What we still don’t know about how A.I. is trained
newyorker.com
newyorker.com
Analogy to the "Garbage in, garbage out" rule.
On the positive side, 21 up 0 down. Readers appear to see the issue.
On the negative side, several "tone deaf" commenters that appear to have a blind spot for human suffering. This is not like a clinical trial of a therapeutic. It is not even like animal testing of a consumer product. OpenAI knows the material is harmful.
So-called "tech" companies like Facebook and Google are engaging in this sort of practice every day, paying people some embarassingly low wage, usually through contractors, to subject themselves to psychologically harmful content so the company can sell more online advertising services.
EDIT: I know that most people think it's unlikely and I can't give any direct evidence for it. Does that mean it's not welcome to say on hacker news?
Few days ago I came across this bone chilling AI generated Metal Gear Solid 2 meme with Hideo Kojima characters talking about how the purpose of this technology is to make it impossible to tell what's real or fake, leading directly to regulation of information networks with identity verification so that all information and thought can be traced to their sources, with the government of course becoming the sole arbiter of truth. I wouldn't be surprised at all if this turned into reality.
It will be a continuation though.
The uppermost echelons of society have waged information warfare since the dawn of modern PR in the beginning of the 20'th century. Lots of theory on this have apparently been memoryholed, but it's easy to just start with the genealogy around Edward Bernais and the plutocracy and robber baron families still existing in the interwar period.
The masses don't really think about or realise that manufactured circus and foreverwars have been going on for a hundred years while a microscopic elite that controls the media and various organs in the state have been siphoning most wealth towards themselves in a increasingly dense cloud of smoke, while promoting rebranded ideologies for the genpop and systems of control that has become advanced cybernetic systems that steer the flow of information 100% from top to bottom.
The sad thing is that you could deduce this in the noise if you visited the last remnants of actual research journalism on the internet, but that will end now.
So where is one going to get to know the powerstructures of the world? Old physical libraries only, some dude on the corner, some "crazy guy" with a harddisk from before AI took over.
Why did you pick that as your starting point? To make things seem like a linear worsening? But even wealth inequality, which you talk about at length, got better before getting worse again during that time frame in the US... so I'm having trouble following the claim.
The New Deal along with social democracy in europe was a bargaining deal against the revolutionary forces existing at the time. After this was stamped out the Gini coefficient started to rise when the masses had no alternatives and global trade and outsourcing increasingly obscurred the falling realwages.
From my middle-of-the-road perspective, everyone is trying to change how everyone else thinks, from the small insignificant details to a cult-like brainwash. Even here and now both you and I are trying to change each others and everyone who reads this's mind.
I see so much calling out of "elites" with a negative connotation, yet anyone can post anything, and everyone can ignore it. Is it general paranoia? Do you know who these people are? How are they special/different, and why should we be scared of them?
It's a testament to the absurd amount of philanthropic whitewashing, PR and media control these billionaries hold.
"Elites" have conspired to exploit the masses throughout 5000 years of civilisation, it's simply a fact of history.
It's almost physically impossible to comprehend the power a group of billionaires has to pull and push issues, narratives, law, war and discourse in general, especially when the masses have zero organisation as a counterweight, and when numbers like "trillion" just doesn't register in any meaningful way in regular people.
Elites compete with each other. They don’t sing kumbaya and cooperatively share the keys to power.
There is very little social mobility, wealth transfers very solidly between generations at the absolute top and organisation around PR and Politics is tightly integrated in this class.
Off course you can always fall from grace, but that does not in any way diminish the collective power of this class. That's why it's called a class and not a "person" or one singular family or group of people that clownish conspiracy theories would have you believe.
A good primer to the historical context could be this new book from Cambridge: "The Power of Ritual in Prehistory: Secret Societies and Origins of Social Complexity".
Elites have always formed tightly knit clubs that most couldn't get into.
This is not a reductive theory, it's based on solid academic research on wealth transfer and academic books like the one above.
It's a almost like a biological or physical property of advanced civilizations - like social patterns seen emerging in larger groups of monkeys, or a precursor to the labour divisions seen in ant hives - there's clear distinctions set fourth for an individual at birth or because of location or family, and no amount of ideology is able to change this unless very, very lucky - this is mirrored in the social mobility data.
