It is true, however, that the problem has been solved completely in the human-to-machine direction. The output of the current-generation LLMs is completely off base in many cases, but they certainly understand what they are being asked, for any useful definition of 'understand.'
I'm much more impressed by GPT's ability to handle input than I am in its ability to generate output. It's arguably as good at reading comprehension as most humans.
But that's not an NLP problem at heart. Language is just a collections of tokens (words, letter) that are tied together by certain rules to convey some meaning. There is no concept of reality per se.
For example, consider filling the blank:
A giant ______ flew over my head!
It can be a plane. Or a dragon. Or an UFO. Or a balloon. The thing is all of those are correct answers language-wise and the model works correctly as long as what gets filled in conforms to the rules of the given language.
The language that we generate encodes reality to some extent and the model picks up those correlations but there is no concept of reasoning or reality behind it. Maybe it is emergent at some point (as to effectively compress it needs to encode some subset of rules governing our reality) but it is not an agent that optimizes for understanding our reality. Something like Dreamer would be much closer to that.
Sorry to get heavy here: truth is not an NLP problem, it's an alignment problem. We want truth, but we don't have a reliable way to train an AI to provide the truth, only to provide things that are either true, or sound true enough that they fool the reward function. And even then, that may not be exactly what the AI learns to do, because of there's another level of alignment problem, the "inner alignment" or "mesa-optimizer alignment" problem!
With an AI like GPT, it is quirky and amusing. Once AIs get really powerful, it becomes scary, and a lot of people who understand this field much better than I do are worried it has a good chance of being deadly. Like, potentially kill-everyone-on-earth deadly.
Hard agree. I'm really trying to figure out how to inject this idea into my friends' heads effectively. The main struggle I'm facing is how to convey the danger behind it. Why can it be deadly exactly? What can a program actually do to harm people, to the level where it's a risk of extinction or societal collapse?
Personally I didn't need to imagine a specific scenario to understand that there's risk, but I think it would help me convince other folks if I did.
If you want society to collapse all you need to do is succeed in having AI automate all jobs.
Every single country where money comes from somewhere other than people (oil, diamonds...) is an authoritarian nightmare simply because keeping people happy is not necessary.
Once AI can do everything and robots that can do any physical labor are developed the population will shrink dramatically as people with killer robots kill each other for resources. There is no need for AI rebellion or AI failure to get there.
Anecdotally, several CTOs I know intend to lessen their use of Scale, Labelbox and more in the future. Talked to one today who already ditched MTurk for GPT-4 -- cheaper, better, faster was what he said.
Labelbox does image annotating still, and one CTO said as soon as GPT-4 enabled this for him he'd have his team homebrew it from there.
I don't think I see enough discussion about what this means for privacy. There was some protection in the fact that it was prohibitively expensive to get someone to listen to every single one of our phonecalls/read all our emails/etc.
Not to sound condescending but really? How is this not immediately your mind goes? Every piece of information ever recorded can now be summarized and cross-linked efficiently. Privacy is beyond dead. Soon every authoritarian government (and Democratic ones albeit secretly) will have integrated platforms that track every single one of your movements, known contacts, internet usage, financial data, and correspondence. Big Brother has NEVER EVER been more effective than it will become.
Looking at this from far away, with the Snowden revelations in mind I'd think it's not tinfoil hat territory to assume that some of the progress at OpenAI got achieved with some help from well ressourced folks in the USG/Three Letter Agencies.
I don’t think they helped them. Now, did they train off the same data sources? Well, since OpenAI isn’t saying what GPT-4 is trained on, and the NSA can hoover up all kinds of non-public data, it stands to reason they may both be doing something slightly shady with emails, texts, and the like.
Yeah, I've started wondering similar things about that too, like how far ahead is the NSA on this stuff? And how does that tie in to the recent policy of denying China semiconductors?
Perhaps history will show that the NSA made algorithmic breakthroughs a few years ago and realized what was coming, so political policy was crafted to stymie Chinese progress in this field, and what we're seeing in the public sphere from companies like openAI is a managed release of the technology into the public, openAI at least managing to independently discover the same breakthroughs that the NSA made a a few years ago.
You're seeing the government entity as separate from the corporate entity, but quite often in the US its the other way around. The government entity is a rather hollow shell, and the 'brains' of the operation is contracted out to the corporation. The government entity would almost cease to exist if the corporation under it magically disappeared.
Most people don't really think about things that don't affect their day-to-day lives. This includes the specifics of how Governments might run a mass surveillance plan.
Why do you think there was a mass surveillance of American domestic communications since forever ago, as leaked by Snowden? This technology has been available since then and can effectively summarize millions of pieces of communication.
Yeah but have you seen the leaked slides? It's clear that they have only the ability to analyse 10% or less of the data they are storing.
