Was that implementation confirmed to work or was it a bug?
I saw people saying it only loads the weights necessary, but all the weights are necessary (unless the architecture is heavily modified). Does it stream or load only the currently necessary layers in a way?
On start up it does load a fraction of the weights, so RAM usage is low, but as you run inferences on the model, it tends to eat up the required amount of RAM. This was from my initial testing.
It mmaps the data so the page any particular bit of data is on is only loaded into memory if something on that page is accessed (at least that is my understanding).
Looks like it will be reverted because it introduced some regressions and as far as I understand, there was no magic in the end, but there was a ton of drama around it.
Does enabling to run LLMs at home really strengthen democracy? People may believe that, but it potentially enables AI to run on any device with a low footprint for surveillance purposes as well, and governments don't even have to pay for the development...
> Does enabling to run LLMs at home really strengthen democracy?
Well, I can't speak about strengthening democracy for sure, but what I see is that AI will become a significant means of digital production, and we wouldn't want those means to be in the hands of a few people.
> but it potentially enables AI to run on any device with a low footprint for surveillance purposes as well,
Yes, a low barrier for inference cuts both ways. That is a risk.
> democratising the means of digital production is the difference between a dystopian authoritarian future and an empowered (and admittedly chaotic) humanity
While I'm generally in favor of redistribution of wealth, power and knowledge, this doesn't seem an insight so trivial that it can be asserted without underpinning to me, nor congruent with history. Industrialization was essential or at least catalytic to democracy, which happened without public control over production production means (I'm ignoring the Marxist view in this).
A low barrier to entry for surveillance means citizens can more easily watch their governments too.
As long as surveillance tech requires big resources, governments will have the edge. Miniaturize it, decentralize it, and it means citizens and public interest organizations can hold the government accountable more easily (this is, notably, what OSINT is all about).
It depends on how you think technology relates to democracy.
If you think this is the modern-day equivalent of a printing press, it's vital to avoid this being in the hands of only one or two multi-billion-dollar corporations with a history of morally questionable moves.
If you think privacy rights and freedom from surveillance are critical to democracy, you want your smart speaker or video doorbell to do all its processing on board, without uploading recordings to the cloud. And you want your neighbour's doorbell to do the same.
And frankly, nobody goes after Microsoft Word when someone writes Xi Jinping satire using it, or ICE buy it - because people know it's a tool rather than a platform.
On the other hand, if you think ML technology is going to produce the next wave of fake amazon reviews and forum spam - you might be a lot less eager to see it in the hands of everyone.
If you think society will be slow to adapt to things like deepfakes; and that cops and corporations will immediately start using things like facial recognition technology - well, then wider access to ML might seem to have a great many downsides.
The washing machine is one of the biggest game changer in the world, allowing people to suddenly have, like a noble used to have, the equivalent of a servant to do a tremendously costly chore. Without the slavery part.
I think AI has the same potential: filling administrative work, walking the complexity of an information overloaded world, keeping up to date with laws you can't ignore but have no way to know about, planning logistics that our modern life impose on us...
Suddenly, like the CEO of today, every one will be able to have a secretary.
Yes, that can make a huge different.
Don't think about AI as the thing we have today. GPT the nokia 3310 of AI. Amazing, but very basic.
Think that one day, somebody will come out with the iphone of AI.
You're just extending the lines in a way you like to imagine. Kinda like wishing for flying cars when we can't even make them navigate 2D city streets autonomously.
Flight autopilot is easy if you only have to fly from one point to another. The hard part is dealing with emergencies. We are far from being able to build an autopilot that could react as well as the human crew did during US Airways flight 1549.
So you're comfortable giving a black box secretary powers, including accessing and using remote APIs on your behalf, while you have zero insight or control over its decisions? Or will you be second guessing it every step of the way rendering it little more then a fancy autocomplete?
I don't think you realize the amount of risk involved in giving these opaque probabilistic systems control over data and compute. They are not people replacements and assuming so and acting on it will create more problems than it solves... Because extending thee lines to wrap around your needs is wishful thinking and there is a lot more complexity and hidden unintended consequences than assumed when saying chatbots capable of using APIs will be dependable personal assistants, which is the point I'm trying to make.
It is hard for people to comprehend the rate of change. Over the short term, it feels like we are on a gentle slope upward with respect to change. But the reality is the slope is near vertical. Even your analogy is a huge understatement. It will be more like Nokia today, and iPhone next week.
OP is saying GPT is the equivalent of an old Nokia feature phone. However cellphones really blew up once Apple figured out how to make smartphones broadly appealing with multi touch (prior to that you had things like Windows CE, Blackberry and Palm, but those were extremely more limited than what the iPhone delivered).
