cheesecakegood

I feel like we need to take a page from the past. Apprenticeships with lock-in. Sponsorships. And so on.

Imagine if even one billionaire dedicated themselves to be a modern “patron of the arts” on a more down-to-earth level. Imagine if people started to run “adopt-a-student” type frameworks for individuals. If fresh grads found it easy to get “grants” for things. And so on.

The first job is often a bit “take a chance on me” in its paradigm, but companies these days don’t want bets, they want sure things. Unfortunately this is to some extent unrealistic, but also it’s true that cutting out entry level roles entirely does cut down on uncertainty. It cuts down on upside too, but that’s more invisible.

It’s almost impossible to prove that you aren’t one of these people though. People like me get hurt by this, collateral damage.

I think it was Karpathy that suggested he preferred “simulated reasoning” but sadly it didn’t catch on.

Popular yes; based on good quality research… no, not really. It’s really pretty unfortunate timing with the AI boom, because the field of education research is not very high on the science quality totem pole. And the field as a whole was and is also slow to incorporate neuroscience findings, too.

To be frank, I think the real risk is still just… war. A big enough war where one side goes “no holds barred” in the cyber sphere will be a rude wake-up call. And we can’t do non-proliferation the same way we do with nukes. Otherwise as you say, the small and medium size things just happen sporadically. In an (existential or fully escalated) wartime scenario between countries, you get all the systemic risks hammered at once.

Yeah, the pity is that Amazon almost went this direction. They have a set of filters! They just don’t work. Not only do they clearly not want you to use them (the filters prematurely reset if you do virtually anything), but their coverage is weird and again, they don’t even work. You’d think Amazon themselves would have basic, valid data but they just don’t care.

As an example of how this stacks up, in the app, I filter air fryers to be red. Easy, right? One option, two sponsored, three more options, an autoplay video ad for a non-red one, then another sponsored link, “customers frequently viewed carousel with more sponsored links, one of which is not red, (so this is a full phone screen and a half with zero real results), one more result, two sponsored links, three more results, and then a non-red deep fryer, which is not even an air fryer to begin with. How does a silver deep fryer end up in these results at all?

The shittiness continues. A sponsored link to a toaster oven (not an air fryer either), another sponsored link, a real result, a used air fryer, silicone liners (not air fryers), a carousel of “trending now” including four of five items that are not air fryers either (and a low-rated red air fryer I haven’t seen yet… which makes me think they are outright lying about it being a ‘trending’ item), another carousel with “brands related to your search”, ALSO sponsored, two more results, another near full screen autopay video ad, three sponsored links for air fryer accessories, two real results for more accessories, then two more real red air fryer results, though one is only available used.

I’m exhausted just typing that out. It’s soul-crushing. A good two-thirds of total screen real estate in any given session is just pure ads. And I typed “red air fryer”, so don’t show me a red air fryer silicone liner! That’s not what I asked for. If you must, label it in a carousel that says “matching accessories, that’s both user friendly and appropriate!

The library ‘ryp’[1] now exists that does the same, the data lives in an arrow format so you can write R inside Python without re-writing to disk

[1] https://github.com/briscverse/ryp

My personal opinion for the last two years or so has been that current AI agents are forever going to be highly limited so long as they don’t possess a real “memory” process. Right now they just have absurdly big working memories, and a few hacky ways of making the equivalent of Post-It notes to future iterations, but no true integration of memory into a new future self. Meaning their “learning” is fundamentally kneecapped to one specific and imperfect modality.

Something about this is deeply funny to me:

> In an iterated prisoner's dilemma game with communication, agents all settle upon the same strategy and they all defect at the same time, tanking their overall rewards.

It’s not always consistent, but humans have a higher capability of self-awareness. It’s kind of telling that these Claudes don’t seem to consider this pretty obvious failure mode.

Overall I think this all makes me appreciate humanity a little more. Sometimes the truculent dev who stubbornly refuses to go with the flow produces very valuable insights, as a small example, discovering things the status quo thought unlikely.

