Really glad this article is so high up because we should be talking *primarily* about regulating Data and not "Algorithms" or "Compute" when it comes to governing "AI."
If commoncrawl didn't exist, there would be no ChatGPT
If ImageNet didn't exist there would be no Tensorflow
etc...
Go look at it, all of the value, the actual information, the vectors, matrices, and weights and hyperparameters...all are derived from a single source: Large scale concatenated tables of data. This isn't your daddy's GOFAI - this is only possible because of internet scale data that is relatively trivial to access and simple to transform.
I've read calls for regulating (or bombing (wtf?!)) GPU clusters to prevent skynet from happening and it's all very misguided and shows how little is understood about how AI systems are actually built.
We should be regulating and managing DATA collection, DATA processing, DATA privacy etc... all of the other parts including the hardware and math will be obsolete or replaced totally over the next decade, but you still need observation of human behavior in some form. Norbert Wiener told us all this in 1950.
That process of computationally observing humans - aka metrics, monitoring etc... that comprises the global surveillance economy, should be the thing regulated. All the other problems stem from that.
The fact that surveillance capitalism underpins literally TRILLIONS of dollars and the entirety of modern society tells me that data collection/processing regulation at the scale required is going to be difficult at best.
regulation around data already exists, and there will likely be more coming in the future :)
In fact, I recall seeing an article about OpenAI and Microsoft having to pay a large fine for some illegal data scraping practices or similar, but I'm unable to re-find it.
I'm curious if you feel strongly about the 'open internet' as well, i.e. that platforms and content should be publicly available by default (as opposed to e.g. Elon's recent move to restrict the viewability of tweets by not logged in users).
To me, these datasets don't come from some nefarious violation of privacy (surveillance) on the people who've written or otherwise authored their content, but from an aggregation of public content that was previously uncontroversial.
If these data were pulled from private content, it would be a different story - although I'm sure that many companies TOS allows the use of your 'private' content to train their models.
Going forward I wonder if we may see an 'AI opt-out' for your data regulated (I'm not sure if the new EU act stipulates this)
All personal data collection should be opt-in only
So if you have data on someone that is attributable to their person, then the only legal way to store that is with the approval and consent of the individual who created it for the explicit and narrow purposes the data is being used for.
So for example if you use any personal data for scoring in recommendation systems, then that specific use case must be agreed to by the user. If you use personal data to cluster people into affinity groups that are then shown different content than other groups, that explicit use case must be agreed to by the user. etc...
The enforcement mechanism is simple: Attributable personal data that is found to be stored without consent must be:
1. Immediately deleted
2. All previous revenues derived from the data will be transferred to the user in question
That rule + enforcement mechanism should put concrete boots on everyone collecting data.
this blogspam seems like it could be half-generated by AI. The actual article is interesting though, and reminded me of the one silly scene from Parasite, where the poor guy kept hitting the lights with… his head? Or whatnot
I was on the open AI website and I was looking for training jobs but I couldn't find anything. The article mentions a guy who works for open AI training the AI
I've been thinking we need a OSS source for fine-tuning and RLHF data. Like an npm, cargo, pip or whatever for RLHF data for any topic. From the original Verge article:
> Only the companies that can afford to buy this data can compete, and those that get it are highly motivated to keep it secret.
In the same way that you want the source code of your software to be accessible to you to edit, it seems reasonable to want your AI to be accessible to you. In a world where models are open, that means having the data to train them that is open.
Hugging face may act as a repo as I described, but I'm thinking about the platform infrastructure around it.
Companies like Scale and Surge have platforms that allow 100,000+ individuals to annotate responses from LLMs. Right now, those annotations are going into walled gardens.
As an analogy, I'm thinking something more like Wikipedia. It isn't just a dataset. It is a browser, editor, etc. that is layered on top of the data. Wikipedia isn't just the data stored in giant flat files.
I’ll add an odd tertiary problem HN folks might find interesting.
