice thing about a central bank is that it doesn’t actually need to have all that cash in the vault. It can simply create IOUs out of thin air, and when it runs out of cash money, it can actually print more cash to cover the IOUs it has issued.

Central banks have the power to ruin the entire economy with excessive money creation. Regular banks do not have this power.

Sometimes, due to mismanagement or corruption, a central bank may create excessive amounts of money, leading to hyperinflation and severe economic problems. In 1920’s Weimar republic (Germany), people would use wheelbarrows for grocery shopping — not because they needed help carrying groceries, but because they needed help carrying money. In 2008 the Zimbabwean central bank’s money printing reached the point where they had to print 100 trillion dollar banknotes. Funnily enough, these banknotes eventually became valuable as collectors’ items.

A central bank has this power, because its IOUs are backed by its literal ability to print more banknotes. A regular bank does not have any power comparable to this. When bank A wants to send money to bank B, bank need B needs to receive some kind of assets from bank A before it will agree to settle the transaction. Theoretically, this asset might be cash. In practice, this asset is often in the form of central bank deposits. In some cases this asset may be an IOU created by bank A out of thin air. This practice is referred to as unsecured lending between banks. Sometimes a bank is in financial trouble and no other banks is willing to lend to it. In this case the bank might lend money from the central bank (”lender of last resort”).

Even though bank A can sometimes settle transactions to other banks by using IOUs created by itself, this ability is highly constrained in practice. If bank A was trying to transfer suspiciously large amounts, other banks would quickly start to demand more trustworthy assets to settle these transactions — they would no longer accept IOUs created by the suspicious bank. This is in stark contrast to central bank’s special power to create money out of thin air.

In conclusion

This article has illustrated the following facts:

A bank deposit is merely an IOU.
A bank loan is merely an exchange of 2 IOUs.
Anyone can create IOUs out of thin air.
When a customer redeems an IOU at a bank, the bank typically needs to hand over another form of money — the kind that it cannot create out of thin air.

Corosim – a tool for modelling the COVID-19 epidemic in Finland

Sat, 16 May 2020 00:00:00 GMT

Corosim is a project I worked on during my time in Futurice. Corosim combines historical estimates and model predictions for the COVID-19 epidemic in Finland. This means you can use Corosim to get some insight towards questions such as ”how many Finns have been infected so far” or ”when will the epidemic peak”.

Our model is a classical epidemiological model (deterministic SEIR). By default the model is initialized with the latest historical estimate & you can easily configure an earlier timepoint as the starting point for the model. You can also tune parameters which impact both historical estimates and model predictions. You can also set action markers to model the effects of different policy changes or other events. All of this configuration is easily discoverable and modifiable directly in the user interface. You can tinker with parameters and see the results immediately.

Historical estimates

Historical estimates ~~are~~ were updated daily based on data provided by Helsingin Sanomat. Automatic updates have been turned off and Corosim is now frozen in time. Latest available historical estimate is frozen to 26.09.2021 (although the historical estimates ”curtain” is currently dragged to 28.07.2020, you can drag to change this timepoint). As Corosim is a fully static web app, there was no backend involved in data updates. Instead, the static web app was updated at build-time on a 6-hour timer.

We would like to emphasize that historical estimates have not been created with the model. Historical estimates are based on historical (past-looking) observations, not (forward-looking) predictions. Although the model is a legitimate epidemiological model, our historical estimates are merely simple ”back of the napkin” type calculations based on confirmed cases, confirmed deaths, and hospitalization data. For example, the estimate for the number of infected is based on the number of confirmed cases, but is also affected by various parameters, such as the percentage of undetected infections, length of the incubation period, how long individuals remain infectious, and so forth. We refer to the source code for details on historical estimates.

Corosim is unique in the way it combines model predictions with historical estimates. Other Coronavirus modelling websites begin the simulation from a theoretical ”day zero”, which is useful to simulate how a theoretical epidemic might spread in a theoretical environment. However, these ”day zero” simulations are not that useful in simulating how the current, real-life Coronavirus epidemic might spread. We are in the middle of the epidemic — long past day zero. Doesn’t it make sense to start the simulation from the most recent estimate of the current situation? That’s what Corosim does.

We have open sourced Corosim on GitHub under the MIT license. The README has instructions on how to customize Corosim for different countries or local areas. Corosim has already been customized for at least Germany and Columbia.

Model predictions

Epidemic models can be divided into compartmentalized models and agent-based models. Generally speaking, agent-based models are often stochastic and simulate individual agents (”Jack meets… rolls dice… 6 people today”) whereas compartmentalized models are often deterministic and simulate groups of people rather than individuals (”today we have 1000 infected… increases by 10%… tomorrow we have 1100 infected”). Agent-based models are often simple on math side, but require significant efforts in software engineering and computational power. Compartmentalized models require more math (typically differential equations), but they are simple to write in terms of software engineering, and they don’t require a lot of computational power. Agent-based models are closer to reality, which makes them easier to extend to different use cases. Corosim implements a compartmentalized model.

Corosim uses Gabriel Goh’s implementation of a SEIR model (Susceptible → Exposed → Infected → Removed). This is a classical infectious disease model, commonly used to this day in the front lines of research. For example, the Finnish health authority THL uses a similar SEIR odel for their official Coronavirus forecasts. One key difference between Corosim and THL’s model is that THL’s model is initialized to a theoretical ”day zero”, whereas Corosim is initialized to the latest historical estimate. Another key difference is that THL’s model is divided into age groups, Corosim’s model is not. These are certainly not the only differences between these models — unfortunately THL has not published their entire model, so we are unable to provide a thorough comparison. Most of the discussion around models in Finland seems to revolve around parameter choices, rather than models themselves.

