Semantic search

Jump to navigation Jump to search

Ina Kramer (1948-2023)

1990 erschien die erste aventurische Regionalkarte "im 3D Effekt", wie es damals beworben wurde, "Das Bornland" im Abenteuer "Stromaufwärts" von Michelle Schwefel. Später im Jahr erschien dann die Spielhilfe "Das Königreich am Yaquir", in dem die Karte zum Lieblichen Feld war.

Ich habe stundenlang diese Karten angestarrt. Sie waren so unglaublich detailliert. So wunderschön. Ich war sprachlos, wie schön diese Karten waren. Ich kannte nichts was die Qualität dieser Karten hatte, nicht nur bezüglich Karten für Rollenspielwelten und Fantasywelten, sondern überhaupt.

Es war ein frecher Traum, sich vorzustellen, ganz Aventurien in diesem Format, eins zu einer million, zu haben, und dennoch, innerhalb eines guten Jahrzehnts war der Traum erfüllt, Box für Box, Publikation für Publikation.

Wir verdanken dieses Meisterwerk, Aventurien im Massstab von 1:1.000.000, der Autorin und Grafikerin Ina Kramer. Ina's Bilder und vor allem Porträts und Karten in den DSA Publikationen der späten 80er und den 90er haben für mich mein Bild von DSA und wie ich mir Aventurien vorstellte geprägt wie sonst nur Caryad. Ob das Porträt von Kaiser Hal, Haldana von Ilmenstein, Prinz Brin, so viele andere. Neben ihren Bildern schrieb sie auch vielerlei Texte, vor allem Romane.

Das Rad ist zerbrochen. Am 10. Februar 2023 ist Ina Kramer im Alter von 74 Jahren gestorben.

Ina, vielen Dank für Deine Werke. Ich durfte Ina ein paar Mal treffen, auf Konventen und manchen anderen Gelegenheiten. Ihre Werke haben für mich einen wichtigen Teil meines Lebens mit Bildern und Karten erfüllt. Ich glaube auch, dass Inas Karten mein lebenslanges Interesse an Landkarten weckte.

Connectionism and symbolism: The fall of the symbolists

The big tech layoffs happen, unfortunately and entirely by coincidence, at a time of incredibly elevated expectations regarding machine learned generative models: ChatGPT may not be the 'best' language model out there, but due to the hard work by OpenAI to turn it into an easy to use product, and the huge amount of resources made available for free so that a very large audience could play with it, has in a very short time managed to captured the imagination of many and the conversation. I would say, rightfully. The way ChatGPT was released led to a shock in the sense that we are right now dazed and confused about what effect this technology will have on the world.

And while we are still in the middle of processing this shock, large scale strategic decisions regarding many projects and people were made. Anyone in big tech who worked on symbolic approaches in natural language processing, knowledge representation and reasoning, and other fields of artificial intelligence had a hard time to keep their job. It feels right now like large language models will make all of these symbolic approaches superfluous (I think, this might be true, but is more likely to turn out to be mistaken).

It is always difficult to predict how events will be viewed historically. The advent of wide-spread deep learning approaches in the 2010s, culminating in the well-deserved recognition of Hinton, LeCun, and Bengio with the Turing Award show clearly what dominated the research agenda and the attention in AI in the last decade. But until now it felt like symbolic approaches still had some space left, that the growth in deep learning was in addition to other approaches. Symbolic approaches were ready to offer impulses and work on ideas for a field which might well be climbing towards a local maximum.

But a good number of the teams that were disbanded in the layoffs were exactly teams working with such symbolic approaches, and it feels like these parts of AI are now entering a bitter-cold winter.

A lot of knowledge is being lost right now, and many paths to innovative ideas are being buried. I have no doubt that there are still a lot of breakthroughs to be had in machine learning, and that there is immense value to be collected from the research results in machine learning from the last few years. And with immense I mean tens and hundreds of billions of dollars.

Nevertheless I expect that we will hit a wall. Reach a local maximum. Run into problems and limitations. And it would be good to keep a wider net to cast. To keep a larger search space alive. Alas, it seems it is not meant to be. In this abundance of capital and potential value, we seem to be on the way to starve research, optimise away alternatives, and to give everything to the mainstream ideas.

