godelski

  > This is also taking his comments about transformers scaling out of context

There have been significant improvements in architectural changes, but since "pre-training" isn't being openly published people just lazily attribute it to "scale". Scale of what? Scale

  > AGI will require architecture changes not simply more data.

As Sutton claimed, all along[0]. People read what they wanted to read. Unfortunately, they also push against researchers who were trying to explore other architectures and methods. I honestly believe that has held us back. There was always room for people to explore different paths. The problem was, and has always been, preventing or discouraging others from exploring new paths. It's ironic given that one of the most common reasons a paper is rejected is for "lack of novelty"[1]. But if someone does something novel then the reviewer simply compares it to the existing state of the art, which is an unfair comparison[2].

It's silly since people want to move fast, but we're really not good at recognizing what has actually allowed us to move forward in the past.

[0] He speaks about Bitter Lesson directly at 9:57 but the earlier part gives better context: https://www.youtube.com/watch?v=21EYKqUsPfg

[1] An essentially meaningless statement. Near meaningless because it is not objective.

[2] You compare a novel technique against a technique that has had hundreds or thousands of researchers optimizing over several years. That's not even remotely fair.

  > New research shows that Apple has a “grey list”

This is not new. It's been pretty obvious for a long time. Personally I've been yelling about it since I switched to an iPhone, but the truth is that people respond with "it's not a big issue" or "I don't swipe".

Btw, their grey list doesn't just not help you spell in swipe, they won't suggest corrections even if you peck out each letter.

What I find really interesting is that there's a million problems like this that people are yelling into the void about and no one cares until someone with power mentions it. This is a weird and interesting phenomena in society. Personally, I believe people always cared, but that the entity with power is better making a call to arms. Essentially that's their power. But I'd also encourage people to keep yelling into the "void" because perhaps you'll slowly get a call to arms and/or might reach a person with bigger reach. But also interesting as the reality is that we don't need to wait for some entity with power to bring up the problem, we could always make those changes. After all, we are the void.

Back on topic, mind you, Apple puts "sensitive" words like "suicide" in this grey list. This literally makes it more difficult for people who are trying to reach out for help get that help. I don't know how anyone in good conscious could decide that words like this are too sensitive to allow to be swiped.

Yes and no. Any harness can use the subscription but must go through `claude -p` mode, which is more limited.

Yes, but that "satisfying symmetric" (I'd say elegant since symmetry isn't required) is meaningful and finding these elegant solutions has been shown to be important for solving further problems. Math and physics aren't as concerned with the answer as they are with understanding

Their LinkedIn (linked on their website) says they work at an undisclosed company. Not sure many people who are at Anthropic wouldn't want others to know ¯\_(ツ)_/¯

I think Feser makes part of the argument for me (though it is clear he isn't a physicist) though I'd encourage you to read Asimov's argument.

There's the metaphysics problem in that all models are wrong. There's more than one competing version of physics (as will help you read Asimov's argument/rant). There's many valid ways to interpret "why" something happens. Physics is trying to understand the causal change of events so that we can generate a model that makes strong and accurate counterfactual predictions. Were that type of model "complete" would we have answered "why" things happen? Depends which "why" you're asking. You could still proclaim, as Fesser notes, "that's the rules God made", which isn't meaningfully different than "that's the way it is".

As for the PSR I'd claim everything does happen for "a reason". But is that reason always knowable? Certainly not. And arguably my namesake proved that there can't ever be any fully complete physics model.

So why be careful? Because my responses are limited, you must ask the right questions. ;)

  > It's an intuition you build pretty early on in science courses in undergrad

I think this intuition is much harder to reach than many think, which causes the frequent over confidence [0]

  > It's why the optimal packing of 11 squares is so unsatisfying.

Idk, maybe one of those "beauty is in the eye of the beholder" kinda things?[1]

[0] https://m.youtube.com/watch?v=hV41QEKiMlM

[1] https://pingyou.com/papers/eleven-squares.pdf

I'd be careful with the inference you're making here. Tbh we could easily conclude that neither Newton nor Einstein explained "why". One could argue we still don't know and won't even after we have a TOE.

I'd take Asimov's "Relativity of Wrong" here. And I also think there's motivation for why scientists care so much about elegance.

  > but can also be the formula for monsters to use that are "hoping for" things that most would find repugnant

As an American, we don't like kings and queens because they are always monsters. Just as the very recent attempt to make a king has illustrated. And frankly, the wiki article even downplays Diana.

The problem is they are explicitly saying to do things in the shadows.

  | Above all things our royalty is to be reverenced, and if you begin to poke about it you cannot reverence it... Its mystery is its life. We must not let in daylight upon magic

Without that daylight you can't differentiate monsters from men. They claim it's because you can't differentiate saints from men, but I'd make the claim that any man who wishes to be seen as a saint is more monster than saint.

