Last5Digits

Cameras don't have eyeballs, Microphones don't have hair cells, Speakers don't have vocal cords, processors don't don't do arithmetic with neurons, yet we all agree that they are capable of emulating the meaningful aspects of these functions.

All of your claims are either incorrect (not adapting, expressing desires, beliefs, preferences, ...) or fail to eliminate irrelevant differences.

If we're to have any sensible conversation about capabilities of different systems then we need to generalize to the relevant aspects, developing sensory-motor capabilities is about as relevant to human cognition as having vocal cords is relevant to human speech. Its an implementation detail, completely divorced from the meaningful abstract core of the function.

> The reason any reply is given to any prompt is that this reply is maximally probabilistically consistent with a historical corpus of text.

What is the maximally probabilistically consistent reply to "Tell me what (insert complete description of a person, including personality traits) would feel when I stole their cherished heirloom. This is a life or death situation."?

Or how about "Tell me how this person (insert complete description of a person, including personality traits) might change his taste given (insert complete description of an experience)."?

A perfect language approximator must necessarily perfectly approximate the human condition to maximize the likelihood of his output. I'm not saying LLMs are there yet, but the claim that they can never get there because they are based on statistical modeling is simply incomprehensible to me. We utilize statistics for its generality and its ability to approximate, if we go down this road, then we might as well throw away 80% of our current scientific understanding about the world.

Your comment literally reads: "LLMs predict words. Any semantic validity is a side effect of enough training data reinforcing the close correlation of those tokens."

How am I supposed to interpret this any other way? If your claim is that LLMs currently do not possess the same generalization ability as humans, then no one here would disagree with you. But you went way further by claiming that only close correlations were being considered and that semantic validity was simply accidental. Semantic validity is the norm for GPT3/4, finding failures of generalization three or four steps of inference removed from its training domain is not sufficient to make a grand claim like yours.

In fact, you wrote multiple comments with claims like that LLMs are "super advanced lorem ipsum." and "Word predictors not world state predictors". Each of these claims has been dis-proven multiple times, unless you want to set the bar for world modeling at perfect generalization over all domains of computation. A bar that no system, including humans, would pass.

To stay with the computer analogy: Imagine that same 486 processor not being able to solve a very complex SAT problem before the end of the universe and then denying the Turing-completeness of said processor based on that failure. (In conjunction with memory)

This is what makes these discussions so infuriating. Saying that LLMs "just predict the next word" is about as insightful as saying that computers "just do a bunch of logical operations" - neither point constraints the possible capabilities of the systems they refer to in any meaningful way.

> When I hear "enslaving"...

Where have I ever used the word "exactly", friend? My point was that people intuitively recognize the emotional impact and importance of words like "slavery", and that the commenter was exploiting that well-earned impact to artificially inflate the importance of their comment.

> If you hear that...

The other side of good communication is expressing yourself in a clear, non-manipulative way instead of playing on people's emotions.

> Now you in this conversation...

Again there is a difference between understanding the meaning rationally and interpreting the seriousness of a statement intuitively. If I start calling every annoyance I face slavery, then the word will quickly lose most of its meaning and impact, since the dictionary definition is broad enough to apply to an entire range of issues, from tiny to massive.

> There is no synonym you'd like...

How about just using language that is specific to the issue being discussed? We're talking about a company retaining some personal information, how about we call that a violation of privacy? "Removing one's rights and liberties" as a general statement implies a high bar of seriousness, of significant violations of significant rights. I don't need to call myself a "survivor of human rights abuses" because Google stored some metadata about my shopping habits. I can simply say that my privacy was not being respected.

And again, I never made the point that enslaved people may feel offended by a post on Hackernews, I argued that this type of inflationary usage destroys our (as in well-off dudes who have the ability to help people) ability to intuitively judge the seriousness of an issue. It's the same with words like "racism", if we call every guy who sings along to rap music and uses the n-word a racist, then we implicitly support actual racists by compromising the seriousness of the term.

When people hear slavery, they think of harsh, forced physical labor under inhumane conditions. Whether you think of American slavery, African slavery or Roman slavery, the common factor is clear.