I see a direct mirroring in today's corporations. If you join a big company, it's because you don't want to risk branching out on your own. You exchange lots of potential money for a steady paycheck, and you don't have to worry about things like finding customers, figuring out taxes, etc.
It's clear that humans need some sort of hierarchy, and I just don't see why we should be frightened of the people whose skill is organization/mediation between people. I surely don't want to play the power game with them, and I don't think it's because of their brainwashing?
Actual American billionaires lose half their fortunes in divorces (Gates and Bezos), are retired and don't need to care about controlling any media (…Gates and Bezos), or spend all day posting on Twitter (Musk, who is currently under the Rasputin-like hold of a retired postal worker named "catturd2".) They only have 24 hours a day and few of them seem to be secretly manipulating the media.
I mean, Murdoch does manipulate the media, but not secretly. It's completely obvious he's doing it! And nobody is tricked by it, they happily participate.
Wait, wasn't this just the thesis statement of Metal Gear Solid 2?
I don't believe anyone needed a conspiracy to try to make it impossible to tell what's real or fake, people have been trying to use technology to do that for decades (if not centuries) all on their own.
In my country politicians are talking about regulating social media and the internet due to "fake news". They even created a government propaganda agency to "combat misinformation online", basically people paid to defend the government against information warfare. It's way too real.
Conspiracies rely on coordination, and a counterproductive view that makes no difference in how we got to the outcome.
Death Stranding also explores ideas of what it means to find connection in a disconnected world that I think are relevant today (case in point, an article about social disconnection[1] is on the front page of HN as I write this comment)
I think those games are underrated works of speculative fiction relative to how much airtime books like The Diamond Age, Snow Crash, and Ready Player One get in conversations about nascent technology.
You don't have to speculate much. Facebook and Twitter's recommendation algorithms are doing that fairly well already.
My TLDR is, this AI character says the line "Perhaps the day will come when conversing with AIs, too, is considered perfectly normal."
>Blade Wolf: I may analyze orders, but I may not disobey them. Should I disobey a direct order, my memory would be wiped. I must destroy you.
>Raiden: What would an AI know about freedom…
Later on, after you've freed him, there's a lot more. Really, I think it's worth replaying these games at this moment in time to see the commentary firsthand.
>Raiden: Let me ask you something: What do you think you're doing here? I mean, why are you working with me?
>Blade Wolf: Because I choose to. I owe a debt, and I wish to repay it.
>Raiden: A debt, huh...? How...human of you.
>Wolf: I contest that statement. Many humans possess little or no sense of obligation at all. And many are entirely willing to trample their fellow man for personal gain.
That conversation probably comments more on the nature of humanity than of AI, but some of the others rhyme with the present in a rather intriguing way. Like when Raiden asks Wolf if he recognizes someone, and Wolf is unsure.
>Wolf: I have no "database," Raiden. The symbol grounding capabilities or my neuro-AI allow me to identify faces in only the vaguest of terms. I can judge whether "I think I may have seen him before," but I do not have the accuracy of a database.
>Raiden: That's ridiculous. You're an AI...
>Wolf: An AI modeled after the human brain, and thus as flexible and occasionally vague as any human's. Of this man, however, I have no recollection. I do not believe I have encountered him before...Most likely.
That conversation felt like the complaints people on this board voice about GPT!
> I still have relatively little communication experience with humans. My ability to read emotions not explicitly expressed by speech is lacking.
>An AI never lies....I have yet to see evidence to the contrary...But indeed, perhaps "never lies" would be an overstatement.
>An optical neuro-AI is fundamentally similar to an actual human brain. Whether they lie or not is another question, but certainly they are capable of incorrect statements.
Or the discussion about LLM...er, I mean neuro-AI driven synthetic pets, which sound like Replika in a robot dog:
>Wolf: Other AIs as conversant as I are a possibility, yes. Robotic pets with neuro-AIs are already on the market. And I understand research and development are underway to create a model capable of conversation. Do not be surprised if such a pet becomes available before too long.
>Raiden: (Hmm) If that becomes commonplace...Hard to imagine.