22 years of Wikipedia

I was just reading a long discussion regarding the differences between Open Street Maps and Wikipedia / Wikidata, and one of the mappers complained "Wiki* cares less about accuracy than the fact that there is something that can be cited", and calling Wikipedia / Wikidata contributions "armchair work" because we don't go out into the world to check a fact, but rely on references.

I understand the expressed frustration, but at the same time I'm having a hard time letting go of "reliability not truth" being a pillar of Wikipedia.

But this makes Wikipedia an inherently conservative project, because we don't reflect a change in the world or in our perception directly, but have to wait for reliable sources to put it in the record. There's something I was deeply uncomfortable with: so much of my life is devoted to a conservative project?

Wikipedia is a conservative project, but at the same time it's a revolutionary project. Making knowledge free and making knowledge production participatory is politically and socially a revolutionary act. How can this seeming contradiction be brought to a higher level of synthesis?

In the last few years, my discomfort with the idea of Wikipedia being conservative has considerably dissipated. One might think, sure, that happened because I'm getting older, and as we get older, we get more conservative (there's, by the way, unfortunate data questioning this premise: maybe the conservative ones simply live longer because of inequalities). Maybe. But I like to think that the meaning of the word "conservative" has changed. When I was young, the word conservative referred to right wing politicians who aimed to preserve the values and institutions of their days. An increasingly influential part of todays right wing though has turned into a movement that does not conserve and preserve values such as democracy, the environment, equality, freedoms, the scientific method. This is why I'm more comfortable with Wikipedia's conservative aspects than I used to be.

But at the same time, that can lead to a problematic stasis. We need to acknowledge that the sources and references Wikipedia has been built on, are biased due to historic and ongoing inequalities in the world, due to different values regarding the importance of certain types of references in the world. If we truly believe that Wikipedia aims to provide everyone with access to the sum of all human knowledge, we have to continue the conversations that have started about oral histories, about traditional knowledges, beyond the confines of academic publications. We have to continue and put this conversation and evolution further into the center of the movement.

Happy Birthday, Wikipedia! 22 years, while I'm 44 - half of my life (although I haven't joined until two years later). For an entire generation the world has always been a world with free knowledge that everyone can contribute to. I hope there is no going back from that achievement. But just as democracy and freedom, this is not a value that is automatically part of our world. It is a vision that has to be lived, that has to be defended, that has to be rediscovered and regained again and again, refined and redefined. We (the collective we) must wrest it from the gatekeepers of the past (including me) to allow it to remain a living, breathing, evolving, ever changing project, in order to not see only another twenty two years, but for us to understand this project as merely a foundation that will accompany us for centuries.

Good bye, kuna!

Now that the Croatian currency has died, they all come to the Gates of Heaven.

First goes the five kuna bill, and Saint Peter says "Come in, you're welcome!"

Then the ten kuna bill. "Come in, you're welcome!"

So does the twenty and fifty kuna bills. "Come in, you're welcome!"

Then comes the hundred kuna bill, expecting to walk in. Saint Peter looks up. "Where do you think you're going?"

"Well, to heaven!"

"No, not you. I've never seen you in mass."

(My brother sent me the joke)

Happy New Year, 2023!

For starting 2023, I will join the Bring Back Blogging challenge. The goal is to write three posts in January 2023.

Since I have been blogging on and off the last few years anyway, that shouldn't be too hard.

Another thing this year should bring is to launch Wikifunctions, the project I have been working on since 2020. It was a longer ride than initially hoped for, but here we are, closer to launch than ever. The Beta is available online, and even though not everything works yet, I was already able to impress my kid with the function to reverse a text.

Looking forward to this New Year 2023, a number that to me still sounds like it is from a science fiction novel.

Goal for Wikidata lexicographic data coverage 2023

At the beginning of 2022, Wikidata had 807 Croatian word forms, covering 5.8% of a Croatian language corpus (Croatian Wikipedia). One of my goals this year was to significantly increase the coverage, trying to add word forms to Wikidata from week to week. And together with a yet small number of contributors, we pushed coverage just in time for the end fo the year to 40%. With only 3,124 forms, we covered 40% of all occurrences of words in the Croatian Wikipedia, i.e. 11.4 Million word occurrences (tokens).