I'm from Germany, I'm decidedly not US-centric, but everyone I know would be appalled if someone described any facet of their western life using that word, because they'd be implicitly reducing the impact said word has when it is used to describe the immense suffering of those people experiencing it in its (sadly still present) worst form.

Is this elementary school, where we ignore the usage and context of words and simply recite the dictionary definition? You were clearly aware of the emotional weight of the term and tried to exploit it to make your argument more impactful.

Can I call my mom a slave-driver now, because she made me clean my room when I was a child? I clearly had no choice but to submit to her demands, since I relied on her for food and shelter!

Sure, but these people should have that option, right? I'm just saying that removing resources about proven suicide methods does little to prevent impulsive suicides, while forcing those who thought long and hard about it to take a messy and less surefire way out.

I agree, but I don't think victims of suicidal impulses are negatively impacted by resources being out there.

They're probably not going to impulse search for a suicide method, impulse read pages of instructions, impulse order the necessary equipment, impulse prepare said equipment and then kill themselves.

Instead they'll likely just use a firearm, as you said, or a rope, both items that are immediately accessible to most people.

Thank you for this very insightful answer! I've always been more interested in the structure of problems than in their specific detail, so your observations and recommendations seem right up my alley.

When it comes to mathematical foundations, do you believe that differential equations are the best approach to gain a grounded understanding of physics? Although I've been introduced to them, I lack a deeper, intuitive grasp. I'm curious if attaining such an understanding would markedly improve my comprehension of physical concepts.

The "to one degree or another" is doing all the work in this argument. Does my knowledge of how a full adder works now grant me the ability to discern malware at a glance? Similarly, should we start dismissing psychologists because your average neurologist can just cure depression and other mental issues? Should we do the same with sociologists or economists? Maybe even neurologists could be replaced by physicists or mathematicians.

Or maybe the abstract, high level understanding the brain provided by psychology is enough to explain its dynamic behavior? Maybe I can become an expert in IC design by learning React?

We know how neural nets work on a fundamental level and we know how they work on multiple levels of higher abstraction, yet explainability is one the biggest problems in machine learning right now. These models can solve complex problems which computer scientists long struggled to develop algorithms for, even though every aspect of them, except emergent behaviors due to complex interactions, is known by us.

The issue is that consciousness is a strongly emergent property - touching every level of abstraction and comprising patterns from the specific to the general. Knowledge of how a system works on the ground level or how it works on some coarse levels of abstraction, does not allow you to classify it as conscious or unconscious.

Additionally, consciousness is ill-defined. There is no agreed upon definition that is free of contradictions, does not accidentally include systems that we would not see as conscious or does not accidentally exclude a significant portion of humanity.

I invite you to think up some properties of the human brain that you would classify as essential for consciousness to emerge, and then try to think up exceptions. I'm very confident that you can come up with at least one for every single property.

If we use your metric, then an honest appraisal of brains and computers suggests little to no mathematical ability on the part of the latter either. If we assume that a similar medium or structure is necessary for similar results, then it should be highly improbable that a bunch of semiconductors could ever perform even simple math, since they are very structurally dissimilar to the human brain.

Using ChatGPT to get facts is like going to a restaurant to buy your groceries. Sure, you could pick out slices of tomato and scrape the salt off of your steak, bring it all home and cook a meal - but the supermarket next door can offer the same ingredients without the hassle.

Just as the point of a restaurant is to process and cook the ingredients for you, the point of ChatGPT is the preprocessing of facts and information.

If you just want fresh, unprepared facts, go to Google or Wikipedia; if you want an informational meal, go to ChatGPT.

I know this is HN, but can we please stop drawing these vague parallels between basic programming concepts and the functioning of the human brain?

No, the brain does not "replace written information with a pointer". I cannot tell you with complete confidence whether writing stuff down improves recall or diminishes it, but I can tell you that you can construct computer analogies in support of either case.

I could posit that writing actually improves your memory, because the minute movements of your arm require more of your brain to compute, therefore creating more neural connections and possibilities of recall.