The clinical dynamics in this model are an elaboration on SEIR that simulates the disease’s progression at a higher resolution, subdividing R into mild (patients who recover without the need for hospitalization), moderate (patients who require hospitalization but survive) and fatal (patients who require hospitalization and do not survive). Each of these variables follows its own trajectory to the final outcome, and the sum of these compartments add up to the R in SEIR.

Note that the model is a simplification of reality in many ways:

All hospitalizations are assumed to occur immediately after the infectious period.
Individuals who are recovering in home or hospital are assumed to be completely isolated.
Hospitalization duration is assumed to be the same for regular ward, ICU, and fatalities.
Icu capacity is just a visual indicator, exceeding capacity has no effect on fatalities.
All fatalities are assumed to come from hospitals. In reality, many fatalities come from nursing homes, which means that this model overestimates hospitalization and ICU counts. Note that the model does not necessarily underestimate fatalities; the fatality rate can be adjusted to take into account all deaths, regardless of where they occur.

Attribution

Corosim is a fork of Gabriel Goh’s Epidemic Calculator.

Source code available on GitHub.

Key differences between Corosim and Epidemic Calculator:

Historical estimates. The original Epidemic Calculator initiates the simulation from a theoretical ”day zero”. Corosim initiates the simulation from the latest historical estimate. Estimates ~~are~~ were updated daily.
Corosim is tailored to the current situation in Finland. In addition to Finnish historical data, all the parameter default values have been chosen based on latest scientific research (as of 2020/05), and specific to Finland when applicable. For example, the most crucial parameter in this model is R. It’s constantly changing and it’s specific to the population which you are trying to model (meaning, the R for Italy will be different than the R for Finland).
User-facing states are different (e.g. infected vs. infectious). The old Epidemic Calculator is a great educational tool about the progression of epidemics in general, but our focus was on practical real-world questions related to this epidemic right now. The states we have chosen to visualize are relevant for practical questions, such as ”how many people are infected” or ”do we have enough health care capacity”.
Multiple action markers. The old Epidemic Calculator only has a single action marker, labeled ”intervention” and it can only reduce the transmission of the virus, not increase it. What if you wanted to model the effect of stopping an intervention? How about modelling multiple policy changes? You can do those things with Corosim. Action markers can be added, deleted, dragged, renamed, and their effect on disease transmission can be tuned.
Scenario outcome summary. The old Epidemic Calculator does not have an easy way summarize an outcome. If you want to compare two different strategies, you need to manually zoom out and eyeball the peak, fatalities, etc. Corosim provides a scenario outcome summary of the most crucial metrics.
Small changes to the model itself: hospitalizations go to hospital without delay and fatalities are affected by the same hospitalization time as recovering patients. These simplifications were motivated by a desire to make the parameterization easier to communicate to the end user. For example, in the original Epidemic Calculator there is a parameter labelled ”Length of hospital stay”, but it actually affects only those patients who eventually survive, not those who eventually die. It would have been difficult to convey the meaning of this parameter to end users, if we would have maintained it as it were. We noticed similar issues with parameters ”Time to hospitalization” and ”Case fatality rate”. In the original Epidemic Calculator, case fatality rate actually affects the hospitalization rate, but this effect is hidden from the end user.
Various design and UX improvements (real dates, more tooltips, reduced clutter, etc.)

A comparison of voting schemes' security properties

Wed, 02 Oct 2019 00:00:00 GMT

Voting is a uniquely difficult problem with seemingly contradictory requirements. On one hand, in order to prevent coercion and vote-buying, we must ensure the secrecy of the vote. On the other hand, in order to prevent manipulation of the results, we must ensure that votes are counted correctly. How can we achieve both goals at the same time?

I explore these issues in my Master’s thesis,

”A framework for comparing the security of voting schemes“.

Abstract:

We present a new framework to evaluate the security of voting schemes. We utilize the framework to compare a wide range of voting schemes, including practical schemes in real-world use and academic schemes with interesting theoretical properties. In the end we present our results in a neat comparison table.

We strive to be unambiguous: we specify our threat model, assumptions and scope, we give definitions to the terms that we use, we explain every conclusion that we draw, and we make an effort to describe complex ideas in as simple terms as possible.

We attempt to consolidate all important security properties from literature into a coherent framework. These properties are intended to curtail vote-buying and coercion, promote verifiability and dispute resolution, and prevent denial-of-service attacks. Our framework may be considered novel in that trust assumptions are an output of the framework, not an input. This means that our framework answers questions such as ”how many authorities have to collude in order to violate ballot secrecy in the Finnish paper voting scheme?”

Results of the comparison:

You can download the thesis below. I’d be happy to receive any comments or corrections. I will fix any discovered errors and I pledge to keep both updated and historical versions available on this site. I have released the thesis under a permissive license (CC BY 4.0), so you are free to do what you want with it.

v-2019-10-09 ← latest version submitted to University
v-2019-10-01

PayPal sends phishing emails to its customers

Tue, 10 Sep 2019 00:00:00 GMT

I rarely use PayPal for anything, so imagine my shock when PayPal sent me an email asking me to “review my recent transactions now”. Oh my god, has someone hacked my PayPal account and is instigating fraudulent transactions?! I should definitely review those transactions right now by following this link in the email and then inserting my username and password to whichever page opens up. Although text in the email indicates it’s linking to paypal.fi, clicking on it actually points to paypal-communication.com. That may sound risky, but I’m willing to do anything to keep my account safe.