It is becoming increasingly common for “undergraduate research opportunities” to in fact just be a job annotating data. Researchers need some custom annotated dataset. Undergraduates want “research experience”. So researchers offer an opportunity to undergrads who think they will be doing meaningful research but instead are just mechanical turkers. No pay, perhaps a promise of research credit later…
It’s disingenuous to students but also disingenuous to companies looking for some meaningful research experience on resumes that list it.
This was also a problem a decade ago. I asked a professor for research opportunities and ended up segmenting a bunch of images. That was it. No credit for that work, either (or pay).
13 comments
[ 3.3 ms ] story [ 47.1 ms ] threadIf commoncrawl didn't exist, there would be no ChatGPT
If ImageNet didn't exist there would be no Tensorflow
etc...
Go look at it, all of the value, the actual information, the vectors, matrices, and weights and hyperparameters...all are derived from a single source: Large scale concatenated tables of data. This isn't your daddy's GOFAI - this is only possible because of internet scale data that is relatively trivial to access and simple to transform.
I've read calls for regulating (or bombing (wtf?!)) GPU clusters to prevent skynet from happening and it's all very misguided and shows how little is understood about how AI systems are actually built.
We should be regulating and managing DATA collection, DATA processing, DATA privacy etc... all of the other parts including the hardware and math will be obsolete or replaced totally over the next decade, but you still need observation of human behavior in some form. Norbert Wiener told us all this in 1950.
That process of computationally observing humans - aka metrics, monitoring etc... that comprises the global surveillance economy, should be the thing regulated. All the other problems stem from that.
The fact that surveillance capitalism underpins literally TRILLIONS of dollars and the entirety of modern society tells me that data collection/processing regulation at the scale required is going to be difficult at best.
regulation around data already exists, and there will likely be more coming in the future :)
In fact, I recall seeing an article about OpenAI and Microsoft having to pay a large fine for some illegal data scraping practices or similar, but I'm unable to re-find it.
To me, these datasets don't come from some nefarious violation of privacy (surveillance) on the people who've written or otherwise authored their content, but from an aggregation of public content that was previously uncontroversial.
If these data were pulled from private content, it would be a different story - although I'm sure that many companies TOS allows the use of your 'private' content to train their models.
Going forward I wonder if we may see an 'AI opt-out' for your data regulated (I'm not sure if the new EU act stipulates this)
So if you have data on someone that is attributable to their person, then the only legal way to store that is with the approval and consent of the individual who created it for the explicit and narrow purposes the data is being used for.
So for example if you use any personal data for scoring in recommendation systems, then that specific use case must be agreed to by the user. If you use personal data to cluster people into affinity groups that are then shown different content than other groups, that explicit use case must be agreed to by the user. etc...
The enforcement mechanism is simple: Attributable personal data that is found to be stored without consent must be:
1. Immediately deleted 2. All previous revenues derived from the data will be transferred to the user in question
That rule + enforcement mechanism should put concrete boots on everyone collecting data.
https://www.theverge.com/features/23764584/ai-artificial-int...
this blogspam seems like it could be half-generated by AI. The actual article is interesting though, and reminded me of the one silly scene from Parasite, where the poor guy kept hitting the lights with… his head? Or whatnot
> Only the companies that can afford to buy this data can compete, and those that get it are highly motivated to keep it secret.
In the same way that you want the source code of your software to be accessible to you to edit, it seems reasonable to want your AI to be accessible to you. In a world where models are open, that means having the data to train them that is open.
Companies like Scale and Surge have platforms that allow 100,000+ individuals to annotate responses from LLMs. Right now, those annotations are going into walled gardens.
As an analogy, I'm thinking something more like Wikipedia. It isn't just a dataset. It is a browser, editor, etc. that is layered on top of the data. Wikipedia isn't just the data stored in giant flat files.
It is becoming increasingly common for “undergraduate research opportunities” to in fact just be a job annotating data. Researchers need some custom annotated dataset. Undergraduates want “research experience”. So researchers offer an opportunity to undergrads who think they will be doing meaningful research but instead are just mechanical turkers. No pay, perhaps a promise of research credit later…
It’s disingenuous to students but also disingenuous to companies looking for some meaningful research experience on resumes that list it.