bloaf

This has been studied. There is a retest effect on most "cognitive performance" tests: scores increase on the first and second re-test, then level off. (i.e. the test scores for the 3rd and 4th time you take the test end up about the same)

Tutoring and extensive exam-specific prep can boost scores. They are not necessary for getting a high score.

That is an article which

1. Makes claims about performance 2. Provides no benchmarks

The only concrete answer to "why have swap if I have enough RAM" is "well, if you run out of RAM..."

The drive to be differentiated is a way more powerful force than people appreciate. When big oil was deciding what to use as an anti-knocking agent (tetraethyllead or ethanol) the two biggest concerns that tipped the decision in favor of lead were:

1. Ethanol might become a competitor to oil. 2. Lead could be patented, licensed, and used to differentiate their gasoline from competitors.

You must not believe any statistics if you think sampling 85% of a 110k population can't resolve a score like this to 1%. Your original question, which I have been addressing, was about generalization to the population (i.e. what the margin of error quantifies) not nebulous "other areas."

I would expect a re-test to have slightly higher averages overall because all the kids now have experience with the class of questions, whereas only some did the 1st time.

So if you're worried about it, do the math. My napkin calc puts the margin at around +-0.3, enough to easily separate 3% from 5%.

> Seems like it could be within a margin of error (3 v 5%).

There is no margin of error when your sample is the entire population.

They increase speed, which buys you time to work on harder problems. I bet a nonzero number of 2nd tier students could get a 1st tier score if given an extra 10 mins on each section.

These guys claim a 50% speedup with a similar approach:

https://www.reddit.com/r/StableDiffusion/comments/xmr3ic/spe...

I love that Nation States uses the "Bus Surprisal Index" as the measure of unexpected death rate.

https://nationstates.fandom.com/wiki/World_Census

So on the subject of library science, I think a really valuable use for machine learning would be the automatic derivation of classifications in systems like UDC [0]. That would open the door to easier traversal of this absolutely massive dataset.

[0] https://en.wikipedia.org/wiki/Universal_Decimal_Classificati...

This particular case is special because it uses 100% library functions. Typically you're composing your own functions, so you just... put breakpoints in your functions.

If you want logging, I've added an example of how to auto-log the composed functions to the gist.

No. I did that for people having a first exposure to this non-idiomatic currying/function chaining.

In actual code I would have put most of those on one line.

I have pretty mixed feelings about this. Personally I find it much easier to debug code that:

1) fits entirely on my screen and

2) doesn't involve much state modification

Every intermediate variable is a chance for me to miss some modification (e.g. it was passed to a func that modifies its arguments) and consequently misunderstand what is happening.

I've been experimenting in Python with the function chaining style of coding enabled by the toolz library. So while not at all idiomatic, the example in the original article would come out as something like this:

https://gist.github.com/ZeroBomb/8ac470b1d4b02c11f2873c5d4e0...

I would say that function-chaining this example would constitute over-engineering, but I have found that writing in this style has really helped me express pretty complex function composition in a way that is still concise without using a bunch of intermediate variables.

The HP laptop I was running Fedora on had some kind of low power mode on the CPU that would cause linux to crash and require a hard-reset to recover. That meant no idling/sleeping/hibernating without a kernel crash.

Full body spinners are still a thing, see Gigabyte and Captain Shrederator. Unfortunately, their performance tends towards "middling" due to control and reliability issues.

The full body spinners struggle to keep spinning, as the forces of impact tend to ping-pong them around the arena. As long as their opponent can tank a hit and has a halfway competent driver, the spinners tend to end up corralled in a corner.

The latest season has an absolutely incredible match between Gigabyte and Hypershock.

It was all most likely a reference to this:

https://longbets.org/1/

I personally think Kurzweil still has a shot at winning it.

I was not implying qualia were magic, insofar as there can be monistic descriptions of qualia. My criticism was that the example appeared to be steering the conversation into a crash course with dualism by invoking qualia-like explanations of compositionality when it was unnecessary to do so.

I believe the "hard problem" is far easier than the dualists' interface problem. It can be explained by viewing consciousness not as the driving force behind our thoughts/feelings, but rather an after-the-fact log the brain keeps for itself. Qualia are therefore distinct from the immediate physics of signals reaching the brain; they are instead the brain's own shorthand description of the impact those signals had on the brain.

Compositionality isn't a categorical stumbling block because the machines are actually getting incrementally better at it. The Winoground paper the article references explicitly says that score on their compositionality benchmark does in fact scale with training dataset size, suggesting that while it is more difficult for AI to discover compositionality during training, there is no reason to think it is impossible

This just fundamentally feels like a bad hill to die on. Compositionality feels like it is:

A) Something AI is currently known to be bad at

B) A matter of degree, not a categorical stumbling block

C) Vague enough of a concept that AI skeptics will continue to complain about it even after the field has moved on

Your example feels less like a description of "compositionality" and more a description of "qualia." It feels an awful lot like dualists trying to carve out a place for magic that no artificial process can reach.

