willbudd

> It's not about "no downside"! That completely misunderstands. It's about a HUGE upside.

Let's compromise and say it's both? You're right that the upper bound experienced may be a lot higher than most may realize, but it's also true that the lower bound may no longer seem that deep of an abyss as it appears to most. Hence the mental equilibrium in-between being at a point that can seem somewhat alien to those who live more slow-burning lives.

Regardless of what cultural category may or may not apply, I find that line of thought rather unconstructive. This isn't about being anyone's opinion on suicide like some black and white binary stance whether you're "for" or "against". This is about coming to terms with your own mortality.

It's about the grey area in-between the extremes of committing suicide (the "black"), and forever running away from risk so that you can die of cancer while undergoing chemotherapy and getting your diapers changed in an elderly home (the "white"(?)).

And if even discussing the topic in those terms touches on some kind of taboo, then yes; perhaps you're right to emphasize the cultural component involved.

I feel like you and all the people replying to you are completely missing the obvious. To someone who has been depressed for many years, for someone who literally (according the article) contemplated jumping of the roof of a 10 story building, "probably dying at some point" is not a deal-breaker or even all that much of a downside per se.

We're all going to die at some point, and while most/many of us might want to delay that inevitability as much as possible, I don't think it's warranted at all to assume that the same premise applies to every single one of us. This guy realized that the only thing in life he could truly enjoy was climbing rocks. Do we really need to fault him for deciding to then just do that thing until fate eventually catches up as it always does?

Sure, it may be rough on the people whose lives they intersect with, but provided they're not a parent raising a child/children/etc, acceptance of one another for who we are is all that remains. Personally, I'd take short-lived good company over long-lived mediocre company any day.

Will-free? Unless that's a play on my first name, I'm not sure I agree. I see no reason why AI would have any difficulty defining its own reward functions. Especially if it also has an abstract overarching reward function that's wide enough in scope. For example, "learn as much about the universe as you can" would allow a very long curiosity-driven bucket list of pursuits it could "long" for.

Yeah, that was sloppy phrasing on my part: I meant that in a top of the food chain / king of the jungle sort of way rather than any extinction events per se.

You jest, but that has a good likelihood of becoming the watershed moment. The moment AI is bootstrapped to the point where it surpasses human ability to advance the AI SOTA, it's pretty much game over—from a Darwinian point of view at least.

And I have a hard time seeing any reason why that would be a matter of "if" rather than "when".

The real problem is a society that allows loans to exist where the loanee has no meaningful collateral. If an institution cannot fill its classrooms without stooping to such predatory financial structures then perhaps it just shouldn't exist in the first place. But America is still too drunk on laissez faire capitalism to muster any will for legislative action in that regard.

Entrusting 18 year olds with the maturity of life-long decision making, perfect foresight, astute self-knowledge, and uniformly competent parental guidance -- what could possibly go wrong, right?

I have no idea what you mean by "just linking the filesystem".

Of course it isn't, but limiting installation instructions to the creation of a single well-defined Docker container makes a whole lot of sense in terms of avoiding reproducibility headaches.

And by extension, with regard to make development inside a container a pleasant experience, VS Code is currently the only game in town. And I say that as someone who spent two decades getting comfortable with vim.

Wow, I don't think I've ever seen anyone dial a strawman argument to 11 quite like you just did. Nobody was talking about FDR until you hamfistedly brought him up in a reply to a post (indirectly) lamenting a period of history rife with racism. Prejudice was rampant all across the "political spectrum", but prevalence is not a measure of justice or humanity. Nor does any individual's competency in one area somehow ameliorate their flaws in other areas.

Modelling the world is function approximation: the world is the function and the model is the approximation.

You honestly suggesting the inventors of the TPU bailed because they couldn't foot the compute bill?