In other words the jump ahead from GPT to broadly useful AI is akin to the jump from the old Nokia feature phone to iPhone. Don’t let the current limitations of GPT color what you think AI will be generally capable of within the next decade.
> Suddenly, like the CEO of today, every one will be able to have a secretary.
I think something that was majorly lost in the past 60 years is the secretary. I agree with you that it’d be great if everyone had one - the working relationship between a secretary and the person they support can be amazingly productive.
>The washing machine is one of the biggest game changer in the world
The biggest game changer was the fossil fuel that enabled that washing machine, and all the other modern devices we rely on.
That same substance is used to as chemical feedstock for fertilizers and more than double yields, and more than double the amount of "arable" land. If we were to run out, half of humanity would likely starve.
The relevant distinction isn’t being a tool it’s how it is built. If you use a bunch of copyrighted material in the design of your tool and some of the material leaks into your answers that’s very different from building Microsoft Word.
The copyright question is of course an important one.
But I was thinking of the tool-vs-platform distinction in the context of the censorship question raised by Midjourney blocking images of Xi Jinping.
'Platforms' like Facebook and Github often censor things like porn, terror recruitment, and political satire fearing reputational damage and loss of advertisers from having such content on their domain, next to their logo - whereas offline tools like MS Word aren't subject to the same fears.
So people who think political satire is a particularly important form of speech might prefer an offline tool for their greater censorship resistance, irrespective of their opinion on issues of copyright.
Much of the fear of AI is really a fear of concentrated increase in power of those controlling it. Decentralizing models mitigates that risk by spreading the power amongst a more diffuse population.
I don't see how it should be any different from the PC revolution. We've all got a 1970s supercomputer in our pocket now, for better or worse. But it's certainly preferable to only the richest and most powerful having one.
Computers can’t think for themselves. It’s more like if everybody had a trained police dog - one of those dogs that will bite and not let go, would you feel safer walking around?
“Hey OpenChatGPT, find a list of every person who has ever wronged me and constantly harass them with death threats from behind seven proxies”
Spot on. Broadly speaking, if you want peace you have to prepare for war. If you don’t distribute and democratize power, those who typically control it in concentrated form will weaponize and wield it (power tends to corrupt, absolute power corrupts absolutely and all that jazz).
You may not feel the need to construct a strong defense because of your disposition and mental model, but you should because of the motivations and mental models of those building a strong offense.
A take as old as time. History books are always written by the victors.
No, and I frequently encounter this analogy about AI, but it's completely absurd. AI is not a nuke. I don't see the point of even entertaining the idea of it.
Google and Microsoft (read: corporations) don't have nukes. But if they could and did, yes, why not Joe Schmoe? A little deterence for those days when your Google account suddenly stops working. Just send a stern letter to Mountain View.
More seriously: Corporations have fiduciary obligations only to the shareholders. Civilization disrupting technology can not, must not, be trusted to sole custodianship of such decision making regimes.
Russia and China suck, but it's still better that only a few countries have nukes instead of millions of people having 1 each :) If everybody had nukes we'd stop living in big cities :)
I believe you've touched on the crux of the matter. Corporations can not only censor, but also train AI models utilizing biased data, granting them significant influence over shaping the AI discourse among the general populace. This warrants serious contemplation and discussion.
Not exactly the point but current smartphones are comparable to 90s supercomputers in terms of FLOPs - an iPhone 14 can hit 2 TFLOPs whereas ASCI Red achieved 2.38 TFLOPs in 1999. Now the comparison is muddied by the fact that supercomputer GFLOPs were fp64 while modern GPUs are optimised for fp32 (with machine learning going for fp16 and now fp8), and the iPhone probably can’t maintain 2 TFLOPs before getting thermally throttled. But otherwise matching the 90s is far more impressive than the 70s.
> Of course, my referral to Graham in this context uses “smart” in a rather limited sense
Yes, limited in the sense that smart = can make computers do stuff.
> if it’s scammy, it will attract the kind of people who’d play in moral grey areas for personal profit; if it’s bureaucratic-authoritarian, yup, you know who you’ll find there
You don’t think smart people run scams or that only idiots want to be authoritarians?
The premise of this is ridiculous. The smartest people I know personally and that I'm aware of generally are not attracted by AI. Person 1 was interested in fundamental physics (dark matter) but had to leave physics due to lack of job opportunities. Person 2 is Terrence Tao. Person 3 is Ed Witten, etc etc.
This reflects the hubris of the tech world that overestimates its own intelligence....
Maybe I don’t have the required level of literacy but I understand « The smartest people I know personally and that I'm aware of generally » as the union of the « know personally » and « aware of generally » sets.
