Sam Altman, OpenAI's CEO, testified before the Senate on May 16, pointing out the potential harms of generative AI and asking for his industry to be regulated. Seeing a CEO do the right thing...
Sam Altman, OpenAI's CEO, testified before the Senate on May 16, pointing out the potential harms of generative AI and asking for his industry to be regulated.
Seeing a CEO do the right thing against his own interests is unusual, to say the least. I'm squinting at this article and trying to figure out what's his angle, and I haven't figured it out yet. A friend of mine opined that the real goal is to shut down open-source LLMs which pose a strategic threat to OpenAI, but Altman repeatedly spoke against that.
I don't think it's against his own interests. I'm thinking he thinks of his interests differently. OpenAI gets a lot of hate on Hacker News for not being open enough. It's a very libertarian...
I don't think it's against his own interests. I'm thinking he thinks of his interests differently.
OpenAI gets a lot of hate on Hacker News for not being open enough. It's a very libertarian crowd, at least when it comes to stuff they want to play with.
On the other side you have Rationalists predicting doom (or at least, not ruling out doom) and yet... there is often libertarian sentiment over there too?
It seems like if you think something is an existential risk, that means it's dangerous enough that it should probably be regulated, somehow. For example, it seems hard to justify a libertarian or egalitarian approach to nuclear weapons or virus research? It's easy to justify limiting work on very dangerous things to qualified experts operating under state control.
Other things are dangerous, but not in same league as nuclear weapons. Big business tends to do well in regulated markets where they're selling something restricted but not illegal. Going in that direction results in a structure like banking where government and industry work to try to keep fraud, money-laundering, and financial crises somewhat under control. Banks do a lot of day-to-day financial rule-making and enforcement, unless the state stops trusting them or they screw up, in which case the state gets a lot less friendly.
OpenAI has been consistent that AI is potentially dangerous. They started out as a nonprofit researching AI safety. Things have changed somewhat, but this is part of why they don't publish as much as other people would like. (And also, they don't want to help the competition.) Hoping to be an established, trusted part of a regulated industry isn't anti-business, it's just not libertarian.
In the meantime, Altman says the right things to not get anyone too upset, and OpenAI keeps doing what they're doing. They're making rules and putting limits on what people can do with their API's, and they just need to get the state to say that their approach is sensible.
[...] we are likely to eventually need something like an IAEA for superintelligence efforts; any effort above a certain capability (or resources like compute) threshold will need to be subject to an international authority that can inspect systems, require audits, test for compliance with safety standards, place restrictions on degrees of deployment and levels of security, etc. Tracking compute and energy usage could go a long way, and give us some hope this idea could actually be implementable. As a first step, companies could voluntarily agree to begin implementing elements of what such an agency might one day require, and as a second, individual countries could implement it. It would be important that such an agency focus on reducing existential risk and not issues that should be left to individual countries, such as defining what an AI should be allowed to say.
[...]
We think it’s important to allow companies and open-source projects to develop models below a significant capability threshold, without the kind of regulation we describe here (including burdensome mechanisms like licenses or audits).
The play could be one of regulatory capture or to sell the government on the idea that OpenAI alone is suited to handling the dangers of AI, thus limiting competition from startups.
The play could be one of regulatory capture or to sell the government on the idea that OpenAI alone is suited to handling the dangers of AI, thus limiting competition from startups.
Introducing PaLM 2 […] […] From the linked page: It looks like Bard is generally available, but not in Germany based on complaints in Hacker News? “You can now collaborate with Bard in Japanese...
[…] today we’re introducing PaLM 2, our next generation language model. PaLM 2 is a state-of-the-art language model with improved multilingual, reasoning and coding capabilities.
[…]
We’ll be making PaLM 2 available in four sizes from smallest to largest: Gecko, Otter, Bison and Unicorn. Gecko is so lightweight that it can work on mobile devices and is fast enough for great interactive applications on-device, even when offline. This versatility means PaLM 2 can be fine-tuned to support entire classes of products in more ways, to help more people.