If you’ve ever been to the internet, you probably recognize this email as a phishing attempt. You would be wrong. This is a legitimate email from PayPal.^[1] Because PayPal, apparently, wants to normalize phishing emails and risky user behavior. But don’t take my word for it. Take their word instead:

How to spot fake, fraudulent, spoof, or phishing emails

Emails from PayPal will always address you by your first and last names or by your business name.

Silly me. Here I was, thinking that the way to spot a fake email is by looking at the sender field (and trusting that a massive financial institution is able to set up DKIM). No, what was I thinking, that’s way too complicated. Instead, let’s just tell our users to literally trust anyone who knows their name.

TracedSVG image placeholders

Wed, 21 Aug 2019 00:00:00 GMT

You need image placeholders on your website in order to stop the page from jumping around when images load. A single color placeholder will do the job, but if you want a polished look, you need something which resembles the actual image to make the transition smoother. And the placeholder image needs to be inlined to HTML so that it can be rendered immediately.

Some options for fancy placeholders:

Blur-up of low-quality JPG version of the original image
TracedSVG
SVG using shapes

Medium popularized the “blur up” placeholder technique and it seems to be the most common choice at the moment. Personally I don’t enjoy looking at blurry images, but I guess many people like it. The advantage of SVGs is that you don’t have to blur them, since a tiny (kb) SVG will render smoothly at any (pixel) size. This means that the placeholder looks sharp even if it’s a tiny file displayed as a huge image. One way of using SVG as placeholders is recreating the original image using geometric shapes. I know this sounds weird, but it looks amazing. This website uses a different method to create the SVG: tracing the outlines from the image. One advantage of tracedSVG over the “shapes” method is that we can paint the output SVG with the same colors as the rest of the website, so it blends into the site better.

Below is an autogenerated gallery which contains all images from my blog posts. Hover over the images to see the placeholders.

If it weren't for those meddling cryptographers!

Thu, 15 Aug 2019 00:00:00 GMT

This is a story about how you should not respond to a vulnerability disclosure.

Back in 2015 a man-in-the-middle (MITM) vulnerability was discovered in an Australian internet voting system called iVote®. Things were heating up because the discovery was made while the system was used by hundreds of thousands of people in a real election. The New South Wales Electoral Commission (NSWEC) didn’t take kindly to meddling cryptographers. They responded in more than few words. This blog post is a commentary on that response.

Just to clarify the severity of the vulnerability: it allowed an attacker to tamper with votes, but they needed a privileged network position to do so (for example, the operator of a WiFi hotspot or an Internet Service Provider could perform the attack). (The vulnerability was caused by loading an analytics JS from a poorly configured external web server. Due to the poor configuration, a MITM would be able to downgrade the encryption, break the weakened encryption, and then serve any JS instead of the real JS file.)

Let’s begin.

This response concerns allegations made during the NSW 2015 state election regarding the security and integrity of the Commission’s iVote®internet voting system by researchers Dr Halderman (University of Michigan – USA) and Dr Teague (University of Melbourne).

“Allegations”. I can see we’re off to a good start here. The NSWEC then goes on a rampage about how their authority is not being respected:

Teague and Halderman sent a report […] to media outlets and CERT Australia […] The correct agency for Drs Teague and Halderman to have reported this incident was in fact the NSW Electoral Commission […] CERT Australia is a federal government agency, it is unable to deal with a state government agency such as NSW Electoral Commission.

Let me get this straight: the national Computer Emergency Response Team was not the correct agency to report this emergency, because this was only a state level emergency, not a federal one. If you are wondering how things would have turned out if the researchers had contacted NSWEC only, wonder no more:

ABC contacted the Commission […] did NOT advise the Commission they were operating under an embargo […] the Commission understood that publication of a story on the content of the report was imminent and not conditional on remediation of the system. Realising this and knowing the potential concern such a story would have on public confidence, particularly if the report’s proposed remediation action had not been undertaken. The Commission knew it had no option but to remove the Piwik link to the iVote® system immediately. Therefore the removal of the link was done expeditiously and without the benefit of an independent risk assessment being performed.

Here’s a free risk assessment for you: The downside of removing that one line of code is you lose analytics. The upside is MITM attackers can’t manipulate votes on your internet voting system. NSWEC is straight up telling us that without the media pressure they would have left the vulnerability unfixed while they would conduct their “independent risk assessment”.

Then the ABC, without advising the Commission, ran the story […] This news story, gave a false impression of the iVote® system’s operation to the public […]

Yeah, false how exactly? They don’t say. Oh, and HOW DARE the media report on this without our permission!

Drs Teague and Halderman were aware that the related Piwik website was used only to provide statistics regarding the iVote core voting system’s operation and was not essential to its main function of receiving votes.

That’s a clever bit of misdirection right there. Yes, the NSWEC intended the JS dependency (“Piwik website”) to be used only for collecting analytics. That’s what they want to highlight for the reader. Not the fact that a MITM attacker could leverage the same dependency to manipulate votes.

The vulnerability was found by Drs Teague and Halderman using a free internet based certificate test service […] The relatively poor certificate test result […]

What a curious choice of words. Relative to what? The test result — an F — was poor relative to other possible test results, namely A, B, C, D and E. I gotta hand it to them, they’re technically right. Everything is relative after all.