In most understandings of English, "and" is used as the conjunction operator https://en.wikipedia.org/wiki/Logical_conjunction (set intersection). If you were expressing a disjunction (a union), you'd use "or" or "as well as" or something similar. Obviously it's possible to eventually puzzle out from context what OP meant, but I made the same mistake as the person you're replying to when reading the comment quickly.
This seems a reach to me. I think your confusion is not unfounded and is not uncommon, but I think it's also extremely (and moreso?) common to read the sentence as the author intended. It took me 0ms to puzzle out the meaning because the intended meaning was the natural assumption of my brain.
Besides that, it seems a bit absurd to me to say that "and" and "as well as" have entirely distinct meanings. They're not totally interchangeable, but there is overlap.
Thanks! This is the sense in which I had intended the sentence. I wish that I knew Terrence Tao and Ed Witten! The smart people I know personally I wanted to keep anonymous....
Also consider: if LLMs weren't as big as they were, would the people pursuing this still be the "smartest engineers?" Would their efforts still be as brilliant? If you answered no to any of those, why not?
It still doesn't occur to me that the 1-3% smartest people in the world would be attracted to LLM/AI just because they're intelligent. The most intelligent people I know are variously uninterested in AI because AI threatens many of their beloved interests in which the confluence of their intelligence can be expressed, i.e creative activities like music, composition, visual art, interactive works like games, and writing.
Perhaps I'm outing myself as a so-called unintelligent person here, but I don't see how AI is a threat to creative hobbies of mine like music and writing. AI isn't going to stop me from sitting down at the piano or writing a short story. There will always be people in the world better at those things than me, so what difference does it make if an AI may or may not be better at them too?
And if I can't tell the difference between a human generated piece of art and an AI generated one, then I find that fascinating, not threatening. It says a lot about 1) your own tastes and perceptions, 2) the meaning of art in general, and 3) what exactly it means when activities previously considered solely human are mastered by programs, such as what happened in the past with chess engines, etc.
I think these are all exciting developments, and I look forward to seeing how art will evolve in the way the emergence of chess engines changed strategy for humans moving forward.
> AI isn't going to stop me from sitting down at the piano or writing a short story. There will always be people in the world better at those things than me, so what difference does it make if an AI may or may not be better at them too?
The threat is that it will no longer be viable to make money from creative works. A completely trivial and laughable concern if you don't live in a society that believes you have to "earn a living".
> The threat is that it will no longer be viable to make money from creative works.
Although I see a comparison to previous technologies replacing human jobs, I'm willing to put that point aside and focus solely on AI.
Will AI generated content be cheaper? Will AI content be more interesting and engaging than anything a human could create? If there is no economic demand for human generated artwork, then should it still be viable to make money from it? Is there reason to believe the people who remain in the art market won't be able to make more money by occupying niches that AI art doesn't serve, or by selling to "hipsters" or "luddites" who are anti-AI art? It used to be the only way to make real money through art was having a patron. Perhaps there will be wealthy people who will still be willing to patronize human artists.
Historically, we are in a special place in history where average people can make a living through creative work, but it's not at all an easy industry to break into, even today, and it has always been competitive and everchanging, just as art itself. Economics aside, people have always produced art for intrinsic purposes, and they will continue to do so even when the profit motive is gone.
I don't disagree with any of that. The GP said they didn't understand why people would be afraid of AI art. The reason is that many people pay rent by producing art, and that will no longer be possible for many of them in an age of (good) AI art
> are you able to name many great artists in history who died rich
That's not relevant. There are tons of working artists today and there have been for decades. But if a person can't pay rent by selling their art, things suddenly change for millions of people.
It makes it so that the actual intellectually interesting analysis of art, when viewed through the lens of an LLM, becomes a cognitively simple to understand puzzle of matching prompts to effects. There's not a lot of interesting conversation to ask about color theory or composition.
The interesting stuff is asking what is an artist trying to convey or explore within the nuances of the medium they are working with: why does neon genesis rebuild movies use live action camera techniques and what does that do to the overall themes of the work? Why does one painting use the texture of the paint instead of the color theory of the paint as a message conveyor? How is the use of reflective or matte surfaces communicating an environment in an installed art piece?
Furthermore, of the very smart people I know, the intellectually stimulating stuff is kind of bizarre, and avenues of finding this stuff or funding this stuff is being drowned out by comparatively understimulating works. Artists usually fund their more unusual pursuits with the commercial work that MidJourney/Stable Diffusion targets.
> There's not a lot of interesting conversation to ask about color theory or composition. The interesting stuff is asking what is an artist trying to convey or explore within the nuances of the medium they are working with
For some, this is an important aspect of art appreciation, and I would expect those people to continue to gravitate toward human produced art. For other people who don't have that background, and don't think about those things, the difference won't be relevant to them, and they'll judge the end product based on their own personal and arbitrary criteria. What's interesting about an artwork to one person makes it banal to another, and no amount of academic study or criticism will change peoples' tastes.