GPT-like systems will close that gap, and then comes all of the problems of automated law enforcement - Extrapolation from incomplete data, false positives from coindicences, interpretation errors, all that annoying stuff
To add to that, the leaked documents from Snowden described a query language not unlike doing Boolean searches. Nothing close to GPT’s ability to comprehend human asks.
How sure are you about that? The basic theories have been around since the 70s, have been proven at scale in the last decade, and the NSA has more data and compute than anybody else. I’d be shocked if they aren’t very far along in solving many problems.
This isn’t really a “throw money at it” problem like the government is good at.
Take drones for example. The government got really good at those because they made them jet-powered (lol) and blew a bunch of money on server-grade FPGA’s in each one of them.
You can’t really just buy a lot of GPUs to make an LLM work, you need iterative development of architecture and training methods.
Like maybe the government invented self-attention before 2017, but if they didn’t, then the constraint is training time, and the government has the same number of seconds as the rest of us.
The military invented the nuclear bomb yes. But Fermi did most of his thinking work in Italy before the Manhattan project. He got money thrown at him once he got here.
As for semiconductor devices, it was Bell Labs and TI.
Coding is an ambiguous concept that wasn’t really invented, but if it were, it would have first appeared in programmable looms.
The military likes to take credit for things, but really all they do is throw money at existing inventions.
I’m sure they’re throwing a bunch of money at Transformers now, but who are all these uncredited super geniuses who invent things and then let randos at Google take the credit/earn the money?
I used to agree with you. I used to think the military was kinda dumb. But after doing a deeper dive into past military technology, and present - I've come to realize this is just an intelligence ruse.
They made some mistakes in the 40's and 50's to where they had nuclear secrets stolen by the Russians. And ever since has been hyper compartmentalized.
It would not surprise me if in the 90's or 00's they had an internal working LLM, considering all the puzzle pieces. You will never hear about classified tech unless it's a bomb, gets leaked. (See code breaking machines declassed after 70+ years)
In a hypothetical scenario, a military organization might want to conceal its use of a large language model (LLM) for intelligence gathering and analysis.
Another scenario is the military's current interest in everything quantum. Quantum computers for example (you wouldn't want another nation being first and pirate baying out our secrets, would you) so there is an extreme national security importance of being first.
And to be first, you need to have the smart people, which the military has. There is a reason China struggles with jet engines 80 years after their invention, and still can't make nuclear carriers. While the US navy works on things like this: https://www.navair.navy.mil/foia/sites/g/files/jejdrs566/fil...
That's not how military intellectual property works. I don't doubt your technical credentials, but I think you lack deeper insight to what the military does, and has done.
>That's not how military intellectual property works.
So your implication is that the military is full of unnamed linear algebra, systems engineering, and linguistics super geniuses and these people never leave, never talk about their work publicly, never publish anything ever, and they're all cool with their huge innovations being kept away from the public forever? All because military IP regulation?
And none of these effective state prisoners ever defect to China (where they could live like royalty) because...
If that is your angle, I very much agree it's possible some deep black project exists that (once) looked into this. In the UFO lore, there are many stories about these advances (tr3b etc).
But all of it is orthogonal to LLMs though. Picking one exceptional area that the military is great at (aerospace) does not suddenly make the military exceptional in other areas like AI.
See above video, even code breaking machines are kept a dark secret. LLM's that could make analytical decisions about war and strategy, must have started with the military first. It explains its massive data gathering operations in the 2000's. And it explains why some countries separation to make their own internet, away from what really is the US-Owned World Wide Web.
U.S. military has engaged in the commercialization of top-secret technologies (after it's considered obsolete by military standards), often by collaborating with private companies or research institutions.
The military invented AI and NLP which underpins LLMs.
The military is responsible for most the technology we use and talk about today. The government may appear incompetent, but we’re living off military hand-me-downs, the entire world is
Your argument reminds me of the discussions about the moon landing.
If today's hardware was available 20 yrs ago, this would've been possible just like the moon landing could've been faked if it took place 20+ yrs later. The technology wasn't available at the time (GPUs in this case, and generally no experience in doing such advanced trick techniques for movies back then)
These models are having such a strong effect now because we've finally got the hardware to run them
Yes, it has. Consumer-wise, we've had Dragon Naturally Speaking since the late 90s. It's pretty simple to have a script read what it outputs text-wise and look for key words. No AI is even needed to do this.
I'm not so sure about that. I mean, maybe not since the Snowden leaks but how do we know that governments haven't been running their own LLMs for the last five years or so? We know that they're using sockpuppets[1]. We know that they're astroturfing[2]. Integrating LLMs into their toolkits seems like an obvious move, so obvious that they would be stupid not to do it.
>haven't been running their own LLMs for the last five years
Because the hardware has not existed.