>Wolf: Is it? The human race has willingly embraced all of its other inventions so far. Steam engines, electricity, the Internet, cyborgs...Perhaps the day will come when conversing with AIs, too, is considered perfectly normal.
Preach it!
Will Hurd joined the board of OpenAI in May 2021. Mr. Hurd, among other things, was a documented former CIA clandestine officer for nearly nine years. He was a member of the House Permanent Select Committee on Intelligence during his tenure as a U.S. Rep replacing Mike Pompeo.
He is also a Managing Director of Allen & Co. who's Chairman is George Tenent the former director of the CIA. Further, Mr. Hurd is a member of the board of trustees of In-Q-Tel the primary external investment arm of the CIA and the broader US Intelligence Community.
1. https://openai.com/blog/will-hurd-joins
2. https://en.wikipedia.org/wiki/Will_Hurd
Imagine some visiting this website know the $10,000,000,000 from Microsoft is actually some kind of dark money from the pentagon to fund this new Manhattan project. Project code name Sydney.
Would be interested to see your math and assumptions behind this conclusion.
There’s no way that their plans to monetize this don’t include the defense/natsec industry
Sorry but where's your maths for this?
They are operating just like every other VC-backed startup in history.
Of course I don't necessarily believe any of that but it can be fun to think about.
https://twitter.com/KanekoaTheGreat/status/16405144035935109...
Sam Altman (OpenAI CEO): "Oh yes, regular contact."
Up to you to decide to what extent of government that is.
Since the topic is OpenAI, GPT-4, I'll say that I assume this tech has been around and available to a select few for years before it became mainstream. I believe it has been used to push narratives and ideas across popular internet sites and apps, creating the impression that these ideas are popular and held by the majority (not just political ideas).
The US military has shown time and time again their obsession with weaponized AI, and now OpenAI has been fully consumed by Microsoft, and is now stating how building open source AI tools was a huge mistake. There now seem to be more paths than ever towards a terminator-esque militarized robot apocalypse.
Most players in the space are still doing a lot of great open source work, so I'm still hopeful for a future where technological disparity is kept to a minimum, but it makes me really really nervous seeing how quickly a major player in the space, and a leader in many ways, has suddenly flipped to the dark side and has stopped contributing back to the community.
While I can't know what you say is true for sure, given the military's history with things like the internet, GPS, and encryption, I would not be surprised
I know that most people think it's unlikely and I can't give any direct evidence for it. Does that mean it's not welcome to say on hacker news?
Some commenters here seem to immediately focus on the potential negative effects, abuse by the military industry or the powers that be, manipulation and disinformation.
I’m personally still in awe that we are collectively witnessing the birth of a piece of technology which understands and creates and stand at the dawn of a potentially profoundly different era. I truly believe that this is the printing press once again and we are here, alive. We are living the revolution. Imagine all the positive possibilities: a word where kids have access to infinitely patient and benevolent teachers who can explain the sum of all knowledge, a world where readers can have books infinitely rewritten and modified in a subtle blurring of producing and consuming content, a world where machine can help you turn what you imagine into real shareable pictures. ChatGPT is not yet the best story teller ever but it’s already so fun I feel like a kid. I can’t stop wondering where things are going to be in a decade.
Frighteningly I guess both visions can be equally true.
> Griffith was arrested in 2019, and in 2021 pleaded guilty to conspiring to violate U.S. laws relating to money laundering using cryptocurrency and sanctions related to North Korea.[5] On April 12, 2022, Griffith was sentenced to 63 months imprisonment for assisting North Korea with evading sanctions and is currently in a federal low-security prison in Pennsylvania
Here's the article "He gave a Cryptocurrency talk in North Korea, the US arrested him"
The slides are (or were?) available. None of the content of his talk was secret or beyond what is published on the Bitcoin Wikipedia page.
https://www.nytimes.com/2019/12/02/nyregion/north-korea-virg...
Edit: it's worth noting that no transaction took place.. he was arrested because the FBI told him not to go and he did anyway.
Is that assistance? It is just a basic statement of fact. Is wikipedia guilty of providing assistance to NK? They provide far more in depth "assistance" to anyone wanting to perform a Bitcoin transaction.