Let's not forget that often creators themselves aren't always actively thinking about composition, texture, color, etc, and part of the work for them can be as black box out of their subconscious as the creation of a neural network.
Similar to people being fooled by white wine dyed red, I think we'll also have critics who remark on how absolutely human a work of art has to be due to X/Y/Z, only for it to be revealed to be an AI work. That I think encapsulates a lot of what I look forward to in terms of the debate over what exactly about art is human-specific and what it means for something, human or not, to have artistic capabilities.
I'm just pointing out, in my anecdotal experience, that the smartest people I know don't have interest in AI because of the way AI clashes with the things they find the most intellectual stimulation from. None of what you said really addresses any of that, and I really have no interest in debating the subjectivity of AI art on the internet, as there's nothing about it that hasn't already been said.
I spoke to an artist the other say, he just said, in true artists style, that maybe he can collaborate with the bots, otherwise he doesn’t care about AI art.
I thought it would be upset so I kind of introduced the topic lightly.
I was relieved how much he didn’t care.
I completely agree with you. A button click image will never relate to most people like a real work of art.
I think the same about television. I think we all like Seinfeld because we know he is a real person with real experiences that we can relate too. I think this is important.
AI/ML and LLMs definitely attract a certain cadre of very smart people. I think they enjoy the way that AI/ML sort of inverts the normal programming mindset and the strategies of handling lots of data.
But I also agree that for many other smart people, what these things do is sort of an "obvious" result of doing bulk statistics on lots of data. The particulars can sometimes be surprising, but the aggregate result "thing that confidently generates false data" isn't.
In between, there seems to be a very passionate, and very excited group of people who don't get it and think the AI magic is "thinking". Some of these people may also be smart, but they sure don't have a strong complete vertical understanding of the topic. These people are numerous, and provide lots of conflicting and confusing information at about the level of through of the latest popular science breakthrough article -- setting wrong expectations all along the way.
The very premise that 'smartness' is easily quantifiable is also flawed. Attempting to rank all human minds on some hierarchical scale is going to run into the same problems as with human bodies. Is it better to be able to run the fastest, or to lift the heaviest weights, or to have the best hand-eye coordination when it comes to throwing and catching objects? Will specialization in one area of mental skill inevitably create weaknesses in other areas of mental skill (e.g. the strongest weightlifters are never the fastest runners over long distances)?
Not only that, there's the whole cult of intelligence issue, which attracts people who are very skilled at presenting themselves as uber-intelligent even if their greatest talents like in the realm of social engineering and bureaucratic manipulation (many highly successful academics fall into this category). This is a kind of 'smartness', but not necessarily the kind that helps one grasp complex maths.
What we call 'smartness', or 'actual skill' is the product of some degree of native intelligence and the amound of time and energy spent developing that intelligence in a particular area via diligent effort.
I would add that some of the stupidest things I've seen people do were actually done by people who were, by any objective measure, extremely smart - either from hubris or general otherworldliness....
That's also a trainable skill, learning how to learn new concepts quickly. Here is where many people report LLMs like ChatGPT are a huge time-saving boost.
This might also be an either-or situation - i.e. some people specialize in the ability to quickly grasp a wide variety of knowledge at a relatively shallow level, others specialize in in-depth knowledge of a particular subject, and it's impossible to be really good at both.
I don't know how many of the tech world now the math/phys world. We had a core set of classes in college around that, but I couldn't compare with computing at the time. After years in juggling with semi-advanced CS/FP ideas, whenever I get into "old" science, it get sobering fast.
Let's fix this by taking the superlative out of the title. I've done so now.
The trouble with reacting to provocations like this, even if the indignation is genuine (it's basically always genuine) is that it leads to much less interesting discussions.
>Paraphrasing Paul Graham, “the most ambitious challenges attract the smartest people”.
What is said here could almost have the same logical outcome as saying "the most challenging problems have no brilliant solution until one of the smartest people come along".
Nothing about attraction whatsoever, factored out instead.
Still seems pretty sensible when it works.
Now pick 3 top mathematical wizards.
Would it have been possible to achieve that status unless they had already accepted (and overcome!) challenges that were too much for those who came before?
And maybe still have their hands full with remaining challenges which were already accepted before AI came near enough to be considered within reach?
It should be most obvious to those having the best numerical abilty, to figure that more than just 3 of the smartest people will be needed with the growing number of challenges that nobody has even tried to overcome, sometimes not even recognize.