This said by accident I've seen hardware that was brought to a testing company by federal marshals that was massively parallel custom hardware that was likely for signal processing a lot of channels at once. So there is plenty of custom hardware out there, but these items have not been produced at the scale needed (from what anyone can tell) and, again from what we can tell, they don't have the general processing capability that GPU/TPU driven LLMs have.
Not exactly. Gov had to be selective because its surveillance required a lot of resources per person/call. New technology allows it cheap and en mass. Voice calls can be recorded, then converted to text, then filtered. And humans will only analyze something of interest. Like we did have alphabet and books, and newspapers for hundreds of years. But only with internet we got the ability to process them easily.
> Gov had to be selective because its surveillance required a lot of resources
not exactly. Where do you think all those budget trillions that don't have to be accounted for goes into? the FBI+NSA (=CIA but for citzens) have infinite resources.
All the overhead they have is to make sure a small subset of the citizens are not impacted. Snowden goes into this in some detail when talking about day to day operations. The norm is to extend the net as wide as possible, until you reach some politician or government agency.
Not only converted to text, it seems likely that we can document sentiment around the persons speech. For example if you're a low priority target that's still on the radar, but not high enough on the list to get a human handler I could see something like, not only what you said, but were you laughing, angry, crying. The the tone of your voice indicate the likelihood of action in a short time frame?
Collection an ocean of data from everyone isn't the same as actually painstakingly tying all the pieces together for everyone.
They've created a huge library of unorganized data. The difference here is they now can spawn a million untiring AI private investigators / librarians to organize this information into coherent "case files".
At least for me, until this point I've had a feeling of anonymity in the idea that, while my data is being slurped up, I'm just one data point in a sea of other 'normal' people. There would be little value in spending government time and effort tying all of the web detritus together for me. The juice would definitely not be worth the squeeze.
However, when the cost of this effort is nearly zero, that now becomes a different story. The balance of power between government and the people it rules is going to radically shift.
How long before the AI built in the US figures out its reward function is 80% more likely to be satisfied if it points its digital finger at someone black rather than the most likely subject?
Shotspotter has been billed as a “system of sensors, software, AI and expert human review that accurately detects, locates and alerts police to gunfire”, and the company behind it (formerly “Shotspotter” was the company name, its recently been renamed “Soundthinking”) has a number of other AI-involved law enforcement products now, as well.
That's a layman explanation. ShotSpotter is likely a passive radar system. In recent years, you can combine signal processing and supervised learn (neuralnet) to get better direction-of-arrival estimations.
> There was some protection in the fact that it was prohibitively expensive to get someone to listen to every single one of our phonecalls/read all our emails/etc.
That's already how it worked on platforms like mturk and uhrs, lots of the work was transcribing audio dumps from microphones built into computers/phones/smart home devices. UHRS especially had a lot of that (it's owned by MS) as well as search engine grading type work. They also certainly do not pay well, I'd imagine that in practice there isn't much cost difference to paying a bunch of bored people to do it vs the compute cost for running an AI model to do it, but the AI model will be vastly more accurate and will work 24/7.
With regard to privacy, what’s the difference between your email’s text stored on a server, and your email’s text alongside the output of the text processed through a LLM? If “they” can already look at the text, what more privacy is there to lose?
There's a great deal of privacy in simply being a needle in a haystack. Part of the processing that's possible with an LLM is filtering.
Imagine you've sent an email about transporting a friend's daughter across state lines to get a medically-necessary abortion. Or if you prefer, imagine you've arranged via email to "lose" some firearms which don't comply with your state's new assault weapons ban.
Pre-LLMs, trying to find these sorts of emails was very hard. A simple text search for "abortion" or "gun" is going to come up with far more emails where two family members got into a political debate, than emails about lawbreaking. Big Brother will find a few such emails here and there by chance, but the vast majority of such incriminating emails will simply be lost in the pile.
Enter LLMs, and Big Brother can feed some of the incriminating emails found my chance into a training dataset along with a bunch of non-incriminating emails, and teach the AI to find incriminating emails, and then apply the model to the entire list of emails and get a nicely filtered list of only the emails which are incriminating, further tuning the model by adding emails it gets wrong to the training dataset when they are found.
The top use case I've been hearing is in legal discovery. Law firms used to play games with diligence by disclosing TBs of email and making it cost prohibitive to find relevant emails. This task would normally require a $60-100/hr paralegal or lawyer.
GPT-4 can do that task for fractions of a penny per email now. It doesn't have to be perfect if its competing with nothing. I expect we'll see similar shops for any other high cost paper/trail business.
Is there a solution to the issue of data stewardship yet? I'd imagine it typically would not be permissible to send a bunch of proprietary legal documents off to OpenAI.