[…]
PaLM 2’s improved multilingual capabilities are allowing us to expand Bard to new languages, starting today. Plus, it’s powering our recently announced coding update.
From the linked page:
We’re launching these capabilities in more than 20 programming languages including C++, Go, Java, Javascript, Python and Typescript. And you can easily export Python code to Google Colab — no copy and paste required. Bard can also assist with writing functions for Google Sheets.
It looks like Bard is generally available, but not in Germany based on complaints in Hacker News? “You can now collaborate with Bard in Japanese and Korean, in addition to US English. We have also expanded access to Bard in all three languages to over 180 countries.”
Since March, we've been previewing the PaLM API with a small group of developers. Starting today, developers can sign up to use the PaLM 2 model, or customers can use the model in Vertex AI with enterprise-grade privacy, security and governance. PaLM 2 is also powering Duet AI for Google Cloud, a generative AI collaborator designed to help users learn, build and operate faster than ever before.
We’re already at work on Gemini — our next model created from the ground up to be multimodal, highly efficient at tool and API integrations, and built to enable future innovations, like memory and planning. Gemini is still in training, but it’s already exhibiting multimodal capabilities never before seen in prior models. Once fine-tuned and rigorously tested for safety, Gemini will be available at various sizes and capabilities, just like PaLM 2, to ensure it can be deployed across different products, applications, and devices for everyone’s benefit.
Here is the Palm 2 technical report. Apparently it’s been trained on more non-English language data than previous versions, and on parallel multilingual data to improve language translation.
They claim similar reasoning ability to GPT-4 and better translations than Google Translate.
One interesting thing about Bard is that unlike ChatGPT, it seems to know about recent news. When playing with this, though, I found that it will return old news if you ask about something...
One interesting thing about Bard is that unlike ChatGPT, it seems to know about recent news. When playing with this, though, I found that it will return old news if you ask about something specific, so I recommend adding “with dates” to your query. (And if I asked it to add dates later, it made them up.)
For general knowledge questions, it will often summarize the Wikipedia article and link to it.
I asked for a weather forecast for a city in California and the first time it wasn’t very good. (It made up a much higher chance of rain.) I couldn’t reproduce that bug, though.
I think Bing is supposed to do similar things, but I don’t have a convenient way to access it.
Others commented on that, but I find it incredible how people are going from "wow, can a computer even do that?" to "this AI is dumb because it forgot that some chairs have three legs". I mean,...
Others commented on that, but I find it incredible how people are going from "wow, can a computer even do that?" to "this AI is dumb because it forgot that some chairs have three legs". I mean, sure, AI gets things wrong, but the transition from amazement to nitpicking was surprisingly short. And I'm not even talking about technically minded people, these are my very regular friends.
A friend of mine decried GPT because it failed to generate the equivalent to a short college essay on command when the very fact that it got even remotely close is astounding to me.
Why are we so quick to call a computer dumb because it errors when the very fact that it often doesn't is mind-blogging in itself?
Guess 1: regular people are apathetic towards AI because linguistic models are just how they always imagined computers worked, or should work in the first place.
Guess 2: apathy is a good way to mask fear.
Guess 3: because the models output something that closely resembles something made by humans, we tend to expect from it all the other human properties and capabilities it does not have.
I feel it's because we all get bombarded by stories and hype and fears of how powerful AI is and how its going to take over, and how experts say the government needs to step in and license AI to...
I feel it's because we all get bombarded by stories and hype and fears of how powerful AI is and how its going to take over, and how experts say the government needs to step in and license AI to prevent a catastrophe, and how it's going to take over the entire job market, and how companies are planning mass layoffs due to replacing workers with AI, and all that crap.
And then you use ChatGPT in something you know enough about and suddenly you see that it's a bit of a facade, it responds with something that looks plausible at first glance but the more you look in to it is useless. It's a counter-reaction to all the sentiment of how amazing it is. It's both a very impressive piece of technology that is only going to get better, and at the same time it's garbage that will lie to you knowingly.