The talk around KL divergence is oversold. People talk about it as if it’s not a benchmark, but at its core it is in the same neighborhood! You get a different KL divergence number on different “calibration datasets”, so in other words it’s data-dependent. It is NOT a universal guarantee about the fundamental divergence of a model.

I wonder if this is enough to push people back onto Luna with their comparative price drop

Also 6-9pm Pacific I think is (coincidentally) peak so it hits the ‘after work hobbyists’ still, which is I suspect is their current main audience.

I guess it’s not unexpected per se, but the peak hours are solidly Chinese daytime work hours with an extended lunch break, and hitting a bit of Europe’s morning. But no American peak hours, unless you count after-work hobby coding (6-9 pm Pacific?) which tells you indirectly a bit that American adoption is not a major driver of revenue for them. Or that the after-work ‘hobbyist’ is the main audience, which is also plausible.

When I visited the Netherlands, they have a fast food chain you would love called FEBO that basically takes this to the extreme: they have a giant vending machine-like wall with small glass food warmers cabinets displaying burgers, fries, etc all freshly cooked, one per door, so you literally just put in your payment info and grab the food already right there in front of you!

Ah yes, just as there’s famously no such thing as Russian hackers (for example) given effectively total impunity to scam, defraud, blackmail, etc any company, so long as it’s not located in Russia. No direct harm! Oh wait…

The thing about your own country, especially the more democratic it is, is that there are brakes in the system. A lot of the control mechanisms are indirect, and thus slow and occasionally prone to failure, but the people do have the ultimate say. What you’re doing is looking at failures of the braking system and concluding that brakes don’t even exist! Faulty logic in the extreme.

Now personally, I don’t believe boycotts work, but I’m not going to be using it in either case. Also I don’t think xAI (or Musk for that matter) actually is ready to handle that degree of scrutiny that thus far they haven’t been exposed to. If xAI thinks that they’ve already experienced it, they have another thing coming.

Now, admittedly, I’m not a major voice mode user for any of the apps really but it’s been interesting to see people realize in real time how controlling the length of response is an inherently difficult problem in voice conversations.

There’s a reason that us humans have to use a lot of nonverbal cues in order to judge how long our responses should be, when to bail early, when someone wants to jump in briefly, beyond simply the context of the question. We even regularly alter content on the fly based on how we view the reception. Voice modes don’t have any of that context short of outright interruptions. In the meantime, some kind of response length parameter/slider would be helpful, but I think that’s a nontrivial addition in the LLM design space.

I’m curious how you were juggling this before, was it just a happy coincidence the verbosity of the replies matched your preferred pacing, or you would aggressively interrupt at times, or the model actually did a good job at conversational pacing?

It’s pretty telling that almost all of the bullet points in the system prompt that was posted for Grok have to do with preventing criminality and CSAM generation. No other provider has this same issue at that scale.

The first-order-thinking reaction is “oh cool, look how they don’t want it to happen” but the second-order reaction is “why does this company have such a problem when others don’t?” It’s their own tactics. If you want the “good” of 4chan-like behavior, turns out you get the bad too.

It’s sorta my own fault, yes - I do indeed have Windows 10 on dark mode mostly because I like the dark toolbar and to a lesser extent UI, but notably that’s just because the OS almost never requires extensive reading. Wish Windows let me mix and match, but alas.

The Crandall Printing Museum! Yeah he used to do live demonstrations and everything, to the degree where he would cast a type piece right there in front of you from the liquid metal! Super cool guy, crazy knowledge about even tiny details. Sadly the original near-one man show passed away about a decade ago, but apparently the family has some role still in running it (I think the son runs the tours now?) since they just barely saved it all from being shut down and sold off, so it’s still around! https://www.thecrandall.org/collection https://youtu.be/p9ikfO0MHQE if you wanted to get a glance at what they had/have. Even a linotype machine too!

Really fond memories of my visit, you can tell when someone is passionate.

It’s not 100% bad though, what people often don’t realize is that lighter weight packaging for example has a fairly decent impact on emissions!

Another dry sarcasm victim. If you check the link, it’s clear they were talking about a supposed contemporary concern about health effects from trains, the implication being that today people are also irrationally afraid of tech.