What I'd really love to implement is a way for GPT-4 to answer questions based on a corpus of "all our Confluence pages plus random other sources of documentation." Like with the legal document issue, it's a bit of a nonstarter right now given the proprietary nature of corporate documentation.
AFAIK the Azure APIs provide suitable data usage requirements. One of the most fascinating aspects of the AI world is that we've made extraordinarily expensive brute force search a valuable tool.
Mostly meaningless unless every other part of the supply chain exists to turn them into A100s. Being that supply chain is very difficult and expensive to extend, it's something that can take years to build out, even in a hurry.
Very interesting. Until the day OpenAI has a problem in their systems and the entire world grinds to a halt. Or they put outrageous new prices. Which apparently never happened in other fields, seems.
"At first, humans accept the deteriorations as the whim of the Machine, to which they are now wholly subservient, but the situation continues to deteriorate as the knowledge of how to repair the Machine has been lost"
Replace "the machine" eith "the market" and it describes some people today.
This space will divide into many competitors, and eventually a Linux-like information-magnet will win the whole thing. Eventually there will be robustness..
Back when computers were first going mainstream, there was discussion about taxing them so as to provide for the folks whose jobs would be lost to them --- never went anywhere, but this is a discussion which we need to circle back to.
Such as the political arrangements to distribute the gains of high yield farmer equipment, fully automatized factories, high frequency trading bots? It is not going to happen.
What is going to happen are private robotic armies making sure private owners remain private owners.
And then, we will go back to times where people were not citizens by default and had fewer rights.
I’m not worried about rogue AI taking over the nukes. I’m worried the same people who think it’s a great idea to charge so much for insulin that people start dying are the ones who will be using AI to hurt people.
Hell, give me a slightly evil AI run amok over any pharma CEO doing their job.
We are fucked. I have no hope in humanity managing this technology responsibly and no hope in my future. The months since ChatGPT's release have been some of the worst in my life, mental-health-wise.
Dude let me share my plan: go build something for yourself, a modest bootstrapped business. If you want to keep being an employee, you'll see your job and other people jobs getting more and more replaced by AIs. Bosses hyped to put AI in the coffee machine. Clients asking to add ChatGPT to their wordpress site.
Either you go build something you want, an island where code doesn't talk back, or leave tech altogether.
I am saddened that I don't even recognise this place anymore. It's not Hacker News anymore, it's AI news. It's starry eyed engineers jumping over each other ready to sell their metaphorical soul. Even I, the Luddite, can't seem to talk about anything else than this bloody thing.
My current plan is to build a small business and retire in the middle of the woods somewhere. Do some Lisp coding while the rest of the world is dancing around their new idol.
/rant, send me an email if you wanna rant about it as well, and discuss your concerns.
Great to see this tech and the money invested in it being used to take low-paying jobs away from people with limited options, instead of something like drug discovery or cancer biology.
In this case it is academic research released to the public, which is why anybody knows about it. This is a fairly good thing compared to any alternative I can think of.
This assumes those people really do have no other way to contribute. I don’t believe that’s the case. Do you?
I believe people can contribute in many different ways. When technology enables us to get my work output without me, that frees me up to produce other things for society.
The issue is not can people contribute in some other way, but can they convince someone else to pay them a living wage for doing so, which is going to prove progressively more difficult as this technology advances.
> that frees me up to produce other things for society
The problem is that it is a disruption for everything because at its core it is a machine for the replication of skill and technology. A concept that has never existed prior with any other technological disruption.
"Climbing the skill ladder is going to look more like running on a treadmill at the gym. No matter how fast you run, you aren’t moving, AI is still right behind you learning everything that you can do."
from a more in depth view I wrote up here describing the rapidly shrinking innovation, disruption and adaption cycles
If you look at the table, the GPT-4 model has better correlation with the expert ensemble than the crowd does, but only on some criteria. The GPT-4 model is closer for all of the ethics questions, but the crowd is closer for the utility level and economic impact questions.
I agree, but there are questions about GPT-N successors.
Surprisingly (to me) many people think that GPT-N will never exceed human level intelligence because it was trained on the internet. I think that argument is obviously wrong.
Another is that I am sure a large chunk of people will never concede that the AI is smarter than them. Literally never, no matter how smart the bot gets. I mean, probably a lot of people think they are as smart as anyone else. They won't agree that someone else is smarter than them, and they certainly won't agree that some bot is smarter than them. It's also a loaded assessment, like they will think that if they agree to that, then they are also implicitly agreeing to cede their personal agency to the bot.
Another possibility is that GPT-N successors that surpass human level cognition will be banned by regulation, like some drugs or nuclear explosives or bio weapons. They could even be pre-emptively banned at some level below human level, and maybe it would never be publicly acknowledged that it's technically possible to go above human level.
GPT4 is already better than most if not all humans in some metrics.