It's probably good to go into a post like this with some basic assumptions about the context: works in/around/with SV "good new boys" network, easy 6-figure software jobs and weird jockeying around securing vast sums extracted from inbred billionaire weirdos and distributed for blood by fast-talking VC weirdos.
The outsider perspective on this is to fancy yourself a smart inventor who's going to outwit/outsmart/outplay the competition with your own company, find product market fit. This factory makes software widgets, and striking gold with such ephemeral products is as hard as striking gold with making sounds (which is the magic that the pop music industry makes happen regularly, somehow). (So, smart enough to write and release quality, reliable software to the public, or build a team capable of doing this).
Demand is so high for any software that becomes popular, that proxies for value are accepted with a "we'll worry about profit later" kind of attitude. So this justifies focusing on a hyped technology even if it isn't real. But if the hyped technology is real, even better.
A rational agent, selling their intellectual labor, understanding the current lay of the land, will see that the subject of "improving AI and integrating AI into applications" is going to be in high demand for a period of time, which means higher income.
I can explain why a higher income is desirable, if you like.
Definitely getting crypto vibes here: "the real hackers are the ones staying up late improving Merkle tree implementations!" Probably read something like that three years ago or so, likely on Medium as well.
The narratives people tell themselves to self-justify their involvement are weird.
> You can find interesting problems in many banal situations. You only have to look.
Unfortunately (unrelated to AI and co), this is also a problem; many engineers write overcomplicated solutions for simple / banal problems, because they get bored otherwise.
Then give your engineers more interesting problems. Or give them more freedom. Or give them days where they can work on whatever they want. There are plenty of options to correct that course.
Generally what's needed to climb career ladders are social skills and networking skills. Coincidentally many smart people are not neurotypical and struggle with that.
The problem is the "concept" of meritocracy. Many smart people may face significant amounts of adversity or may due to their circumstances be stuck in poverty.
One example of that are people with adhd, autism, asperger's etc, many of whom score high on iq tests. But due to their handicaps they still struggle significantly more than neurotypical people because they are a minority.
> Generally what's needed to climb career ladders are social skills and networking skills.
And also, just the will to climb. Smart people often look at the big picture and come to view the rat race as pointless, thus staying on a lower rung of the ladder for life.
Child prodigy, IQ assessed between 250 and 300, contemporary, classmate, and friend of Norbert Wiener (father of cybernetics, also a child prodigy), and ended up ... making a life's work of studying bus transfers. As in the tickets, not commute patterns themselves.
IQ of 250 is a meaningless statement. IQ of 200 means “better than 1/10 billion”. There aren’t enough humans to make a statement like that meaningful, and an IQ of 250 is positing something like “smarter than all humans living in the solar system centuries from now”.
Maybe when we make AGIs in bulk we’ll be able to talk about analytical capacity on some more meaningful scale than IQ, but for now, numbers like you gave attached to a human are just silly hyperbole.
This article reminds me of SSLeay by Eric Young which openssl is based on. Without his effort we'd all still be digging in the dirt with text in the clear. Open source is in the debt of a few gifted altruistic hard working people.
ITT: angry people saying fellow brilliant hackers jart and ggerganov are actually not that smart. I'm surprised jart even keeps posting here when we're so hostile and toxic.
Like Paul Graham, I'm in my 50s. Unlike Paul Graham, I've never been confident that I can identify what a 'smart' person is. The only thing I know for sure is that it rarely has anything to do with what people like to call 'intelligence'.
My problem and shame has been the false negatives I have. I meet people with great potential all the time - who have done more, or know more, or think more deeply about something than others. And it's apparent to me - they're going to do great stuff.
My problem is that there are also plenty of people who don't trigger that sense in me, but that have gone on to do really great stuff all the same. So I try to be humble about my assessments.
I agree that whatever separator line there is, it's not 'intelligence', maybe more like intelligence * drive * longevity, and I may be keying in on the first factor but not others.
105 comments
[ 1.7 ms ] story [ 190 ms ] threadThere is the assumption that it will be written about.
There is the assumption that "solving" is the appropriate response.
Combined, they describe a constrained perspective. A volume which, compared to the alternative, is infinitely small.
So consider that the next time you feel a riddle's tug, smarty.
I saw people saying it only loads the weights necessary, but all the weights are necessary (unless the architecture is heavily modified). Does it stream or load only the currently necessary layers in a way?
Most of the drama is contained in this thread, if you like reading this sort of stuff: https://github.com/ggerganov/llama.cpp/pull/711#issuecomment...
Well, I can't speak about strengthening democracy for sure, but what I see is that AI will become a significant means of digital production, and we wouldn't want those means to be in the hands of a few people.
> but it potentially enables AI to run on any device with a low footprint for surveillance purposes as well,
Yes, a low barrier for inference cuts both ways. That is a risk.
