None of the things you listed mean much when it comes to ethno-nationalism. There are a lot of Chinese that love American brands but still hate the US for nationalist reasons.
Among my relatives I would say all are anti-US. About 5-6 of them vehemently so and want the war to start immediately.
I grew up in China and if you think there’s less propaganda in China compared to the US I don’t know what to tell you.
It's heavy and the extra encumbrance is not great for general purpose computing, though I never felt the need for extra screen real estate while I'm coding, watching youtube, sending emails etc.
That said I still use my AVP regularly, it's a great home theater system for 4k HDR content that's portable for travel. I've owned other headsets like the valve index but the novelty of 3d gaming wore off pretty quickly.
VR gaming is in a weird state because a lot of devs still haven't gotten over the hill of trying to design everything almost exactly the same as PC games despite the fundamentally different paradigm.
For an example of what I mean, compare a game that actually embraces the medium like Gorilla Tag to the endless numbers of "PC FPS, but in VR with nausea-inducing stick movement" shooter games that don't.
This is kind of the visual equivalent of asking an LLM to count letters. The failure is more related to the tokenization scheme than the underlying quality of the model.
I'm not certain about the specific models tested, but some VLMs just embed the image modality into a single vector, making these tasks literally impossible to solve.
I think a lot of the confusion on whether LLMs can think stems from the fact that LLMs are purely models of language and solve intelligence as a kind of accidental side-effect.
The real problem that an LLM is trying to solve is to create a model that can enumerate all meaningful sequences of words. This is just an insane way of approaching the problem of intelligence on the face of it. There's a huge difference between a model of language and an intelligent agent that uses language to communicate.
What LLMs show is that the hardest problem - of how to get emergent capabilities at scale from huge quantities of data - is solved. To get more human-like thinking, all that is needed is to find the right pre-training task that more closely aligns with agentic behavior. This is still a huge problem but it's an engineering problem and not one of linguistic theory or philosophy.
What we feed these huge LLMs is not just language, but text. and an enormous amount of it. The transformer is an arbitrary sequence to sequence modeller.
Think about what is contained (explicitly and implicitly) in all the text we can feed a model. It's not just language, but a projection of the world as humans see it.
GPT-3.5 Instruct Turbo can play valid chess at about 1800 ELO, no doubt because of the chess games described in PGN in the training set. Does Chess suddenly become a language ability because it was expressed in Text ? No
Chess is a great example because it highlights the subtle difference between LLMs and agents. What GPT3.5 does is not quite playing chess but creating realistic chess moves that a human might make.
An LLM could play chess though, all it needs is grounding (by feeding it the current board state) and agency (RF to reward the model for winning games)
3.5 Instruct (different model from regular 3.5 that can't play) can play chess. There's no trick. Any other framing seems like a meaningless distinction.
The goal is to model the chess games and there's no better way to do that than to learn to play the game.
>all it needs is grounding (by feeding it the current board state)
The Model is already constructing a board state to play the game.
>agency (RF to reward the model for winning games)
Predict the next token loss is already rewarding models for winning when the side they are predicting wins.
And when the preceeding text says x side wins and it's playing as x side then the loss is rewarding it to do everything it can to win.
I agree different goals and primary rewards led to this ability to play and with it , slight manifestations(GPT can probably modulate level of play better than any other machine or human) but it is nonetheless playing.
The explanation is perfectly sensical, just too complex for humans to understand as the model scales up.
The thing you're looking for - a reductive explanation of the weights of a ANN that's easy to fit in your head, does not exist. If it were simple enough to satisfy your demands, it wouldn't work at all.
Yet, when a master player makes a decision what move to play, they often have concrete reasons for it, that they discuss in after game analysis. They evaluate some advantage or chances higher than others or some risks greater than others and calculate specific sequences ahead to be sure to solve a subproblem correctly and base their decision on that.
I was writing an SVG editor web app but didn't know anything about the fabricjs api. Here's a portion of my prompts with chatgpt (april 2023 version, the date at the top is the share date). The initial code came from a different chatgpt thread.
I finished the app in 2 days, with a third day for css/visual styling. I previously might have hired someone to do this or tried to figure it out myself, which would have taken about a month.
At the end of this thread chatgpt kind of goes off the rails a bit and fails to center the uploaded image. I think this is because it can't execute the code and see its results, and can only take blind guesses at what the problem might be.
I still had to write about 10% of the code myself, but it's about 10x faster for me to use chatgpt. I think I'd use chatgpt even if it were slower, because I prefer "thinking in natural language" vs "thinking in code"
I think this approach isn't ideal because you're representing pixels as 150x150 unique bins. With only 71k fonts it's likely a lot of these bins are never used, especially at the corners. Since you're quantizing anyways, you might as well use a convnet then trace the output, which would better take advantage of the 2d nature of the pixel data.