I think it's a hard sell to say that nothing better than human is allowed.
But the existential concern comes from supposed exponential takeoff. To me it should be easy to convince people that something 10 or 100 times faster or smarter than humans should not be allowed.
Weirdly you don't see people talking about regulating autonomy which is also part of it.
I find the argument that Eliezer Yudkowsky makes with the super slow aliens to be very compelling especially in the context of fully autonomous AI that people have stupidly designed to imitate human (animal) characteristics like survival instincts. I suspect that that regulators will ban extremely high performing AI. But unfortunately the prediction is that this can't be contained which means it's quite probable that militaries will cause the end of the world just like we have been expecting but in a new way. Since they will likely be excluded from the ban at least secretly.
Probably we both think similar things, but I disagree with one thing.
I think it's weird that you say "I think it's a hard sell to say that nothing better than human is allowed." in combination with "To me it should be easy to convince people that something 10 or 100 times faster or smarter than humans should not be allowed."
To me, those situations are way too close to each other (human level vs. 10x 'smarter') to be able to say one is a "hard sell" and the other is "easy to convince people." I mean maybe you are exaggerating your certainty to make a point, but I don't think it's at all realistic to say that one will be hard and the other will be easy.
From reading the paper, GPT-4 also outperformed the researchers themselves in many categories, despite the researchers being the ones who created the dataset being used to perform the comparison.
In other words, the metrics are biased in the researchers’ favor — so GPT-4 would have beat them even more often (probably a majority of the time based on the numbers), if someone else had created the guidelines and golden labels.
“ Employing Surge AI's top-tier human annotators at a rate of $25 per hour would have cost $500,000 for 20,000 hours of work”. That’s a wrap for Surge AI
Lots of immediate business from companies needing humans to spin up their models though... but as LLMs get more advanced it's anyone's guess what will happen here.
So, uh, GPT-4 outperforms at labeling. What is that labeling used for?
"Employing Surge AI's top-tier human annotators at a rate of $25 per hour would have cost $500,000 for 20,000 hours of work, an excessive amount to invest in the research endeavor. Surge AI is a venture-backed startup that performs the human labeling for numerous AI companies including OpenAI, Meta, and Anthropic."
What could go wrong? Using GPT-4 to perform labeling used by OpenAI in order to train...uh, wait.
Think about it, how many millions of articles are posted online produced by OpenAI's GPTs to date... Good luck clearing out the training data for GPT-5.
True human content will get gradually scarce. We steer it for sure for our posts, but it is still GPTs that do the heavy lifting.
OpenAI's own classifier fails to detect GPT-4 generated text at the moment.
That's because beyond the 'As an AI language model' and a few key words it can be nearly impossible to detect GPT-4 especially if any prompt is used to intentionally keep it from being detected.
Human like text is a solved problem. There is no more getting better at detecting AI written text, there is only classifying more humans incorrectly at this point.
>This breakthrough saved the researchers over $500,000 and 20,000 hours of human labor.
BTW, this is interesting. There is a lot of noise about AI carbon footprint. Now imagine how much humans would eat and fart for 20.000 work hours. It's about 10 man/years. Assuming 8h / 5d / 50 weeks schedule.
This is really interesting result. Immediate and direct application of LLMs, with significant financial benefits. I think LLMs will drive tremendous productivity increase.
150 comments
[ 2.0 ms ] story [ 203 ms ] threadI'm much more impressed by GPT's ability to handle input than I am in its ability to generate output. It's arguably as good at reading comprehension as most humans.
For example, consider filling the blank:
A giant ______ flew over my head!
It can be a plane. Or a dragon. Or an UFO. Or a balloon. The thing is all of those are correct answers language-wise and the model works correctly as long as what gets filled in conforms to the rules of the given language.
The language that we generate encodes reality to some extent and the model picks up those correlations but there is no concept of reasoning or reality behind it. Maybe it is emergent at some point (as to effectively compress it needs to encode some subset of rules governing our reality) but it is not an agent that optimizes for understanding our reality. Something like Dreamer would be much closer to that.
With an AI like GPT, it is quirky and amusing. Once AIs get really powerful, it becomes scary, and a lot of people who understand this field much better than I do are worried it has a good chance of being deadly. Like, potentially kill-everyone-on-earth deadly.
Personally I didn't need to imagine a specific scenario to understand that there's risk, but I think it would help me convince other folks if I did.
If you want society to collapse all you need to do is succeed in having AI automate all jobs.
Every single country where money comes from somewhere other than people (oil, diamonds...) is an authoritarian nightmare simply because keeping people happy is not necessary.
Once AI can do everything and robots that can do any physical labor are developed the population will shrink dramatically as people with killer robots kill each other for resources. There is no need for AI rebellion or AI failure to get there.
