3PS

This feels like the opposite to me? The MoT architecture looks like the ideal that the Bitter Lesson alludes to - just take all of your data in all of your formats (audio, image, text, action, video) and dump it all into a single shared latent space. Then let the model sort things out, with just enough structure to handle the different requirements/output formats needed (e.g. autoregressive stuff for sequence modeling/prediction, diffusion stuff for generation).

This is also the case with Google Fuchsia, just replace 9P with FIDL. I'm really hoping Fuchsia doesn't end up just being vaporware since it has made some very interesting technical decisions (often borrowing from Plan 9, NixOS, and others.)

Different bracketing.

"If it sounds too good to be true, it usually isn't (true)"

vs

"If it sounds too good to be true, it usually is (too good to be true)"

I think you're technically correct here: if you just have a bunch of Merkle trees where each one tracks the hash of the previous block, it would be accurate to refer to it as a blockchain even if you're not bothering to implement any of the distributed consensus algorithms that cryptocurrencies are actually known for. It's probably not the first thing that would come to mind, but it is a correct way to use that word.

> One way of thinking about a blockchain is to think of it as a shared datastructure to keep databases in sync. Any time you want to distribute your database over more than just a single central place, in a cryptographically secure way, you're probably going to re-invent a blockchain to do it.

Even more specifically, a blockchain is for when you want Byzantine fault tolerance, i.e. you don't trust one or more of the actors involved. This is the main distinguishing feature of blockchains IMO, the reason we have proof of work, proof of stake, etc. It's also the main thing I saw people getting wrong when using blockchains during the earlier waves of cryptocurrency fever; most proposals for blockchains did make sense as distributed public ledgers, but didn't really need the extra computational overhead because only trusted parties were adding blocks to begin with.

I think both you and GP are correct, but in different ways.

It's true that the English language has a very large number of phonemes... but accents tend to regularize/restrict these phonemes. For example, a typical bilingual speaker of Indian English and Hindi will replace instances of the /æ/ phoneme (as in "blast" or "fast") with another phoneme like /a:/ (as in "father"). Which isn't that unusual since /æ/ is pretty uncommon among languages.

Other rare English phonemes include the dental fricatives, i.e. the "th" sounds in "ether" (voiceless) and "either" (voiced). Speakers of Indian English often replace this with a dental stop, a "t" sound (voiceless) or "d" sound (voiced). (Note that Devanagari has a _lot_ of stops, so this is one place where it cannot be cleanly encoded into the Latin alphabet without diacritics.)

So overall: while I think Devanagari can't encode e.g. American English, it can actually do a pretty solid job of encoding Indian English, but not the other way around.

For prices specifically I think it's fair to say that inflation only goes in one direction, but for larger market trends, IMO the key here is _habit building_.

Many things were technically feasible pre-pandemic but not done habitually: remote work, streaming movies instead of going to the theater, ordering delivery instead of dining out, and so on. The pandemic forced many people to change their habits and get over any initial inertia (e.g. investing in a WFH setup or home theater). The result is that when the world returned to normal, the markets didn't: consumer habits had already moved on.

Broadly summarizing.

This is OK and fair use: Training LLMs on copyrighted work, since it's transformative.

This is not OK and not fair use: pirating data, or creating a big repository of pirated data that isn't necessarily for AI training.

Overall seems like a pretty reasonable ruling?

> But fish is too much like bash in syntax, meaning that I just think of it like bash until I have to type "(foo)" instead of "$(foo)", or "end" instead of "fi"

Note that fish does also support bash's "$(foo)" syntax and has for a few years now.

Nit: that's not what negative reinforcement means. Negative reinforcement is about removing a negative stimulus, like inducing someone to go to a desirable website by improving their initially bad text contrast whenever they go there.

In this case, jumpscaring yourself would just be considered punishment (or "positive punishment").

Quiver was absolutely indispensable when I did a category theory course a few years ago. The UI was clean, intuitive, and featureful. Compared to banging one's head against Tikz, it's absolutely no contest.

It doesn't always work, but I like the SQLite model: the core offering is free and open source, but enterprises can pay for things like

* Professional support, including on-prem hosting when applicable

* Additional features that enterprises care about (encrypted databases, SSO)

* Compliance documentation/certifications