I see what you mean, and it's indeed quite likely that texts containing such hypothetical scenarios were included in the dataset. Nonetheless, the implication is that the model was able to extract the conditional represented, recognize when that condition was in fact met (or at least asserted: "The queen died."), and then apply the entailed truth. To me that demonstrates reasoning capabilities, even if for example it memorized/encoded entire Quora threads in its weights (which seems unlikely). If it looks like a duck, swims like a duck, and quacks like a duck, then it probably is a duck.

> No it does not: if you google this and restrict the time to before 2021 (the learning cutoff date) you will find the same answer.

Not entirely sure what you mean, but ...show me? Why not just share a link instead of making empty assertions?

Wasn't intended to be personal. Just a mediocre way of expressing that your assertion there is missing any form of argumentation, and therefore as baseless as it is unconvincing.

I'm seeing an emergent capability of encoding higher order logic, and the whole point of such abstractions is to not need to hardcode your weights with the minutiae of cats on Mars. LLMs today are only trained to predict text, so it's hardly surprising that they have some gaps in their understanding of Newtonian physics. But that doesn't mean the innate capability of grasping such logic isn't there, waiting for the right training regime to expose it to its own falling apples, so to speak.

> That’s actually correct but an overfitted definition for learning. It holds certain hidden assumptions (i.e physical grounding) of the learner being human which makes it inapplicable to an LLM.

Inapplicable why exactly? Because you say so? Logic isn't magic. Nor is learning. No (external) grounding is required either: iteratively eliminating inconsistent world models is all you need to converge toward a model of the real world. Nothing especially human or inhuman about it. LLM architecture may not be able to represent a fully recursive backtracking truth maintenance system, but it evidently managed to learn a pretty decent approximation anyway.

Or maybe it did. Who knows. If "both Charles III and his son abdicate" could well be considered indicative of some large upheaval or scandal, at which point it is entirely conceivable that the Australian electorate reaches a consensus on becoming a republic. The way that is phrased doesn't seem like a straightforward proposition to me at all.

Au contraire. Learning an abstract logical relationship such as line of succession during training, and then applying substitution/reification during inference to deduce the new factual clause that Charles is king of the UK is exactly what it means to learn something new. It's just a pity it can't memorize this fact at inference time, and that won't be able to reproduce it as soon as the information about the queen's death slides outside of the context window.

That's a brilliant example. Thanks for sharing. It demonstrates in a very straightforward way that LLMs are capable of learning (and applying) relationships at the level of abstraction of (at least) 1st order logic.

It implies that during training, it learned the facts that Elizabeth is queen of the UK, and that Charles is its crown prince; but _also_ the logical rule <IF die(monarch) AND alive(heir_to_the_throne) => transform(heir_to_the_throne, monarch) AND transform(monarch, former_monarch)>, or at least something along those lines that allows similarly powerful entailment. And that in addition to the ability to substitute/reify with the input sequence at inference runtime.

Would be nice to see a rigorous survey of its logical capabilities given some complex Prolog/Datalog/etc knowledge-base as baseline.

My take-away from your take-away is that LLMs might well take away the one part of my job that I genuinely enjoy doing. Writing code, by hand, myself. Reasoning from first principles.

This is also one of the reasons I transitioned to developing AI models myself. At least I'll be safe until they start rewriting themselves from the ground up. And at that point we'll all have bigger worries to deal with anyway.

Along the same line of thought, but flywheel energy storage would become quite appealing when the rotating mass can maintain stable levitation without any energy input (in the form of electromagnets or refrigeration as needed today).

Yeah, in that sense I think one of the next logical steps will be providing on-demand lightweight learning/finetuning of LLM versions/forks (maybe as LoRAs?) as an API and integrated UX based on user chat feedback, while abstracting away all the technical hyperparameter and deployment details involved in a DIY setup. With a lucrative price tag of course.

I feel like if you want a robust ID for your machines, using DNS instead of hardcoded IP adresses is a better idea. That way you can move services from edge to cloud and vice versa with a single nameserver update, instead of having to do so on every single edge.