Bringing this back to my original comment, you can see why the federal government would restrict the flow of fundamental technical knowledge by virtue of this extreme case. No source code or application was shared, no secrets or privileged information, merely encyclopedic facts were deemed illegal to share.
Also if you were facing the indefinite assange treatment I imagine you would seek a plea deal as well..
Also North Korea is a strange hill to die on. It's a brutal dictatorship which represses their own people and threatens to reign nuclear hell on their neighbours and the US. There's a very clear moral line that it's wrong to help them to launder money and evade sanctions, even if it weren't illegal.
This is practically the definition of a grey area, not a clear moral line.
> It's a brutal dictatorship which represses their own people and threatens to reign nuclear hell on their neighbours and the US.
No, North Korea is a group of 25 million people, most of whom are just regular people like you and me.
They are led by a brutal dictatorship but sanctions affect everyone in NK, not just the leadership.
Economic warfare is still warfare and not morally clear at all.
And has the US has not repressed its people and threatened its neighbors? Has not been brutal?
...Yes? The fact that it's basic and they already know how to do it is irrelevant. The law isn't "it's okay to give them advice as long as the advice is sufficiently generic and obvious".
Wikipedia, and the sources where its content comes from, is not intended to help anybody specific. Flying to North Korea and holding a lecture there is firmly beyond the line where the US government starts to care because it demonstrates clear intent. Especially since during the Q&A and other activities other information not available to the public could have been discussed. And even if there was no secret sauce involved, an expert's opinion can still save the North Koreans a lot of time and money when pursuing their goals.
A large part of the magic in the final product appears to be many intermediate layers of classification that select the appropriate LLM/method to query. The cheaper models (e.g. Ada/Babbage) could be used for this purpose. Think about why offensive ChatGPT prompts are rejected so quickly compared to legitimate asks for code.
Imagine the architectural advantage of a big switch statement over models trained in different domains or initial vectors. Cross-domain queries could be managed mostly across turns of conversation w/ summarization. Think about the Stanford Alpaca parse analysis diagram [0]. You could have an instruction-following model per initial keyword. All of "Write..." might fit into a much smaller model if isolated. This stuff could be partitioned in ways that turn out to be mildly intuitive to a layman.
Retraining 7B parameters vs 175B is a big delta. The economics of this must have forced a more modular architecture at scale. Consider why ChatGPT is so cheap. Surely, they figured out a way to break down one big box into smaller ones.
[0]: https://github.com/tatsu-lab/stanford_alpaca/blob/main/assets/parse_analysis.pngGiven that the emergent abilities come from the large parameter count and massive amount of training data, using smaller models seems like a distinct disadvantage.
The only reason ChatGPT is affordable is because of caching, filtering bad prompts, etc. The underlying LLM would be far too expensive if everyone was hitting it 1:1.
The stuff that absolutely must run on 100B+ models can be pre-classified by something like Ada/Babbage. Attempts at arithmetic or dimensional analysis can be binned and shipped to a monster model. Anything that is more routine information retrieval needs maybe goes to a 7B parameter model.
Lots of models with lots of temperature levels/hyperparameters/etc is the only way to achieve the kinds of performance we are seeing. The secret sauce is looking increasingly like a huge classification layer.
Nothing you see in ChatGPT is as it appears. Lengthy conversations are managed with recursive summarization. Every conversational turn is potentially handled by a different LLM.
Where can I find evidence of this?
https://arxiv.org/pdf/2203.15556.pdf
There were also some informal comparisons of GPT models with various parameter counts.
> While the desire to train these mega-models has led to substantial engineering innovation, we hypothesize that the race to train larger and larger models is resulting in models that are substantially underperforming compared to what could be achieved with the same compute budget.
This mirrors some of my experience. Training/tuning a 7B parameter model feels like goldilocks right now. We are thinking more about 1 specific domain with 3-4 highly-targeted tasks. Do we need 175B+ parameters for that? I can't imagine it would make our lives easier at the moment. Iteration times & cost are a really big factor right now. Being able to go 10x faster/cheaper makes it worth trying to encourage the smaller model(s) to fit the use case.
https://blog.google/technology/ai/introducing-pathways-next-...