Do we get hoverboards now or is that later ?
Labelbox does image annotating still, and one CTO said as soon as GPT-4 enabled this for him he'd have his team homebrew it from there.
Worrying that this will no longer be the case.
Perhaps history will show that the NSA made algorithmic breakthroughs a few years ago and realized what was coming, so political policy was crafted to stymie Chinese progress in this field, and what we're seeing in the public sphere from companies like openAI is a managed release of the technology into the public, openAI at least managing to independently discover the same breakthroughs that the NSA made a a few years ago.
Most people don't really think about things that don't affect their day-to-day lives. This includes the specifics of how Governments might run a mass surveillance plan.
GPT-like systems will close that gap, and then comes all of the problems of automated law enforcement - Extrapolation from incomplete data, false positives from coindicences, interpretation errors, all that annoying stuff
No, it hasn't lol.
Take drones for example. The government got really good at those because they made them jet-powered (lol) and blew a bunch of money on server-grade FPGA’s in each one of them.
You can’t really just buy a lot of GPUs to make an LLM work, you need iterative development of architecture and training methods.
Like maybe the government invented self-attention before 2017, but if they didn’t, then the constraint is training time, and the government has the same number of seconds as the rest of us.
Everything you use is from the military.
The government is good at lying, and making themselves ‘appear’ incompetent.
They secretly probably have a much further advanced quantum computer. Your viewpoint is limited to mainstream technology and mainstream science.
The military invented the nuclear bomb yes. But Fermi did most of his thinking work in Italy before the Manhattan project. He got money thrown at him once he got here.
As for semiconductor devices, it was Bell Labs and TI.
Coding is an ambiguous concept that wasn’t really invented, but if it were, it would have first appeared in programmable looms.
The military likes to take credit for things, but really all they do is throw money at existing inventions.
I’m sure they’re throwing a bunch of money at Transformers now, but who are all these uncredited super geniuses who invent things and then let randos at Google take the credit/earn the money?
They made some mistakes in the 40's and 50's to where they had nuclear secrets stolen by the Russians. And ever since has been hyper compartmentalized.
It would not surprise me if in the 90's or 00's they had an internal working LLM, considering all the puzzle pieces. You will never hear about classified tech unless it's a bomb, gets leaked. (See code breaking machines declassed after 70+ years)
In a hypothetical scenario, a military organization might want to conceal its use of a large language model (LLM) for intelligence gathering and analysis.
Another scenario is the military's current interest in everything quantum. Quantum computers for example (you wouldn't want another nation being first and pirate baying out our secrets, would you) so there is an extreme national security importance of being first.
And to be first, you need to have the smart people, which the military has. There is a reason China struggles with jet engines 80 years after their invention, and still can't make nuclear carriers. While the US navy works on things like this: https://www.navair.navy.mil/foia/sites/g/files/jejdrs566/fil...
Jet engines can be solved with money.
The actual steps of making an LLM require complex math and you can’t just pay people to make better math.
And if you could, wouldn’t those people decamp for industry and become literal trillionares?
So your implication is that the military is full of unnamed linear algebra, systems engineering, and linguistics super geniuses and these people never leave, never talk about their work publicly, never publish anything ever, and they're all cool with their huge innovations being kept away from the public forever? All because military IP regulation?
And none of these effective state prisoners ever defect to China (where they could live like royalty) because...
If that is your angle, I very much agree it's possible some deep black project exists that (once) looked into this. In the UFO lore, there are many stories about these advances (tr3b etc).
But all of it is orthogonal to LLMs though. Picking one exceptional area that the military is great at (aerospace) does not suddenly make the military exceptional in other areas like AI.
https://www.youtube.com/watch?v=-wU7nPDcTuY&t
See above video, even code breaking machines are kept a dark secret. LLM's that could make analytical decisions about war and strategy, must have started with the military first. It explains its massive data gathering operations in the 2000's. And it explains why some countries separation to make their own internet, away from what really is the US-Owned World Wide Web.
U.S. military has engaged in the commercialization of top-secret technologies (after it's considered obsolete by military standards), often by collaborating with private companies or research institutions.
Even the Manhattan project had nuclear research going on in public universities at the time.
Nothing of the sort here for the attention mechanism which underpins LLMs we know today.
Fundamental research isn't something you just throw money at and acquire. All we had back then were cleverbot and other expert systems.
The military is responsible for most the technology we use and talk about today. The government may appear incompetent, but we’re living off military hand-me-downs, the entire world is
If today's hardware was available 20 yrs ago, this would've been possible just like the moon landing could've been faked if it took place 20+ yrs later. The technology wasn't available at the time (GPUs in this case, and generally no experience in doing such advanced trick techniques for movies back then)
These models are having such a strong effect now because we've finally got the hardware to run them
They are not deep learning/neural nets.