This kind of reminds me of dalle-1 where the image is represented as 256 image tokens then generated one token at a time. That approach is the most direct way to adapt a causal-LM architecture but it clearly didn't make a lot of sense because images don't have a natural top-down-left-right order.
For vector graphics, the closest analogous concept to pixel-wise convolution would be the Minkowski sum. I wonder if a Minkowski sum-based diffusion model would work for svg images.
Thank you for the suggestion. A couple of ML engineers with whom I've spoken after publishing the blog also suggested that I should try representing x and y coordinates as separate tokens.
In pixel space a convnet uses pixel-wise convolutions and a pixel-kernel. If you represent a vector image as a polygon, the direct equivalent to a convolution would be the Minkowski sum of the vector image and a polygon-kernel.
You could start off with a random polygon and the reverse diffusion process would slowly turn it into a text glyph.
A doctor is a lot more than just a black box taking the patients' descriptions and measurements, and running regressions on them. Doctors can touch, feel, understand, comfort in ways that our sensors or tensors (hah) can't.
Same applies for a teacher too, in various other aspects. Reducing important professions into statistical models is exactly the kind of crappification that the author's talking about. The logical conclusion of perfect sensors and tensors is not here, and the lacking substitutes along the way will be profit-driven, not solution-driven.
Sure they can, but many don't either because of lack of time, or quite frankly, because many doctors are bad at their job. And even in the best case scenario we will never be able to provide doctors to 100% of the population. For many people the choice won't be AI or a (free) caring, passionate doctor who has time to understand you and answer your questions, it's AI or nothing.
Same with teaching. A lot of people simply don't have access to teachers, and if even the ones that do, might not have teachers that have the time and knowledge to actually teach what they want to learn.
The richest countries in the world cannot even produce enough competent doctors and teachers to fill their current needs. A world that produces enough skilled human doctors to meet everyones needs is even more science fiction than a world with skilled AI doctors.
> The richest countries in the world cannot even produce enough competent doctors and teachers to fill their current needs.
They can easily produce enough doctors, they just don't. A couple of reasons for this: schools inflate the amount of education required so they they can make more money and doctors go along with it (and a crazy amount of licensing requirements) to prop wages up by keeping the supply of doctors artificially low.
You could be an ICU nurse with 20+ years of experiences. Want to make a jumpt to becoming a doctor? You have to start ALL the way from the beginning of med school as if you an 23 year-old humanities major who decided to go to med school. Your 2 decades of hands-on medical experience counts for exactly nothing in the eyes of medical schools and certification boards. Does anyone really believe this is a good way to run things?
Speaking from one of the formerly rich countries (UK), we treat our doctors and teachers incredibly shabbily - long hours, low pay, terrible conditions. It's frankly a miracle that anyone over the last 15 years has gone into either profession.
Fix the low pay and terrible conditions and yeah, you'll easily produce enough doctors and teachers, but late-stage capitalism isn't going to do that...
If the UK where to offer doctors the best pay and working conditions in the world, it could fix the UK doctor shortage, but only by 'stealing' doctors from other countries and making their situation even worse. To the best of my knowledge there aren't many empty slots at UK medical schools due to no one wanting to be doctors.
It's 'easy' for any one richer country to fix their problems simply by outspending and buying up resources from a 'poorer' country (in fact some people claim the UK's problems are due to other countries buying up all the UK doctors and nurses), but that doesn't solve the global problem
> there aren't many empty slots at UK medical schools
Also underfunded and treated shabbily (like all the educational establishments in the UK.) I should have been clearer, I suppose, and said that just improving conditions for the existing doctors and teachers is a stopgap, what's actually needed is a burning out of the hideous policies of the last 12 years and a solid return to a more socialist approach to government.
The richest countries in the world choose not to produce enough competent doctors because of capitalist incentives, not because it's actually impossible.
By itself it's an argument for both; the argument for "we can't have more doctors" is "we want some of those people to do other things besides doctoring".
The fallacy you are falling victim to, which is common in these debates, is comparing an LLM teacher to a human teacher as a 1-1 replacement, when really you need to be comparing an LLM teacher to what a child has today outside of access to a human teacher: static books and today's internet + search engines.
It's very easy for me to see how an "LLM teacher" developed and trained specifically for that purpose could be of HUGE value over that status quo. That doesn't mean that the child's human teacher goes away, only that they now have access to a new amazing tool at home as well.
Unfortunately most doctors don’t have the time to be that hands on and at the end of the day are just taking your symptoms and comparing it to their flesh database of illness.
There is a lot of value from just having help diagnose/triage people with illness. Certainly not a replacement but definitely a complement to get access to healthcare to more.
> The logical conclusion of perfect sensors and tensors is not here, and the lacking substitutes along the way will be profit-driven, not solution-driven.