That's not my reading.

It sounds to me like transparency does always increase accountability full stop. It just turns out that accountability to lobbyists is a greater driver of legislator behavior than accountability to voters.

The problem is therefore neither accountability nor transparency, but the disproportionate influence of lobbyists.

You know about the Spanish Inquisition, right? Have you heard what Catholics say about it? They say that according to their records it wasn't that bad[0] and that the numbers are deliberately inflated by anti-Catholic sentiment.

It would be great if there were some other record keeper so we could ensure the church didn't "lose" any records that made them look bad.

But I'm sure you would have stopped any 3rd party back then who was setting out to transcribe the inquisitions' trials. After all, they were just giving the church yet another platform to spread their intolerance.

[0] https://www.theguardian.com/world/2004/jun/16/artsandhumanit...

I don't think SQL gets replaced until we move to a different database paradigm. For example, I could imagine a function database model [0] getting paired with a "grammar of graphics" type of metadata for the records to create much more concise query/aggregate declarations, especially where time series are concerned.

[0] https://en.wikipedia.org/wiki/Functional_database_model

I'm willing to tolerate a little friction to setup syncing, but keepass has friction at almost every level.

Their importer is powerful... but I had to manually configure it to import from lastpass. Just bake that in ffs.

There is no "generate password" button on the main application screen despite that being the #2 use for a password manager behind autofilling. Instead the main screen has "find", "find entries" and "search" buttons which are all slightly different despite sounding completely redundant.

To get a browser plugin, you have to pick from a disturbingly long list, each with pros and cons, each requiring a few config steps, and none of which work as well as the cloud competition. I had random disconnections from keepass and many autofill failures.

Finally, you get to syncing, and the friction is bigger than you expect. It is not enough to set up a LAN folder; your solution needs to be able to resolve conflicts if you ever use more than one device simultaneously, otherwise you run the risk of losing changes. The most common sync solution is dropbox because that is the only mainstream cloud offering with first party support for linux, which is fine, but can still cause conflicts. The recommended way to handle this is quite convoluted[1]. Just bake that in, FFS!

[1] https://keepass.info/help/kb/trigger_examples.html#dbsync

It would regularly fail connect to the browser plugins when I tried it (despite successfully connecting previously), and the browser plugins failed to autofill more often than not.

There were a ton of usability headaches in the core application too. The UI was quite unintuitive.

It may be technically great, but that's just one dimension of greatness.

In that same vein, I really like the Alan Tutorial channel on youtube. Starts off as slightly off comedy, and subtly builds to horror.

I've had way more issues on commodity systems than any PC I've built. Every sub-$1500 laptop I've ever bought has had some form of hardware issue. Anything from CPU incompatibilities with OS low power states to the machine I'm currently typing on which regularly BSODs due to the absolute dumpster fire of Realtek wireless drivers.

I'm more curious about Allulose[1] which is claimed to have anti-hyperglycemic effects by reducing glucose absorption in the intestines.

[1] https://en.wikipedia.org/wiki/Psicose

https://pubmed.ncbi.nlm.nih.gov/32708827/

> I walk into a library or a bookstore sometimes and see how many of these things are on display, and I feel that same feeling, that there’s so much content you could drown in it all. It makes that howling feeling more intense, as my own contributions to the culture are rendered insignificant.

This is the singularity. The typical definition of the singularity involves "unpredictability," but that is just a symptom. The cause of the singularity is AIs generating new information faster than we can consume or act on it.

Of course in art there is no point in exceeding the rate of human consumption; as others have pointed out here we face a future in which AI continuously generates art tailored to each individual. As a result, this doesn't quite rise to the level of "unpredictability" envisioned by singulatarians.

All fields will eventually face the same reckoning as this author has spelled out for art, and some (such as circuit or AI design) will introduce positive feedback loops that will not only put the relevant engineers out of work, but also cause the field to advance so rapidly that we humans will have no chance of catching up, let alone predicting what the AIs will develop next.

I think this is the missing piece. Excel is great at user input.

Getting the same level of copy-paste, drag-drop, drag-auto-complete, "tweak until it looks right before I move on" data-entry experience is non-trivial.

The dangerous things about excel, imho, are:

1. Too many different places for functionality to hide. Are these numbers updated by external links? VBA? Pivot tables? Equations? Some addin?

2. Difficult to version control. To capture all the functionality its not enough to just strip the VBA code into git, you need to track a lot of other config variables related to the features in 1.

3. Crap accretion. Complex spreadsheets have a tendency to accumulate broken external links, redundant named ranges, and extraneous scratchpad sheets as people copy data around. This makes troubleshooting harder just by increasing the amount of noise.