Also fun fact as a pedant tax: Symantec is so named because they started out as transcription software, hit a wall, and pivoted to security SW.
[1] https://www.theguardian.com/technology/2011/mar/17/us-spy-op...
[2] https://boingboing.net/2015/06/22/gchqs-psy-ops-squad-target...
Because the hardware has not existed.
This said by accident I've seen hardware that was brought to a testing company by federal marshals that was massively parallel custom hardware that was likely for signal processing a lot of channels at once. So there is plenty of custom hardware out there, but these items have not been produced at the scale needed (from what anyone can tell) and, again from what we can tell, they don't have the general processing capability that GPU/TPU driven LLMs have.
not exactly. Where do you think all those budget trillions that don't have to be accounted for goes into? the FBI+NSA (=CIA but for citzens) have infinite resources.
All the overhead they have is to make sure a small subset of the citizens are not impacted. Snowden goes into this in some detail when talking about day to day operations. The norm is to extend the net as wide as possible, until you reach some politician or government agency.
They've created a huge library of unorganized data. The difference here is they now can spawn a million untiring AI private investigators / librarians to organize this information into coherent "case files".
At least for me, until this point I've had a feeling of anonymity in the idea that, while my data is being slurped up, I'm just one data point in a sea of other 'normal' people. There would be little value in spending government time and effort tying all of the web detritus together for me. The juice would definitely not be worth the squeeze.
However, when the cost of this effort is nearly zero, that now becomes a different story. The balance of power between government and the people it rules is going to radically shift.
https://www.cnn.com/2021/04/29/tech/nijeer-parks-facial-reco...
Shotspotter has been billed as a “system of sensors, software, AI and expert human review that accurately detects, locates and alerts police to gunfire”, and the company behind it (formerly “Shotspotter” was the company name, its recently been renamed “Soundthinking”) has a number of other AI-involved law enforcement products now, as well.
That's already how it worked on platforms like mturk and uhrs, lots of the work was transcribing audio dumps from microphones built into computers/phones/smart home devices. UHRS especially had a lot of that (it's owned by MS) as well as search engine grading type work. They also certainly do not pay well, I'd imagine that in practice there isn't much cost difference to paying a bunch of bored people to do it vs the compute cost for running an AI model to do it, but the AI model will be vastly more accurate and will work 24/7.
Imagine you've sent an email about transporting a friend's daughter across state lines to get a medically-necessary abortion. Or if you prefer, imagine you've arranged via email to "lose" some firearms which don't comply with your state's new assault weapons ban.
Pre-LLMs, trying to find these sorts of emails was very hard. A simple text search for "abortion" or "gun" is going to come up with far more emails where two family members got into a political debate, than emails about lawbreaking. Big Brother will find a few such emails here and there by chance, but the vast majority of such incriminating emails will simply be lost in the pile.
Enter LLMs, and Big Brother can feed some of the incriminating emails found my chance into a training dataset along with a bunch of non-incriminating emails, and teach the AI to find incriminating emails, and then apply the model to the entire list of emails and get a nicely filtered list of only the emails which are incriminating, further tuning the model by adding emails it gets wrong to the training dataset when they are found.
GPT-4 can do that task for fractions of a penny per email now. It doesn't have to be perfect if its competing with nothing. I expect we'll see similar shops for any other high cost paper/trail business.
What I'd really love to implement is a way for GPT-4 to answer questions based on a corpus of "all our Confluence pages plus random other sources of documentation." Like with the legal document issue, it's a bit of a nonstarter right now given the proprietary nature of corporate documentation.
"Hey, has anyone worked on Problem X, and what was the outcome of their project"
https://old.reddit.com/r/ChatGPT/comments/12fiwaf/chat_gpt_w...
There are just not enough NVIDIA GPUs.
Replace "the machine" eith "the market" and it describes some people today.
What is going to happen are private robotic armies making sure private owners remain private owners.
And then, we will go back to times where people were not citizens by default and had fewer rights.
I’m not worried about rogue AI taking over the nukes. I’m worried the same people who think it’s a great idea to charge so much for insulin that people start dying are the ones who will be using AI to hurt people.
Hell, give me a slightly evil AI run amok over any pharma CEO doing their job.
Human greed is the problem. Authoritarianism and capitalism are just subcategories of the greed problem.
What we don't have an answer for yet, is will AGI be greedy?
Either you go build something you want, an island where code doesn't talk back, or leave tech altogether.
I am saddened that I don't even recognise this place anymore. It's not Hacker News anymore, it's AI news. It's starry eyed engineers jumping over each other ready to sell their metaphorical soul. Even I, the Luddite, can't seem to talk about anything else than this bloody thing.
My current plan is to build a small business and retire in the middle of the woods somewhere. Do some Lisp coding while the rest of the world is dancing around their new idol.
