cabalamat

> Age moves in one direction so it has a simple merge function:

Better to use date of birth as that doesn't change at all.

>I don't think it has a 'power' switch that you can flip to cut power.

There should be.

And an entirely manual brake.

It may well be, but when I try to download the pdf it fails.

> Negative results are super important.

I agree, and this is often overlooked. Knowing what doesn't work (and why) is a massive help in searching for what does work.

My understanding of AM and Eurisko (having looked into them a decade or so ago) was that their source code hadn't been published, and that there was a dispute as to what their capabilities actually were and how much was exaggeration by Lenat.

I don't know if that's still the case. I do think that it would be worth creating systems that mix the ANN and GOFAI approaches to AI.

A subsidy for building houses in areas where they are likely to be struck by natural disasters (flooding in Florida, earthquakes in California) does not strike me as sensible.

> UK a consultant earns £100k pretax and pays approximately 30% net tax rate and 60% marginal rate.

According to https://www.moneysavingexpert.com/tax-calculator/ , someone in the UK who earns £100k takes home £67,049:

>Earn £100,000 in 2023/24 and you'll take home £67,049. This means £5,587 in your pocket a month. Over the year you'll pay £27,432 income tax and £5,519 in national insurance.

Locality of behaviour is, on the whole, a good thing.

I say "on the whole" because there are always counterexamples to any maxim. The article realises this and mentions 2: DRY and SoC.

Writing good software is often a matter of taste.

Am I the only person for whom this looks like horrible, over-complex, write-only spaghetti code?

> if they wanted to ban behavioral advertising outright they should have and could have done that instead of what they have done

Yes they could, so the question is why didn't they?

I suspect the answer is that there are lots of factions within the EU, some want to ban it and some don't and what they've done is a compromise.

This is true for most READMEs on github projects. Is it really too hard to write a paragraph giving the context?

> If it senses another (somewhat famous) protein, PD-L1, the immune cell self destructs

So why don't pathogens evolve to display PD-L1 on their surfaces, to make immune cells harmless against them? Or do they?

> that's trying to establish an initial niche

Which suggests to me that while kagi might be good now, if it gets popular it may undergo enshittification in an effort to monetise its popularity.

> I think Haskell will be one of the "good ones" amongst target languages, eventually

Another advantage of Haskell is that its side-effect-free nature makes it harder for a rogue AI to put in malicious code, thus helping to solve the alignment problem.

It may be that languages and OSes will be specifically designed to put everything in a sandbox (or multiple layers of sandboxing), which will also be helpful when humans write code.

> while less-used languages tend to produce worse results

I figured that might be the case.

> Haskell may or may not need a manually-tuned LLM to generate solid results.

One possibility would be to write some Haskell programs, that illustrate paradigms in the language and train the LLM on them. Then get the LLM to solve similar problems, and automatically test how good its solutions are on grounds of solving the problem (passing test cases), performance (time and space used) and how easy it is for humans to understand the programs.

Then use those solution to train the LLM to be better at Haskell, and give it another set of slightly harder problems to solve.

In this way it would learn programming pretty much as humans do, by solving incrementally harder problems.

How easy would it be to make it output programs in a functional language, such as Haskell? it would be nice if it could be made to produce functions that are guaranteed to be free from side effects.

In the long term I can imagine that the output languages of code generators like this might be ones specially designed to work well with AI code generators.

Why not just define the function headers in Python? It's less verbose.

Even if these services were free, I would still prefer BitTorrent. Why?

- It has a wider selection of content. I don't have to faff around looking for which content is available on which service.

- There is no DRM.

- Do these services even work on Linux? I know BitTorrent does.

> It’s also unnecessary as a sales tool because what would “pro-corporate” mean in the context of a book social network?

Does whatever maximizes Amazon's short-term profits, probably.

> except XML feeds were then replaced by simpler files

Yes. XML was basically superceded by JSON, which was far less verbose and clunky.

They were essentially the same idea as WASM -- a portable machine code that runs on web browsers -- but a quarter of a century earlier.

I suspect Java applets fell out of fashion because they could only manipulate a rectangular area of the screen and not the DOM. Did the DOM even exist in 1995?

And yet the same ridiculous fragmentation doesn't prevent Python from having a GUI that works on Python.

Also you can build a drone AI and cheaply deploy it to 1000 drones. Whereas each human pilot must be expensively trained.

Defence is c. 2% of the world economy and probably something similar for carbon emissions.

>It wasn't always like this as you remember, and that's why piracy died for a time.. you could just have one streaming service, and you're good. But people got greedy. They saw what Netflix had, and wanted it for themselves.. not realizing what that would mean

The problem is -- from the content companies' point of view -- that if there is one streaming service and multiple content creators then the steaming service will make the lion's share of the money.

So the big content companies all want to have their own streaming service.

This of course drives people back to the pirates.

As far as I can tell it isn't stigmatised in the UK. Or at least no-one has ever told me I'm a bad person for copying stuff.

In a reasonable world, if you've paid the copying fee, then downloading content would be legal. But no, the content corporations want to take a double dip.

Seems a bit odd that the Windows version has a GUI but not the Linux one. Building a GUI on top of Tk (like Python does) would probably not be hard.

In terms of the 4-document model, I find a lot of software systems (libraries and frameworks), while they are good at providing tutorials and an API reference, they are a lot poorer at providing an Explanation.

What I want to know when learning a new software system are what are the main classes/concepts in it and how they relate to each other.

> it certainly resonates with me!

I recently started getting into Javascript 2D game programming.

MDN has a nice simple straightforward tutorial[1] on how to make a Breakout game in pure Javascript. Having read this tutorial, it all fits my brain. It's simple and I understand it.

There are also lots of frameworks you can use, e.g. Excalibur, which also has a breakout game tutorial[2]. This does not fit my brain: there are vast number of classes to learn, and how everything fits together is not obvious at all.

While I was reading the MDN tutorial I was thinking to myself how I could easily build a framework to automate a lot of the stuff. (No doubt many others thought the same!) If I did build such a framework I would understand it well. It would fit my brain. But would anyone else understand it? Possibly not.

I suspect that what makes a framework easy to learn is, above everything else, good documentation.

[1]: https://developer.mozilla.org/en-US/docs/Games/Tutorials/2D_...

[2]: https://excaliburjs.com/docs/getting-started/

> So this week, we're introducing a general-purpose tool-building factory factory factory, so that all of your different tool factory factories can be produced by a single, unified factory. The factory factory factory will produce only the tool factory factories that you actually need, and each of those factory factories will produce a single factory based on your custom tool specifications.

Looks like someone's been doing too much enterprise Java progamming!