We have been advised that the likelihood of someone intercepting online votes using this approach is similar to that of a malicious postman replacing a postal vote.

That’s a false equivalence if I’ve ever seen one. A rogue postman will be able to replace a few votes at best. MITM attacks on a computerized voting system scale quite differently: the effort to replace a single vote with this attack is roughly the same as the effort to change a lot of votes. For example, a large Internet Service Provider (or a rogue employee) that executes this attack would be able to change all of the votes sent via their connections, compared to just a few votes that a postman would be able to change.

Should the attack […] have actually occurred the Commission would have reasonably expected that our verification service would have alerted affected voters who would have contacted the Commission. Some 1.7% of electors who voted using iVote® also used the verification service and none of them identified any anomalies with their vote.

More than a year after the election it turned out that this claim was false. A staggering 10% of verification attempts had failed to retrieve any vote. It is unclear whether the NSWEC was aware of the successful verification rate and lied, or was oblivious to the successful verification rate and lied.

The NSWEC then goes on to reveal the “hidden agenda” of Teague and Halderman:

Teague and Halderman are advisory board members of the US based anti-internet voting lobby group Verified Voting […] Halderman and Teague are anti-internet voting activists […] Halderman has supported a statement […] Teague has endorsed […] She has also said […] Halderman has also undertaken activism […] Halderman is well known for his dramatisation of security issues […] Teague and Halderman typically do not disclose to the public their affiliation with US based anti-internet voting lobby group Verified Voting when making media statements on this subject.

In other news, anti-heroin activists are critizing a strain of heroin. They are claiming that this particular strain is bad, when in fact they oppose all strains of heroin! We have uncovered their hidden agenda by looking at their public affiliations, board memberships in anti-heroin organizations, and statements to the media made over a number of years, all of which they attempted to hide from the public eye.

The attacks on Teague and Halderman continue:

Some of the statements they have made appear to fall outside their core areas of research. In particular they recommended in “The Conversation” that electors “stick with an old-fashioned paper ballot”. The Commission is not aware of any research done by Drs Halderman and Teague which assesses the comparative risks of internet voting against paper voting for NSW elections. We therefore believe this statement is more likely a strongly held personal view rather than a product of peer reviewed and evidence based research, either conducted by them or other reputable researchers.

Your Honor, the witness may be an expert in electronic voting systems, an expert in paper-based voting systems, and sure, she may have implemented some voting systems specifically in Australia, but let me ask the court, has she published any peer-reviewed studies on the comparative risks of internet voting against paper voting in New South Wales specifically? I rest my case.

The Commission’s view is that internet voting is not a problem for academic cryptographers to solve, but rather an evolving technology requiring a broad range of technical and electoral skills and engagement with electoral stakeholders.

Although the entire article is a comedy goldmine, this one takes the cake. We want a broad range of technical and other skills, except for technical skills in cryptography, which we don’t want, because they’d just reveal how horribly insecure our internet voting system is. And remind me again, what are problems for cryptographers exactly, since internet voting isn’t one of them?

This reminds me of a quote by another Australian politician, Malcolm Turnbull:

The laws of mathematics are very commendable, but the only law that applies in Australia is the law of Australia.

I don’t know what it is with Australian politicians. Do they all think math just goes away if you outlaw it or pretend it doesn’t exist? If a tree falls in the forest and no cryptographers are around to hear it, is the New South Wales iVote® system still insecure?

How to set up email

Thu, 08 Aug 2019 00:00:00 GMT

Mankind has eradicated polio and sent a Tesla to orbit, and yet we haven’t figured out how to reliably send messages over the internet. It’s a weird world, I know, but it’s the one we’ve got. This article offers some practical advice to maximize your chances of receiving and delivering email. Let’s start with the most obvious choice for email and then look at other options.

Why you shouldn’t use Gmail/Outlook

If you use Gmail/Outlook, you will not receive 100% of your legitimate incoming email. Some of the email will be bounced with a notification to the sender, some will be accepted and placed in your spam folder where you have 30 days or so to catch it, and some will be simply accepted and dropped in a black hole, never to be seen again.

When I’ve talked about this in the past, a common reaction has been “what, none of my email has ever disappeared”. I have two counter-arguments to this statement: first of all, how would you know? Secondly, when you actually measure delivery rates, you don’t find a magical email unicorn that delivers 100% of legitimate email. That’s the sending side. On the receiving end it’s possible to get to 100%, just not with aggressively spam-filtering services like Gmail. There’s also many other reasons to avoid Gmail, but we’ll ignore those for now.

Why you should rent a custom domain

Even if you use Gmail as a host, rent a custom domain. You want to minimize the risk of losing access to your email addresses. If you are using a custom domain and Gmail decides to ban you for life because you accidentally clicked on a Google+ link, or any other reason that may or may not make sense, at least you can migrate your email addresses onto a new host.

One of the perks of renting a custom domain is that you can generate unique on-the-fly addresses. For example, if you need to give up your email address for pizza delivery, you might use pizza139432@yourdomain.fi. If you ever start receiving spam into that address, you can simply close the address. Plus you get to know which service sold your contact info. But the absolutely best part about on-the-fly addresses is that you don’t need spam filtering, which means you get to receive 100% of your legitimate incoming email (if you run your own server or if you use a host which allows you to turn off spam filtering). This may sound like a chore, but it’s not. You can set up a catch-all mailbox which receives email from all unregistered addresses under your domain. Some software also allows you to edit the sender field on the fly (for example, Thunderbird).