The way I think about it is that once technology progresses pass a certain threshold, people's expectation goes from "not at all" to "all the way". CGPGrey's video about self driving car had a...
The way I think about it is that once technology progresses pass a certain threshold, people's expectation goes from "not at all" to "all the way". CGPGrey's video about self driving car had a comment about how people's attitude went from "it will never happen" to "why isn't it here yet".
The threshold could also be very small though. LouisCK had another bit about a guy complained about the wifi on their flight being down, despite only found out that's a thing just a moment ago.
Link to the LouisCK bit (since it's a classic): https://www.youtube.com/watch?v=me4BZBsHwZs And there is a longer version of the bit from his special, Hilarious, too:...
I think it's what happens when you try to use something for real. I used MidJourney to create pictures for my blog. I'm happy with what I got, but the pictures I used weren't the first ideas I...
I think it's what happens when you try to use something for real.
I used MidJourney to create pictures for my blog. I'm happy with what I got, but the pictures I used weren't the first ideas I tried. I would try an idea, find that it couldn't really do it, and then come up with another idea, until I found a prompt that gave me an image I liked enough to use.
And that's for a blog, where the picture is just conceptually related and doesn't need to match too closely. Someone trying to make something specific will be more picky. Illustrating a story wouldn't work if you cared about making a bunch of drawings with the same characters in them, drawn the same way.
So this isn't about "is the computer great at art?" (yes/no), or "is the computer better than a human?" (Which human?) It's how well it does at some task you're trying to get it to do.
Also, prompting is a wide-open text interface. People can try anything and see it if it works. They're going to try all sorts of different things, with no real guidance about what it can actually do, and get different results.
Most of my classmates talk fairly openly about asking ChatGPT to answer whatever schoolwork comes up, a group project I was a part of a few weeks ago decided to tell ChatGPT to make them a poem...
Most of my classmates talk fairly openly about asking ChatGPT to answer whatever schoolwork comes up, a group project I was a part of a few weeks ago decided to tell ChatGPT to make them a poem about the Nordic Mythology to spruce up our project, an idea which several other people also had. The poem had pretty bad rhymes in multiple occasions but this was for philosophy (myths of creation + rituals) and the philodophy teacher isn't the best when it comes to this kind of stuff. Some of the more technical teachers said they already look into programs that detect if a student is using this type of program, and one has even managed to make it impossible for us to copy the text of an article that we needed to summmarize/explain, presumably because of said chat progams.
Funnily enough, today many of my classmates were complaining that these programs can give them multiple different answers to math/chemistry questions and when my classmates then ask the machine if the answers it provided were correct, the machine was saying it's own answers were wrong. I have to guess that means most of it's source data on homework is the incorrect answers then, lmao.
I think it’s actually a reasonable tool to use when learning in school, but more so as a research buddy (that you can’t fully trust) than as a way to write everything for you.
I think it’s actually a reasonable tool to use when learning in school, but more so as a research buddy (that you can’t fully trust) than as a way to write everything for you.
I hadn't heard of Scale AI before, but they seem to be San Francisco-based startup and they claim in a blog post that they're supplying a system that uses a large language model to the US military:
I hadn't heard of Scale AI before, but they seem to be San Francisco-based startup and they claim in a blog post that they're supplying a system that uses a large language model to the US military:
[...] I am proud to be delivering this world-changing capability to the United States government with Donovan - the first LLM deployed on a classified network for the XVIII Airborne Corps - ingesting over 100,000 pages of live data to enable actionable insights across the battlefield. The sheer volume of inputs from various formats is a key challenge facing analysts and operators across the federal space. [...]
AI Zombies Lately I've been thinking and anticipating the imminent and inevitable near-future introduction of AI-resurrected "zombies": deceased individuals whose past conversations are used to...