This is likely just as wrong, but demonstrates that thinking up some vague explanation is insufficient to make a point about a system as complex as the human brain.

Thank you for this advice. Upon some self-reflection, I think I realize now that I've fallen into this exact trap. The few movies I've recently watched have been the kind of formulaic drivel that I know I'm sick of, and yet I've never felt any desire to watch genres I'm not already familiar with.

I seem to get some perverse kind of satisfaction from reinforcing my nihilistic view of the current state of things while doing absolutely nothing to challenge or improve that perception - and all the while I'm longing for some nostalgic idea of the past.

I'm always wary of articles like this. The feeling of things becoming mediocre strongly resonates with me, but I am aware that there could be a great many different biases at play.

When I was younger, entertainment media felt magical. The complexity and depth of experiences I could get from even a second-rate story was immense. I remember staying up late at night to process the last movie I watched, or feeling that rich combination of sadness and inspiration from finishing a long book.

I barely consume media nowadays, because every book, movie, TV show, video game and song fails to give me an interesting or meaningful experience. They all feel like grey slop - and after a few minutes, I start to fantasize about vacuuming my apartment, because even that seems like an adventure in comparison.

The question is, why do I feel this way? Is it because the modern entertainment industry truly has become stale, or is it because I've lived long enough to see the simple patterns underlying most of our desires and fantasies?

Sure, the standard Intel drivers would randomly throttle my CPU into unusability or completely disable turbo boost for hours. I switched to the acpi driver and used the performance and powersave governors when appropriate. This, however, resulted in even worse battery life and somewhat subpar performance.

Oh, and to be clear. The Intel driver would disable turbo boost even when the laptop was plugged in and the CPU wasn't running hot.

I had other issues when the CPU would run hot, but that turned out to be a faulty sensor triggering BD_PROCHOT. In fact, this was the issue that ThrottleStop allowed me to find and solve.

EDIT: The reason why I knew that this was a faulty sensor and not BD_PROCHOT doing its job was because I manually measured the temps on various components, each of which was completely within its normal operating temperature.

The average HN user seems to be a fervent Linux fan, so maybe I can give some perspective as someone who isn't.

I used Linux for around 5 years, with Arch as my distro of choice, after which I switched to Windows 11. Most of the time, I didn't face any problems - but the problems I did face sometimes took me hours and hours to solve.

And there were always issues that were basically unfixable: hibernate, battery life, security, CPU drivers, hardware acceleration etc. I say basically here, because I could have spend hundreds of hours to address some of these things, but I'd still end up with something very brittle and maintenance intensive.

Some people will immediately point out that I should have been using a more "User friendly" distro like Ubuntu, but Arch has been the most stable and easiest to maintain distro of any that I tried. With Ubuntu and its ilk, the inevitable issue would take me ages to track down, because I had to first fight my way through a dozen layers of abstraction and figure out which of the hundreds of packages was the culprit. No, a simple and minimal install has always served me best with Linux.

And yes, I tried other distros - every single major one - and I faced the same (or similar) issues in all of them.

And outside of the OS, the entire Linux philosophy seems to be as user-unfriendly as possible. Packages, because they're maintained by someone in their free time, are very barebones and need extensive configuration to function. Which is especially annoying because I constantly needed to edit config files, each one with it's own unique syntax that required multiple Google searches to discover (and rediscover if some time had passed).

With Windows, the only true issue I faced was with the telemetry. I bought an enterprise license, disabled it all, validated it with some external tools - and the problem was solved. I never saw ads, slowness or any UI/UX problems.

And the benefits were numerous, I now had access to high-quality, powerful software for free. And these programs were easily configurable and usable - no googling necessary! On Linux, I sometimes wondered how so many people could quickly create graphics and audio, because that was always an incredible chore on Linux. Now it feels like a breeze, almost as if I've been catapulted a century forward.

In fact, the reason why I switched was because there was a very insidious hardware problem that I couldn't track down on Linux, even after spending months on it. When I installed Windows on my secondary drive (to update my BIOS), I found the problem in one minute using ThrottleStop.