> democratising the means of digital production is the difference between a dystopian authoritarian future and an empowered (and admittedly chaotic) humanity
The usage here means make accessible to everyone.
As long as surveillance tech requires big resources, governments will have the edge. Miniaturize it, decentralize it, and it means citizens and public interest organizations can hold the government accountable more easily (this is, notably, what OSINT is all about).
If you think this is the modern-day equivalent of a printing press, it's vital to avoid this being in the hands of only one or two multi-billion-dollar corporations with a history of morally questionable moves.
If you think privacy rights and freedom from surveillance are critical to democracy, you want your smart speaker or video doorbell to do all its processing on board, without uploading recordings to the cloud. And you want your neighbour's doorbell to do the same.
And frankly, nobody goes after Microsoft Word when someone writes Xi Jinping satire using it, or ICE buy it - because people know it's a tool rather than a platform.
On the other hand, if you think ML technology is going to produce the next wave of fake amazon reviews and forum spam - you might be a lot less eager to see it in the hands of everyone.
If you think society will be slow to adapt to things like deepfakes; and that cops and corporations will immediately start using things like facial recognition technology - well, then wider access to ML might seem to have a great many downsides.
I think AI has the same potential: filling administrative work, walking the complexity of an information overloaded world, keeping up to date with laws you can't ignore but have no way to know about, planning logistics that our modern life impose on us...
Suddenly, like the CEO of today, every one will be able to have a secretary.
Yes, that can make a huge different.
Don't think about AI as the thing we have today. GPT the nokia 3310 of AI. Amazing, but very basic.
Think that one day, somebody will come out with the iphone of AI.
- reading boring documents and giving me the gist of it
- classifying my emails
- checking for typos and problems in my writing
- writing the boiler plate of documents
- translating doc
- doing basic search on a topic, such as legal obligations that haven't change in the last 2 years
It's cheaper than my secretary, faster, and has roughly the same error rate for those tasks. I have to double check her work anyway.
With plugins, people are already making stuff like booking a plane, another thing my secretary does.
It will expend.
I don't think you realize the amount of risk involved in giving these opaque probabilistic systems control over data and compute. They are not people replacements and assuming so and acting on it will create more problems than it solves... Because extending thee lines to wrap around your needs is wishful thinking and there is a lot more complexity and hidden unintended consequences than assumed when saying chatbots capable of using APIs will be dependable personal assistants, which is the point I'm trying to make.
After playing around with ChatGPT for a few hours I have a hard time seeing how someone could objectively call it "basic".
But compared to what we have today, it's very basic.
Same for chatgpt.
The washing machine has reached a tipping point, we hit diminishing returns from now on.
The AI is in the nokia phase. It's at the primitive stage of development.
You lost me at iPhone for AI. I think I know what you mean but it can be expressed better.
In other words the jump ahead from GPT to broadly useful AI is akin to the jump from the old Nokia feature phone to iPhone. Don’t let the current limitations of GPT color what you think AI will be generally capable of within the next decade.
I think something that was majorly lost in the past 60 years is the secretary. I agree with you that it’d be great if everyone had one - the working relationship between a secretary and the person they support can be amazingly productive.
The biggest game changer was the fossil fuel that enabled that washing machine, and all the other modern devices we rely on.
That same substance is used to as chemical feedstock for fertilizers and more than double yields, and more than double the amount of "arable" land. If we were to run out, half of humanity would likely starve.
But I was thinking of the tool-vs-platform distinction in the context of the censorship question raised by Midjourney blocking images of Xi Jinping.
'Platforms' like Facebook and Github often censor things like porn, terror recruitment, and political satire fearing reputational damage and loss of advertisers from having such content on their domain, next to their logo - whereas offline tools like MS Word aren't subject to the same fears.
So people who think political satire is a particularly important form of speech might prefer an offline tool for their greater censorship resistance, irrespective of their opinion on issues of copyright.
I don't see how it should be any different from the PC revolution. We've all got a 1970s supercomputer in our pocket now, for better or worse. But it's certainly preferable to only the richest and most powerful having one.
“Hey OpenChatGPT, find a list of every person who has ever wronged me and constantly harass them with death threats from behind seven proxies”
Yeah that will make the world better.
It works both ways. Attack and defense.
You may not feel the need to construct a strong defense because of your disposition and mental model, but you should because of the motivations and mental models of those building a strong offense.
A take as old as time. History books are always written by the victors.
More seriously: Corporations have fiduciary obligations only to the shareholders. Civilization disrupting technology can not, must not, be trusted to sole custodianship of such decision making regimes.
It's literally the dictatorship of the majority over all the other minorities.