AI Zombies
Lately I've been thinking and anticipating the imminent and inevitable near-future introduction of AI-resurrected "zombies": deceased individuals whose past conversations are used to train LLMs and voice models, and their images used to animate new portraits. Just a portrait one can talk with on a screen.
People will use AI Zombies to talk with deceased loved ones. I anticipate that this will explode in popularity: who doesn't want to not just hear their deceased loved one's voice but have an intelligible conversation? Talk to grandma about your garden, ask for a parent for advice, tell your brother about your upcoming travel plans — and hear those words, not just read them.
I'm undecided if this will be socially healthy, however. A critical component of the human experience is accepting that everyone, ourselves included, must leave this mortal plane eventually. But maybe AI Zombies—which I know is a crass term—can become a new step in the human grieving process.
Language models can already be used to make fictional characters you can talk to. (I wrote a blog post about this.) The AI is the writer. Someone could do that and decide they don't like the...
Language models can already be used to make fictional characters you can talk to. (I wrote a blog post about this.) The AI is the writer. Someone could do that and decide they don't like the writer. They could delete some bad dialog and regenerate, or hire a different writer.
When creating a fictional character based on someone else, living or dead, I think giving the character a different name would be healthier, to avoid indulging in willful misunderstanding. If that doesn't appeal, maybe they're doing it for the wrong reasons?
If someone does this without being clear on what they're really doing, it seems kind of like a modern version of idol worship? It does seem likely that it will happen, somewhere.
A dystopic thought I had is that Facebook will be uniquely suited for this application. They have all the personal data of individuals for training, their user demographics are the older ones,...
A dystopic thought I had is that Facebook will be uniquely suited for this application. They have all the personal data of individuals for training, their user demographics are the older ones, memorialized accounts are already a thing, status updates, comments and daily photos are interaction modes with low enough fidelity to suspend disbelief and are all within the capability of current AIs. Maybe the fiction of how the deceased are settling down on "the other side" could be an acceptable form of make-believe.
Most of my relatives aren't much like their Facebook profiles, thank goodness. In some cases it would be more of an anti-memorial of their most political side. In other cases, they rarely post.
Most of my relatives aren't much like their Facebook profiles, thank goodness. In some cases it would be more of an anti-memorial of their most political side. In other cases, they rarely post.
ggml seems to be becoming a more standard format for local LLMs, soon to support MosaicML. There is still a lot of breaking changes, but that's to be expected with any bleeding edge technology....
ggml seems to be becoming a more standard format for local LLMs, soon to support MosaicML. There is still a lot of breaking changes, but that's to be expected with any bleeding edge technology. I'll take having to redownload a model or run a conversion script every so often if it means not committing to a bad standard.
In demonstrations this week and during a presentation at the company’s annual developer conference on Wednesday, executives showed off the new version of Google Search. The system does generate its own answers but checks them for accuracy against real websites. It also posts those links directly next to the generated text, making it easy for people to click on them. For questions that are about sensitive topics like health, finances and hot-button political issues, the bot won’t write an answer at all, instead returning a news article or links to websites.
[…]
Google is also adding tags to images in Google Search if it knows that they are computer-generated rather than actual photos, and is making it easier to see when an image first appeared on the web. Misinformation researchers already use Google’s reverse image search to check whether an image is old or new and where it came from, but the new tools are meant to make it easier for regular search users to do the same.
What the heck is this article even talking about? I've never heard any of these arguments being made in the LLM spaces.
That’s true – but why call the errors “hallucinations” at all? Why not algorithmic junk? Or glitches? Well, hallucination refers to the mysterious capacity of the human brain to perceive phenomena that are not present, at least not in conventional, materialist terms. By appropriating a word commonly used in psychology, psychedelics and various forms of mysticism, AI’s boosters, while acknowledging the fallibility of their machines, are simultaneously feeding the sector’s most cherished mythology: that by building these large language models, and training them on everything that we humans have written, said and represented visually, they are in the process of birthing an animate intelligence on the cusp of sparking an evolutionary leap for our species.