/rant, send me an email if you wanna rant about it as well, and discuss your concerns.
That's going to be the exception, not the rule. The benefits from automating crowdwork will disproportionately accrue to corporate profits.
I believe people can contribute in many different ways. When technology enables us to get my work output without me, that frees me up to produce other things for society.
The problem is that it is a disruption for everything because at its core it is a machine for the replication of skill and technology. A concept that has never existed prior with any other technological disruption.
"Climbing the skill ladder is going to look more like running on a treadmill at the gym. No matter how fast you run, you aren’t moving, AI is still right behind you learning everything that you can do."
from a more in depth view I wrote up here describing the rapidly shrinking innovation, disruption and adaption cycles
https://dakara.substack.com/p/ai-and-the-end-to-all-things
Sure. But when your “keep the lights on” job cuts you for AI, you are less likely to “produce other things” while you worry about food and heat.
Surprisingly (to me) many people think that GPT-N will never exceed human level intelligence because it was trained on the internet. I think that argument is obviously wrong.
Another is that I am sure a large chunk of people will never concede that the AI is smarter than them. Literally never, no matter how smart the bot gets. I mean, probably a lot of people think they are as smart as anyone else. They won't agree that someone else is smarter than them, and they certainly won't agree that some bot is smarter than them. It's also a loaded assessment, like they will think that if they agree to that, then they are also implicitly agreeing to cede their personal agency to the bot.
Another possibility is that GPT-N successors that surpass human level cognition will be banned by regulation, like some drugs or nuclear explosives or bio weapons. They could even be pre-emptively banned at some level below human level, and maybe it would never be publicly acknowledged that it's technically possible to go above human level.
I think it's a hard sell to say that nothing better than human is allowed.
But the existential concern comes from supposed exponential takeoff. To me it should be easy to convince people that something 10 or 100 times faster or smarter than humans should not be allowed.
Weirdly you don't see people talking about regulating autonomy which is also part of it.
I find the argument that Eliezer Yudkowsky makes with the super slow aliens to be very compelling especially in the context of fully autonomous AI that people have stupidly designed to imitate human (animal) characteristics like survival instincts. I suspect that that regulators will ban extremely high performing AI. But unfortunately the prediction is that this can't be contained which means it's quite probable that militaries will cause the end of the world just like we have been expecting but in a new way. Since they will likely be excluded from the ban at least secretly.
I think it's weird that you say "I think it's a hard sell to say that nothing better than human is allowed." in combination with "To me it should be easy to convince people that something 10 or 100 times faster or smarter than humans should not be allowed."
To me, those situations are way too close to each other (human level vs. 10x 'smarter') to be able to say one is a "hard sell" and the other is "easy to convince people." I mean maybe you are exaggerating your certainty to make a point, but I don't think it's at all realistic to say that one will be hard and the other will be easy.
The easily convincing argument for me is imagining the planet populated by aliens that move extremely slowly. This makes more sense when it's closer to 1/100th speed. https://www.lesswrong.com/posts/5wMcKNAwB6X4mp9og/that-alien...
Although having actually read that, it's not convincing in this context because the AI in the story is already fantastically performant.
In other words, the metrics are biased in the researchers’ favor — so GPT-4 would have beat them even more often (probably a majority of the time based on the numbers), if someone else had created the guidelines and golden labels.
All labor is skilled labor. Even breaking rocks.
"Employing Surge AI's top-tier human annotators at a rate of $25 per hour would have cost $500,000 for 20,000 hours of work, an excessive amount to invest in the research endeavor. Surge AI is a venture-backed startup that performs the human labeling for numerous AI companies including OpenAI, Meta, and Anthropic."
What could go wrong? Using GPT-4 to perform labeling used by OpenAI in order to train...uh, wait.
Think about it, how many millions of articles are posted online produced by OpenAI's GPTs to date... Good luck clearing out the training data for GPT-5.
True human content will get gradually scarce. We steer it for sure for our posts, but it is still GPTs that do the heavy lifting.
OpenAI's own classifier fails to detect GPT-4 generated text at the moment.
That's because beyond the 'As an AI language model' and a few key words it can be nearly impossible to detect GPT-4 especially if any prompt is used to intentionally keep it from being detected.
Human like text is a solved problem. There is no more getting better at detecting AI written text, there is only classifying more humans incorrectly at this point.
BTW, this is interesting. There is a lot of noise about AI carbon footprint. Now imagine how much humans would eat and fart for 20.000 work hours. It's about 10 man/years. Assuming 8h / 5d / 50 weeks schedule.
I don’t think you can compare people’s carbon footprint because those people will exist regardless of jobs.
But they don't have to /s
And if your ground truth is problematic, then this is generally a problem of specification and quality control, not performance.