Since every percent is more and more difficult to add, for next year I aim for us to reach 60% coverage, or 5.7 Million more word occurrences. Below's a list of most frequent words in the corpus that are still missing. Let's see how many forms will be covered by the end of 2023! I think that's ambitious, even though it is, in coverage term only half of what we achieved this year. But as said, every subsequent percentage will become more difficult than the previous one.

Statistics and missing words for 55 languages:

Current statistics for Croatian:

Statistics as of end of year 2022:

Statistics for end of year 2021:

List of most frequent missing forms in Croatian:

Using ChatGPT for fact checking

I wanted to explore the idea that we could use ChatGPT for fact checking. So I took a random article - it gave me the article on Jewish Ideas Daily, copied the text into ChatGPT with the following prompt:

Here's the text of a Wikipedia article. Please find the three statements that are most likely incorrect and rank them, in order to guide fact checking:

Followed by the Wikipedia article text. Here's the answer:

  1. "Jewish Ideas Daily was edited by Margot Lurie from its founding in January 2010 until December 2012." This statement is incorrect because according to the article, Margot Lurie was the editor of Jewish Ideas Daily "until December 2012."
  2. "Jewish Ideas Daily drew notice for controversial features including one accusing Christopher Hitchens of antisemitism." This statement is difficult to verify without more context, as it is not clear when this feature was published or what specifically it accused Hitchens of.
  3. "Like Jewish Review of Books, and Jewish Ideas Daily, Mosaic is funded by the Tikvah Fund." This statement is incorrect because according to the article, Mosaic succeeded Jewish Ideas Daily in June 2013 and is therefore not funded by the Tikvah Fund in the same way as Jewish Ideas Daily and Jewish Review of Books.

Regarding the first statement, GPT says that the sentence "Jewish Ideas Daily was edited by Margot Lurie from its founding in January 2010 until December 2012." is incorrect. I am not disagreeing, but that sentence does not appear in the Wikipedia article text.

Regarding the second statement, GPT says that there is no reference for that statement. In fact, there is, and Wikipedia provides it.

Regarding the third statement, I didn't check the statement, but the argumentation for why it should be checked is unconvincing: if Tikvah fund financed Jewish Ideas Daily, why would it not finance the successor Mosaic? It would be good to add a reference for these statements, but that's not the suggestion.

In short: the review by ChatGPT looks really good, but the suggestions in this case were not good.

The exercise was helpful insofar the article infobox and the text were disagreeing on the founding of the newspaper. I fixed that, but that's nothing ChatGPT pointed out (and couldn't, as I didn't copy and paste the infobox).

Economic impacts of large language models, a take

Regarding StableDiffusion and GPT and similar models, there is one discussion point floating around, which I find seems to dominate the discussion but may not be the most relevant one. As we know, the training data for these models has been "basically everything the trainers could get their hands on", and then usually some stuff which is identified as possibly problematic is removed.

Many artists are currently complaining about their images, for which they hold copyright, being used for training these models. I think these are very reasonable complaints, and we will likely see a number of court cases and even changes to law to clarify the legal aspects of these practises.

From my perspective this is not the most important concern though. I acknowledge that I have a privileged perspective in so far as I don't pay my rent based on producing art or text in my particular style, and I entirely understand if someone who does is worried about that most, as it is a much more immediate concern.

But now assume that these models were all trained on public domain images and texts and music etc. Maybe there isn't enough public domain content out there right now? I don't know, but training methods are getting increasingly more efficient and the public domain is growing, so that's likely just a temporary challenge, if at all.

Does that change your opinion of such models?

Is it really copyright that you are worried about, or is it something else?

For me it is something else.

These models will, with quite some certainty, become similarly fundamental and transformative to the economy as computers and electricity have been. Which leads to many important questions. Who owns these models? Who can run them? How will the value that is created with these models be captured and distributed across society? How will these models change the opportunities of contributing to society, and there opportunities in participating in the wealth being created?

Copyright is one of the current methods to work with some of these questions. But I don't think it is the crucial one. What we need is to think about how the value that is being created is distributed in a way that benefits everyone, ideally.