What the heck is this article even talking about? I've never heard any of these arguments being made in the LLM spaces.
Yeah, that's making up an explanation rather than researching how the term "hallucination" was actually used in the history of machine learning. (Which is rather like what chatbots do.) I tried...
Yeah, that's making up an explanation rather than researching how the term "hallucination" was actually used in the history of machine learning. (Which is rather like what chatbots do.)
I tried doing a few searches and wasn't able to find a good answer, but the term "hallucination" used in machine learning is at least five years old.
Also, here's a paper from around that time that talks about how generating hallucinations could be useful for training data.
I wonder if the popularity of DeepDream had anything to do with it? It's an early image generator that creates images that look rather psychedelic. Someone even wrote a paper based on it:
Altered states of consciousness, such as psychotic or pharmacologically-induced hallucinations, provide a unique opportunity to examine the mechanisms underlying conscious perception. However, the phenomenological properties of these states are difficult to isolate experimentally from other, more general physiological and cognitive effects of psychoactive substances or psychopathological conditions. Thus, simulating phenomenological aspects of altered states in the absence of these other more general effects provides an important experimental tool for consciousness science and psychiatry. Here we describe such a tool, which we call the Hallucination Machine.
Yeah, I really feel like we've reached the 'peak crypto' stage of articles about AI, where very few people actually know what it's doing but because it's so popular everyone CERTAINLY has opinions...
Yeah, I really feel like we've reached the 'peak crypto' stage of articles about AI, where very few people actually know what it's doing but because it's so popular everyone CERTAINLY has opinions on it, both unrealistically optimistic and pessimistic.
It's gotten to that part of the culture war over it where the venture capitalists hyped it too much and now people are just sick of it, so instead of writing fluff pieces or doomering over how it'll kill us all, these media outlets start writing trite takedowns for the clicks. Just like crypto, it rarely has an actual association with the merits or demerits of the tech, just whatever their viewers want to hear.
I don't think it makes sense to be outraged about "walling it off inside proprietary products" when open-source models are such a major player in this field. Just recently we saw a paper from a...
I don't think it makes sense to be outraged about "walling it off inside proprietary products" when open-source models are such a major player in this field. Just recently we saw a paper from a Googler worrying that commercial megacorp language models have no strategic advantage over free and open-source versions.
If I may respond to the title: no, it's not time to worry. This technique only works if: You spend time training the model to build a personalized profile of you You cooperate with the researcher...
If I may respond to the title: no, it's not time to worry. This technique only works if:
You spend time training the model to build a personalized profile of you
You cooperate with the researcher to focus on simple, easy to parse phrases. Jumping your brain between topics only produces gibberish.
You are physically within an MRI machine
Could a super villain use this in a TV setting? Maybe. Is it in any way feasible in real life? I don't think so.
Oops. I wrote a response to a now-deleted comment, so I'll reply to myself instead. It was definitely an interesting study! While it's not currently that useful as a communication assistance tool...
Oops. I wrote a response to a now-deleted comment, so I'll reply to myself instead.
It was definitely an interesting study! While it's not currently that useful as a communication assistance tool (due to the MRI/heavy training requirement), it seems like this same methodology could be applied to other scanning technologies (EEG or PET). More optimistically, electrodes planted under the scalp might allow for portable use. Even if the sensitivity is lower, you could train it for an hour a day from home and begin to build a strong personalized model.
From there, it's easy to imagine hooking that up to a speaker or computer. Instead of slowly moving a cursor, as some BCI tools have done, you might be able to just think about the link you wish to click. That could be life changing for disabled individuals.
I agree, this is an unexpected yet somewhat obvious application for LLMs. Language is so universal, and so core to our human experience, that there's likely thousands of novel applications that are just waiting to be thought up.
Sam Altman, OpenAI's CEO, testified before the Senate on May 16, pointing out the potential harms of generative AI and asking for his industry to be regulated.