Quite possibly both; governments only switched to universal education instead of having 12 year olds in factories because it was good for the economy, even if some of the lessons are supposed to be good for the (for lack of a better term) "soul".
If the OpenAI pricing is indicative, reading 16k token of medical history and giving a 4k token response will cost $1.50 on GPT-4 and 6.4¢ on GPT-3.5-turbo.
The lower of those two is roughly what someone at the UN abject poverty threshold will spend in 48 minutes 30 seconds on "not literally starving to death".
>spend in 48 minutes 30 seconds on "not literally starving to death".
I don't get that. In the poorest people get by on the equivalent of about $1 per day and not many starve. In fact the only part of the planet where the population is booming at the moment is sub Saharan Africa.
Chat GTP and similar will presumably have free tiers.
(Assuming you mean literally an order of magnitude: 0.64¢ is, judging by Amazon.com, less than the bulk price of a single sheet of unused printer paper, or two thirds of a paperclip).
[0] 3 or anything equivalent to it, given 4 obviously wasn't
I was making the point that this is new tech, not available to us at all a few mere years ago, so assuming a constant cost when making predictions is difficult. Assume inference prices will go down, not up.
Ah, in that case we're on the same page. I'm expecting at least a factor of 1000 to be possible given the apparent higher efficiency of the human brain vs current computers, which is of course terrifying given how good and cheap the various creative AI already are, while also seeming like a prerequisite for robotic/car AI to be all three of "good enough", "fast enough", and "within the limited power budget".
I understood the "for free" as "with very low marginal cost"-- and no matter how you socialize healthcare/education, that's not something that humans can match.
Why the "if"? Doctors are systematically bribed by gigantic medical corporations to write prescriptions for their addictive and lethal medication. Pain killer addiction (opioid addiction) kills thousands.
>A doctor for every person, a teacher for every child, available any time and for free.
I'm sorry... I'm supposed to trust my healthcare and child's education to a piece of software whose primary feature is its ability to effectively hallucinate and tell convincing lies?
And assuming AI is at all effective, which implies valuable (which implies lucrative,) you expect services built on it to remain free?
Live performance (orchestra and operas) were for rich only. Beautiful paintings were for the noble and churches. Porcelain was something needed to be imported from another continent. Tropical fruits were so expensive that people rented them.
Now we have the affordable versions of them for everyone in developed countries, and the middle class in developing ones. Yes, often we just got inferior, machine-made or digital copies, but I personally prefer something inferior than nothing.
You're comparing the value of AI versus a human being with the knowledge and skill necessary to earn a medical degree to the value of hearing Mozart live or seeing the Mona Lisa in person to Youtube and JPEGs, as an argument in favor of AI?
>but I personally prefer something inferior than nothing.
Say that again when your AI physician prescribes you the wrong medication because it hallucinated your medical history.
> You're comparing the value of AI versus a human being with the knowledge and skill necessary to earn a medical degree to the value of hearing Mozart live or seeing the Mona Lisa in person to Youtube and JPEGs, as an argument in favor of AI?
Yes, and I think it's a pretty good analogy.
> Say that again when your AI physician prescribes you the wrong medication because it hallucinated your medical history.
I personally prefer something inferior than nothing. I just said it again.
When your human doctor prescribes the wrong medication, would you reach the conclusion that the world would be better without human doctors?
The fact is simple. Professional diagnosing is such a scarce resource that people buy over-the-counter drugs all the time. It's not AI vs doctors; it's AI vs no doctor.
When a human doctor prescribes the wrong medication, it's a mistake. One doesn't conclude the world would be better without human doctors because human beings are capable of thought, memory, perception, awareness, and when they don't make mistakes - and most don't most of the time - it's the result of training and talent.
Meanwhile, AIs don't possess anything akin to thought, memory, perception or awareness. They simply link text tokens stochastically. When an AI makes a mistake, it's doing exactly what it's designed to do, because AIs have no concept of "reality" or "truth." Tell an AI to prescribe medication, it has no idea what "medication" is, or what a human is. When an AI doesn't make a mistake, it's entirely by coincidence. Yet humans are so hardwired with paredolia and gaslit by years of science fiction that such a simple hat trick leads people to want to trust their entire lives to these things.
>The fact is simple. Professional diagnosing is such a scarce resource that people buy over-the-counter drugs all the time. It's not AI vs doctors; it's AI vs no doctor.
That's not a fact, it's your opinion, and I'm assuming you've got some interest in a startup along these lines or something, because I honestly cannot fathom your rationale otherwise. You're either shockingly naive or else you have a financial stake in putting poor people's lives in the hands of machines that can't even be trusted to count the number of fingers on a human hand.
I have no doubt the future you want is going to happen, and I have no doubt we're all going to regret it. At least I'm old enough that I'll probably be dead before the last real human doctor is put out to pasture.