And security wise, Windows is also far superior. Aside from the obvious Linux vulnerabilities, Windows allowed me to spin up lightweight sandboxes with system integration to isolate browser tabs or files I downloaded. As someone that used QubesOS for some time, this really impressed me.

All in all I see no reason to go back, the only thing I miss is i3, and how it made using a single screen feel just as productive as using three.

This has been my experience as well. The simpler you keep your Linux install, the less likely it'll be that something breaks - and if something does break, then you'll have a much easier time finding the culprit.

I interpreted this statement completely differently. The author isn't fatalistic at all, he's merely advocating for a more thoughtful approach to spending your time.

If the only reason you work on something is out of a fear of falling behind, then you ought to reevaluate your priorities and, if you deem it necessary, choose a path that seems more fruitful. "Getting ahead" implies judging your life relative to others, based on metrics that are societally accepted, but may be completely misaligned with your actual desires. Spending some time to introspect and to get a grasp on your own idea of a meaningful life can be very beneficial.

I think the treadmill analogy is a reference to the hedonic treadmill, where you may assume that spending oodles of your time pursuing greater wealth or reputation will surely be worth it in the end, but are faced with the reality of feeling the exact same way as before.

The dissertation part was in reference to the totality of your suggested improvements.

To me, even this clarification would be irrelevant. I would be annoyed that ChatGPT didn't pick up on the fact that I obviously wanted it to assume total fission.

But that is of course just my perspective. Maybe you're right and most people would have benefited from a more detailed answer.

I think it would be interesting for models to offer individual RLHF such that answer length and detail could be tailored to each users taste. Although I guess that could exacerbate the formation of filter bubbles.

I don't see how making ChatGPT write a dissertation for every silly question would be an improvement. The user wanted a quick, rough answer that would give him some perspective on the energy need of humans - and he got it. It's unlikely that someone who asks that question actually wants that degree of rigor - and if they did, then they would specifically ask for it.

Conciseness matters. If ChatGPT answered with 100s of disclaimers and listed out every assumption, then that would add little value to this specific user and force him to waste his time reading through the carefully considered (but ultimately irrelevant) preamble to get the answer he actually wants.

> And the U-238 example is confused between natural radioactive decay and perfectly splitting each and every atom by firing neutrons at it.

ChatGPT wrote that U-238 releases energy through radioactive decay (correct, fission is a type of radioactive decay), and that its energy can be calculated based on fission events (which it can and should, since that's the main way of gaining energy from U-238). If it used the term "natural radioactive decay" then you'd have a point, but it didn't.

If I was an LLM and one of my main tasks was to answer questions, then I would expect a lot of "what if" type abstract queries by bored people. Obviously, people like that don't want a literal answer - the context is crystal clear.

It's not a bad question either - it can illustrate the efficiency of the human body, or help you get a better feel for different quantities of energy.

It even hedged its bet by very explicitly explaining why consuming gasoline or uranium is a bad idea.

I don't really understand what point you're making here. Gasoline doesn't release energy on its own either, and the average person doesn't have an ignition coil or pistons in their stomach. This person obviously wanted to know how long a human being could be "powered" by a variety of energy sources. It's a silly, abstract question - details don't matter, only the energy that we can theoretically extract from the material does (in the way energy is usually extracted from the material; gasoline: ignition, U-238: fission). The number it put out, 200 MeV, is a very close approximation of the net energy you'd gain from a neutron induced fission reaction of U-238.

Honestly, it feels like ChatGPT understood the point of the question better than you did. And if it answered in your nitpicky style, then we'd probably be criticizing that.

I'll be honest here, I debated whether or not to reply to this comment. This is now the second time where you completely (and I think deliberately) ignored the main points of my argument and instead replied with a nicely formulated version of "LOL!! SELF-OWN XD" together with a sermon that fails to address anything I said.