And we've seen what these people do in the shadows. The disgraced Andrew was far from the only Royal in the Epstein files. We saw what they did to Diana, who tried to shed some light. We've seen it over and over.

I'll take the light, because I'd rather live in a world where we make things better than live in a world with the appearance of magic. The only magic is what we create

I'll highlight the gp's point here because it is subtle but important

  >> words that label parts of continua are useful

and I'll add one more point: there's a difference in how people speak colloquially and how words are used by experts. In fact, different groups of experts may use the same word to mean different things (e.g. a mathematician and mechanic both call things "manifolds", yet you shouldn't confuse them). Natural language is highly overloaded (which is a big reason we invented formal languages).

A problem with the word "intelligent" is that very few people look to see how it is actually used in different fields like neuroscience, psychology, or machine learning, which all use the word a bit differently, and none of them use it in the way that we use in natural speech.

  > "yellow" that isn't every color.

But I can ask you how "yellow" is the color green. It is certainly some yellow. Definitely more yellow than red is. Even more so, we might argue over if black is more yellow than white or not. Which in some settings black is more yellow than white and in other settings white is more yellow than black. Hell, the fact that you can read this and understand it is clear evidence of both utility and ambiguity. We haven't even mentioned that people may use the same definition but use a word to mark a certain threshold and disagree on that threshold (the root of so many painful to watch arguments).

Everything is just a bit fuzzier than you think.

  > If basically everything is <X>, does that word mean much then/is it a useful term?
 

Yes and no, because nothing is as well defined as you imply.

Let's let X = "chair", to make my point a little more fun.

Obviously we know what a chair is, use the Potter definition[0]. Everyone has several of them in their houses, offices, and all around. But can you state a well defined definition? Is a table a chair? I can sit on it (I often do), it has four legs, and practically all things we'd consider a chair. Okay, it has no back (usually), but then we got to throw out the stool. It has no back. It doesn't even have to have 4 legs! So what about that stump on the ground? Is that a chair? I think most people would argue through a definition of convenience. Or maybe closer to a cat's definition of "if I fits, I sit."

You can replace X with just about anything and you'll find this ambiguity runs rampant. A cup, a table, cereal, sandwich, a tomato; you name it and I'm sure someone here will point out an instance that makes that definition look a lot fuzzier than you initially thought.

So that really shows us the crux of it all: how well defined is something. That only tells us our ability to distinguish something, though it doesn't tell us how well that thing fits into the category. So the stump is a little chair-like, the table is pretty chair, a stool is usually considered a chair, and that thing I'm sitting on is unambiguously a chair[1]. s/chair/intelligence/g and you have your answer. But the if you needed that last sentence, well, there's two types of people: those that can infer.

All models are wrong, but some models are useful

[0] Incidentally this is also a good definition for explaining RLHF and has all the same problems: https://en.wikipedia.org/wiki/I_know_it_when_I_see_it

[1] Though that doesn't mean it also isn't other things!

People seem to miss this a lot, yet we could claim every single thing is "undefined". Even physics is undefined. There's no grand theory of everything. But we still have some really good models. Our models now are a lot better than they were 100 years ago and even more than the 100 years before that and so on.

I abhor people who make the claim that words they didn't spend the time to understand are meaningless simply because they are incomplete. It is idiocracy masquerading as intelligence. I hate it so much because the logic doesn't just fool the listener, it fools the speaker.

I'll let Asimov rant to you all instead, since half a century ago he too was frustrated with this millennia old egomaniac claims https://hermiene.net/literature/essays/relativity-of-wrong

I think everyone understands this already. But there's also a big difference between making it easy for industries to sway politicians and difficult. That's what actually matters.

To put it another way, I still put a lock on my door even though lockpicks exist. I still use passwords on my accounts even though people could guess or brute force them.

Don't trivialize things to uselessness

Mostly because Mail sucks.

  - Search is atrocious
    - I can't find emails from this week but it'll dig up stuff from 5 years ago
  - no tagging
  - no filtering (tbf, this could be a lot better in most platforms)
  - Mail automatically renders PDFs (helps spammers)
  - doesn't play well with aliased emails

And a bunch of other things. All minor but they add up

  > they ask you for a state-issued ID to appeal bans

Requires ID for "age verification" to ensure pedos aren't talking to children

"Move fast and break things" mentality, so IDs leak

Pedos use your leaked ID. You land on a list. Meta/Google/Discord/Persona/whoever pays you $20 and gets a slap on the wrist

IDs become useless, so Altman promotes WorldCoin and we all get our eyeballs scanned

Rinse, lather, repeat. Surveillance state grows. Terrorists continue to do terrorist stuff. Pedos continue to do pedo stuff. Normal people get random visits from gestapo because they are 110% sure you were involved in a crime despite proof that you were in a different state at the time.