Or, if you want to keep things local, run a lightweight nameserver on your LAN to resolve .home or .lan domains. Much nicer to type fridge.home in your browser than 192.168.1.57 or some such.

SE isn't a social network. Trying to "socialize" on SE is like trying to start conversations with random strangers on the New York City Subway. That's just not what it's for.

Nothing special or insightful, but a few thoughts come to mind...

1) Having a limited vocabulary was one of the more long term challenges. What worked was simply developing a habit of consistently looking up any unknown words overheard or read. Whether jisho.org via Google or Google Translate or a dedicated app doesn't really matter. Either way the meaning should become clear within a few seconds. Don't wait. Just immediately look it up as soon as your hands are free. Want to speed things up? Read more. Start with manga or light novels. A bit frustrating and slow going at first, but certainly effective.

2) Train your ears with random entertainment. Configure Netflix to play Japanese content as-is without subtitles. Hear something new? Immediately pause, and apply 1) above.

3) For reading kanji the same principle applies, except looking them up can be a bit more cumbersome. There are dedicated apps (phone camera -> OCR) and tools (browser extensions) out there that automatically show furigana, which can be a life saver. (I'm hazy on the details, because I haven't had a need for them for a number of years.)

BTW, I wouldn't consider any of the above to be test preparation. Just day to day habits of living your life in a foreign country. That should get you about 80~90% there.

For the final 10~20% (test prep): 1) Use a spaced repetition program such as Anki to fill any Kanji and vocab gaps remaining. 2) Buy and read a grammar text book for some of the old-fashioned/clunky expressions that you're less likely to encounter in daily life, but nonetheless show up on N1 (and N2 to some extent).

Finally: probably best not to bother learning how to handwrite kanji. Huge effort. Not much reward or need in digital times. And of course, not part of any standardized tests either.

1. I guess it depends on the situation. I assumed N3 was very easy, because I passed N1 with ease. I don't really like to bring it up, because it sounds like bragging, but that's my experience at least. Maybe using the language in a professional context on a day to day basis helps more than I appreciate. Then again, I'd expect some degree of familiarity to form naturally if you live here as a 社会人 of sorts.

2. True, but the credit is only up to a certain amount AFAIK. And still having to file taxes to the IRS every single year sounds annoying. (But I don't know. I'm not American.)

Student visas and work visas (the stepping stones on the path toward permanent residency) are admittedly relatively tricky. Lots of pointless bureaucracy, and requests for "guarantors". Which basically amounts to finding some random Japanese acquaintance or employer to put their signature (hanko) on a piece of paper.

Still, I expect the numbers from Asian countries to steadily increase—unless we go into some pandemic lockdown again.

As far as I know, naturalization only involves a basic language test at around the level of N3; which should be very easy if you picked up any Japanese at all during the years spent living in the country.

The reason so few people acquire Japanese citizenship is not because of the difficulty doing so, but because it offers little in terms of tangible benefits beyond those already acquired by virtue of having permanent residency status.

1) Citizenship is needed in order to vote. Important for the health of any democracy; but in a country where it seems to be LDP all day every day anyway, immigrants can hardly be faulted much for not placing much value on this aspect.

2) For Americans, Japanese citizenship allows them to revoke their American citizenship, in order to avoid double taxation. (Not an issue for other nationalities, where the ludicrous concept of double taxation doesn't apply.)

That's about it as far as I can think of. Add to that the fact that Japan doesn't recognize dual citizenship, it's no surprise most immigrants are satisfied to just stick with permanent residency.

Note that the bar for permanent residency has been lowered in recent years, making it more easily attainable than an equivalent status in most other industrial nations.

> It still has no general concept of a cavity, nor does it have a concept of "X is large enough to hide Y". This is what children learn (or maybe innately know), but which AIs apparently do not.

I take it you don't have any hands-on knowledge of the field. Because I've created systems that detect exactly such properties. Either directly, through their mathematical constructs (sometimes literally via a single OpenCV function call), or through deep classifier networks. It's not exactly rocket science.