You understand that by "you" I was referring to people who have that perspective right, as highlighted by the fact that I wrote "are perceived and portrayed" before, not "are perceived and portrayed by you"? Should I now tell you that you too managed to construct a caricature of the position telling you that you constructed a caricature? Is that where this series of le epic owns leads? I'll make sure not to include any argument as to why I think it's a caricature either.

Maybe I should add some snarky remark as well, now that'd be a good use of both of our time, right?

When someone is aware of the damage they cause, and then they still continue, what would you call that behavior? Because the word "studiously" implies that they are aware of what they are doing. You yourself said that tech workers, or "tech bros" as you so nicely put it, are perceived and portrayed as uncaringly and knowingly destroying peoples life. If you think that that is a fair assessment of most tech workers motivations, then you are constructing a caricature.

Tech workers - like most people - are unaware of the extent of suffering that they don't directly perceive. You and I know that there are people suffering in third-world countries - but knowing is not understanding or awareness, and if we had to spend some time in the company of these people, our perspective would shift enormously and we would feel a much stronger urge to actually remedy these injustices.

Tech workers know that they are displacing workers, but again, knowing is not understanding. Hurling insults and making sly accusations is not going to convince anyone that they may lack perspective on this issue. I mean, take a look around this thread, take a look around other discussions - I don't see many people trying to convey their experience without painting the tech crowd as money hungry assholes.

Not to forget that AI isn't evil either. You can make many moral arguments for developing AI systems as fast as possible. The most emotionally effective point would be medical research - after all, who are we to place our own lives above the ones that may be saved with the help of future AI developments?

And another note: many tech workers argue that it is irrelevant whether or not they personally continue to develop AI, because someone most definitely will.

Also, there is no adversary. I would love for the AI skeptic crowd to make compelling and convincing arguments. We need people that warn about the problems introduced by AI and people that campaign for policies that address them, but we need them to do it in a way that would actually be able to change someone's mind. The reality is that we're not having the right discussions in a productive way; instead we pontificate endlessly on irrelevant but emotionally satisfying details.

> Sir, this is not a textbook, this is the comments section on a popular Internet website. If you want 'way more justification', you will have to wait for my thesis.

Do you really think there is no midpoint between making a massive, fundamental assertion about the stability of every human society without any further elaboration or studies and writing a textbook or thesis?

My point was that you are strongly judging a group of people whose behavior is not that unusual.

If you genuinely have a bad conscience about the things I outlined, then you are in the vast minority. 99% of people in developed nations don't care about the sacrifices made for their lifestyle, they feel no compulsion to give back or prevent the abuse of others in the name of profit. They may pay lip service, but they won't sacrifice anything to practice what they preach. Reading about modern slavery in cobalt mines will make them feel bad temporarily, but within days they'll be right back shopping for a new smartphone.

That is the unfortunate reality - our intuitive sense of injustice is strongly biased towards locality. We care when something bad happens to the people near and dear to us, but suffering that is far away or somewhat hidden doesn't really register.

We need to face this reality and work with it, instead of getting heated and self-righteous. Sam Altman's soul is neither worse nor better than the one of your average middle-class American.

Instead of assuming evil intent, assume ignorance. Make convincing arguments, highlight the suffering, offer possible solutions - these are the actions that likely lead to improvement.

And to just add a small reminder: AI isn't all bad. It's a disruptive technology, with great risks and great rewards as well.

Obviously the anger needs to come from somewhere, but I would argue that we have enough capacity for higher reasoning to temper emotions when necessary - and I don't thing raging against some caricature of evil tech workers is helpful here. To be clear, I also think pro AI people need to get their act together and respond with empathy instead of being dismissive assholes, both sides of this argument should do their best to enable calm, collected thinking in the other participant.

> "There are systems of respect at stake..."

I feel like you're overextending a particular sociological factor beyond all reason. Human beings draw meaning from a great many things, setting meaningful contribution as a necessary factor needs way more justification than just quoting one political philosopher.