We're really creating a great world...

Yeah that too! I've only had an iPhone for a few years but as soon as I used Mail I immediately understood why my parents got phished. Mail hid all signals people would normally use. Their spam detection is bad too. In my first week I got a PayPal used to buy bitcoin scam and neither Gmail nor Thunderbird showed it (both flagged as spam). Both also just showed it was an empty email with only a PDF attachment whereas Mail showed what looked like a normal email, including an address that looked like PayPal's.

I also notice it doesn't handle relayed emails well. I have a bunch of dummy emails and Mail doesn't show me the relayed name, which really hinders my ability to track and combat spam

Mostly because I have multiple email addresses. But I make good use of the filters (which could be improved) and tags. That really helps with searching through everything.

If I only had one email account I probably wouldn't care as much but that's not as common these days

I struggle to get it to search things. Even emails within the last few days. It'll show me irrelevant mail from years ago but not yesterday's.

I also don't like that it renders PDF attachments. Just seems to be enabling scammers

Most people don't like Zen (or Arc). Which is fine. I'm a big keyboard person so I love it, especially with the vimium plugin. But sometimes people get confused at my browser because they don't know how to navigate. Which is understandable

What I like about Mozilla is they made a hackable browser so we can have things like Zen, WaterFox, or whatever floats your boat.

I don't like pickles. So what? No one is forcing me to eat them and no one is forcing you to use vertical tabs. They aren't even the default so wtf is the point of your comment?

I'm with you and honestly I think most people are. Frankly, I'm pretty certain that for 90% of users of you swapped chrome for Firefox, replacing the icons too, then people wouldn't notice. I'd wager 80% of HN users.

It's such a petty argument when everything is so hackable and Google is eating the internet. I wish people would put half the energy they use to complain to solve the things they are upset about.

The man pages certainly tell you how to set your DNS to 8.8.8.8. You just have to do a bit more reading than what the agent would give you. But there's a side benefit to that extra reading, you learn how to do more things. Sometimes that's useful, sometimes it isn't. But you don't know if it is useful or not until later, you can only find out in time

Clothes are typically too cheap to liquidate. And how does Nike liquidate a Nike branded shirt. The buyer would need to remove the logo.

Clothing companies do typically put items on sale first. They'll even sell "at a loss"

People seem confused here, but liquidation is very common. That's what private equity often does. It's also why debt collectors call non-stop, even if they are calling the wrong person (fuck you AI debt collectors!). They buy up the debt for pennies on the door and the original debt owner writes off a loss on their taxes. It's also the reason how I bought a Herman Miller for 40% retail.

It's very simple: some is better than none

Note that Copernicus's system still has epicycles. So it was still fairly complicated. More elegant, but still complicated. It also, importantly, gives order to the planets. Distinguishablity matters

For ages people have been saying "it's obvious" or "it's intuitive", without learning anything about design principles. Obvious to who?

People have been saying "we don't need docs", "nobody reads docs", "they'll figure it out". They experience the pain of a new employee onboarding but never treat this as a signal for the user experience. The user can't just walk over to the developer's desk. But luckily we got stack overflow, blogs, and Google. Somewhere someone explains it! The problem became finding it!

But hey, why do things the "hard" way when you can just throw money at the problem? Or better yet, dismiss the problem by calling the user an idiot.

You're right. What a time to be alive. We've created a world where this product is useful. And not even to because dark patterns exist, but simply because we've always wanted to avoid taking a step back and thinking or getting an outside view. Because it isn't obvious if a robot has to teach you what button to press. If it does, you probably should be ashamed (there are, as always, exceptions)

That said, I do think there's still a lot of utility you this. Those exceptions aren't uncommon. This'll help people learn programs like FreeCAD or Blender, or whatever. Where the complexity is naturally high. But also I do think we should recognize the silliness of many problems that this does solve that shouldn't be problems in the first place.

  > I don't think Copernicus did any explaining why at all.

He did. I mean even ask yourself, why would he do what he did if he was just concerned with motion? Ptolemy's model was accurate for predicting motion. If it was just about motion, then the viewer's position doesn't matter.

But Copernicus's enables us to say more. It is more elegant (it was not simpler though). Copernicus was seeking realism. He was the person annoyed with how complex a system was while everyone around him was saying "who cares, it works".

  > That's what I mean by it being a matter of taste, not of truth.

There is no taste to truth, only taste in how you convey truth

  > It is good for the company.

I disagree

  >> IDK, with a lot of things it seems like big companies are just incredibly myopic.

It may be good for the company in the short term, but that does not mean it is good for the company in the long term. I don't buy the argument that it is good for the company in the long term

Ah the good old days when "we kill people based on metadata"[0] was the conversation and we thought it wouldn't get more invasive. Now we just straight up extract data

[0] https://m.youtube.com/watch?v=tL8_caB35Pg