We should move towards decentralising power (ideally up to the individual level), not strengthening democracy.
Democracy is just a tool we use to decide what to do with the collective power we have.
And enabling everyone to run LLMs definitely helps everyone to become more independent of governments and large corporations.
I don't believe that comment was necessary.
Yes, limited in the sense that smart = can make computers do stuff.
> if it’s scammy, it will attract the kind of people who’d play in moral grey areas for personal profit; if it’s bureaucratic-authoritarian, yup, you know who you’ll find there
You don’t think smart people run scams or that only idiots want to be authoritarians?
This reflects the hubris of the tech world that overestimates its own intelligence....
This Medium blog post is not the PR person for the "tech world".
What about “I want to thank both you and the original article author”? Should it be “or” instead of “and”?
This seems a reach to me. I think your confusion is not unfounded and is not uncommon, but I think it's also extremely (and moreso?) common to read the sentence as the author intended. It took me 0ms to puzzle out the meaning because the intended meaning was the natural assumption of my brain.
Besides that, it seems a bit absurd to me to say that "and" and "as well as" have entirely distinct meanings. They're not totally interchangeable, but there is overlap.
Also consider: if LLMs weren't as big as they were, would the people pursuing this still be the "smartest engineers?" Would their efforts still be as brilliant? If you answered no to any of those, why not?
And if I can't tell the difference between a human generated piece of art and an AI generated one, then I find that fascinating, not threatening. It says a lot about 1) your own tastes and perceptions, 2) the meaning of art in general, and 3) what exactly it means when activities previously considered solely human are mastered by programs, such as what happened in the past with chess engines, etc.
I think these are all exciting developments, and I look forward to seeing how art will evolve in the way the emergence of chess engines changed strategy for humans moving forward.
The threat is that it will no longer be viable to make money from creative works. A completely trivial and laughable concern if you don't live in a society that believes you have to "earn a living".
Although I see a comparison to previous technologies replacing human jobs, I'm willing to put that point aside and focus solely on AI.
Will AI generated content be cheaper? Will AI content be more interesting and engaging than anything a human could create? If there is no economic demand for human generated artwork, then should it still be viable to make money from it? Is there reason to believe the people who remain in the art market won't be able to make more money by occupying niches that AI art doesn't serve, or by selling to "hipsters" or "luddites" who are anti-AI art? It used to be the only way to make real money through art was having a patron. Perhaps there will be wealthy people who will still be willing to patronize human artists.
Historically, we are in a special place in history where average people can make a living through creative work, but it's not at all an easy industry to break into, even today, and it has always been competitive and everchanging, just as art itself. Economics aside, people have always produced art for intrinsic purposes, and they will continue to do so even when the profit motive is gone.
The only artists I can think of that died with at least some money had sponsors that encouraged their work.
So I’m not sure much changes here.
If AI is supposed to stop us all from working, then those people will just be better off.
That's not relevant. There are tons of working artists today and there have been for decades. But if a person can't pay rent by selling their art, things suddenly change for millions of people.
The interesting stuff is asking what is an artist trying to convey or explore within the nuances of the medium they are working with: why does neon genesis rebuild movies use live action camera techniques and what does that do to the overall themes of the work? Why does one painting use the texture of the paint instead of the color theory of the paint as a message conveyor? How is the use of reflective or matte surfaces communicating an environment in an installed art piece?
Furthermore, of the very smart people I know, the intellectually stimulating stuff is kind of bizarre, and avenues of finding this stuff or funding this stuff is being drowned out by comparatively understimulating works. Artists usually fund their more unusual pursuits with the commercial work that MidJourney/Stable Diffusion targets.
For some, this is an important aspect of art appreciation, and I would expect those people to continue to gravitate toward human produced art. For other people who don't have that background, and don't think about those things, the difference won't be relevant to them, and they'll judge the end product based on their own personal and arbitrary criteria. What's interesting about an artwork to one person makes it banal to another, and no amount of academic study or criticism will change peoples' tastes.
Let's not forget that often creators themselves aren't always actively thinking about composition, texture, color, etc, and part of the work for them can be as black box out of their subconscious as the creation of a neural network.
Similar to people being fooled by white wine dyed red, I think we'll also have critics who remark on how absolutely human a work of art has to be due to X/Y/Z, only for it to be revealed to be an AI work. That I think encapsulates a lot of what I look forward to in terms of the debate over what exactly about art is human-specific and what it means for something, human or not, to have artistic capabilities.
I thought it would be upset so I kind of introduced the topic lightly.
I was relieved how much he didn’t care.
I completely agree with you. A button click image will never relate to most people like a real work of art.
I think the same about television. I think we all like Seinfeld because we know he is a real person with real experiences that we can relate too. I think this is important.