You might be thinking “hey, I can do the same thing with Gmail’s + suffix”. No you can’t. You can’t turn off Gmail’s spam filtering. The main purpose of on-the-fly addresses is to serve as a spam filter with a 0 false positive rate. If you use on-the-fly addresses together with aggressive spam filtering, what’s the point?

How to rent a custom domain

Choose a domain registrar that isn’t known for accidentally losing domains and isn’t intentionally scamming its customers. Choose a top-level domain that is not plagued by incompetence (.io) and doesn’t get seized for political reasons (.com, .eu). As a Finnish citizen I went with .fi, one of the more reliable top-level domains out there.

Incoming mail

By use case:

Personal: if you set up your own server to receive email, you will be able to receive 100% of email sent to you. You can get pretty close by using a commercial email host which doesn’t do aggressive spam filtering. I’m not sure if any email host allows you to completely turn off spam filtering.
Business: you obviously can’t use on-the-fly addresses because your business needs to have stable, public email addresses. You should not set up your own server. Just live with the crappy spam filtering by commercial email providers.

Outgoing mail

By use case:

Personal: Gmail and other large providers have pretty good deliverability. The key is sending your email from a server which is sending a lot of email with a very low percentage of spam in it.
Transactional: Postmark provides excellent deliverability for transactional email.
Marketing: Do not mix your marketing email with your transactional or personal email. Your marketing email is not going to get good deliverability and it’s not supposed to get good deliverability. If you use a personal/transactional email platform to send marketing email, you also hurt the deliverability of other peoples’ important mail.

Email client recommendations

Beware of traps. For example, BlueMail is a highly rated Android app made by a company that is not selling any services publically. Furthermore, the developers behind BlueMail prefer to remain anonymous. You know, good samaritans who like to spend their days anonymously polishing an email client that they just give out for free. It’s still in the app store, even though it’s been caught sending users’ email data to BlueMail servers with no explanation why and also caught sending users’ email login credentials to BlueMail servers.

I recommend using Thunderbird for desktop and K-9 Mail for Android. I don’t know if there are trustworthy clients for the iPhone. Sorry.

Email host recommendations

ProtonMail offers custom domain catch-all for 6.25€/month, although you have to pay more if you want to send emails.
Fastmail offers custom domain catch-all for $5/month, although you have to manually set up addresses to send emails.
Migadu offers custom domain catch-all for 4€/month, although you have to manually set up addresses to send emails. This company might be one person’s overgrown hobby project, which represents an additional security risk compared to ProtonMail and FastMail.

Disclaimer: This site is not monetized in any way. There are no advertisements and no affiliate links. I do not get paid for these recommendations in any way.

In conclusion

Do you enjoy roleplaying sysadmin on your free time?

YES → Set up your private server to receive e-mail and piggyback off a large corporation to send e-mail.
NO → Use Migadu/Protonmail/Fastmail.

Furthermore,

Get a custom domain.
Use an e-mail client that doesn’t steal your private data (Thunderbird/K-9 Mail)
If you need to send transactional e-mails, use Postmark.
If you need to send marketing e-mails, use a platform that’s intended for marketing e-mails.

Don't send email from your own server

Mon, 01 Jul 2019 00:00:00 GMT

It’s fine to run your own server to receive email, but save yourself a headache, don’t try to send any. This is a cautionary tale about not believing things you read on the internet. It’s also a story about a decentralized communication protocol that was widely adopted by the entire world, only to be slowly suffocated to death by a giant adtech company. But first and foremost, it’s a story about me trying to send email to my mom.

Prologue

It was a cold morning in 2016. After two decades of using email like a normal person I had finally had enough. I was going to set up my own server. And I was going to do it perfectly, so I wouldn’t have to deal with these problems that were caused by other people running my email server. What problems, you ask? Two words: spam filtering.

One time I would reply to my mom’s email — a Gmail address — and a week later she would call to check that I’m alive because she never got my email. Then I would call my email service provider, which at the time was my ISP, and they would tell me that Gmail blacklisted them for a few days because someone was using their servers to send spam. ”But don’t worry”, they would say, ”it works now”. Sure, maybe it works today. And maybe tomorrow my mom has to worry again because someone didn’t set up bounce notifications. (Fun fact: later this ISP just closed my email address because they were ending B2C services in anticipation of a buyout.)

Have you ever registered to a service and wondered why you had to wait 2 hours for the confirmation email? This used to be a common issue for me, and I now know it was caused by an anti-spam method called graylisting. Graylisting means that incoming email from unfamiliar servers is bounced with a note that says ”please try again later”. The sending server then tries again. And again. And again. After a few hours the receiving server says ”ok I guess you’re not a spammer since you went through all this trouble” and the email finally lands in the inbox. Uh, ok, so… first we wrapped optical fiber around the entire planet so we could send near-instant messages across the world. Then someone looked at this system and said “I know how we can make it better: let’s create a protocol that delays messages for hours!” Progress, I guess?

My brand new email server

The first thing you need to know about setting up an email server is that it’s not just the original email protocol. It’s a whole dance. If you want the cool kids to play with you, you need to configure a bunch of different things like reverse DNS, DKIM, SPF, HELO, and about 12 other protocols which I’m actively trying to forget and will not be enumerating at this time. After that you need to sacrifice a goat and perform a rain dance. If your email is still not delivered you might as well pray to the rain gods, because the cool kids (Gmail and Outlook) sure as hell are not going to tell you what you did wrong.