Seeing a CEO do the right thing against his own interests is unusual, to say the least. I'm squinting at this article and trying to figure out what's his angle, and I haven't figured it out yet. A friend of mine opined that the real goal is to shut down open-source LLMs which pose a strategic threat to OpenAI, but Altman repeatedly spoke against that.
I don't think it's against his own interests. I'm thinking he thinks of his interests differently.
OpenAI gets a lot of hate on Hacker News for not being open enough. It's a very libertarian crowd, at least when it comes to stuff they want to play with.
On the other side you have Rationalists predicting doom (or at least, not ruling out doom) and yet... there is often libertarian sentiment over there too?
It seems like if you think something is an existential risk, that means it's dangerous enough that it should probably be regulated, somehow. For example, it seems hard to justify a libertarian or egalitarian approach to nuclear weapons or virus research? It's easy to justify limiting work on very dangerous things to qualified experts operating under state control.
Other things are dangerous, but not in same league as nuclear weapons. Big business tends to do well in regulated markets where they're selling something restricted but not illegal. Going in that direction results in a structure like banking where government and industry work to try to keep fraud, money-laundering, and financial crises somewhat under control. Banks do a lot of day-to-day financial rule-making and enforcement, unless the state stops trusting them or they screw up, in which case the state gets a lot less friendly.
OpenAI has been consistent that AI is potentially dangerous. They started out as a nonprofit researching AI safety. Things have changed somewhat, but this is part of why they don't publish as much as other people would like. (And also, they don't want to help the competition.) Hoping to be an established, trusted part of a regulated industry isn't anti-business, it's just not libertarian.
In the meantime, Altman says the right things to not get anyone too upset, and OpenAI keeps doing what they're doing. They're making rules and putting limits on what people can do with their API's, and they just need to get the state to say that their approach is sensible.
Governance of superintelligence (OpenAI)
[...]
This seems broadly sane to me. Curious what the experts will say. Thanks for posting this!
The play could be one of regulatory capture or to sell the government on the idea that OpenAI alone is suited to handling the dangers of AI, thus limiting competition from startups.
Introducing PaLM 2
[…]
[…]
From the linked page:
It looks like Bard is generally available, but not in Germany based on complaints in Hacker News? “You can now collaborate with Bard in Japanese and Korean, in addition to US English. We have also expanded access to Bard in all three languages to over 180 countries.”
Here is the Palm 2 technical report. Apparently it’s been trained on more non-English language data than previous versions, and on parallel multilingual data to improve language translation.
They claim similar reasoning ability to GPT-4 and better translations than Google Translate.
Here is Bard’s experiment updates page.
One interesting thing about Bard is that unlike ChatGPT, it seems to know about recent news. When playing with this, though, I found that it will return old news if you ask about something specific, so I recommend adding “with dates” to your query. (And if I asked it to add dates later, it made them up.)
For general knowledge questions, it will often summarize the Wikipedia article and link to it.
I asked for a weather forecast for a city in California and the first time it wasn’t very good. (It made up a much higher chance of rain.) I couldn’t reproduce that bug, though.
I think Bing is supposed to do similar things, but I don’t have a convenient way to access it.
Others commented on that, but I find it incredible how people are going from "wow, can a computer even do that?" to "this AI is dumb because it forgot that some chairs have three legs". I mean, sure, AI gets things wrong, but the transition from amazement to nitpicking was surprisingly short. And I'm not even talking about technically minded people, these are my very regular friends.
A friend of mine decried GPT because it failed to generate the equivalent to a short college essay on command when the very fact that it got even remotely close is astounding to me.
Why are we so quick to call a computer dumb because it errors when the very fact that it often doesn't is mind-blogging in itself?
I feel it's because we all get bombarded by stories and hype and fears of how powerful AI is and how its going to take over, and how experts say the government needs to step in and license AI to prevent a catastrophe, and how it's going to take over the entire job market, and how companies are planning mass layoffs due to replacing workers with AI, and all that crap.