AI physician prescribes you the wrong medication because it hallucinated your medical history.
The big question is will that happen more or less often than it does with human doctor? Human doctors 'hallucinate' stuff all the time, due to lack of sleep, lack of time, lack of education and/or just not caring enough to pay proper attention to what they are doing.
>Human doctors 'hallucinate' stuff all the time, due to lack of sleep, lack of time, lack of education and/or just not caring enough to pay proper attention to what they are doing.
No, they don't. If that happened anywhere near all the time, we would never have given up alchemy and bloodletting, because there would be no reason to trust medicine at all, and yet it works overwhelmingly well most of the time for most people. Meanwhile, AIs hallucinate by design.
Hallucinations are an engineering problem and can be solved. Compute per dollar is still growing exponentially. Eventually this technology will be widely proliferated and cheap to operate.
> Hallucinations are an engineering problem and can be solved.
I'd like a little more background on that claim.
As far as I've been able to tell from my understanding of LLMs, everything they create is a hallucination. It's just a case of "text that could plausibly come next based on the patterns of language they were trained on". When an LLM gets stuff correct, that doesn't make it not a hallucination, it's just that enough correct stuff was in the training data that a fair amount of hallucinations will turn out to be correct. Meanwhile, the LLM has no concept of "true" or "false" or "reality" or "fiction".
There's no meta-cognition. It's just "what word probably comes next?" How is that just "an engineering problem [that] can be solved"?
I agree it's more than a simple engineering challenge, but I do so because it is not entirely clear if even humans avoid this issue, or even if we merely minimise it.
We're full of seemingly weird cognitive biases: Roll a roulette wheel in front of people before asking them the percentage of African counties are in the UN, their answers correlate with the number on the wheel.
Most of us judge logical strengths of arguments by how believable the conclusion is; by repetition; by rhyme; and worse, knowledge of cognitive biases doesn't help as we tend to use that knowledge to dismiss conclusions we don't like rather than to test our own.
How is that bias weird? It has a straightforward explanation - the visual system has an effect on reasoning. This, as well as other human biases, can be analyzed to understand their underlying causes, and consequently mitigated. LLM output has not discernible pattern to it, you cannot tell at all whether what it's saying is true or not.
That's what happened with the internet, which was supposed to be the new Library of Alexandria, educating the world, liberating the masses from the grip of corporate ownership of data and government surveillance, and enabling free global communication and publishing.
It's almost entirely shit now. Instead of being educated, people are manipulated into bubbles of paranoid delusion and unreality, fed by memes and disinformation. Instead of liberation from corporate ownership, everything is infested with dark patterns, data mining, advertising, DRM and subscriptions. You will own nothing and be happy. Instead of liberation from government, the internet has become a platform for government surveillance, propaganda and psyops. Everyone used to have personal webpages and blogs, now everything is siloed into algorithmically-driven social media silos, gatekeeping content unless it drives addiction, parasociality or clickbait. What little that remains on the internet that's even worth anyone's time is all but impossible to find, and will eventually succumb to the cancer in due time.
LLMs will go the same way, because there is no other way for technology to go. Everything will be corrupted by the capitalist imperative, everything will be debased by the tragedy of the commons, every app, service and cool new thing will claw its way down the lobster bucket of society, across our beaten and scarred backs, to find the bottom common denominator of value and suck the marrow its bones.
But at least I'll be able to run it on a cellphone. Score for progress?
We've seen time and time again that "the people" prefers centralised, paid, convenient management of complexity.
If the majority of "the people" prefers paying a subscription for watching movies or listening to music, I doubt they'll make the effort to learn how to tune and run private LLMs locally, for medical aspects or otherwise. Not when there will be major companies spending billions on marketing more convenient options.
Guillermo Rauch's essay [1] still rings true: it's hard to forego efficiency (though it works just as well for convenience).
But I haven't seen any good argument as to why an AI teacher that is even cognitively equivalent to the real deal (and surpasses a human in everything a machine can) won't just become an intellectual worker. I'm not saying this eliminates the need for education, but it certainly erases its major component and that is to be competitive on the job market.
I don't know where this idea comes from that we can get more from language models then what we put inside. Thinking we can process any amount of data and get a competent surrogate mind out of it borders on magical thinking.
>Getting out what we put in is kind of par for the course in education, yes?
Yes, you put in a person + knowledge and get out an educated person. Is it reasonable to expect to put in GPU + text and get somehow a competent actor, however narrowly we define competence (maybe if we define actor narrowly enough)?
The foreign buyers tax, the empty home tax, and now the total ban on foreign buyers introduced this year has had zero effect on home prices. It's time to stop blaming foreigners.
Among my relatives I would say all are anti-US. About 5-6 of them vehemently so and want the war to start immediately.
I grew up in China and if you think there’s less propaganda in China compared to the US I don’t know what to tell you.