So I did the whole dance thing. Then I used a bunch of services like DKIMValidator, mail-tester and Glockapps to validate that my server is operating correctly. I fixed every tiny thing that they indicated may be affecting my deliverability negatively. I continued until there was nothing left to fix. Despite these efforts, my personal email wasn’t getting past spam filters. This was unfathomable to me: I’m the only person sending email from this dedicated domain and dedicated IP, I only send personal email, and somehow, those emails are being labelled as spam.

Maybe someone has gained access to my server and is using it to spam without my knowledge? I know this is not the case because I’ve configured DMARC and Gmail is sending me DMARC reports which show 1 outgoing email from my server to Gmail on the days that I email my mom. I’ve also periodically checked that my domain and IP are not on any spammer blacklists.

Rain gods, do you hear my prayer?

Both Outlook and Gmail provide forms to report deliverability issues. Gmail didn’t answer my prayers. Outlook answered that I’m “not eligible for mitigation”. They wouldn’t say why, but they were kind enough to provide some helpful tips, like “ensure your email lists are up to date”. I tried to explain to them, again, that I mostly email my mom and I don’t have any email lists because I’m not a spammer. They replied “after reviewing” my case that I should “ensure that the unsubscribe process is visible”, among other tips which they thought would be helpful to people who try to send personal emails to their loved ones.

The cool kids also offer some interactive tools to monitor email deliverability and help troubleshoot issues. Gmail’s offering is called Postmaster tools whereas Outlook offers Junk Mail Reporting Program and Smart Network Data Services Program. I joined all of these. None of them show me any data, because I am not a ~~spammer~~ large volume sender.

Gmail offers this article for those who are interested in improving their deliverability to Gmail. The title of the article is ~~Spammer~~ Bulk Sender Guidelines, which gives you some indication of whose deliverability Gmail wants to improve.

Here is a quote from Migadu, a small email provider:

In some cases recipient servers intentionally rejected correct emails just because we are a low volume sender. Ironically that is how an ideal sender should be.

Wait, what? Why would anybody do that? AWS docs seem to confirm this:

If you use dedicated IP addresses, it is your responsibility to maintain your sender reputation by sending consistent and predictable volumes of email […] you must warm up those addresses by sending an amount of email that gradually increases every day […] Once your dedicated IP addresses are warmed up, you must then maintain a

consistent sending pattern […] Most internet service providers (ISPs) only track the reputation of a given IP address if they receive a significant volume of mail from that address.

I’m sure you’ve noticed the pattern already, but I’m going to spell it out anyway: Gmail and Outlook are not delivering clearly legitimate email from low-volume senders. Deliverability tools, mitigations and guidelines are available for ~~spammers~~ large-volume senders only. This is pretty much what you might expect in a dystopia where the largest email providers are huge adtech companies.

I give up

Over the course of 2 years that I was actively using my own server to send email from the same domain and same IP, there were periods of time when my emails were landing in Gmail’s inbox. And then there were periods of time when my emails were placed in spam, sometimes outright bounced, and sometimes received and dropped without even landing in the spam folder. You could never know what Gmail would do with your email. Outlook was worse. During those 2 years I was never able to land an email to Outlook’s inbox. Not even to my girlfriend, who frequently emailed me back from the very same address. Yes, Outlook’s spam filtering is somehow even worse than Gmail’s.

But did you do the thing?

Yes I did do the thing. Yes I did join DNSWL. Yes I did configure return-path header to match reply-to header. Yes I did tell my server it’s been naughty when it’s been telling other servers it’s from 127.0.0.1. Yes I did configure SPF 3.0, and that isn’t even a real thing, and you wouldn’t even have noticed that I just made it up, because email deliverability is a dystopian hellscape.

For some reason many people are in disbelief when I tell this story. A common response is “did you do this one thing where you twirl counter-clockwise while you are shouting the rain god’s name and…” YES I DID, and I also tried twirling clockwise, just in case that would work, and also why is this a thing? Just for the sake of argument, let’s suppose that I did do something wrong. Let’s suppose that it is possible to deliver emails to Outlook and Gmail from your personal email server. If it’s this difficult to achieve, does it seem like a worthwhile endeavour?

I’m not the only one with this problem. Periodically a thread about email deliverability blows up on Hacker News and a lot of people are ranting about the same issue. Here’s one. These threads also seem to bring personal email evangelists out of the woodwork, saying things like “It’s easy to get your email delivered. Just do this one thing…” Don’t believe everything you see on the internet, even when it’s repeated by many reputable sources. These people mean well, but they are not presenting any evidence, which leads me to the obvious conclusion that they don’t actually measure their deliverability. They just feel like their email is delivered. When the occasional mail is lost, it’s easy to assume that the intended receiver simply chose not to write back. And you know what, maybe one of them actually has the secret sauce to deliver email, IP from the holy maiden IP block, a domain touched by a lucky rabbit’s foot, whatever. I ask again: does this seem like a worthwhile endeavour?

Disclaimers: The dystopian hellscape joke is stolen from John Oliver. Gmail’s spam filtering almost always let me email my mom, the real trouble was emailing new contacts.