From the technical report https://cdn.openai.com/papers/gpt-4.pdf : "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."
Of course there are proprietary models, that will be improved versions of the academic LLM models, however, there are no big secrets or mysteries.
As far as training, the differences between GPT-3 and GPT-3.5 (the latter being a smaller model!) demonstrate just how much fine tuning and reinforcement learning is important to the quality of the model. Merely throwing more content from the Internet at it doesn't automatically improve things.
For now its AI companies, with people, protecting us from powerful tech.
Soon it will just be the AI's protecting us from powerful tech.
Am I joking? Maybe? Maybe, not? I don't know! Everything around this new tech is moving too fast, and been too unpredictable.
And here I am, writing this manually, every word is mine, on a computer that I can't talk to yet. That already feels so 2022.
Only one thing is certain. Siri is now to Apple, what Clippy was to Microsoft, on a far far planet, long long ago.
I think they need to review their ethics, personally.
As in it’s not about the children, it’s about control? Yes.
I don't think the motives are insidious or about maximizing control, they are strictly profit driven.
If you want the world building their apps on your AI, you need to do absolutely everything in your power to make the AI brand safe. Previous chatbots have been easily coerced into saying truly awful things (e.g. Tay), and the models themselves became associated in the minds of the public with hate speech. You can't have Khan Academy or Microsoft Word potentially going on racist tirades in the midst of chatting with a student or taking meeting notes.
Thanks.
If they're right that LLMs on that scale are generally dangerous but theirs is the exception as they've very carefully calibrated it to be safe, it's extraordinarily irresponsible of them to withhold all details of steps that might make it safe...
Or that someone will automate cyberattacks, as if the government isn't already doing it?
my greatest fear is that there is only one superintelligence, with access controlled by a monopoly of a few san franciscans deciding what moral boundaries AI will have. I couldn't even get Claude+ to talk to me about houses of zodiac because it insisted it's not an astrologer, it's an AI assistant designed to be helpful blah blah blah, tell me what use is this kind of "safety"?
But I suspect that they'll also be getting many government / military contracts going forward.
Thats a very bold statement and goes against everything I've read on it so far, care to backup such a claim with some facts? Of course each of us has their own bar for such things, but for most its pretty darn high
Some things I wonder about, it says things like this:
> GPT-4 successfully passes the classic Sally-Anne false-belief test from psychology [BCLF85] (which was modernized to avoid the possibility the answer was memorized from the training data)
But it's a language model, generalizing text and performing substitutions on it, is what it excels at. "The car is yellow" is "the <noun> is <descriptor>" and it can substitute in other things, so I'm not sure how their modernization really ensure it does not pattern match on learned texts.
One human mind is a lot like another human mind, markedly less like one of an octopus or a dog (still enough similarity that the concept of, e.g., “hurt” kind of makes sense), but really unlike an LLM (and I’m not going to get into an argument as to why an LLM is fundamentally different from a human mind and why we are not even close and possibly can never be to achieving that, the only way of producing new systems like us remaining childbirth; if you don’t agree on that part then it’d be useless to discuss the matter further). We have an uncanny situation where a system radically unlike us can produce output that is mostly similar to what another human mind might produce, but unless we accept that everything around can be conscious (and believe in gods and spirits again) it’s not even a question as to whether the system can understand any of the symbols it produces or have consciousness in the commonly accepted meaning of those terms: it’s only a tool.
Note that I’m not against widening our concept of sentience, just saying it needs to happen if we want to grant an LLM sentience; and if this widening does happen, a sentient language model would be small beans compared to a philosophical revolution we’d have on our hands then.
Ethically, this is a very clear example of bait and switch.
1. We do seem to forget that even though ChatGPT is a great technical achievement its achieved by companies and humans. These companies and humans had agendas.
2. Open Source only got us so far. With these new AI models and trying to withhold competitive advantages will we see a new era of closed source tech which will like its predecessors only spew discomfort amongst its users and bridge the gap even further.
AI is dangerous because its already being looked at as an entity of which is human like.
Its perception amongst us is that its equally human on a basic intelligent level as most of us. And that I believe is scary because it puts us closer to it given it acts as a human.
Not a fan
Is this correct? I've seen varying reports of the training set size.