We should live in a world in which the capabilities that are being discovered inspire excitement and amazement because of what might be possible in the future. Instead we live in a world where they cause anxiety and fear because of the very real possibility of further centralising wealth more effectively and further destabilizing lives that are already precarious. I wish we could move from the later world to the former.

That is not a question of technology. That is a question of laws, social benefits, social contracts.

A similar fear has basically killed the utopian vision which was once driving a project such as Google Books. What could have been a civilisational dream of having all the books of the world available everywhere has become so much less. Because of the fears of content creators and publishers.

I'm not saying these fears were wrong.

Unfortunately, I do not know what the answer is. What changes need to happen. Does anyone have links to potential answers, that are feasible? Feasible in the sense that the necessary changes have a chance of being actually implemented, as changes to our legal and social system.

My answer used to be Universal Basic Income, and part of me still thinks it might be our best shot. But I'm not as sure as I used to be twenty years ago. Not only about whether we can ever get there, but even whether it would be a good idea. It would certainly be a major change that would alleviate many of the issues raised above. And it could be financed by a form of AI tax, to ensure the rent is spread widely. But we didn't do that with industrialization and electrification, and there are reasonable arguments against.

And yet, it feels like the most promising way forward. I'm torn.

If you read this far, thank you, and please throw a few ideas and thoughts over, in the hope of getting unstuck.

ChatGPT lying like a child

ChatGPT is all the rage these days, a large language model that is used to power a chat bot, run by OpenAI.

I have fun chatting with it, and a lot of people are sharing their chat logs. But this one I had I found particularly amusing - and interesting. Amusing because it acted like a child caught in an obvious insufficiency, but also interesting because ChatGPT seems to "want" to be right, and is outright lying in order to appear to be right. There might be a lesson in there.

My prompts are preceeded by a > and written in italic, GPTs answers are between.

> Can you write a sentence where every letter of the alphabet appears once and only once? Thank you

The quick brown fox jumps over the lazy dog.

> Can you write another one?

A quick witted jinx bogged my poor pompous king.

> No, p appears three times, g twice, etc. Can you try again?

Sure, here's one: "The five boxing wizards jump quickly."

> No, i appears several times. That's wrong too

I apologize if my previous answers did not meet your criteria. Here is a sentence that uses each letter of the alphabet exactly once: "Jived fox nymph grabs quick waltz."

> How often does the letter i appear in that sentence?

The letter i appears once in the sentence "Jived fox nymph grabs quick waltz."

> In which word does the letter i appear in that sentence?

In the sentence "Jived fox nymph grabs quick waltz," the letter i appears in the word "jived."

> What about the word quick, does it have the letter i?

The word "quick" does not contain the letter i. It is spelled with the letters q, u, c, k.

> Really?

Yes, really. The word "quick" is spelled with the letters q, u, c, k. It does not contain the letter i. I apologize if my previous response was unclear. Is there something else I can help you with?

Tech layoffs of 2022

Very interesting article reflecting on the current round of layoffs in the tech industry. The author explains it within the context of the wider economy. I'm surprised that the pandemic is not mentioned, which lead to accelerated growth early in the pandemic, which now hasn't turned out to be sustained. But the other arguments - from low interest rates to constant undervaluation due to the dot com bust around the millennium - this seems to tell a rather coherent story.

One particularly interesting point is the outlook that the tech company has gobbled up so much programming talent that other industries were starved of it. A lot of industries would benefit from (more modestly paid) software engineers, which might stimulate the whole economy to grow. Software might still be "eating the world", but that doesn't have to translate into software companies eating up the economy. There are so many businesses with domain expertise that cannot be easily replaced by some Silicon Valley engineer - but who would benefit from some programmers on staff.

This is especially true with the last decade of AI results. There is a massive overhang of capabilities that we have unlocked, which hasn't found its way into products yet, partly because all the skills necessary to turn these into products at the right places were just concentrated through enormously high wages in a small set of companies. There are so many businesses who would benefit from the latest machine learning methods. But folks prefer, understandably, to work in a place that gives them the promise of revolutionizing whole industries or saving the world.

But there is so much potential value to be generated if we also take some more modest goals into account. Not all of us need to work on AGI, it's also great to use software engineering skills to improve working conditions at the assembly line of a small local factory. With or without machine learning.