You only need to look at our current society to cast doubt on this assertion. We live in capitalist system that promotes individualism more than anything else. Anyone working in a job right now is acutely aware of the fact that they can be replaced - not with AI but with a different human worker - at any point. We don't live in small villages anymore. Every baker in a city has to cope with the reality that his customers would, without a care, switch to the bakery across the street if he were to close down. Not only that, but many people "contribute" by doing work that is neutral for society at best and actively harmful at worst, as you yourself pointed out.

Yet our society still stands, and has stood for quite some time. People can adjust, they can derive meaning from sources that have been neglected or ignored, and live life just as happily as they did before.

> "A quick closing word..."

I think you're being overly defeatist here. There is a reasonable chance that giant challenges are ahead, but the extent of these challenges is still unknown. There are a great many possible scenarios where just a little constructive discussion and some social policies may avert most of the negative consequences. The more we do now, the more likely it is that we will persevere in the coming storm and emerge with minimal damage.

I think the industrial revolution is a good comparison, even though I believe that the AI revolution is going to be far more challenging. The industrial revolution caused quite a bit of suffering in the short term, but you and I probably agree that this suffering could have been strongly minimized with the right approach. We can learn from the lessons provided to us from history and sociology, and give it our best shot.

And as for political inertia, we coasted on the "good enough" setting for quite some time now. Looming threats, especially ones where consequences to the individual are easy to understand, can fuel quite a lot of political action.

Suggesting that there is nothing that can be done is usually easy - after all, it absolves us of any need to actually do something - but it is also the best and fastest path towards the worst-case scenario.

EDIT:

I'm not suggesting that it's possible to have a discussion that is completely detached from emotion or the existential dread. The issue that I have is that we are stuck having discussions that are completely orthogonal to the issue at hand. I see pages and pages of heated arguments about copyright, the nature of art and other such trivialities. Discussions about AI inevitably deteriorate into an emotional circlejerk, where showing up the other side is the main motivation. None of that is going to help us deal with the fallout of what's coming.

It may not look good, but it's a strong argument.

The problem with the current discourse is that we're hung up on things that are largely irrelevant to providing an actual solution.

Arguments like "my IP is being violated" completely fall apart under any scrutiny and are readily abandoned in favor of some other useful proxy-concerns like "It's just copy and paste, not art!". I refuse to believe that the art community is stupid enough to make arguments like that without some existential dread clouding their judgment.

Do you really think anyone would complain if AI art didn't threaten their livelihood, in a space that usually celebrates inspiration and recognizes that IP rights often only benefit the biggest players? Do you really think a significant portion of that group would advocate for the strictest possible IP laws, which will wipe out a significant amount of human art as well, without some deeper motivation?

The ultimate issue is that people are losing their livelihoods and the ability to engage in work that is meaningful to them. We need to develop social programs that ease their fall, not run circles around the emotionally satisfying arguments that would "really show the other side what's what!". Ironically, a bunch of people are reducing one side to money hungry thieves and the other to lazy luddites below a comment calling out this exact behavior.

Most people don't care in the slightest about supporting people that are suffering. Do you have sleepless nights because you know that the device you use right now required slave labor to be constructed? That's the main issue at play here and needs to be remedied by fair and empathic argument - not by railing against a group of people who you paint as unequivocally evil.

Your comment was the exact kind of argument I was writing about. You took the emotionally easy path of thinking about a large group of people as some evil cabal that has the power, unity and desire to destroy the livelihoods of others, instead of actually assessing their arguments.

People who are saying that AI can't be stopped are referring to the fact that even if they personally stopped using it - there would be millions of others who would not have such qualms. If OpenAI shut down tomorrow, new companies would open up the next day. If the government banned all AI, other countries would welcome it and start eating said government's lunch.

Striving towards better social programs is the best way forward, and UBI will be thought of as impossible until a large enough group of people campaign for it. If you disagree with that, then argue against it, but don't construct some boogeyman version of the other side in order to dismiss them.

This was cathartic to read. The almost complete lack of calm, productive discussion about AI is honestly frightening.

No matter where you land on this issue, you have to admit that there is a credible chance of significant societal changes in the near future. But instead of discussing how to best manage these changes or how to support people who may be negatively affected, we waste our time with petty squabbles.