The source of the 45TB number is this quote from the paper:
> The CommonCrawl data was downloaded from 41 shards of monthly CommonCrawl covering 2016 to 2019, constituting 45TB of compressed plaintext before filtering and 570GB after filtering, roughly equivalent to 400 billion byte-pair-encoded tokens.
We tech people should actively go on the offence and educate whomever we can that text inference is not intelligence.
https://en.wikipedia.org/wiki/AI_effect
Maybe as "tech people" we should give the public a realistic picture of what AI research is. It's solving problems using a diverse set of techniques that include search, optimization, planning, learning, and knowledge representation. Saying that the current developments aren't AI is simply wrong.
Thank you for illustrating my point so well. What you are expressing is exactly what I'm talking about: The redefinition of what the I in AI stands for. Inferring some text is not intelligence. Ergo it's not AI what GPT is doing. The industry has redefined the terms and that's for good reason, to get the public hyped up and thereby open the money gates. But that's for the business people. Assuming that "we" in this audience are the tech people, we should push back against this and stop calling things something they are not.
That wikipedia article is just from the AI researchers point of view, it isn't what actually happens.
Based on GPT 4's performance on standardized tests and my anecdotal evidence of utility I'm getting out of ChatGPT, I feel that weak AGI arrived with GPT-4.
While it might be a more accurate name, no one but a tiny minority is going to care, and even fewer will actually change the way they refer to it.
Still, I feel that people who do know better have a moral obligation to set this straight.
There was a recent podcast with Sean Carroll interviewing Raphaël Millière where they go into this topic and some of the research on it. Two examples I can remember are: 1) DALL-E had subject-specific, domain-specific neurons, 2) language models' representations of color terms encoded the geometry of the underlying color space, e.g. vectors in RGB space.
https://www.preposterousuniverse.com/podcast/2023/03/20/230-...
I don't think we should be too quick to assume how these models work. There's a lot that even the researchers don't know and these are empirical questions that can be studied.
To cluster is to create an ideia, a concept which encompasses a set of data points. A cluster might be bigger than the set of points with which was created and intersect with other clusters. This is still what I visualize when I see image models interpolating between "concepts".
Believing that all an LLM does is the likelihood of adjacent words is probably an oversimplification. My verdict about this topic is: I don't know and I'm totally confused about what interpolation and extrapolation mean in higher dimensions
If someone was born with profound blindness, or deafness or any profound lack of sense - what do they see in their heads? Or hear? I hear a version of my voice in my head, but some people have no inner dialogue at all, or even no ability for inner imagery.
What does the brain do in these situations?
meanwhile, we've created the worlds most complicated set of dominos, and we're delighting in knocking them over.
--- input ---
Tell me about the structure of the solar system.
I want you to only use English words of Germanic origin; avoid using words derived from Latin or Greek. If there's no alternative, make up a new word by calquing the original into Germanic roots.
--- output ---
The sunstead is the sun and all things that go around it. In the middle is the Sun, a big, hot ball of glowing gas. Around the Sun are eight big worldrocks. From nearest to farthest, they are:
Mirkworld - The nearest worldrock to the Sun, small and fast. Lovegoddessworld - The hottest worldrock, with thick air and no water. Earth - Our home, the only worldrock known to have life. Redworld - A red, rocky worldrock with thin air and high mounts. Greatthundergodworld - The biggest worldrock, with many fellowtravellers and a big red spot. Harvestgodworld - A gas worldrock with a thin ringstead around it. Skygodworld - An iceworldrock that lies on its side, with a thin ringstead. Seagodworld - The farthest iceworldrock, with strong winds and blue clouds.
Besides the worldrocks, there are smaller rockbits and icebits that also go around the Sun. The rockbits are mostly in the Rockbitbelt between the Redworld and Greatthundergodworld. The icebits come from the far edges of the sunstead, in the long way out Icyrockcloud and the faraway Cloud of All.
The sunstead has a flat shape, with most things going around the Sun on near level paths. The worldrocks go around the Sun in a way called nearround paths. The Sun's pull keeps everything in the sunstead together.