Argument framing with "This time is different"

Mon, 22 Apr 2019 00:00:00 GMT

There’s an old saying that extraordinary claims require extraordinary evidence. For example, if two individuals are debating whether the sun will rise tomorrow or not, the burden of proof is on the party claiming the sun will not rise — and they would need some pretty spectacular evidence to convince anyone. However, when we move outside toy examples, it’s often hard to differentiate extraordinary claims from those which are merely ordinary. This opens up a new rhetoric tool in debates: framing the opposing argument as extraordinary to magically shift the burden of proof from your shoulders. This way you might convince people of your position without actually providing evidence to support it.

A recent McKinsey Global Institute study showed 60% of all jobs could easily see 30% of their work fall to machines. But what if we’re wrong about AI? What if it’s not the end of work but the beginning of a massive job boom unlike anything we’ve ever seen in history? I know. I know. This time is different. I’m sure the loom weavers in 19th century England thought the same. So did the stone masons, whose secret knowledge of how to put together glorious temples disappeared and yet today we create even more amazing buildings that stretch up into the endless sky. Every generation imagines that it’s radically different from the last but what if it’s not?

The quote above is Daniel Jeffries on Hackernoon arguing “Why AI will bring an explosion of new jobs”. Not only did he frame the opposing argument with “this time is different”, but he also did it in a clever way where it was seemingly accepted by everyone.

CGPGrey articulates a nice counterpoint to job creation by technological progress:

Imagine a pair of horses in the early 1900s talking about technology. One worries all these new mechanical muscles will make horses unnecessary. The other reminds him that everything so far has made their lives easier — remember all that farm work? Remember running coast-to-coast delivering mail? Remember riding into battle? All terrible. These city jobs are pretty cushy — and with so many humans in the cities there are more jobs for horses than ever. Even if this car thingy takes off you might say, there will be new jobs for horses we can’t imagine.

But you, dear viewer, from beyond 2000 know what happened — there are still working horses, but nothing like before. The horse population peaked in 1915 — from that point on it was nothing but down. There isn’t a rule of economics that says better technology makes more, better jobs for horses. It sounds shockingly dumb to even say that out loud, but swap horses for humans and suddenly people think it sounds about right.

Unfortunately, CGPGrey accepts the premise that his claim is extraordinary. He’s arguing that this time is different. There’s no need to do that. In addition to horses, we used to employ pigeons to send messages. We employed mules to carry heavy loads over distances. The demand for pigeons and mules and horses in the job market is a fraction of what it used to be. In fact, human employment seems like an anomaly in the grand scheme of things. All other species have already been displaced in the job market by technology. To say that this one species will not be displaced is like saying… this time is different.

How can we have the opposing sides of an argument successfully framing the other as extraordinary? And they were both right, kind of. It’s just that when you spin a narrative around an argument, you have a lot of freedom in what kind of narrative you spin, and some things (like a particular claim being extraordinary) may be true within one particular narrative, but not necessarily within a different narrative. Here we had 2 narratives around the same topic: one narrative made it sound like it would be extraordinary if technology would make humans obsolete, whereas the other narrative made it sound like the opposite was true.

So what’s the takeaway from all this? Since it’s possible to frame both sides of an argument with “this time is different”, maybe we shouldn’t give too much weight to those magic words. Let’s put more weight to actual evidence instead. In the case of technology displacing human jobs, we know very well which jobs are disappearing right now (“evidence”), and we have pretty good arguments why many jobs will disappear in the future (more evidence).

Everybody agrees that many jobs are disappearing permanently — the dispute is whether new jobs will always continue to replace the lost ones. Where’s the evidence to support that claim? The people holding this position are not providing any. They’re just spinning a narrative and asking if this time will be different — but we just saw how you can do the same trick with the opposite position. Perhaps it’s time we start worrying about those jobs already?

Stop bikeshedding AI ethics

Sat, 06 Apr 2019 00:00:00 GMT

A self driving car is going down a narrow path when suddenly a baby and an elderly person appear on the road. There is no time to stop and not enough room to turn so the car must choose which of them to run over.

You’ve probably heard a variation of this dilemma many times in many places. But you know where you didn’t hear it? At driving school. That’s right. The institution solely dedicated to training people how to drive safely on the roads didn’t think it was important to teach you how to choose between running over a baby versus running over a grandmum. Do they think that humans will instinctively make the right choice? Let’s be honest: this item was never on the agenda of any meeting at any driving school.

If we are put in a life-or-death situation where we have to make a split second decision, we will likely prioritize our own wellbeing and that of our next of kin over the wellbeing of strangers. Random factors will likely affect our decision a lot as well. We’re not going to produce a balanced, nuanced ethical analysis of the situation in the blink of an eye. And that’s fine. There are more important things to teach at driving school, like how to recognize those pesky symbols that they put up on the side of the road, when it’s your turn to drive through an intersection, or why you should slow down in conditions of reduced visibility. Likewise, there are more important things to get right in a self driving system, like I don’t know, basic object recognition would be a good start.

So if this is not the most important ethics question of the decade, why do people keep bringing it up? Self driving cars are scary, they will bring major change to society, and some serious ethical dilemmas along with that. We want to discuss those ethical questions, maybe even prevent some of the worst catastrophies that may happen, but that’s too difficult. So we talk about something easy instead: baby or grandmum?

This is called bikeshedding. The origin story for this term is quite funny:

Parkinson noticed that a committee whose job is to approve plans for a nuclear power plant may spend the majority of its time on relatively unimportant but easy-to-grasp issues, such as what materials to use for the staff bikeshed, while neglecting the design of the power plant itself, which is far more important but also far more difficult to criticize constructively.

