solresol

The second case mentioned in the article is interesting because no other case about employee choice had gone to merits before.

Thus it might be the first example of AI winning a case that sets some kind of precedent.

At the very least it shows that AI can put together coherent legal arguments even when there are no close precedents or case law it can refer to. This was something a lot of lawyers didn't believe would be possible for AI to do.

In some fields, all of those statements are true already.

In a large number of few fields, all of those statements will be true in the near future (10 years maximum).

There's no reason that the Humboldtian model should be the right structure for the next century's universities, just as it wasn't the right structure for the 18th century's universities.

I got banned for violating terms of use apparently, but I'm mystified as to what I rule I broke, and appealing just vanishes into the ether.

Has anyone else got examples of reasoning traces where it uses a word in a way that no human being would?

I just proved that constraint solving problems can be encoded as p-adic linear regression problems[+], and that therefore we can use machine learning optimisation techniques to get exact answers.

So of course no journal or conference is in the least bit interested, and I'm now reformatting it for another obscure low-tier journal that no-one will ever read.

Otherwise:

- automating the translation of a Byzantine Greek work that has never been translated into English before. https://stephanos.symmachus.org

- also preparing evidence for a case against the university I sometimes work for.

[+] Linear regression, but instead of minimising the Euclidean distance, minimise the p-adic distance - https://en.wikipedia.org/wiki/P-adic_valuation

Stephanos of Byzantium wrote what we would call an encyclopedia of people and places. Most of it has been lost, but shortened versions still exist. I've set up a bot to scan a version of what we have, translate it, extract out the proper nouns and try to figure out what was lost. https://stephanos.symmachus.org/ I'm also linking it to the translation of Pausanias https://pausanias.symmachus.org that another bot is doing.

Also, trying to finish a PhD on machine learning when you want to minimise a p-adic loss.

[delayed]

> The volume of cargo carried by sailing vessels in the old days was orders of magnitude lower.

Surprisingly, no, it wasn't. I'll slightly fudge the numbers and talk in terms of proportion of world trade that was carried by ocean-going vessels (because if you double the population then it's reasonable to talk about doubling the number of ships).

The world economy was very globalised in 1913. That level of globalisation in trade wasn't matched again until the 1990s.

We're only a little more global now than we were in the age of sail.

The British navy and merchant fleet was a wonder of its era.

Writing a course for a customer on how to use Claude Code well, especially around brownfield development (working on existing code bases, not so much around vibe-coding something new).

If the "outcompeting" is possible because of Chinese government subsidies, then it's important to protect local industry from unfair competition.

It's similar to the logic behind anti-trust actions against monopolists. If the playing field isn't level, then the USA government steps in to level it.

(Whether BYD is subsidised or not is another question, but the above is the logic of protecting local industry.)

Indeed. The obvious counter-example to the claim is "rainbows" which were definitely the topic of heated scientific argument for hundreds of years (and non-scientific ones before that).

I think of it as trying to encourage the LLM to want to give answers from a particular part of the phase space. You can do it by fine tuning it to be more likely to return values from there, or you can prompt it to get into that part of the phase space. Either works, but fiddling around with prompts doesn't require all that much MLops or compute power.

That said, fine tuning small models because you have to power through vast amounts of data where a larger model might be cost ineffective -- that's completely sensible, and not really mentioned in the article.

I have been wrestling with this too. I only see two options: no tech university or AI wrangling university.

https://solresol.substack.com/p/you-can-no-longer-set-an-und...

A beautiful implementation of the most exotic sorting algorithm I'm going to teach this semester:

https://bead-sort.arithmetic.guru/

Oh, and also a PhD on p-adic machine learning, but no one will remember that.

I'm not sure what's wrong with me, but I just wasted several hours wrestling codex to make it behave.

Here's my workflow that keeps failing: - it writes some code. It looks good a first glance - I push it to github - automated tests on github show that there's a problem - go back to codex and ask it to fix it - it does stuff. It looks good again.

Now what do I do? If I ask it to push again to github, then it will often create a pull request that doesn't include stuff from the first pull request, but it's not a pull request that stacks on top of the previous pull request, it's a pull request that stacks on top of main.

When asked to write something that called out to gpt-4.1-mini, it used openai.ChatCompletion.create (!?!!?)

I just found myself using claude to fix codex's mistakes.

Trying to measure how well LLMs can make scientific hypotheses, and more generally, execute on the scientific process (as part of a pivot in my PhD).

At the time when they need to be making connections and needing patronage you want PhD students to be showing up integrity and honesty problems and asking awkward questions about the powerful people in the community?

Of course that's what should be happening, but the incentives aren't pushing in the right direction for it currently.

It seems to be the season for scraping SEC filing archives and running them through AI bots to extract information. I made this:

https://boards.industrial-linguistics.com/

It only looks at information about directors on company boards, and unlike your much better project, I don't have any clear idea how to commercialise it.

(Australia)

I received my notice from the land valuer general yesterday (for unimproved land value).

Looking at building costs for my area and what other houses are going for in the street, the valuation looks spot-on.

I tried... it started with the idea was that log loss might not be the best option for training, and maybe it should be a loss related to how wrong the predicted word was. Predicting "dog" instead of "cat" should be less penalised than predicting "running".

That turns out to be an ultrametric loss, and the derivative of an ultrametric loss is zero in a large region around any local minimum, so it can't be trained by gradient descent -- it has to be trained by search.

Punchline: it's about one million times less effective than a more traditional architecture. https://github.com/solresol/ultratree-results

When I last looked up the literature, Keto diet was one of the least effective interventions.

That is, if you follow it, I'm sure it works.

But the vast majority of people drop out of keto diets very quickly. So it's lousy advice and an unsuccessful intervention.

It's a bit like saying to a patient "you gotta sacrifice -- you should doing 3 hours a day of cardio". If they do follow through with it, it will work. But the vast majority of people won't be able to maintain doing that.

I administered the CDT to ChatGPT and got Claude to diagnose what was wrong with the "patient" based on the results.

There are signs of pre-frontal cortex damage or early stage dementia.

Writing up a paper for my PhD on machine learning with non-Euclidean loss functions.

Taking over the affairs of one of my elderly relatives now that she can't manage by herself.

A project with a medical insurer for adjudicating insurance claims using LLMs.

I wrote about this phenomenon 4 years ago: https://blog.ifost.org.au/2020/09/something-funny-is-about-t...