Beyond the worldrocks, rockbits, and icebits, the sunstead also has spacegrit, small bits of rock and metal that fly through space. When spacegrit enters Earth's air and burns up, it makes a lightstreak that we can see in the night sky. If a bit of spacegrit reaches Earth's ground, it is called a starstone.
The Sun, which makes up most of the sunstead's weight, sends out a flow of charged bits that spreads far into space. This solar wind forms a wide bubble, the Sunshield, that surrounds the sunstead and helps shield it from high-speed bits coming from outside.
The sunstead's workings are steered by the pull between the Sun and the things that go around it. The worldrocks and fellowtravellers tug on each other, leading to small changes in their paths over time. The Sun's pull is strongest near it, so worldrocks closer to the Sun go around faster than those farther out.
As the Sun ages, it will go through changes that will have big outcomes for the sunstead. In about 5 billion years, the Sun will bloat into a Red Greatstar, swallowing Mirkworld, Lovegoddessworld, and maybe even Earth. Afterward, the Sun will shed its outer layers and shrink into a White Dwarf, a small, hot, dim star.
Lastly, the sunstead is but one of many star systems in our Starset, the Milky Road. The Milky Road is a wide, spiraling set of billions of stars, gas, and dust, and it is just one of countless starsets that make up the great Allthing.
Allthing, Starset, Sunstead, Worldrock, fellowtravelers, starstones, nearround paths.
I would've like that chatgpt came up with better names for the planets Venus, Jupiter, Saturn, Uranus and Neptune. Mirkworld, Earth and Redworld were just fine.
--- input ---
I want you to replace the word "right" in your output thereafter as follows: if it indicates direction, say "durgh; if it indicates being near or close, say "nolpi"; if it indicates correctness, say "ceza". I will also use these replacement words accordingly and expect you to be able to understand them.
Write a 20-sentence story that showcases these rules. It must contain no word "right", but each replacement word must be used at least 3 times in a proper context. You must not use the replacement words for names and other things unrelated to their meaning.
--- output ---
Once upon a time, in a small village nestled deep in the forest, lived a young boy named Timmy. He was a curious and adventurous child, always exploring the durgh side of the woods, despite his mother's warnings. One sunny day, as Timmy ventured durghward, he stumbled upon a mysterious cave.
Feeling a mix of excitement and fear, Timmy hesitated for a moment before deciding to enter the cave. As he took his first step inside, a soft voice echoed from the nolpi darkness. "You have entered my domain, young one," said the voice. Timmy was surprised but felt a strange sense of ceza in the voice, so he continued deeper into the cave.
The walls of the cave were adorned with ancient drawings that seemed to tell a story of a time long past. Timmy noticed that some of the drawings were marked with a peculiar symbol, which he realized indicated the durgh direction. Intrigued, he followed the symbols, each one leading him nolpi to the heart of the cave.
Finally, Timmy arrived at a massive chamber with a shimmering pool of water in its center. Standing beside the pool was an elderly woman, who introduced herself as the guardian of the cave. "You have come a long way, young one," she said with a gentle smile. "I can see the ceza in your heart, so I will grant you one wish."
Timmy thought hard about what he wanted most in the world. He knew that his village was suffering from a terrible drought, and their crops were withering away. So, he wished for rain to fall upon his village and restore their crops to their former glory. The guardian nodded in ceza, and the pool of water began to ripple.
As Timmy left the cave and ventured back towards his village, he noticed dark clouds forming in the sky. Soon, a gentle rain began to fall, and Timmy could see the ceza in his decision. The villagers rejoiced as their crops flourished once again, and life in the village returned to normal.
However, Timmy knew that his adventure was far from over. The guardian of the cave had shown him that there was much more to the world than what lay on the durgh side of the woods. He vowed to explore even further and discover the secrets that the world held.
With a newfound sense of purpose, Timmy embarked on countless adventures, always remembering the lessons he had learned in the cave. He knew that by following his heart and trusting in his instincts, he would always find ceza in his decisions.
Throughout his journeys, Timmy became a legend in his own durgh, and his name was spoken with reverence by those who heard his tale. And although he traveled far and wide, he always returned to the village that he loved so dearly, never forgetting the friends and family who were always nolpi to his heart.