Problem is I’m fairly sure the link is wildly overstating its case, and looks like a content farm. Might even be, ironically, AI slop itself! For example, it talks about “railway spine”: “Physicians of the time believed that the jarring motion and vibrations of train travel could shatter the nervous system, causing lasting mental and emotional distress.”

This is bullshit. As just one example of the low trustworthiness of the “article”, railway spine was the result of a train crash, a notably traumatic event, and the symptoms described are in part just PTSD, a real condition and a real concern (not normal rail travel). For that matter railway travel in those days was genuinely unpleasant (lots of vibrations and jarring movement, poorly ventilated cars, and so on) which ironically modern science would probably validate as being some kind of health risk.

I’d actually view the listed example of train paranoia a great example of historical ignorance. People of the past were not stupid, contrary to popular belief. History has some genuine examples of silly hysteria, but these are usually the exception not the rule.

I was able to do something probably few have, which is visit one of those small hole in the wall museums staffed essentially by just a single passionate old guy assembled with his life savings, dedicated to the printing press. Incredible experience.

Anyways, it’s been about 15 years, but the thing that made an impression on me was that it wasn’t just one invention. Gutenberg basically had to assemble like two dozen separate innovations to make it all work.

The type alloy (3 metals!) had to be just soft enough to cast easily, just expandable enough to set well, just hard enough to last. How to cast in quantity and re-melt as needed. The ink had to be just runny enough to apply easily but not so runny it gets everywhere, sticky enough to stick to type, and dry at a good middle speed. The blotter to apply the ink to type. The system of wooden casing to hold the pieces together with the right spacing, that won’t fall apart. How to set up the casing. The screw press needed to have very even pressure. The paper needed to be able to slide and take ink properly without spreading but also not tear. Proper fonts and ligatures for the type. And so on…

IIRC most things stayed pretty similar for a few hundred years, mostly changing stuff like the paper feeding mechanism or tweaking ink application or minor variations in how the lever or screw work. So to be honest, Gutenberg did almost all the hard work himself. It was a big deal because a lot of these smaller inventions had to come together for a narrow purpose all at once (on top of the obvious huge historical effect of cheaper books!)

I'm a certified dark mode hater, so in that light, that's a no-go for me: there seems to be no way to change the theme.

You mean like Commit Mono does? https://commitmono.com/

I wonder if OpenAI knew and deliberately price dropped to front-run the news cycle in their favor.

Except quite oddly, virtually no one has collected precise keystroke information or if they have it certainly is not public. I genuinely wonder if such a dataset exists

Given that their chart cost axes are almost always log-scale, I’ve noticed starting with Fable that the Low and Medium effort settings might actually be worth setting as your default.

I find myself in major disagreement here. The nice thing about humans is we always have context and ongoing internal conversations including about ethics. If you recruit a bunch of hackers to take down a country, not only is the pay an order of magnitude higher, you have to worry about them backstabbing you, leaking your intent to the government, whistleblowing to the press, and so on. It’s not trivial to do that with a group of (especially capable) humans. They will also have differences of opinion with you and coworkers with some regularity.

I try to recruit a bunch of people to attack a country and it’s going to be hard to get people to say yes, and they will definitely ask or find out which country, and wonder about potential retribution. You see this dynamic show up even to some extent among cybergangs, not all targets are equal.

A single private individual wielding a compliant and hyper capable LLM is an entirely different paradigm. They are accountable to nearly no one and often have few brakes. Frequently they may not care about avoiding detection. And the AI itself may be incapable of the same scale of self reflection and brake behavior a human team will.

We may potentially be entering the age of lone wolf cyberterrorism, and some of the same principles and problems apply. When it is easier for single people to plot and carry out high-impact, destructive acts they happen more often. Doubly so if there’s a social contagion. Gun violence isn’t actually a bad analogy here. And do you remember how many corporate sites got defaced in the prime Anon era?

You might enjoy “Motel of the Mysteries” then, a large format illustrated paperback that I grew up with, which showed some hilarious conclusions about 1980s ‘culture’ by some future-millennia archaeologists. They conclude for instance that the TV box is an object of worship because all the architecture in the motel is designed around or points toward it!