And then you use ChatGPT in something you know enough about and suddenly you see that it's a bit of a facade, it responds with something that looks plausible at first glance but the more you look in to it is useless. It's a counter-reaction to all the sentiment of how amazing it is. It's both a very impressive piece of technology that is only going to get better, and at the same time it's garbage that will lie to you knowingly.
That makes sense. The hype is way too much, so people have exaggerated expectations and end up underwhelmed.
The way I think about it is that once technology progresses pass a certain threshold, people's expectation goes from "not at all" to "all the way". CGPGrey's video about self driving car had a comment about how people's attitude went from "it will never happen" to "why isn't it here yet".
The threshold could also be very small though. LouisCK had another bit about a guy complained about the wifi on their flight being down, despite only found out that's a thing just a moment ago.
Link to the LouisCK bit (since it's a classic):
https://www.youtube.com/watch?v=me4BZBsHwZs
And there is a longer version of the bit from his special, Hilarious, too:
https://www.youtube.com/watch?v=G8OM2TdkaHg
I think it's what happens when you try to use something for real.
I used MidJourney to create pictures for my blog. I'm happy with what I got, but the pictures I used weren't the first ideas I tried. I would try an idea, find that it couldn't really do it, and then come up with another idea, until I found a prompt that gave me an image I liked enough to use.
And that's for a blog, where the picture is just conceptually related and doesn't need to match too closely. Someone trying to make something specific will be more picky. Illustrating a story wouldn't work if you cared about making a bunch of drawings with the same characters in them, drawn the same way.
So this isn't about "is the computer great at art?" (yes/no), or "is the computer better than a human?" (Which human?) It's how well it does at some task you're trying to get it to do.
Also, prompting is a wide-open text interface. People can try anything and see it if it works. They're going to try all sorts of different things, with no real guidance about what it can actually do, and get different results.
Most of my classmates talk fairly openly about asking ChatGPT to answer whatever schoolwork comes up, a group project I was a part of a few weeks ago decided to tell ChatGPT to make them a poem about the Nordic Mythology to spruce up our project, an idea which several other people also had. The poem had pretty bad rhymes in multiple occasions but this was for philosophy (myths of creation + rituals) and the philodophy teacher isn't the best when it comes to this kind of stuff. Some of the more technical teachers said they already look into programs that detect if a student is using this type of program, and one has even managed to make it impossible for us to copy the text of an article that we needed to summmarize/explain, presumably because of said chat progams.
Funnily enough, today many of my classmates were complaining that these programs can give them multiple different answers to math/chemistry questions and when my classmates then ask the machine if the answers it provided were correct, the machine was saying it's own answers were wrong. I have to guess that means most of it's source data on homework is the incorrect answers then, lmao.
I think it’s actually a reasonable tool to use when learning in school, but more so as a research buddy (that you can’t fully trust) than as a way to write everything for you.
I hadn't heard of Scale AI before, but they seem to be San Francisco-based startup and they claim in a blog post that they're supplying a system that uses a large language model to the US military:
AI Zombies
Lately I've been thinking and anticipating the imminent and inevitable near-future introduction of AI-resurrected "zombies": deceased individuals whose past conversations are used to train LLMs and voice models, and their images used to animate new portraits. Just a portrait one can talk with on a screen.
People will use AI Zombies to talk with deceased loved ones. I anticipate that this will explode in popularity: who doesn't want to not just hear their deceased loved one's voice but have an intelligible conversation? Talk to grandma about your garden, ask for a parent for advice, tell your brother about your upcoming travel plans — and hear those words, not just read them.
I'm undecided if this will be socially healthy, however. A critical component of the human experience is accepting that everyone, ourselves included, must leave this mortal plane eventually. But maybe AI Zombies—which I know is a crass term—can become a new step in the human grieving process.
Language models can already be used to make fictional characters you can talk to. (I wrote a blog post about this.) The AI is the writer. Someone could do that and decide they don't like the writer. They could delete some bad dialog and regenerate, or hire a different writer.