But I also agree that for many other smart people, what these things do is sort of an "obvious" result of doing bulk statistics on lots of data. The particulars can sometimes be surprising, but the aggregate result "thing that confidently generates false data" isn't.
In between, there seems to be a very passionate, and very excited group of people who don't get it and think the AI magic is "thinking". Some of these people may also be smart, but they sure don't have a strong complete vertical understanding of the topic. These people are numerous, and provide lots of conflicting and confusing information at about the level of through of the latest popular science breakthrough article -- setting wrong expectations all along the way.
Not only that, there's the whole cult of intelligence issue, which attracts people who are very skilled at presenting themselves as uber-intelligent even if their greatest talents like in the realm of social engineering and bureaucratic manipulation (many highly successful academics fall into this category). This is a kind of 'smartness', but not necessarily the kind that helps one grasp complex maths.
What we call 'smartness', or 'actual skill' is the product of some degree of native intelligence and the amound of time and energy spent developing that intelligence in a particular area via diligent effort.
Meanwhile, intelligence research continues to be one of the more fruitful (and reliable) fields of the social sciences
This might also be an either-or situation - i.e. some people specialize in the ability to quickly grasp a wide variety of knowledge at a relatively shallow level, others specialize in in-depth knowledge of a particular subject, and it's impossible to be really good at both.
Maybe to a limited extent, but most of the difference is likely natural ability.
Wisdom is knowing not to put it in a fruit salad.
The trouble with reacting to provocations like this, even if the indignation is genuine (it's basically always genuine) is that it leads to much less interesting discussions.
What is said here could almost have the same logical outcome as saying "the most challenging problems have no brilliant solution until one of the smartest people come along".
Nothing about attraction whatsoever, factored out instead.
Still seems pretty sensible when it works.
Now pick 3 top mathematical wizards.
Would it have been possible to achieve that status unless they had already accepted (and overcome!) challenges that were too much for those who came before?
And maybe still have their hands full with remaining challenges which were already accepted before AI came near enough to be considered within reach?
It should be most obvious to those having the best numerical abilty, to figure that more than just 3 of the smartest people will be needed with the growing number of challenges that nobody has even tried to overcome, sometimes not even recognize.
Whose ambition?
Smart how?
For what outcome? For who?
Would that be even efficient? (there are many other issues to "panic" over other than AI)
The outsider perspective on this is to fancy yourself a smart inventor who's going to outwit/outsmart/outplay the competition with your own company, find product market fit. This factory makes software widgets, and striking gold with such ephemeral products is as hard as striking gold with making sounds (which is the magic that the pop music industry makes happen regularly, somehow). (So, smart enough to write and release quality, reliable software to the public, or build a team capable of doing this).
Demand is so high for any software that becomes popular, that proxies for value are accepted with a "we'll worry about profit later" kind of attitude. So this justifies focusing on a hyped technology even if it isn't real. But if the hyped technology is real, even better.
A rational agent, selling their intellectual labor, understanding the current lay of the land, will see that the subject of "improving AI and integrating AI into applications" is going to be in high demand for a period of time, which means higher income.
I can explain why a higher income is desirable, if you like.
You can find interesting problems in many banal situations. You only have to look.
The narratives people tell themselves to self-justify their involvement are weird.
Unfortunately (unrelated to AI and co), this is also a problem; many engineers write overcomplicated solutions for simple / banal problems, because they get bored otherwise.
Tons of smart people I know are doing mediocre jobs.
The fact someone is tackling hard problems doesn't make them the smartest people in the room, you need a combination of risk taking and smart in life.
And also, just the will to climb. Smart people often look at the big picture and come to view the rat race as pointless, thus staying on a lower rung of the ladder for life.
<https://en.wikipedia.org/wiki/William_James_Sidis>
Child prodigy, IQ assessed between 250 and 300, contemporary, classmate, and friend of Norbert Wiener (father of cybernetics, also a child prodigy), and ended up ... making a life's work of studying bus transfers. As in the tickets, not commute patterns themselves.
Several HN submissions though no substantive discussion: <https://hn.algolia.com/?q=William%20James%20Sidis>
Sidis Archives: <https://www.sidis.net/>
Maybe when we make AGIs in bulk we’ll be able to talk about analytical capacity on some more meaningful scale than IQ, but for now, numbers like you gave attached to a human are just silly hyperbole.
2. <https://news.ycombinator.com/item?id=21078083>
My problem is that there are also plenty of people who don't trigger that sense in me, but that have gone on to do really great stuff all the same. So I try to be humble about my assessments.
I agree that whatever separator line there is, it's not 'intelligence', maybe more like intelligence * drive * longevity, and I may be keying in on the first factor but not others.