Self driving cars do come with seriously impactful ethical dilemmas, but nobody talks about those, because they are really, really hard and you can’t become an “AI ethics expert” by waiving your hands around them. I’ll name one: when is it ethical to bring the product to market? At any point in time we can say that this self driving system would be safer after 1 more year of development, but we can’t just wait forever - we have to draw a line in the sand somewhere and say that “this expected amount of deaths and injuries is acceptable, we can bring the product to market now”.

Problem is that there is a mismatch in incentives when the company decides where to draw that line. Companies have the incentive to bring their products to market too early because much of the cost resulting from those decisions is externalized (you know, to the people who die), whereas much of the gains are internalized (the that the company gains dwarfs any societal value from having self driving cars on the roads at this level of technology). Note that this is not a theoretical problem. We have self driving cars on the roads right now without basic object recognition capability. This is an actual problem causing actual deaths.^[1][2]

How do you solve this AI ethics problem? I have no idea. It’s a hard problem. But what if I were to tell you that the employees who code these AI systems often cycle to work and don’t have any kind of… structure that would protect their bikes from the rain? Won’t anybody think of the ~~children~~ bicycles! Now that’s something we can talk about.

How this blog is made

Fri, 05 Apr 2019 00:00:00 GMT

I wanted to have the freedom to publish my thoughts and the freedom to tinker with their presentation - without being limited by corporate politics or technical constraints. I ended up creating my own blog with GatsbyJS and publishing it via my own domain. Behind that link is a concise comparison of Gatsby against typical website setups, in case you are thinking about starting your own blog. Next I will explain the motivation behind my tech choices, and then go over the actual setup for this blog.

Medium is not a public square

A few years ago I wanted to out a particular company about bad stuff they were doing. I asked Medium if they will protect my freedom of speech when the company inevitably demands the content to be taken down. To my absolute surprise Medium said they will not, they will simply take down my content at the first sight of a legal threat. I decided not to publish the story. I haven’t heard of any examples of Medium actually censoring people, so presumably they don’t do this often. But the Internet is littered with examples of large corporations taking down speech they don’t like. It’s kind of what you’d expect, really: a large corporation will protect your freedom of speech as long as it costs them nothing and nobody powerful asks them not to.

Furthermore, Medium provides absolutely dreadful user experience. If you open a Medium article on mobile, more than half of your screen real estate will be covered by pop-ups. Why are you asking me to subscribe to an author before I’ve even had a chance to glimpse at their article? No I don’t want to get notifications, why would anybody want those from a blogging site? No I don’t agree to your user tracking, even though you only have an “agree” button as pseudo-GDPR-compliance. I just came here to read an article, can I just do that?

It’s absolutely astonishing that Medium has become the de facto blogging platform in recent years.

Wordpress is fine, I guess

If you just want to write, WordPress is not a bad choice. You can self-host. You can tinker a bit. If you want to tinker a lot, it’s much easier to just create your own website. My main issue with CMS’s like WordPress, Ghost, etc. is that while you theoretically own your own content, it’s going to be in a format which will be difficult to move to another platform. Sure you can theoretically do it, and Gatsby actually has source plugins for WordPress and Ghost, but in practice you won’t move. You’ll be stuck with whichever limitations of WordPress annoy you. When you create your website by yourself instead of using a platform, you won’t really have limitations other than the ones imposed by your own tech choices, which you can always change.

Why this blog is statically generated

GatsbyJS is a static web site generator. This means that my blog is generated into HTML,JS,CSS at build time and I don’t have a costly server dynamically generating pages. The main advantages of a static site generator are:

Hosting the blog costs nothing at providers like Netlify
Improve loading times to most places on earth because assets can be delivered by CDN (instead of a single server in a single location, you have many servers in many locations and any request is routed to the nearest one).
No server maintenance is required

I write my blog posts in Markdown. It’s really easy to format and my content will not be married to any platform. For example, if I want to ditch Gatsby in the future and create my own static site generator, I can do that knowing that transforming Markdown isn’t going to give me too much trouble.

Why GatsbyJS specifically

There are many static web site generators out there. I didn’t do a comparison. I like that Gatsby is made for React (instead of reinventing the wheel in that aspect). Their developer community seems active and friendly. But the main thing that attracted me to Gatsby was its promise of delivering performance out of the box. If you have any non trivial website and you want good performance, you have to do a ton of optimizations to get that. The idea that a framework can do those optimizations for you is really appealing. Gatsby does a lot of cool things, but this promise didn’t really turn out to be true (I ended up spending a lot of time optimizing the performance and some things still don’t work as they should. For example, Gatsby generates duplicates of some images, so when an image is used on multiple pages, the user downloads the same image multiple times from different paths without leveraging browser’s cache). Despite its faults, this site now has pretty sweet performance with the help of Gatsby:

The website is visible and functional after only 1 round trip and ~20kB of data.
Internal links are prefetched on hover, giving the UX that subsequent pageloads load instantly.
Autogenerated tracedSVG image placeholders are stylized to create a smooth look and transition as the image loads without the page jumping around.
Also works for users who have JS disabled. A static version of the website is generated, which is rendered on first load, and then the page is hydrated into React (for users who have JS).

On one hand I want to try making my own static site generator after tinkering so much with Gatsby. On the other hand I’m really excited to see what the future holds for Gatsby and where the developers will take it within a few years.