When creating a fictional character based on someone else, living or dead, I think giving the character a different name would be healthier, to avoid indulging in willful misunderstanding. If that doesn't appeal, maybe they're doing it for the wrong reasons?
If someone does this without being clear on what they're really doing, it seems kind of like a modern version of idol worship? It does seem likely that it will happen, somewhere.
A dystopic thought I had is that Facebook will be uniquely suited for this application. They have all the personal data of individuals for training, their user demographics are the older ones, memorialized accounts are already a thing, status updates, comments and daily photos are interaction modes with low enough fidelity to suspend disbelief and are all within the capability of current AIs. Maybe the fiction of how the deceased are settling down on "the other side" could be an acceptable form of make-believe.
Most of my relatives aren't much like their Facebook profiles, thank goodness. In some cases it would be more of an anti-memorial of their most political side. In other cases, they rarely post.
ggml seems to be becoming a more standard format for local LLMs, soon to support MosaicML. There is still a lot of breaking changes, but that's to be expected with any bleeding edge technology. I'll take having to redownload a model or run a conversion script every so often if it means not committing to a bad standard.
Google is changing the way we search with AI. It could upend the web. (Washington Post)
[…]
What the heck is this article even talking about? I've never heard any of these arguments being made in the LLM spaces.
Yeah, that's making up an explanation rather than researching how the term "hallucination" was actually used in the history of machine learning. (Which is rather like what chatbots do.)
I tried doing a few searches and wasn't able to find a good answer, but the term "hallucination" used in machine learning is at least five years old.
Wired wrote an article in 2018 about it: AI Has a Hallucination Problem That's Proving Tough to Fix. However, this was about image recognition and self-driving cars, not generative AI.
Also, here's a paper from around that time that talks about how generating hallucinations could be useful for training data.
I wonder if the popularity of DeepDream had anything to do with it? It's an early image generator that creates images that look rather psychedelic. Someone even wrote a paper based on it:
Yeah, I really feel like we've reached the 'peak crypto' stage of articles about AI, where very few people actually know what it's doing but because it's so popular everyone CERTAINLY has opinions on it, both unrealistically optimistic and pessimistic.
It's gotten to that part of the culture war over it where the venture capitalists hyped it too much and now people are just sick of it, so instead of writing fluff pieces or doomering over how it'll kill us all, these media outlets start writing trite takedowns for the clicks. Just like crypto, it rarely has an actual association with the merits or demerits of the tech, just whatever their viewers want to hear.
LLMs are, shockingly, language models. They're not models for human intelligence. They're good at language tasks like translation, summarization, etc.
I don't think it makes sense to be outraged about "walling it off inside proprietary products" when open-source models are such a major player in this field. Just recently we saw a paper from a Googler worrying that commercial megacorp language models have no strategic advantage over free and open-source versions.
Nature - Mind-reading machines are here: is it time to worry?
If I may respond to the title: no, it's not time to worry. This technique only works if:
Could a super villain use this in a TV setting? Maybe. Is it in any way feasible in real life? I don't think so.
Oops. I wrote a response to a now-deleted comment, so I'll reply to myself instead.
It was definitely an interesting study! While it's not currently that useful as a communication assistance tool (due to the MRI/heavy training requirement), it seems like this same methodology could be applied to other scanning technologies (EEG or PET). More optimistically, electrodes planted under the scalp might allow for portable use. Even if the sensitivity is lower, you could train it for an hour a day from home and begin to build a strong personalized model.
From there, it's easy to imagine hooking that up to a speaker or computer. Instead of slowly moving a cursor, as some BCI tools have done, you might be able to just think about the link you wish to click. That could be life changing for disabled individuals.
I agree, this is an unexpected yet somewhat obvious application for LLMs. Language is so universal, and so core to our human experience, that there's likely thousands of novel applications that are just waiting to be thought up.