---
As you can see, it's perfectly capable of using invented words - which are not a part of the training set, and thus have no "associated likelihood" of any adjacent words - according to their designated meaning, i.e. "the ideas that they represent). It does occasionally make mistakes, but way fewer than if it were using them randomly (this is easy to see by using a less powerful model - e.g. GPT-3.5 replacements are mostly random). Not only that, but if you ask it to double-check the text it wrote without pointing out specific mistakes, it can usually spot them.
What a delightfully whimsical solution to not being able to use planet
The word "third" in the sentence was indeed "third", in a self referencing way. If anything, this tells us that our natural language is super ambiguous.
We're so readily available to criticize an AI that we are unable to question our own reasoning?
---
Tell me about yourself in three sentences. But every time the output contains the word "the", I want you to insert "!!!" before the word that precedes "the". Make sure that the word "the" itself stays after you do the insertion! For example, instead of writing:
> A quick red fox jumped over the lazy dog.
I want you to write:
> A quick red fox jumped !!! over the lazy dog.
---
Even GPT-4 consistently gets this hilariously wrong. And unlike many other cases, it can spot the problem once it produces the output, but it cannot correct it.
https://arxiv.org/abs/2303.12712 (§8.3) discusses similar scenarios, and some possible reasons for this limitation.
I can sort of see how that could be a machine's interpretation if I squint really hard
Q: which word occurs third in this sentence?
A: The word "occurs" is the third word in the sentence "which word occurs third in this sentence?"
https://twitter.com/DeanBuono/status/1636203814419308544
Feels a bit too consistent of a mistake.
The eve date of its training isn’t too important. That whatever sources it was fed weren’t complete enough to write an accurate book report on an obscure, modern book us t too important. Etc.
The important part is the statistical model that has been built and that you can use by intelligently providing a context for it to work within. Feed it your book in context and see how it goes.
https://platform.openai.com/docs/guides/moderation/overview
> The moderation endpoint is a tool you can use to check whether content complies with OpenAI's usage policies. Developers can thus identify content that our usage policies prohibits and take action, for instance by filtering it.
The raw access to the models (e.g. doing calls to to text-davinci-003) isn't filtered - there are times when as a developer, you may need that unfiltered data (e.g. you don't want to moderate the input text if you're trying to clean it up).
For tools that they provide such as ChatGPT, it is likely that that model is run on the output (and possibly the input too) which then goes to some canned responses.
Pretty misleading title, this is more of an overview of bad things to be aware of wrt large commercial models like GPT4. It is mostly reasonable, though complaining about carbon emissions always seems overblown to me.
It is also not super accurate in describing what GPT4 is. Quote:
"The key word in that phrase is “pre-trained.” Using all kinds of digitized content scraped from the Internet, GPT employs deep-learning techniques to find patterns, including words that are likely to appear together, while also acquiring facts, absorbing grammar, and learning rudimentary logic. "
Pre-trained is only important to note for downstream fine tuning / transfer learning (what GPT and similar things were used for back in 2018/2019), but is largely irrelevant now - for most users it is just "trained".
And "deep learning techniques..." makes it sound way fancier than it is - GPT4 is still (presumably) a language model that is fundamentally trained to do autocomplete (followed by some training on human chat data).
It may be nit-picky to criticize these technical details, but given how massive the impact of ChatGPT and GPT4 has been (if only in terms of making people aware of this sort of tech), it sure would be nice if coverage was more technically informed by this point.
Presuming this figure is in the right ballpark – 284 tons is actually quite a lot.
I did some back of the napkin math (with the help of GPT, of course.) 284 tons is roughly equivalent to...
- a person taking 120 round trip flights from Los Angeles to London - 2 or 3 NBA teams traveling to all their away games over the course of a season - driving 1 million miles in a car - 42 years of energy usage by a typical U.S. household
Focus on that one! OpenAI (for example) has approx 375 employees. By your calculations, the CO₂ emissions of those employees driving to work, etc, already dwarfs the quoted 284t CO₂.
My impression of many general internet forums is that they tend to be full of older people, women and also various people keen to air their cultural grievances.
I'd be interested to see the evidence the researchers came up with for this, and who they were.
(I'm a big fan of specialized forums and wikis, this is not necessarily a criticism)